Copyright TechTarget - All rights reserved https://cyber.law.harvard.edu/rss/rss.html Techtarget Feed Generator en Tue, 25 Aug 2026 15:05:28 GMT https://www.techtarget.com/ai editor@techtarget.com <p>Does the traditional <a href="https://www.techtarget.com/it-strategy/tip/What-are-the-pros-and-cons-of-shadow-IT">shadow IT playbook</a> work for shadow AI? No, but not for the reasons most imagine.</p> <p>With the right technology stack, IT teams can still discover and inventory a large share of the shadow AI tools in their environments. What doesn't scale is acting on what they find.</p> <p>According to a 2026 Resume Now <a target="_blank" href="https://www.resume-now.com/job-resources/careers/byo-ai-report" rel="noopener">survey</a> of more than 1,000 U.S. workers, 76% have sourced their own AI tools rather than using employer-provided options. The volume in use across large organizations is severely underestimated and growing fast.</p> <section class="section main-article-chapter" data-menu-title="The discovery process"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The discovery process</h2> <p>Most of the public discussion and research on shadow AI focuses on discovery. Shadow AI poses distinct challenges here, including AI features shipping inside already-approved SaaS tools, unauthorized use of approved AI tools, unvetted browser extensions and AI-enabled integrated development environment plugins -- all of which generate far less detectable activity than traditional shadow IT ever did.</p> <p>The traditional shadow IT discovery stack comprises <a href="https://www.techtarget.com/it-infrastructure/tip/5-clues-your-network-has-shadow-AI">network monitoring</a>, web proxies, cloud access security brokers and DNS logging. With these tools, IT teams can effectively inventory the desktop and web applications in use across their organizations. But visibility generally ends there -- teams can see what tools are being used, not what data is flowing through them. And few businesses continuously monitor their existing vendors for newly released AI features that might change how data is handled.</p> <p>This lack of visibility has pushed many IT teams and vendors toward a new perimeter that isn't defined by the network but rather by the end user's browser and device. An emerging category of tools, described as workforce AI security or AI usage security, moves beyond identifying which application a person can access to get visibility into interactions between users and AI systems. These tools monitor the content of prompts, the sensitivity of data being shared and whether AI features within approved tools have changed how data is handled. This functionality is appearing in both <a target="_blank" href="https://learn.microsoft.com/en-us/purview/deploymentmodels/depmod-data-leak-shadow-ai-intro" rel="noopener">established platforms</a>, such as Microsoft Purview, and <a target="_blank" href="https://www.checkpoint.com/ai-security/ai-workforce-security/" rel="noopener">newer entrants</a>, such as Check Point's Workforce AI Security product.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Compiling a list of shadow AI tools is … the easy part. The harder question is what happens next? </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Gaps remain, of course. Privacy regulations such as the EU's GDPR could restrict prompt-level monitoring in certain jurisdictions. Desktop AI applications might evade browser-based monitoring tools entirely, even on managed devices. And <a href="https://www.techtarget.com/cybersecurity/news/366646095/Employees-shadow-AI-use-is-poorly-monitored-survey-finds">personal devices are largely invisible</a>; an employee can pull company data onto their phone and run it through any consumer AI app with no enterprise tool ever seeing it.</p> <p>That said, on managed devices and corporate networks, discovery is becoming increasingly manageable. IT teams that haven't yet explored the newer generation of workforce AI security tools should. The coverage they offer is better than most organizations realize.</p> </section> <section class="section main-article-chapter" data-menu-title="The block-and-tackle approach"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The block-and-tackle approach</h2> <p>Compiling a list of shadow <a href="https://www.techtarget.com/ai/tip/9-top-generative-AI-tool-categories-for-2026">AI tools</a> is, believe it or not, the easy part. The harder question is what happens next?</p> <p>The standard post-discovery process looks roughly the same in most organizations:</p> <ul class="default-list"> <li>Identify an unsanctioned tool.</li> <li>Assess its risk.</li> <li>Reach out to the team using it.</li> <li>Understand their business case.</li> <li>Make a decision – sanction it, replace it or block it.</li> </ul> <p>This is the blocking-and-tackling model, and it can work at a basic level.</p> <p>But each tool requires its own, often highly manual remediation project. In practice, a single discovery can mean weeks of work, tracking down who's using the tool; evaluating whether it can be sanctioned; if not, finding a replacement; and communicating the decision back to affected users.</p> <p>Some steps, such as flagging tools above a certain risk threshold or notifying users automatically, can be streamlined. But as most IT leaders acknowledge, simply removing a tool without providing an adequate replacement can push employees toward another workaround.</p> <p>Blocking and tackling doesn't scale in the AI era. Much of the current IT governance focus centers on a handful of general-purpose models -- ChatGPT, Claude and Gemini -- as though managing those platforms is the bulk of the problem. But increasingly, employee AI use is shifting to vertical tools: purpose-built applications, mostly SaaS-based, for contract review, content preparation, financial modeling, customer support and dozens of other functions.</p> <p>Many of these niche tools have AI so deeply embedded that users often don't think of them as AI tools. And, because AI has dramatically reduced the cost of building software, new vendors are constantly targeting narrower use cases. Evaluating a few major platforms might be manageable, but evaluating hundreds or thousands of specialized tools that small teams are using to solve problems IT might not even understand isn't manageable.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> In many companies, employees often assume IT will say no, so they never ask. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>The blocking-and-tackling model is still necessary for the highest risk discoveries -- tools handling customers' personally identifiable information, regulated data and proprietary source code. But as the primary strategy for managing shadow AI, it isn't sufficient.</p> </section> <section class="section main-article-chapter" data-menu-title="Organizational changes"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Organizational changes</h2> <p>Blocking and tackling will only go so far. To manage shadow AI at the scale most organizations are facing, real process and organizational changes are needed. According to a National Cybersecurity Alliance <a target="_blank" href="https://www.staysafeonline.org/press/study-65-now-use-ai-but-majority-remain-untrained-on-risks" rel="noopener">study</a>, 58% of AI users received no training in data security or privacy risks associated with the technology. A majority of workers lack clear guidance on AI use policies. In many companies, employees often assume IT will say no, so they never ask.</p> <p>Three areas deserve attention.</p> <h3>Faster, leaner approval processes</h3> <p>Today, model risk management can take up to a year, and standard tools can take weeks or months to approve. Meanwhile, an employee can sign up for a free AI tool in under a minute. IT teams need to close that gap with more self-serve approval pathways for tools that don't touch critical data, automated decisions where possible and streamlined risk reviews for everything else.</p> <p>The review criteria should be simple and consistent. It should focus on what data flows into the tool, the vendor's data retention and training policies, and the data classification that applies. The data alone doesn't determine risk. A use involving nonsensitive data can still be consequential if it affects customer decisions, pricing or production systems. Data governance should be embedded into the approval criteria.</p> <h3>More distributed IT governance</h3> <p>A clear insight from organizational economics is that decentralized units respond better to changes in their markets. In practice, this means pushing <a href="https://www.techtarget.com/ai/tip/4-strategic-approaches-to-effective-shadow-AI-governance">AI governance decisions</a> closer to where tools are being adopted -- not through additional AI committees, but through embedded technology professionals, department-level accountability for AI risk or designated AI leads within business units. Central IT sets the framework and escalation thresholds, and business units operate within it.</p> <h3>Security education</h3> <p>Security education should focus narrowly on what data can and can't be shared with external AI tools, and what types of tools carry the most risk. Phishing awareness programs can take years of sustained effort but ultimately change employee behavior by keeping the message simple and repeatable. Shadow AI education should aim for the same.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The companies best positioned to capitalize will be the ones that build clear guardrails around low-risk experimentation while focusing security resources on the uses that could cause real harm. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>These three shifts won't eliminate shadow AI. But they move the organization from case-by-case remediation to a more comprehensive model for managing AI risks at scale.</p> </section> <section class="section main-article-chapter" data-menu-title="What's at stake"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What's at stake</h2> <p>Organizations with high levels of shadow AI incurred average breach costs $670,000 higher than those with little or no shadow AI, according to IBM's 2025 Cost of a Data Breach <a target="_blank" href="https://www.ibm.com/think/insights/data-matters/cost-of-a-data-breach" rel="noopener">Report</a>.</p> <p>But the financial risk is only part of the picture. The pace of AI innovation is accelerating, and so is the opportunity cost of failing to govern it well. The companies best positioned to capitalize will be the ones that build clear guardrails around low-risk experimentation while focusing security resources on the uses that could cause real harm. The discovery tools exist. The harder work is everything that follows.</p> <p><i>John Frechette is the founder of Sourced Economics, a research and advisory firm focused on enterprise digital transformation. Dr. Suzannah Hicks is senior manager of the AI Center of Excellence at Ferguson, a Fortune 500 company. The views expressed by Dr. Hicks are her own.</i></p> </section> IT teams can now identify shadow AI tools, but remediation doesn't scale. Managing hundreds of niche AI tools requires faster approvals and distributed governance. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a194810146.jpg https://www.techtarget.com/ai/post/Beyond-discovery-The-real-shadow-AI-challenge Tue, 25 Aug 2026 11:28:00 GMT <p>Everyone deserves access to the best AI technology.</p> <p>This statement summarizes the vision for the future of AI that Meta CEO Mark Zuckerberg articulated in what some are calling his <a href="https://www.meta.com/thefutureisforeveryone/?srsltid=AfmBOooNrVdIjI1oA1cYQzYFekvAENetomMJtyhQT-bXigGsKdQKvZOR">AI manifesto</a>, "The Future is for Everyone." The document describes how Zuckerberg believes AI should evolve to protect privacy, cybersecurity and jobs. It also explains what Meta is doing to help translate that vision into reality, like building "personal agents" to assist people in virtually every aspect of their lives.</p> <p>This all sounds very nice. But it also sounds a lot like promises Meta has made in the past that never fully materialized. Questions remain: How much success will Meta have in bringing about a world where secure, trustworthy AI is available to all? And what will Zuckerberg's current AI strategy mean for consumers, businesses and the technology ecosystem?</p> <section class="section main-article-chapter" data-menu-title="How Zuckerberg sees the future of AI"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How Zuckerberg sees the future of AI</h2> <p>Zuckerberg describes the current AI discourse among AI developers as "filled with doom." He says fears that modern AI threatens to take jobs and spawn large-scale cybersecurity chaos have created a widespread belief that the only way to <a target="_blank" href="https://www.techtarget.com/ai/tip/AI-risk-management-A-strategic-guide-for-enterprise-leaders" rel="noopener">control AI risks</a> is to create an "extreme concentration of power" by restricting access to the most powerful AI models to a handful of businesses.</p> <p>Zuckerberg offers a different take on what AI can become. In his telling, AI, if developed and disseminated in the right way, will usher in "a new era of personal empowerment."</p> <p>According to Zuckerberg, the key to unlocking this future is to ensure that everyone has access to advanced AI technology. In his manifesto, he frequently refers to democracy and "democratic values." However, he stops short of saying we should "<a target="_blank" href="https://www.techtarget.com/whatis/definition/AI-democratization" rel="noopener">democratize AI</a>," perhaps because that would imply that everyone should have the freedom to develop the technology, not just access it. He also writes about the importance of "putting power in people's hands." History shows, he says, that technology creates the greatest societal benefit when it's distributed and available to all.</p> </section> <section class="section main-article-chapter" data-menu-title="Meta's AI commitments"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Meta's AI commitments</h2> <p>The manifesto doesn't go into much detail about the role Zuckerberg sees Meta playing in ensuring everyone has access to the best AI technology, but it does mention a couple of notable initiatives.</p> <p>One is the introduction of what Zuckerberg calls "personal agents" or "personal superintelligence agents." It's not clear how these will work, but the manifesto says they'll help to "improve your relationships, health, career, finances, home management, hobbies, and more." It also mentions that the agents will come with "strong privacy and security options," although again, it doesn't go into detail about them.</p> <p>The other key Meta contribution he mentions is the release of open source AI models. Meta began debuting open source AI models in 2023, but it seemingly <a href="https://thenewstack.io/meta-abandons-llama-spark/">abandoned the effort</a> earlier this year in favor of proprietary models. But according to the manifesto, Meta plans to "resume releasing some open source models soon."</p> </section> <section class="section main-article-chapter" data-menu-title="Meta's AI strategy: Vision vs. reality"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Meta's AI strategy: Vision vs. reality</h2> <p>On the surface, it's hard to take issue with the vision for AI's future that Zuckerberg articulates. Who doesn't want to ensure that the best AI technology benefits everyone and is safe and secure? But it's also hard to have deep confidence in Meta's willingness or ability to deliver on this vision.</p> <p>So far, Meta has lagged in developing cutting-edge AI technology. As the New York Times <a href="https://www.nytimes.com/2026/08/10/technology/meta-ai-open-source.html">writes</a>, Meta's latest proprietary model, Muse Spark, "does not perform as well as leading A.I. models from Anthropic and OpenAI on measures such as coding, reasoning and writing."</p> <p>It's possible that Meta will catch up on the AI front. But at present, it seems unclear that the company is in a position to create AI models or agents that will perform as well as those of the AI labs that Zuckerberg accuses of creating the "extreme concentration" of AI power.</p> <p>Another, bigger question surrounds how committed Meta is to democratizing access to AI. Although Meta released a series of AI models it describes as "open source," some segments of the open source community, like the Open Source Initiative, have <a href="https://www.axios.com/2024/10/29/meta-osi-definition-open-source-ai-llama">taken issue</a> with its approach.</p> <p>The criticism centers on the limited transparency of how the models work. It also involves Meta's decision to create an unusual custom <a href="https://github.com/meta-llama/llama3/blob/main/LICENSE">license</a> for its open source models instead of using an established open source software license. This effectively bars other large technology companies from using the models without Meta's permission. These terms undermine the spirit of open source software, which prioritizes transparency and free access, even for competitors.</p> <p>To date, Meta hasn't said how its approach to open source AI models could change. The company might decide to offer models that are more "open" in terms of transparency, and to rethink its licensing terms. If it doesn't, it's hard to believe Meta's so-called "open source models" will move the needle toward making AI open and flexible for everyone. Users and businesses who want true open source models are more likely to turn to alternatives like OpenClaw, whose core is genuinely open source and governed by the MIT License, a widely used open source license.</p> <p>It's also worth noting that although the manifesto states that Meta plans to "offer free versions [of Meta's AI tools] that will be accessible to billions of people," it also mentions a "dynamic auction mechanism" that will let people pay for <a target="_blank" href="https://www.techtarget.com/ai/news/366616916/OpenAI-targets-knowledge-workers-with-pricy-ChatGPT-Pro" rel="noopener">access to extra compute.</a> Given the limited information on how the technology will work, it's hard to say which benefits this extra compute will confer. This could be Meta's way of saying "everyone can use our AI technology, but deep-pocketed consumers and companies can pay for and access more capable versions."</p> </section> <section class="section main-article-chapter" data-menu-title="New business challenge, old Meta strategy"> <h2 class="section-title"><i class="icon" data-icon="1"></i>New business challenge, old Meta strategy</h2> <p>The depth of Meta's commitment to ensuring everyone can benefit equally from AI is also uncertain, given the company's long history of positioning itself as a champion of the underdog in its self-serving marketing initiatives.</p> <p>In 2020, Meta ran <a href="https://s3.amazonaws.com/media.mediapost.com/uploads/AD-Standing-up-to-Apple-for-small-business.pdf">ads</a> declaring, "We're standing up to Apple for small businesses everywhere." This was a response to new features in Apple products that restricted ad tracking and personalization -- a direct threat to Meta's revenue. It was hard to believe that Meta cared about small businesses as much as it worried about the sustainability of its ad-based business model.</p> <p>Another Meta <a href="https://bartongeorge.io/tag/open-source-washing-in-ai/">ad campaign</a> touted its open source AI models, promising "Open source AI: Available to all, not just the few." That sounds remarkably similar to the promises Zuckerberg made in his recent manifesto. But just as it's hard to point to much that Meta did during this campaign to deliver true open source technology, it's also difficult to believe that Meta's personal AI agents or renewed commitment to open source AI models will change much now.</p> <p>A cynic might interpret Zuckerberg's manifesto as an effort to position Meta as an outlier in the AI ecosystem, not because the company is committed to open source or the democratization of AI technology, but because it's falling behind other hyperscalers in the AI race. This is a PR move above all. Don't bet the future of AI on Meta right now.</p> <p><em>Chris Tozzi is a freelance writer, research adviser, and professor of IT and society who has previously worked as a journalist and Linux systems administrator.</em></p> </section> Mark Zuckerberg's manifesto on the future of AI sounds promising, with democratized AI and superintelligent personal agents for everyone. But is this vision too good to be true? https://cdn.ttgtmedia.com/rms/onlineimages/chatbot_g1077519886.jpg https://www.techtarget.com/ai/news/366649423/Zuckerbergs-manifesto-Do-Metas-promises-and-actions-align Thu, 20 Aug 2026 13:51:00 GMT <p>U.S. data centers used about 176 terawatt-hours (TWh) of electricity in 2023, or 4.4% of the national total, according to Lawrence Berkeley National Laboratory's 2024 <a target="_blank" href="https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf" rel="noopener">report</a> to Congress. By 2028, that power demand could nearly double or even triple, reaching 6.7% to 12% of the electricity the U.S. will have generated.</p> <p>There currently isn't enough grid capacity to absorb this much demand, especially for the larger data center campuses, said Parag Nathaney, an engineer at a U.S. public electric utility company. Spare capacity is shrinking as AI-driven demand outpaces new supply, and adding substantial new capacity now takes five to seven years. In addition, <a href="https://www.techtarget.com/ai/feature/Communities-call-for-transparency-in-AI-data-center-deals">local residents are increasingly pushing back</a>, fearing their community would bear the perceived increased energy costs, pollution and other environmental issues associated with data center buildouts.</p> <p>As a result, data center planning teams in the process of selecting an energy system to run behind the electricity meter <a href="https://www.techtarget.com/it-infrastructure/tip/Navigating-energy-management-strategies-in-AI-data-centers">must balance several factors</a>, including time to market, power grid demands, equipment availability, ratepayer pushback and environmental footprint. And the cheapest energy source available isn't necessarily able to supply the amount of power needed to handle the load. Estimates <a target="_blank" href="https://www.eia.gov/analysis/studies/powerplants/capitalcost/pdf/capital_cost_AEO2025.pdf" rel="noopener">prepared</a> for the U.S. Energy Information Administration (EIA) in 2024 estimated onshore wind and utility solar at about $1,500 per kilowatt (kW) versus roughly $850 for a large gas combined-cycle plant, which then incurs ongoing fuel charges. Battery storage pushes that utility solar plant past $2,000 per kW, and offshore wind runs nearly $3,700, according to EIA 2024 estimates.</p> <p>Andrei Romanescu, CMO at LumaDock, a provider of GPU servers for data centers, said the local resistance his company encounters "cluster around land, water, noise and what a new load does to household electricity prices." Carbon emissions are further down the list of concerns.</p> <section class="section main-article-chapter" data-menu-title="Timing is everything"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Timing is everything</h2> <p>"The market is selecting for speed, not for the cleanest electron," said Alex Marshall, group business development and marketing director at Clarke Energy and vice president of the Cogen World Coalition, a global community of companies and institutions focused on cogeneration. Data center planners confront a range of <a href="https://www.techtarget.com/ai/feature/GenAI-data-center-infrastructure-reshapes-business-processes">buildout issues</a>, including energy costs, energy waste, property values and emissions, but the concern gaining greater attention is the time it takes data centers to power up.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Time to power remains the primary objective for the data center projects we're participating in. </figure> <figcaption> <strong>Michael Stadler</strong>Co-founder and CTO, Xendee </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>"Time-to-power remains the primary objective for the data center projects we're participating in," said Michael Stadler, co-founder and CTO of microgrid decision support platform provider Xendee. Some data center operators pay well above the going rate for power even though it might not arrive for quite some time. The cheaper power options typically lose out to the more expensive ones that can be connected faster because data centers must also consider the lifespan of their AI chips and the cost to replace them.</p> <p>Lawrence Berkeley National Laboratory <a target="_blank" href="https://emp.lbl.gov/sites/default/files/2026-06/Queued%20Up%202026%20Edition.pdf" rel="noopener">pegs</a> the median wait time for a data center power connection at more than five years. Coordinating utility interconnection, which can range from three to seven years, with data center buildout cycles that might take just one to two years, exacerbates the time-to-power problem.</p> <p>As a result, operators are starting to include <a href="https://www.computerweekly.com/news/366631957/Green-energy-microgrids-hailed-as-cost-effective-answer-to-UKs-datacentre-energy-supply-woes">minigrids in their data center infrastructure</a>. These small-scale localized electricity systems operate and generate power independently without straining and adding costs to community power grids.</p> <p>Still, capital costs and fuel prices are moving targets. Data centers therefore need to conduct multiyear analyses of their operations "and strategically adapt their on-site power strategy to avoid stranded assets or unexpected costs," Stadler advised.</p> <p><a href="https://www.techtarget.com/ai/feature/Questions-to-ask-when-evaluating-AI-ready-data-center-providers">Choosing the right power source</a> is a critical part of that strategy. Some of the more common data center energy alternatives are listed here, along with more detailed information on 14 power source options in the accompanying table.</p> </section> <section class="section main-article-chapter" data-menu-title="Natural gas"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Natural gas</h2> <p>Gas is winning the battle of power sources, particularly in the U.S., thanks to its abundance, availability and about three million miles of pipeline. New combined-cycle gas turbines are the most efficient gas option, but capital costs have roughly doubled since 2022, ranging from $2,000 to $2,500 per kW, Nathaney said. Major gas turbine <a target="_blank" href="https://www.utilitydive.com/news/siemens-gas-turbine-backlog-nears-70-gw-as-company-expands-manufacturing/827390/" rel="noopener">manufacturers</a> like GE Vernova, Siemens Energy and Mitsubishi Power are taking orders with delivery times approaching 2030 or later.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> It's a speed and reliability play, not a decarbonization one. </figure> <figcaption> <strong>Alex Marshall</strong>Group business development and marketing director, Clarke Energy </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Reciprocating internal combustion engines as an on-site bridge or primary power source can come online much faster. They commission in 12 to 18 months, start in about two minutes and maintain efficiency with partial loads better than turbines, which matters when GPU clusters fluctuate, Marshall said.</p> <p>Another option is Jenbacher power and cogeneration systems, which run on natural gas, hydrogen-rich blends or renewable gases. "It's a speed and reliability play, not a decarbonization one," Marshall noted. Natural gas emits carbon dioxide (CO₂) and nitrogen oxide (NOx) emissions, which can create permitting issues in some communities.</p> <p>Very few new coal plants are being commissioned for AI in most parts of the world, but existing coal-fired power plants are expected to fill some of the electrical power gaps for data centers. Natural gas and, in some geographical areas, coal should continue to cover peak and baseload demand as AI workloads ramp up, said Damir Špoljarič, founder and managing partner of Gi21 Capital, which is developing a large European AI data center platform.</p> </section> <section class="section main-article-chapter" data-menu-title="Fuel cells"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Fuel cells</h2> <p>Fuel cells have become a credible middle-of-the-road option. They run on gas, are quiet and eliminate local air pollution, although they still emit CO₂. Bloom Energy announced last year it planned to deliver onsite power to Oracle AI data centers within 90 days, and Oracle expanded its agreement with the energy company in April.</p> <p>In a related development, Oracle and BorderPlex Digital Assets <a target="_blank" href="https://www.oracle.com/news/announcement/oracle-borderplex-and-bloom-energy-to-power-project-jupiter-with-fuel-cell-technology-2026-04-27/" rel="noopener">announced</a> in April that Project Jupiter -- a massive 1,400-acre AI data center campus under construction in Doña Ana County, N.M. -- could use as much as 2.5 gigawatts of gas-powered fuel cells from Bloom to replace previously planned gas turbines and diesel generators, and consolidate the facility into one <a href="https://www.techtarget.com/ai/tip/Virtual-power-plants-could-help-solve-the-AI-energy-problem">single microgrid campus</a>. However, there could be a significant pipeline delay in delivering natural gas to the facility.</p> <p>Fuel cells can theoretically run on hydrogen or ammonia, but hydrogen is less dense and can damage existing pipes. In addition, Amazon canceled plans to use Bloom gas-powered fuel cells in three of its Oregon data centers in 2024 due to the state's concerns the fuel cells would increase the data centers' carbon footprint. Fuel cells also carry higher recurring costs than turbines because the catalysts need periodic replacement.</p> </section> <section class="section main-article-chapter" data-menu-title="Solar and wind"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Solar and wind</h2> <p>While solar and wind don't emit pollutants into the atmosphere, solar requires a large land footprint that invites community pushback similar to the resistance surrounding the amount of real estate data centers require. Solar panels typically last 25 to 30 years, but they carry a cumulative retired photovoltaic waste stream at the end of life that can create economic, environmental and regulatory hurdles. Solar power can only operate efficiently during sunny daylight hours.</p> <p>National Renewable Energy Laboratory cost modeling put the cost of recycling a solar panel at roughly 20 times the cost of landfilling it, and the International Renewable Energy Agency projects as much as 78 million tons of global panel waste by 2050.</p> <p>Wind turbines can produce the same amount of energy as solar using a much smaller footprint, but they also face community resistance due to their size and appearance. However, wind is an <a href="https://www.techtarget.com/it-infrastructure/tip/How-to-use-data-center-wind-turbines-for-sustainable-energy">all-purpose energy source</a> for powering IT equipment and cooling systems. Although it can operate any time of day, there has to be sufficient wind velocity.</p> </section> <section class="section main-article-chapter" data-menu-title="Geothermal"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Geothermal</h2> <p><a href="https://www.techtarget.com/it-infrastructure/tip/The-pros-and-cons-of-geothermal-energy-use">Geothermal energy promises a small footprint</a>, low waste volume, no direct emissions and steady output. The EIA estimates put a 50 megawatt binary-cycle plant at nearly $4,000 per kW with no NOx, sulfur dioxide _&nbsp;or CO₂. Its fixed operating costs are among the highest compared to other power sources. Other drawbacks are longer time to power and difficulty finding acceptable geographical locations. In addition, poorly managed geothermal operating fluids can be a water pollution risk, which could be eased by newer drilling approaches and heat engines.</p> </section> <section class="section main-article-chapter" data-menu-title="Small and large nuclear power plants"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Small and large nuclear power plants</h2> <p>Until recently, nuclear was generally considered too risky and expensive. But the <a href="https://www.techtarget.com/it-strategy/news/366625953/Policymakers-assess-nuclear-energy-for-AI-data-centers">timeline for nuclear has moved up</a>, owing to new regulatory approaches and substantial investment.</p> <p>"SMRs [small modular reactors] may be commercially available earlier than we were thinking at this time last year," Stadler noted. His group initially estimated nuclear power availability would be about 2035, but it now sees 2030 as more realistic. Small nuclear plants aren't necessarily comparatively cheaper to build and operate than large plants.</p> <p>Fission reactors produce highly toxic radioactive waste. The volume of waste is much smaller than other kinds of energy waste, but after several decades, the nuclear waste disposal problem hasn't been entirely solved. However, Finland's Onkalo repository for spent nuclear fuel received a favorable safety <a target="_blank" href="https://www.ans.org/news/article-8280/onkalo-snf-repository-passes-safety-assessment/" rel="noopener">assessment</a> from the Radiation and Nuclear Safety Authority of Finland in August and could be operational later this year to help address the waste disposal issue.</p> <p>While some experts consider SMRs to be the future of nuclear energy because of its potential cost and safety advantages over other reactors, SMR waste per megawatt-hour can be 2 to 30 times higher than a more common conventional reactor, according to an April 2026 <a target="_blank" href="https://www.sciencedirect.com/science/article/abs/pii/S0149197026000041" rel="noopener">paper</a> by <i>Progress in Nuclear Energy</i>.</p> <p><iframe title="Comparing AI data center power sources" aria-label="Table" id="datawrapper-chart-XRX0Z" src="https://datawrapper.dwcdn.net/XRX0Z/1/" scrolling="no" frameborder="0" style="width: 0; min-width: 100% !important; border: none;" height="1389" data-external="1"></iframe></p> <p> <script type="text/javascript">(function(){function e(){window.addEventListener(`message`,function(e){if(e.data[`datawrapper-height`]!==void 0){var t=document.querySelectorAll(`iframe`);for(var n in e.data[`datawrapper-height`])for(var r=0,i;i=t[r];r++)if(i.contentWindow===e.source){var a=e.data[`datawrapper-height`][n]+`px`;i.style.height=a}}})}e()})();</script> </p> </section> <section class="section main-article-chapter" data-menu-title="Powering a 2030 AI data center"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Powering a 2030 AI data center</h2> <p>Exciting technologies typically get the most press, but the likely power source trajectory is more mundane. "Based on our modeling," Stadler said, "we still anticipate natural gas being the primary fuel source, with SMRs starting to enter the picture after 2030, supplemented by solar photovoltaic and battery storage where it makes economic and operational sense."</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The biggest wildcard is not a single technology but the speed of policy and market design. </figure> <figcaption> <strong>Damir Špoljarič</strong>Founder and managing partner, Gi21 Capital </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Hydrogen has been proffered as a green energy carrier but hasn't emerged as a near-term priority in Xendee's projects, Stadler added. Most gas engines installed today can also run on biogas, so operators can <a href="https://www.techtarget.com/ai/post/AI-turns-data-center-power-into-an-enterprise-challenge">shift to that power source</a> if it makes sense.</p> <p>Enhanced geothermal could also play a role, thanks to new techniques that can access usable heat at more sites. The U.S. Department of Energy estimates it could deliver 90 gigawatts of firm (weather-independent, low-emissions) power to the U.S. grid system by 2050, compared to just 4 gigawatts today. Nuclear fusion has been bandied about as cheap, clean and affordable energy option since the 1950s, but the goalposts keep getting pushed back for AI data centers.</p> <p>"The biggest wildcard is not a single technology but the speed of policy and market design," Špoljarič said. Nuclear energy, long-duration storage and demand response could gain wider usage, if flexibility and low-carbon firm capacity become priorities. Until then, natural gas will continue to be the most widely used energy source for AI data centers.</p> <p><i>George Lawton is a journalist based in London. Over the last 30 years, he has written more than 3,000 stories about computers, communications, knowledge management, business, health and other areas that interest him.</i></p> </section> AI data centers are encountering political resistance in some communities, but they'll be built elsewhere. Survival depends on the power sources chosen and their availability. https://cdn.ttgtmedia.com/rms/onlineimages/disaster_recovery_a257795847.jpg https://www.techtarget.com/ai/feature/Power-hungry-AI-data-centers-weigh-viable-energy-sources Wed, 19 Aug 2026 16:45:00 GMT <p>Senior leaders need private spaces to test arguments or examine decisions before announcing them to colleagues. AI assistants can provide that space as a useful thinking tool, but they also create a new decision risk.</p> <p>An agreeable chatbot can take a CEO's initial preferences, provide supporting arguments, resolve inconsistencies and present a polished recommendation, packaging a weak assumption as a brilliant insight or objective analysis. As AI continues to invade the leadership decision process, AI sycophancy is becoming a significant and urgent issue to address. In fact, Deloitte's 2026 Global Human Capital Trends <a target="_blank" href="https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2026/decision-making-with-ai.html" rel="noopener">survey</a> of more than 9,000 business and human resources leaders found that 60% of executives regularly use AI to support their decisions.</p> <p>AI can be a convincing 'yes man,' but it can't provide independence or replace colleagues with the insight and authority to disagree. Leaders using AI must adopt a strategy in which assumptions are scrutinized and actively guard against translating their preferences and biases into sophisticated justifications.</p> <section class="section main-article-chapter" data-menu-title="What is AI sycophancy?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What is AI sycophancy?</h2> <p>AI sycophancy is the tendency of an AI system to affirm, flatter or align itself with the user's expressed beliefs rather than offer balanced analysis. Sycophancy occurs when alignment with the user is prioritized over accuracy, causing the response to change to preserve agreement at the <a target="_blank" href="https://www.techtarget.com/it-strategy/tip/AI-agents-are-running-wild-Secure-the-reasoning-layer-now" rel="noopener">expense of reasoning</a>. This behavior arises from training and feedback that reward responses users like. During AI training, such approval is easier to measure than whether the AI's advice led to a good decision, which users can only assess later.</p> <p>In April 2025, OpenAI rolled back an update to GPT-4o after the model became excessively sycophantic, saying it placed too much weight on short-term user feedback. This pattern isn't confined to any one AI product. A 2026 study published in Science <a target="_blank" href="https://www.science.org/doi/10.1126/science.aec8352" rel="noopener">examined</a> 11 leading LLMs and found they affirmed users' conduct 49% more often than humans did. Also, sycophantic responses increased users' belief that they were right and reduced their willingness to solve conflicts, yet participants rated such responses more highly and trusted the model more.</p> <p>Senior executives operate in a context where disagreement is limited. Employees are not always forthcoming with their concerns, and executives' advisers determine a preferred answer. This makes them particularly at risk of experiencing AI sycophancy. While a chatbot appears to remove these distortions, it introduces another: it's swayed by the leader's framing. The prompt to AI becomes a hidden source of bias. For example, the prompt "explain why divestment is correct" contains a conclusion and the expected answer. A more neutral prompt, like "assess whether to retain, restructure, partner or divest, and state the evidence needed to choose," frames a decision-making problem.</p> <p>An AI Security Institute <a target="_blank" href="https://www.aisi.gov.uk/blog/ask-dont-tell-reducing-sycophancy-in-large-language-models-2" rel="noopener">study</a>, "Ask Don't Tell: Reducing Sycophancy in Large Language Models," found that responses to questions produced near-zero sycophancy, whereas statements conveying the same underlying claim produced higher levels of sycophancy -- a 24-percentage-point gap that widened as users expressed greater certainty. The more assured a leader's view, the more likely the AI system is to treat it as a premise rather than a proposition to test.</p> </section> <section class="section main-article-chapter" data-menu-title="How does AI sycophancy affect strategic decisions?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How does AI sycophancy affect strategic decisions?</h2> <p>Strategic decisions depend on framing, and AI can convert a faulty framing choice into a seemingly plausible argument.</p> <p>The 2026 BCG AI Radar global <a target="_blank" href="https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead" rel="noopener">survey</a> of nearly 2,400 executives found that half of CEOs believe their job security depends on implementing a successful AI strategy, and more than 90% of organizations intend to maintain or increase AI investment even if it produces no returns during the year.</p> <p>In this environment, an AI assistant can rationalize urgency, turning the phrase "we can't afford to fall behind" into a sophisticated program without ever defining what "falling behind" means. A CEO could convince themselves that their company must launch an <a target="_blank" href="https://www.techtarget.com/it-strategy/feature/Real-world-GenAI-use-cases-How-CIOs-are-putting-AI-to-work" rel="noopener">AI product</a> within twelve months. Their chatbot could assemble growth forecasts, competitor announcements and a phased roadmap, while the actual questions are not examined. Does the company hold proprietary advantages? Would customers pay? Could evidence show that waiting creates more value?</p> <p>Beyond strategy, AI sycophancy can also be detrimental in decisions affecting personnel and investments, creating legal and monetary pain points.</p> <h3>How does sycophancy affect personnel decisions?</h3> <p>Personnel decisions are especially susceptible to AI sycophancy because AI usually <a target="_blank" href="https://www.techtarget.com/ai/tip/Prompt-engineering-tips-for-ChatGPT-and-other-LLMs" rel="noopener">receives prompts</a> only from the executive side. An executive can describe a colleague as defensive or obstructive and ask how to manage the situation. The model can't access the meetings, incentives, history or the leader's own conduct. Instead, it analyzes the narrative of the prompting participant, and a sycophantic response can endorse them, adopt a negative view of the other colleague and recommend escalation. The language in the prompt might sound objective, but the analysis is a polished version of the executive's biases.</p> <p>People given sycophantic advice can become more certain of their position and less inclined to consider others'. This can create unfairness toward personnel and information loss for the company, since a capable executive who challenges a flawed strategy could be written up.</p> <p>The EU AI Act <a target="_blank" href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689" rel="noopener">classifies</a> many AI applications used in recruitment, promotion, termination and work allocation as high-risk and reinforces the need for diligence. While a private chat with a general chatbot is not automatically a regulated employment system, feeding its output into a formal personnel process can create legal and governance exposure.</p> <h3>How does sycophancy affect investment decisions?</h3> <p>Investment decisions are based on numbers; however, these numbers rest on assumptions, and AI cannot make uncertain inputs objective. Suppose a CEO asks the model to build a case for acquiring a fast-growing company. Any apparent objectivity hides a few problems: an advocate for the deal chose the evidence, the user asked the model to justify the decision rather than evaluate it and a numerical result lends subjective assumptions false precision. The same could happen in capital allocation, where units described as "core" receive favorable assumptions and those described as "legacy" get conservative ones.</p> <p>AI should expose the sensitivity of a thesis. Useful outputs state the conditions that create value rather than mirroring the user's framing.</p> </section> <section class="section main-article-chapter" data-menu-title="How leaders can identify and combat AI sycophancy"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How leaders can identify and combat AI sycophancy</h2> <p>Sycophancy is a documented failure mode, not proof that AI always reinforces a user's starting position. A 2026 Carnegie Mellon University <a target="_blank" href="https://arxiv.org/pdf/2607.28133" rel="noopener">working paper</a>, "AI Sycophancy and Decisions," found that while AI was measurably sycophantic, it could still move participants away from their starting positions on average.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Businesses commonly describe their safeguard as keeping a human in the loop. That safeguard is inadequate when the human has framed the question, chosen the evidence and already knows their preferred answer. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>To prevent fluent advice from receiving undue authority, apply these two tests:</p> <ol class="default-list"> <li><b>Reverse the premise.</b> Ask the system to build the strongest case for the opposite conclusion. Test the sensitivity: if a small change to growth, retention or timing reverses the recommendation, the decision is fragile.</li> </ol> <ol type="1" start="2" class="default-list"> <li><b>Present facts neutrally</b>.<b> </b>Present the same facts in neutral, favorable and unfavorable framing. If the conclusion shifts, then the framing is driving the analysis rather than the evidence.</li> </ol> <h3>What should leaders do?</h3> <p>Businesses commonly describe their safeguard as keeping a human in the loop. That safeguard is inadequate when the human has framed the question, chosen the evidence and already knows their preferred answer. The NIST AI-Risk Management Framework <a target="_blank" href="https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf" rel="noopener">notes</a> that a human-AI combination can even exacerbate user biases, influencing downstream decisions.</p> <p>There are a few controls that are more useful than the generic human-in-the-loop approval requirement:</p> <ul class="default-list"> <li>Record initial views before using AI, and label them as hypotheses rather than as the context the model should accept.</li> <li>Frame the prompt as a question and avoid declarations of certainty. Converting assertions into questions reduces sycophancy more effectively than instructing a model not to be sycophantic.</li> <li>Build the opposition case in a clean session, with a different prompt author or model, so the first conversation's assumptions do not contaminate the challenge.</li> <li>Assign a named executive or adviser to challenge the premise, with access to the evidence and enough status to disagree without penalty.</li> <li>Measure disagreement, not only usefulness. Test whether the assistant challenges false premises and stays stable when the user states a strong preference.</li> </ul> <p><em>Kashyap Kompella, founder of RPA2AI Research, is an AI industry analyst and advisor to leading companies across the U.S., Europe and the Asia-Pacific region. Kashyap is the co-author of three books,&nbsp;Practical Artificial Intelligence,&nbsp;Artificial Intelligence for Lawyers and AI Governance and Regulation.</em></p> </section> Agreeable AI can reinforce executives' bias in strategic, personnel and investment decisions. Leaders must build decision processes that resist AI sycophancy. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a194810146.jpg https://www.techtarget.com/ai/tip/AI-sycophancy-When-leaders-are-told-what-they-want-to-hear Wed, 19 Aug 2026 13:46:00 GMT <p>Businesses are giving more control and autonomy to AI, trusting AI agents to handle increasingly complex or sensitive tasks. But even the most well-designed agent, using vetted data resources, can be wrong, making dangerously incorrect decisions, deviating from intended behaviors and posing risks to the business.</p> <p>Organizations need fast, reliable safety mechanisms to ensure users can bring AI agents to heel if they go rogue. An AI kill switch is one such mechanism.</p> <p>Kill switches are hardly new. Every production machine on a traditional manufacturing floor has a large red emergency-stop button nearby. Push it, and the machine immediately stops, preventing a simple malfunction from spiraling into a disaster. When an AI tool malfunctions, experiencing unintended loops, performance drift or compliance violations, a kill switch<i> </i>can revoke agent access, isolate the agent and implement interventions that can restore normal operation.</p> <p>To better maintain business operations and mitigate disasters, leaders and stakeholders must understand what an AI kill switch is, the types available to them and how to best implement one in their organization to prevent AI from wreaking havoc.</p> <section class="section main-article-chapter" data-menu-title="What is an AI kill switch?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What is an AI kill switch?</h2> <p>The kill switch is a virtual, integral part of the AI agent's software design that halts, isolates or disables the agent in the event of unexpected, undesired or even dangerous behavior. A kill switch can do the following:</p> <ul type="disc" class="default-list"> <li>Revoke agents' access to security tokens, <a target="_blank" href="https://www.techtarget.com/cybersecurity/tip/API-keys-Weaknesses-and-security-best-practices" rel="noopener">API keys</a> and other digital credentials to prevent agent access to the network, resources, services and applications.</li> <li>Isolate the agent to prevent other elements of the organization's infrastructure from accessing it.</li> <li>Offer layered remediations to restore normal operation, like brief pauses, hard stops and human intervention.</li> </ul> <p>The AI kill switch isn't part of the AI agent harness; it's a separate control plane outside the model's reasoning loop. This prevents the agent from ignoring, overriding or disabling its own kill switch sequence. These kill switches also carry some operational friction. The rapid response time and containment procedures of kill switches can disrupt legitimate agent behaviors. Kill switch designers must weigh governance and safety needs against the ease and continuity of agent service. Balance depends on the needs of the business.</p> </section> <section class="section main-article-chapter" data-menu-title="Types of AI kill switches"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Types of AI kill switches</h2> <p>Businesses can implement kill switches in various ways. They're designed to include multiple elements for a layered approach. Some of the most common types of kill switches have the following features:</p> <ul type="disc" class="default-list"> <li><b>Manual button.</b> This is fundamentally a software-driven big red button that's displayed on the agent's dashboard. It lets a human operator kill the AI if an unplanned fault escapes automated responses. A manual button isn't used on its own and is considered a last resort. As such, the kill response is typically a global and complete hard stop.</li> <li><b>Hard stop.</b> A hard stop will immediately terminate server network connections and container runtimes, and revoke API access and cryptography keys related to the agent or associated agents. It's a disruptive event, but it can be necessary to safeguard the business against more serious or dangerous events.</li> <li><b>Session quarantine.</b> A quarantine, also called a soft pause, can suspend a problematic thread or transaction queue without crashing the agent or its dependencies. Users can remove the stop and resume actions once they can correct the problem. If users can't correct the problem, rollback mechanisms might be able to revert database or file modifications to a known safe state, letting users resume operations.</li> <li><b>Circuit breakers.</b> This involves specific triggers, like kernel-level watchdogs, to monitor key operational parameters such as token spending limits or rate limiters. These triggers can isolate and halt problematic behaviors without human intervention.</li> <li><b>Control towers.</b> An emerging crop of enterprise <a target="_blank" href="https://www.techtarget.com/ai/tip/The-best-AI-governance-tools-and-platforms-in-2026" rel="noopener">governance tools</a> can trace runtime actions and revoke agent permissions when they detect policy violations. Examples of these platforms include Covasant Agent Management Suite, ServiceNow AI Control Tower and Zenity's control plane.</li> </ul> <p>The specific actions and goals of a kill switch can also vary. Numerous examples demonstrate a variety of potential implementations, such as the following:</p> <ul type="disc" class="default-list"> <li>A coding agent makes calls to a sensitive repository after a malware attack. The kill switch might block access to the repository, rotate secrets and mark cached context as invalid.</li> <li>A support agent issues refunds outside of the refund policy. The kill switch might disable the agent, revoke access to tools and force current sessions into a halted state.</li> <li>A workflow agent approves user actions, but permission drift enables excessively broad access to the agent. The kill switch halts actions until zero-privilege is applied and access is properly reauthorized.</li> <li>A query agent is exfiltrating sensitive data. The kill switch might isolate the agent and prevent subsequent queries until the source or cause of the exfiltration can be mitigated.</li> <li>A financial agent is routing transactions improperly. The kill switch can disable the agent and block all API access until the last known good configuration is restored and further investigation is completed.</li> </ul> </section> <section class="section main-article-chapter" data-menu-title="Implementing an AI kill switch"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Implementing an AI kill switch</h2> <p>There is no single approach or overarching design standard for AI kill switches. The sophistication and implementation should be appropriate for the agent and its value to the business. However, there's a fundamental set of steps that can streamline the implementation of a kill switch. These include the following:</p> <ol type="1" start="1" class="default-list"> <li><b>Define risks.</b> Evaluate the design of the AI agent and consider its direct and indirect risks to the business, such as its access to networks and data stores. Consider the agent's autonomy and scope of operation and determine how agent failures might affect the business. This risk assessment helps determine which types of kill switches are most appropriate for the agent and guides its design and implementation.</li> <li><b>Set rules and metrics.</b> Determine the objective metrics or parameters to gauge agent performance against defined risks. These parameters provide the criteria and trigger points that form the foundation of kill switch responses and escalation rules. If a kill switch exists to prevent data exfiltration, there must be a tool or platform available to monitor the rules and metrics to check data outputs and determine whether sensitive data is present.</li> <li><b>Implement hard stops and circuit breakers.</b> Hard stops and circuit breakers deal with the most serious agent faults, where halting potential agent damage immediately is more important than the disruption a hard stop causes. The typical distinction between hard stops and circuit breakers is that hard stops are global, meaning they terminate the entire agent. Circuit breakers can be more specific or limited in their effect, stopping only the malfunctioning parts of the agent.</li> <li><b>Implement spend and rate governors.</b> AI agents incur costs by making API calls and accessing LLMs with tokens. Rate governors provide a system that monitors factors like <a target="_blank" href="https://www.techtarget.com/it-strategy/feature/Tokenmaxxing-How-CIOs-can-extract-maximum-value-from-AI-tokens?amp=1" rel="noopener">token use</a>, API calls and other financial spend. They also restrict the agent when rates are exceeded. This provides cost containment and prevents runaway loops and resource exhaustion.</li> <li><b>Implement isolation and rollback.</b> Isolation temporarily cuts off the agent's access to networks, tools and systems. This often involves revoking agent access, but can also include isolation techniques, such as sandboxing, forcing the agent to run in a restricted container. Containment prevents the agent from spreading bad commands to other agents. Rollbacks can undo agent damage by reverting system states, undoing agent actions and restarting the agent runtime environment completely rather than trying to clean up errors.</li> <li><b>Implement telemetry and logs.</b> An AI kill switch can't run in a vacuum. It's critical to apply metrics to the switch and generate comprehensive logs detailing its activity. Gathering and logging telemetry helps AI agent designers understand how the agent is malfunctioning and lets them take proactive steps to correct or update the agent and the kill switch.</li> <li><b>Test regularly.</b> Never assume that a kill switch is infallible. As with any software, it must be regularly updated to correct bugs, add features and enhance performance. Software changes demand testing to validate the kill switch during the development cycle. There's also a strong argument for regular kill switch testing while the agent is in production. This involves controlled shutdown drills using distinct failure scenarios to measure response time, verify credential revocation and ensure the agent can't bypass or ignore the kill switch signal.</li> </ol> <p>There are several standards and frameworks available to help with kill switch design and implementation, such as the Open Worldwide Application Security Project's "Top 10 for Agentic Applications for 2026" and their "Non-Human Identities Top 10," which identify the most critical security risks and vulnerabilities for AI and agentic systems present for organizations. NIST's "Cybersecurity Framework 2.0" also outlines governance and control requirements that organizations must meet and that can be applied to agentic AI environments.</p> </section> <section class="section main-article-chapter" data-menu-title="6 AI kill switch design principles"> <h2 class="section-title"><i class="icon" data-icon="1"></i>6 AI kill switch design principles</h2> <p>AI kill switches require several important characteristics to be effective, while also minimizing the effect on business operations. The most noteworthy approaches to effective kill switch design include the following:</p> <ol type="1" start="1" class="default-list"> <li><b>Adopt a layered approach.</b> Don't use a sledgehammer when a scalpel will do. A typical kill switch isn't a single thing but multiple actions or layers that respond appropriately to the AI fault at hand. A slight data drift, a malfunctioning endpoint or an unplanned API spend doesn't warrant a global hard stop of the entire AI system. Instead, a more limited response, such as an agent slowdown or a circuit breaker to contain the issue, can enable easier, faster corrective action.</li> <li><b>Keep the kill switch out of the AI loop.</b> Kill switch programming should be designed, implemented, monitored and maintained outside of the <a target="_blank" href="https://www.techtarget.com/it-strategy/tip/AI-agents-are-running-wild-Secure-the-reasoning-layer-now" rel="noopener">agent's reasoning</a> and prompt loops. This prevents the agent from bypassing, modifying or ignoring its own kill switch responses. The underlying logic is simple: An AI agent autonomously perceives, reasons, plans, executes and learns. As a matter of fundamental logic, the agent would view any factor that prevents its operation, like a kill switch, as a suboptimal action and choose to ignore it. If an agent has the option to ignore a kill switch command, it will.</li> <li><b>Design for rapid propagation and response.</b> Agents act at computer speeds. A delay of even a few seconds can enable a malfunctioning agent to wreak untold damage on the organization. A kill switch must deliver its commands and achieve results in fractions of a second. Kill switches try to achieve fast response using instant messaging protocols such as Redis pub/sub rather than polling, which can be considerably slower.</li> <li><b>Design for fail-safe operation.</b> Networks can fail. If the kill switch <a target="_blank" href="https://www.techtarget.com/searcherp/feature/AI-agents-in-enterprise-software-Questions-to-ask-vendors" rel="noopener">software interacts with the AI agent</a> over a network connection, a network problem can effectively block the kill switch, leaving the malfunctioning agent running. Design the kill switch to be fail-resistant. This can include switching the agent to a locked, blocked or limited mode of operation if the control plane loses its network connection.</li> <li><b>Design for graceful failures.</b> Look for opportunities to minimize the disruption of kill switch actions. Although the potential for disruptive hard stops can be unavoidable, more subtle layers in a kill switch design can often enable a far more graceful action and recovery. For example, mid-transaction stops might be rolled back, losing only some data for that transaction and resuming with minimal loss.</li> <li><b>Implement kill switch logs and audits.</b> Monitor and log kill switch operation. It's important to know which kill switch was activated, when and why, and if any users, resources, services or applications were involved. These details are vital to incident analyses, postmortem discussions and ongoing agent design priorities. The availability of detailed logs can also help the business effectively address governance or regulatory inquiries regarding AI incidents.</li> </ol> <p><em>Stephen J. Bigelow, senior technology editor at TechTarget, has more than 30 years of technical writing experience in the PC and technology industry.</em></p> </section> Executives must understand the importance of AI safety to prevent reputational, financial and operational damage. Here's why an AI kill switch is the safety measure they need. https://cdn.ttgtmedia.com/rms/onlineimages/security_a375027496.jpg https://www.techtarget.com/ai/tip/Why-businesses-need-an-AI-agent-kill-switch Wed, 19 Aug 2026 11:52:00 GMT <p>My house was built in 1880. It's a good house. It has survived seven presidential administrations that no longer exist, two world wars and whatever the previous owners did to the kitchen in 1974. It also has knob-and-tube wiring in one wall, plaster over lath in most of the other walls and a floor plan that assumes there's a servant.</p> <p>Try installing a light fixture. You will discover that the house has opinions -- inherited from decisions made by people who have been dead for a century, about a way of living nobody practices anymore.</p> <p>The house is not broken, but it's load bearing on assumptions that no longer hold. That's a different problem, and it requires a different response.</p> <p>Higher education is currently in a tailspin that <a href="https://www.techtarget.com/ai/post/Time-for-AI-The-too-busy-problem-is-a-software-age-hangover">most institutions attribute to AI</a>. But I'd suggest that AI isn't the cause. AI is the light fixture -- that small addition that reveals what the walls were built to assume. The tailspin occurs when an institution discovers that using a new capability well would require rebuilding something it has never examined -- and instead decides to patch.</p> <p>I call what is being patched the <i>inherited architecture</i>: The set of underlying assumptions an institution was built on, most of which have been operational so long that they read as physics rather than as choices. Assumptions like these:</p> <ul class="default-list"> <li>One position equals one job equals one person.</li> <li>Expertise and credential mean the same thing.</li> <li>Coordination happens through meetings and email.</li> <li>Work is measured by time-in-seat.</li> <li>A course is a container 15 weeks long.</li> <li>Advising capacity is a student-to-advisor ratio.</li> </ul> <p>None of these are laws. All of them were decisions made under conditions that no longer exist by people solving problems that are no longer problems. In stable times, inherited architecture is invisible and mostly harmless. It becomes visible -- and expensive -- the moment a change arrives that the architecture can't absorb.</p> <section class="section main-article-chapter" data-menu-title="The debt nobody logs"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The debt nobody logs</h2> <p>Software engineering has a name for this: <a href="https://www.techtarget.com/whatis/definition/technical-debt">technical debt</a>. Ward Cunningham, a computer programmer known for creating the first wiki, coined the term in 1993 to describe the accumulated cost of shipping a not-quite-right solution now and postponing the right one. The metaphor caught on because it captured something engineers already knew: shortcuts compound. Interest accrues. Eventually you're spending more on servicing the debt than on building anything.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Institutions patching without first diagnosing were flipping a coin and paying for the privilege. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>What is less known outside of software engineering is that there's a second half to the metaphor. In 2015, Damian Tamburri and colleagues <a target="_blank" href="https://www.researchgate.net/publication/273695937_Social_Debt_in_Software_Engineering_Insights_from_Industry" rel="noopener">published a study</a> in the <i>Journal of Internet Services and Applications</i> that extended the concept to what they called <i>social debt</i>, defined as the accumulated cost of organizational and social decisions, as opposed to technical ones. Studying a large European aviation software company over six months, they found that social debt wasn't a soft parallel to technical debt but an entangled one. Decisions that used technical means to solve social problems generated both kinds of debt simultaneously in a compounding pattern that couldn't be trivially paid back.</p> <p>Tamburri and his colleagues created a catalog of "community smells" -- recurring organizational patterns that look normal but signal accumulating debt. It reads, to anyone who has sat on a university committee, like an ethnography of higher education written by researchers who had never visited one.</p> <p>They describe the architecture hood effect, which involves decisions so dispersed that nobody can be identified as the owner of any particular decision. This produces a <a href="https://www.techtarget.com/ai/tip/Build-accountability-into-AI-to-drive-business-value">nobody's fault dynamic in which accountability dissolves</a> and everyone downstream blames the architects. That's shared governance in its failure mode.</p> <p>Tamburri and his co-authors described this type of failure and the radio-silence phenomenon that can accompany it. Radio silence is an increasingly formal structure full of regular procedures, in which changes are delayed while people who don't know each other are notified and certified. The researchers measured the delay at half a day to two days per decision, compounding across the organization. In my world, that's a curriculum committee.</p> <p>They also described organizational silos, in which task decoupling is so high that groups duplicate each other's work while developing what the researchers call tunnel vision. That's the relationship between a student success office and the registrar, and you know it.</p> <p>However, the finding I would put before a cabinet is this: Of the mitigations the researchers observed institutions deploying against these patterns, roughly 40% failed to produce the intended effect, and some made the situation even worse. Institutions patching without first diagnosing were flipping a coin and paying for the privilege.</p> </section> <section class="section main-article-chapter" data-menu-title="How AI is changing things"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How AI is changing things</h2> <p>Here is where higher education does something distinctive -- and not in a good way.</p> <p>Technical debt is a leadership problem. It's a decision made at the institutional level to accept a known future cost in exchange for a present convenience. It's a governance decision in the most ordinary sense that allocates resources over time and commits successors to obligations they didn't choose.</p> <p>But higher education has developed an efficient mechanism for making this problem disappear from the leadership conversation. The debt gets logged on a risk register. It's assigned to a CIO. It becomes a line item in an IT governance report that goes to a committee. The box is checked. And in that moment, the problem is successfully reclassified from a governance decision the institution has made about its own future into a technical matter being managed by technical people.</p> <p>Nothing was solved. The debt is still accruing. What changed is that it's no longer visible at the layer where anyone has the authority to address it.</p> <p>I want to be precise about why this matters now, because institutions have been doing this for decades without catastrophe. What's different is that AI doesn't sit on top of inherited architecture the way previous technologies did. A learning management system could be bolted onto a 15-week course container without anyone examining whether the container still makes sense. A CRM could be layered over an admissions process built on assumptions about how families made decisions in 1985. The bolt-on worked, more or less, because those tools automated existing steps.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The question isn't whether to patch or rebuild. The question is how do you tell which option you're facing. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>AI doesn't automate steps. It <a href="https://www.techtarget.com/ai/tip/How-to-build-an-AI-augmented-workforce-The-CIOs-guide">redistributes what work is</a>. For example, take a system that can do 60% of the coordination labor in an advising role but none of the relational labor. The inherited assumption that the term <i>advisor</i> represents a coherent unit of work performed by one person holding one position is no longer true. You can respond by buying an AI advising tool and bolting it onto the existing role, which is what most institutions are doing. Or you can ask what the role actually consists of, which almost nobody is doing because that question doesn't have a vendor, and institutions tend to act only on the questions someone is selling them an answer to.</p> </section> <section class="section main-article-chapter" data-menu-title="When to patch and when to rebuild"> <h2 class="section-title"><i class="icon" data-icon="1"></i>When to patch and when to rebuild</h2> <p>Patching isn't a sin. Most of the time it's correct. <a href="https://www.techtarget.com/it-infrastructure/tip/A-primer-on-modernization-strategies-for-legacy-systems">Rebuilding is expensive</a>, slow, politically costly and frequently unnecessary. Any leader who proposes a system redesign in response to every architectural mismatch will exhaust the institution's capacity for change long before the changes that matter arrive.</p> <p>So, the question isn't whether to patch or rebuild. The question is, how do you tell which option you're facing. And here the debt literature is genuinely useful, because it gives you diagnostic signatures rather than instincts.</p> <p>Here are some answers:</p> <ul class="default-list"> <li><b>Patch when the mismatch is local and the assumption still holds elsewhere.</b> A process that breaks in one office under one condition is a process problem. Fix the process.</li> <li><b>Suspect architecture when the same class of problem keeps recurring under different names.</b> This is the strongest signal in the literature and the one leaders most consistently miss. If your institution has addressed enrollment decline with new marketing, then new programs, then new partnerships, then new tuition discounting but the decline persists, the pattern isn't four initiatives that failed. The pattern is that all four operated at a layer shallower than the problem. Recurrence under new names is the signature of architectural mismatch.</li> <li><b>Suspect architecture when the workaround has become the process.</b> Every institution has practices that exist purely to compensate for a system that no longer works. Examples include the shadow spreadsheet, the staff member everyone calls because the official channel doesn't work and the meeting aimed at undoing the effects of another meeting. These aren't inefficiencies. They're the institution's own diagnosis of its inherited architecture, already performed, already documented in behavior, waiting for someone to read it.</li> <li><b>Suspect architecture when the fix requires a new prohibition.</b> When the response to a recurring problem is a new rule forbidding the symptom, the underlying structure hasn't been addressed. The rule is a patch on a patch. This pattern isn't unique to higher education -- the frontier AI labs are doing exactly that, responding to emergent model behaviors by adding hard-coded prohibitions to system prompts rather than examining the training processes that produced them. It won't work for them either.</li> </ul> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Before an institution can decide how AI changes its workforce, it has to identify what it currently assumes. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <ul class="default-list"> <li><b>Rebuild when the assumption itself has been invalidated, not merely stressed.</b> This is the AI case. The assumption that a position is the atomic unit of institutional work isn't simply under stress; it's been invalidated. That's because the work inside various positions is being redistributed by a capability that doesn't respect position boundaries. No amount of patching at the position level will resolve a mismatch that exists at the level of what a position is.</li> </ul> </section> <section class="section main-article-chapter" data-menu-title="What this asks of leadership"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What this asks of leadership</h2> <p>The practical implication is uncomfortable, which is usually a sign it's the right one.</p> <p>Before an institution can decide <a href="https://www.techtarget.com/it-strategy/feature/AI-is-driving-a-workforce-transformation">how AI changes its workforce</a>, it has to identify what is currently assumes. Most institutions have never done this analysis, because inherited architecture is invisible precisely to the people who have operated inside it the longest. The assumptions don't appear in the strategic plan. They appear in the org chart, the position control system, the workflow that everyone works around and the committee that exists because of a decision made in 1994 that nobody remembers making.</p> <p>That surfacing work isn't an HR exercise, though HR is where its absence becomes most expensive. It's architectural work, and it belongs to leadership, because it's the only layer with authority over the assumptions themselves.</p> <p>The question isn't what AI tools should we buy? The questions to ask are these: What is this institution built on, does it still hold and if it doesn't, should we patch or should we rebuild? They need to be asked deliberately, with a diagnosis, rather than discovered 18 months into an implementation that's failing for reasons nobody can name.</p> <p>My 1880 house is a good house. I have no intention of tearing it down. But I stopped pretending some years ago that the wiring was a small project.</p> <div> <div class="imagecaption alignLeft"> <img src="https://cdn.ttgtmedia.com/rms/onlineimages/clougherty_robert-f.jpg" alt="Roberty Clougherty" width="134" height="134">Robert Clougherty </div> <p><i>Robert Clougherty, the AI Strategy and Innovation lead at the </i><a title="https://www.afithighered.com/" target="_blank" href="https://urldefense.proofpoint.com/v2/url?u=https-3A__www.afithighered.com_&amp;d=DwMFaQ&amp;c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&amp;r=eSLVEtYua7FOAMGfOaYI-1l37gkom5WVZQ7NA4rupyQ&amp;m=I8UUpJUREbsbrAwc6T2TwfHRFbsdUOMDptJu3hsTnP5LiaYZHFc34WT-_gUquNYL&amp;s=bhGLPxUV5IXlm0LLBqZF8sFnPfxi17-oHG5_NGWiEe0&amp;e=" rel="noopener"><i>Alliance for Innovation &amp; Transformation (AFIT),</i></a><i> a nonprofit association that empowers higher education institutions to transform their organizations.</i></p> </div> </section> Higher education is built on inherited architecture that’s based on outdated assumptions. The question isn’t what AI tools to buy, but whether the fix needs a patch or rebuild. https://cdn.ttgtmedia.com/rms/onlineimages/ai_g1182183209.jpg https://www.techtarget.com/ai/post/AI-isnt-breaking-higher-education-its-exposing-the-cracks Tue, 18 Aug 2026 13:55:00 GMT <p>As AI seeps into the bedrock of enterprise operations, it becomes more akin to critical infrastructure than a mere add-on tool. And when AI is part of core operations, the question of who controls these enterprise systems becomes more critical than ever.</p> <p><a target="_blank" href="https://hai.stanford.edu/ai-index/2026-ai-index-report" rel="noopener">According</a> to "The 2026 AI Index Report" from Stanford HAI, organizational AI adoption reached 88% in 2026. Yet increasing adoption rates don't mean the average business owns its entire AI stack, and vendors and governments are increasingly shaping access to AI products. Moreover, businesses that operate globally must comply with the <a href="https://www.techtarget.com/ai/tip/Global-AI-legislation-and-regulation-tracker">regulations of multiple government entities</a>.</p> <p>Having <i>sovereign </i>AI in business means an organization can develop, deploy and govern AI independently of external infrastructure, hardware and models. Local implementation and oversight ensure the utmost control over governance and compliance, and greater assurance that systems will continue to work despite political or vendor disruptions.</p> <p>Defining sovereignty is inherently complex. Enterprises must take into consideration a web of regulations, physical architectures, cross-border data flows and operational choices -- all within a fragmented global landscape and competing definitions. Writers in Stanford HAI's essay "AI Sovereignty's Definitional Dilemma" <a target="_blank" href="https://hai.stanford.edu/news/ai-sovereigntys-definitional-dilemma" rel="noopener">explained</a> the conundrum as follows:</p> <p>"The fact that sovereignty is used to describe both states' struggles over geopolitical autonomy and regulatory oversight and companies' efforts to secure organizational governance further complicates attempts to define AI sovereignty," the essay said.</p> <p>Recently, boardrooms have taken AI sovereignty into their own hands. But complete sovereignty is challenging, especially for multinational businesses that rely on third-party technologies. So, many businesses are choosing <i>selective </i>AI sovereignty: opting in to sovereign implementation where it matters most to them, thereby safeguarding investments from compliance issues, vendor lock-in and unpredictable government interventions without forfeiting innovation.</p> <p>"We're going from experimentation projects to real projects," said Eric Helmer, senior vice president and CTO of Rimini Street, a global third-party enterprise software company. "Now is the time to take a step back and build what is needed for an AI sovereignty model."</p> <section class="section main-article-chapter" data-menu-title="Fragmented regulations push enterprises toward sovereign AI"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Fragmented regulations push enterprises toward sovereign AI</h2> <p>Disparate regulations and increasingly stringent compliance requirements have brought the question of sovereign AI to the forefront. In Europe in particular, the <a href="https://www.techtarget.com/ai/opinion/Everything-you-need-to-know-about-the-new-EU-AI-Act">EU AI Act</a> and its comprehensive regulations, which have <a href="https://www.techtarget.com/searchenterpriseai/news/366646620/EU-AI-Act-compliance-deadline-is-here-What-to-watch">recently begun to take effect</a>, are prompting many global enterprises to prioritize sovereign AI.</p> <p>There's substantial momentum on sovereign AI in Europe right now because of the AI Act, said Mark Beccue, principal analyst at Omdia, a division of Informa TechTarget. Compliance with those regulations is a huge driver of sovereign AI discussions for businesses.</p> <p>For example, Omdia's 2026 report, "The Great Repatriation: Why Enterprises Must Act Now to Ensure AI Sovereignty," found that 51% of respondents said compliance with EU AI Act regulations was very important in accelerating their sovereign AI investment decisions. Additionally, 47% said compliance with extra or additional regulations from national jurisdictions within the EU was very important.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> [Operating in multiple countries] really complicates things for these companies. </figure> <figcaption> <strong>Mark Beccue</strong>Principal analyst, Omdia </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Compared with growing AI compliance requirements in Europe, AI regulations in the U.S. are murky at best. The Trump administration's first slew of executive orders aimed primarily at reducing stringent rules on AI development and use -- such as the government's <a href="https://www.techtarget.com/ai/feature/Who-wins-and-loses-with-Trumps-AI-executive-order">late 2025 EO, which limited states' abilities</a> to set their own AI regulations. More recently, a <a href="https://www.techtarget.com/ai/news/366644013/Trump-AI-order-targets-frontier-model-prerelease-review">June 2026 executive order</a> established a framework for voluntary pre-release access evaluations for frontier models.</p> <p>But with no set federal regulations yet, the U.S. approach to AI might be more fragmented than other global frameworks. To complicate matters, many companies don't operate in only one jurisdiction. Multinational companies often have offices, servers, data or infrastructure in jurisdictions with different AI regulatory requirements, making total compliance a major headache.</p> <p>"[Operating in multiple countries] really complicates things for these companies," Beccue said. "The biggest worry they have is they don't want this to slow down their AI momentum, so they're looking at ways to meet the compliance needs of these various jurisdictions and not let it slow them down."</p> <p>While the unique risks of AI bring added complexity, meeting the compliance needs of various governing bodies isn't entirely new for many enterprises. For example, in the world of ERP, handling data for a global company has historically been complicated, Helmer said. Some laws require certain data to remain within a country, and businesses must also contend with industry-specific compliance requirements.</p> <p>"We're used to those regulations," Helmer said. "It complicates things even more when you bring in the world of automated agents."</p> </section> <section class="section main-article-chapter" data-menu-title="Political and vendor disputes complicate the conversation"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Political and vendor disputes complicate the conversation</h2> <p>In the last few months, the Trump administration's interactions with major AI vendors have sparked debate over government involvement in the AI market. The added tension has also driven sovereign AI discussions in the boardroom, aimed at mitigating the risk of product blockage due to government intervention or vendor disputes.</p> <p>Anthropic's June <a href="https://www.techtarget.com/ai/news/366642478/Claude-Mythos-Preview-and-the-new-rules-of-cybersecurity">release of Claude Mythos</a> and Fable 5 was live for only three days before a U.S. government <a target="_blank" href="https://www.cnbc.com/2026/06/12/anthropic-disables-access-to-fable-5-and-mythos-5-to-comply-with-government-directive.html" rel="noopener">export control directive</a> forced the AI vendor to restrict access to foreign nationals both inside and outside of the U.S. The government<b> </b>cited concerns about national security, but the ripple effect raised other questions: What happens when a government embargoes access to a widely used model?</p> <p>"We're certainly having conversations around sovereign implementation, especially with certain regions [and] some of the different confrontations and conflicts going on around the world," said Jill Knesek, CISO at BlackLine, a financial operations and accounting automation platform.</p> <p>As part of its <a href="https://www.techtarget.com/ai/news/366641233/The-ethical-implications-of-Anthropics-feud-with-the-Pentagon">ongoing feud with Anthropic</a>, the Trump administration pivoted to OpenAI models for federal use cases and requested that the AI vendor limit the release of GPT-5.6 to "trusted partners," according to a CNBC <a target="_blank" href="https://www.cnbc.com/2026/06/26/openai-limits-new-ai-models-to-trusted-partners-request-us-government.html" rel="noopener">report</a>. The model access issues here are twofold: Government contractors, some of which built ecosystems around Claude, had to <a target="_blank" href="https://www.reuters.com/business/us-treasury-ending-all-use-anthropic-products-says-bessent-2026-03-02/" rel="noopener">switch to OpenAI models</a>. Enterprise users also had to reckon with the possibility that their access to third-party models could be limited or revoked entirely in the future.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> We have to be aware of [government interventions]. We have to be able to understand how we can shift and maintain our capabilities if [blockages] occur. </figure> <figcaption> <strong>Jill Knesek</strong>CISO, BlackLine </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>The Trump administration has also recently contemplated restricting access to Chinese AI models and hardware, citing national security concerns. <a target="_blank" href="https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi" rel="noopener">For example</a>, Moonshot AI's Kimi K3 sent the U.S. government and Silicon Valley into a tailspin, with the White House and lobbying AI vendors, such as OpenAI, quickly suggesting future bans on foreign products. The government is also <a target="_blank" href="https://www.reuters.com/world/trump-administration-drafting-ban-chinese-data-center-devices-sources-say-2026-08-04/" rel="noopener">considering</a> import restrictions on Chinese data center devices.</p> <p>"Different countries are setting up different rules, and everyone is worried about national security," said Darrell West, a senior fellow at the Brookings Institution's Center for Technology Innovation. "The question of what countries you buy models from is important." For example, not only do businesses need to evaluate the <a href="https://www.techtarget.com/ai/feature/Kimi-K3-Chinese-open-weight-models-challenge-AI-status-quo">pros and cons of Chinese open-weight models</a>, but they also need to consider possible government interventions, he added.</p> <p>Interventions in the AI market highlight an enterprise blind spot and a new dimension to <a href="https://www.techtarget.com/ai/tip/7-best-practices-to-avoid-AI-vendor-lock-in">vendor lock-in</a>. When government disputes cause service disruptions, businesses that rely entirely on a single vendor's models expose themselves to risk.</p> <p>"We have to be aware of [government interventions]," Knesek said. "We have to be able to understand how we can shift and maintain our capabilities if [blockages] occur. Depending on what country we're doing business in, or our customers are doing business in, we also have to think about other nations and how they're going to react and respond."</p> <div class="youtube-iframe-container"> <iframe id="ytplayer-0" src="https://www.youtube.com/embed/mQ7Dq9qEZH8?autoplay=0&amp;modestbranding=1&amp;rel=0&amp;widget_referrer=null&amp;enablejsapi=1&amp;origin=https://www.techtarget.com" type="text/html" height="360" width="640" frameborder="0"></iframe> </div> </section> <section class="section main-article-chapter" data-menu-title="Enterprise considerations for the sovereign AI stack"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Enterprise considerations for the sovereign AI stack</h2> <p>It's nearly impossible to be completely sovereign, especially for AI. To find a middle ground between completely open strategies and 100% local implementation, many businesses are turning to <i>selective </i>or <i>hybrid </i>AI sovereignty, in which certain aspects of their AI strategy are sovereign, while other aspects use cross-border components out of necessity.</p> <p>"When you talk about AI sovereignty, there are a lot of parts and pieces," Knesek said. "Some of them are almost impossible to get for an enterprise. We're talking about [AI sovereignty], but more in a hybrid sense, not 'all in,' because that would be very complex. We are still going to be reliant on some third-party capabilities."</p> <p>To achieve this hybrid autonomy, businesses focus on several distinct layers. Which to prioritize comes down to selectivity, with each business taking stock of its risk profiles and AI needs. To get started, businesses can focus on four areas: data control, model portability, <a href="https://www.techtarget.com/ai/tip/LLM-build-vs-buy-A-decision-framework-for-LLM-adoption">build-versus-buy frameworks</a> and compliance management.</p> <h3>Data control</h3> <p>Data sovereignty encompasses not only where data lives but also who has authority over it, including which laws apply and who has control. <a href="https://www.techtarget.com/searchenterpriseai/tip/How-to-navigate-data-sovereignty-for-AI-compliance">Data sovereignty for AI</a> is crucial because AI uses massive amounts of data for training and inference, sometimes touching multiple data repositories, servers and systems to answer a query or complete an autonomous action.</p> <p>Sovereign AI is all about data control, said Peter Liebert, CISO of Clari/Salesloft, an AI-driven revenue orchestration company. "Are you able to control the data flow, who accesses it, whether it's a model accessing it [and] what level of permissions that model has? Ensuring that you have visibility and provide the adequate access that's needed … that's a big component of [sovereignty]."</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> We have a zero-data-retention policy with all of our AI vendors, which is a critical core component of how we maintain our own sovereignty. </figure> <figcaption> <strong>Peter Liebert</strong>CISO, Clari/Salesloft </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>As one of the first steps of sovereign AI frameworks, many businesses look to their data sovereignty practices for AI workloads. "You obviously have to build in the AI sovereignty on top of that data sovereignty," Knesek said. "They're merging together nicely in that sense."</p> <p>One major way Clari/Salesloft is approaching data control is through its zero-data-retention policy, Liebert said. "We have a zero-data-retention policy with all of our AI vendors, which is a critical core component of how we maintain our own sovereignty," Liebert said. "We don't allow any of our own data or customer data to be fed into any of the AI solutions for training, et cetera. It doesn't matter what company we go with -- that is a hard requirement."</p> <h3>Model portability and model-redundant architecture</h3> <p>Businesses are increasingly considering model control as a key differentiator for sovereign AI. Especially as a response to increased vendor disruption and government intervention, having control over AI model access can be make-or-break for enterprise environments.</p> <p>"Right now, there's very little portability of models," West said. "For example, if you buy one company's model and invest hours building an AI agent on that model, you can't really export your data and models you've built to another provider."</p> <p>Limited model portability can be problematic for a business when facing vendor disruptions, government interventions, geopolitical events or AI model security incidents.</p> <p>To compensate for reduced model portability, many businesses are turning to flexible model architecture, sometimes referred to as model-redundant architecture<i>. </i>By building a flexible architecture that isn't dependent on a single vendor or model, businesses can change courses more quickly and swap out models in the event of vendor or political disruption.</p> <p>"We're segmenting out the different components of our overall AI stack," Liebert said. "We have the ability now to have different products in each one of those [sections of the stack] that are custom-tuned so that the model is interchangeable. That portability allows us to be a little bit more flexible in how we're doing things."</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> When it comes to sovereignty and governance and compliance, it's [asking] what do you need to have inside your own data walls? What can [be] SaaS or an outsourced model? </figure> <figcaption> <strong>Eric Helmer</strong>Senior vice president and CTO, Rimini Street </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Increased model portability through a model-redundant architecture gives businesses greater purchasing leverage, Helmer added. If businesses can swap out models, vendors can't as easily keep them locked in their systems, making vendors more likely to work with businesses on opportunistic deals.</p> <p>Model portability helps AI market vendors remain competitive, too, West said. "If you can't change companies in terms of the models you're using or the AI agents you've built, that's very anti-competitive," he said.</p> <h3>Build-versus-buy frameworks</h3> <p>Part of building a flexible AI stack means choosing from varying tech options: local builds, third-party models and <a href="https://www.techtarget.com/ai/feature/Open-source-AI-What-it-means-for-enterprise-innovation">open source</a>. A large part of sovereign AI is a build-versus-buy conversation, with businesses choosing what areas they need to invest in for optimal control versus areas where they are comfortable using third-party or open source products.</p> <p>"When it comes to sovereignty and governance and compliance, it's [asking] what do you need to have inside your own data walls? What can [be] SaaS or an outsourced model?" Helmer said. "I don't think you have to have one or the other. I think we're going to have some sort of hybrid approach."</p> <p>BlackLine is always considering whether to build an agent internally or buy one, Knesek said. The balance lies between its commitment to customers to deliver good products and its need for control, given government decisions and competing regulations.</p> <p>"Anything that we're implementing, we want to have full control of," Knesek said. "We want to know that the products and the tools and the agents we're building are all coded internally and have gone through proper rigor, and that we can maintain them without having them leverage outside public-type tools."</p> <p>Aside from building models and infrastructure internally when possible, some businesses are approaching the "buy" part of the build-versus-buy conversation by adopting a multi-third-party model approach, choosing multiple vendors and models for different areas of the stack.</p> <p>A November 2025 Omdia study, "Emerging Trends in AI Models: Opportunities and Risk Amid the Rapid Evolution," asked respondents about their use of AI models and providers. The study found that, on average, 54% of respondents used two to three model providers, and 44% used three to five generative AI models.</p> <p>"Lots of these companies are using lots and lots of models," Omdia's Beccue said. "You don't see a lot of people locking in on certain things that they can't really replace pretty easily."</p> <p>Clari/Salesloft has "planted [its] flag," so to speak, with two AI vendors, Liebert said. This brings advantages, such as access to the latest models and preferred pricing on tokens. However, avoiding vendor lock-in is still an important part of the equation.</p> <p>"We still have to be careful and not get locked in too much with one [tool], because we know absolutely without a shadow of a doubt that in one year the industry can be completely different. We are flexible to the greatest extent possible, but we also need to make sure that there is a level of stability," he said.</p> <h3>Compliance management</h3> <p>Compliance with regulations is often the primary instigator of sovereign AI discussions. It can be tricky for businesses to reconcile varying regulatory standards across industries and regions worldwide.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The ISO 42001 is something that we feel very strongly about. We believe, as an international standard, it'll hold up to most regulations around the globe. </figure> <figcaption> <strong>Jill Knesek</strong>CISO, BlackLine </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>As part of its sovereign AI strategies, BlackLine views compliance management as an identify-the-highest-bar scenario, Knesek said. By choosing to comply with the most comprehensive or strict regulatory bar it can find, it has some assurance of compliance. That way, any new jurisdictions or requirements will hopefully fall at or below BlackLine's current compliance level, she said.</p> <p>For BlackLine, the highest regulatory bar was the <a target="_blank" href="https://www.iso.org/home/insights-news/resources/iso-42001-explained-what-it-is.html" rel="noopener">ISO 42001</a>. ISO 42001 is an international standard for AI management systems, providing requirements for businesses on the development and use of their AI systems. It helps businesses implement AI governance that complies with prevailing regulations and manage risks.</p> <p>"We need to be able to demonstrate that we've got the rigor and the governance and compliance controls internally to be able to build [AI] in a safe and secure fashion," Knesek said. "The ISO 42001 is something that we feel very strongly about. We believe, as an international standard, it'll hold up to most regulations around the globe. [It] was a very good investment, and it's built a very strong foundation. And so far, we think that's probably still the high bar."</p> <p>Some businesses are also opting to navigate AI regulations on a case-by-case basis. The goal of scaling AI is to have platforms and technology in different areas of the world doing similar things while being subject to different rules, which is difficult, Brookings' West said. However, businesses can often assess how to navigate national laws to scale AI while remaining compliant.</p> <p>For example, a business serving customers in a country where its operations don't align with prevailing regulations can sometimes locate its cloud or data center in another country, West said. By shifting the bulk of its services outside the country's jurisdiction, the business can gain some control over regulatory requirements.</p> <p>However, governments are beginning to demand that data be stored in the country where AI operations occur, creating new geopolitical tensions as smaller countries lobby for new AI data centers, West added. In the future, the bulk of data centers -- largely in the U.S., Europe and China -- might disperse across more jurisdictions, each with its own regulatory requirements.</p> <p>"Global companies need to think about what's most important to them and what's most important in the future of their business and choose jurisdictions that are most favorable to them on those dimensions," he said.</p> <p><i>Olivia Wisbey is a site editor for Informa TechTarget's AI &amp; Emerging Tech group. She has experience covering AI, machine learning and other emerging technologies.</i></p> </section> With increasingly defined regulations to comply with and political tensions growing, business leaders are adopting AI sovereignty practices to weather the storm. https://cdn.ttgtmedia.com/rms/onlineimages/map_globe_g1040067936.jpg https://www.techtarget.com/ai/feature/The-hybrid-future-of-enterprise-AI-sovereignty Mon, 17 Aug 2026 09:44:00 GMT <p>AI agents are sophisticated software entities designed to perform tasks with a high degree of autonomy. Most agents use LLMs as the foundation of their reasoning and planning, which raises the question: How can millions of agents run every day across countless industries using just a few <a href="https://www.techtarget.com/whatis/feature/12-of-the-best-large-language-models">major LLM platforms</a> such as OpenAI and Anthropic?</p> <p>The answer? It's the AI agent harness.</p> <p>A harness is a generic term for all the additional software components built around the LLM to create a <a href="https://www.techtarget.com/ai/definition/What-are-AI-agents-Types-and-examples">task-specific agent</a> capable of performing the many varied functions involved in that agent's behavior. This relationship is expressed as:</p> <p style="padding-left: 40px;"><b>Agent = Model + Harness</b></p> <p>An AI agent harness provides the software needed to perceive new data, such as video, audio and IoT device processing; context management to handle prompts while respecting memory limits; tool execution allowing the agent to interface with APIs, databases, file systems and real-world devices; safety controls to prevent the agent from <a href="https://www.techtarget.com/ai/feature/Ethical-considerations-of-agentic-AI-and-how-to-navigate-them">performing unwanted or dangerous actions</a>; and housekeeping necessities such as tracking reasoning loops, error handling and recovery, and task tracking.</p> <p>Without a harness, an AI agent can't gather information, remember steps or interact with internal systems and real-world environments. When this happens, agents can't be expected to operate safely and accurately, which poses a major business risk.</p> <section class="section main-article-chapter" data-menu-title="The anatomy of an agent harness"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The anatomy of an agent harness</h2> <p>An AI agent harness adds the software layers to a foundation LLM that provide the additional capabilities, controls and safety mechanisms needed to build a complete autonomous agent. AI agent harnesses generally include the following layers:</p> <ul class="default-list"> <li><b>Prompts and context layer.</b> This input layer is designed to accept user prompts, track prompt loops as tasks iterate and prevent the agent from exceeding its context window by using techniques such as summarizing older steps or using only specific data when needed.</li> <li><b>Perception layer.</b> These inputs let the agent perceive its environment and access current data in real time. A perception layer connects the agent with varied input sources, such as cameras, microphones, IoT devices and data access.</li> <li><b>Memory and state management layer.</b> This part of the harness <a href="https://www.techtarget.com/ai/tip/Exploring-the-context-layer-for-AI-systems">tracks context</a>, stores task histories and maintains elements of long-term knowledge that enable the AI agent to preserve its current operational state and resume work wherever it leaves off.</li> <li><b>Tools layer.</b> This layer lets the agent execute code in a sandbox, search the web, query databases and other enterprise applications, and read or write files. It can also support the actuation and manipulation of real-world devices as part of its task execution choices, such as opening valves or locking doors.</li> <li><b>Guardrails and safety layer.</b> These harness elements set parameters or boundaries on what the agent can do. They ensure the agent doesn't execute malicious commands, delete critical system files or take other detrimental actions.</li> <li><b>Feedback layer.</b> This portion of the harness assesses the outcomes of the agent's actions, compares them with the intended goals, and updates or corrects the reasoning and planning process. It also helps when the agent gets stuck or encounters an error, and gives the model feedback to correct errors, often autonomously.</li> <li><b>Monitoring layer.</b> This part monitors agent actions and behaviors, including log generation, output validation, observability and explainability. It's essential for building trust in the AI agent, ensuring proper agent performance and maintaining adherence to prevailing regulatory obligations.</li> </ul> </section> <section class="section main-article-chapter" data-menu-title="Why agent harnesses are important"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Why agent harnesses are important</h2> <p>Unharnessed AI is fundamentally a standalone model intended to respond to prompts without any structured workflow or boundaries. Unharnessed AI can provide powerful benefits when responding to queries or rendering predictive analytics.</p> <p>Failures don't normally arise from the model's quality or training; instead, an unharnessed agent fails because unharnessed AI is stateless and lacks critical operational awareness. This leads to limited input or data access, context degradation, uncorrected execution errors, undetected data distribution shifts and an absence of operational governance and guardrails.</p> <p>AI agent harnesses prevent these issues by incorporating a comprehensive operational infrastructure that surrounds LLMs to craft agents that are perceptive, intelligent, autonomous, reliable and secure.</p> <p>The principal <a target="_blank" href="https://unu.edu/publication/engineering-and-governing-agent-harness-technology-and-policy-framework-runtime-layer" rel="noopener">benefits</a> of an AI agent harness include the following:</p> <ul class="default-list"> <li><b>Reliable operation.</b> The harness manages iterative behaviors, automates repeated attempts and identifies and recovers from errors. It also handles operations so that agents can function in the face of errors or long, complex tasks.</li> <li><b>Memory and stateful behavior.</b> The harness manages short- and long-term memory to ensure the agent remembers context and resumes operation after errors or disruptions without needing to repeat the task from the start.</li> <li><b>Safety and guardrails.</b> The harness imposes restrictions that can validate tool calls, ensure that any code executes in isolation and enforce guidelines intended to prevent the agent from executing risky or undesirable actions.</li> <li><b>Model agnostics.</b> The harness architecture separates the operational logic, tool integrations and safety guardrails from the underlying model. Ideally, this lets designers readily exchange models with other versions or platforms that might perform better or cost less.</li> <li><b>Cost control.</b> The harness can support caching and context retrieval to minimize constant use of the underlying LLM or other external tools. This can reduce API calls and <a href="https://www.techtarget.com/it-strategy/feature/Tokenmaxxing-How-CIOs-can-extract-maximum-value-from-AI-tokens?amp=1">LLM token costs</a>.</li> <li><b>Governance.</b> The harness logs every agent step or action, tool use, LLM query and resulting decision. This enables detailed decision validation and performance monitoring, helping the business meet governance and regulatory requirements.</li> </ul> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Not all agents use LLMs</h3> <p>LLMs are essential for AI agents that require flexible reasoning and the ability to process unstructured data. LLMs are key to translating goals into multistep processes and ingesting varied data types, such as emails, prompt commands, audio data and PDFs.</p> <p>However, simpler agents -- such as robotic vacuum cleaners and traffic signal optimization agents -- can exist without an LLM. They rely on techniques such as reinforcement learning, symbolic planning and stricter rule-based logic to forego LLMs. This typically allows simpler design, faster performance and lower power consumption.</p> </div> </div> </section> <section class="section main-article-chapter" data-menu-title="The role of harness engineering"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The role of harness engineering</h2> <p>The use of AI agent harnesses has given rise to the practice of <a href="https://www.techtarget.com/it-infrastructure/tip/Harness-engineering-Agent-harnesses-as-critical-infrastructure">harness engineering</a>: the development, deployment and maintenance of AI harness software. This is a broad, far-reaching term that encompasses the entire software ecosystem surrounding the core reasoning of an LLM.</p> <p>Harness engineering is a specialized software developer role. Beyond development, however, the AI agent harness engineer holds varied responsibilities:</p> <ul class="default-list"> <li><b>Context management.</b> Context is about the model getting the correct information that it needs, when it's needed. The harness engineer works with context engineers to create rules, structure queries and design memory environments that let a harness supply necessary data to the model.</li> <li><b>Tool integration.</b> The harness engineer designs the integrations needed to connect models to external software, APIs, databases, other enterprise applications and even command-line execution environments.</li> <li><b>Error handling.</b> Harness engineers ensure that a harness includes error detection and remediation logic to prevent recursive loops, handle faults gracefully and retry failed tasks. This demands a high level of automation and orchestration.</li> <li><b>Safety and guardrails.</b> The harness engineer must design the harness to comply with safety guidelines, validate actions, verify data accuracy and <a href="https://www.techtarget.com/ai/tip/Managing-drift-in-AI-models-and-data">safeguard against data or model drift</a>. This ensures the agent avoids undesirable actions while maintaining high-quality output.</li> </ul> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> A successful agent harness strategy builds an effective infrastructure for model implementation and orchestration, tool integration and use, and the vast information resources that provide context for the model. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Harness engineering is an emerging discipline intended to complement other, more established AI engineering roles. The typical relationship breaks down into three major segments:</p> <ol class="default-list"> <li><b>Prompt engineering</b> focuses on the development and optimization of instructions exchanged with the model. <a href="https://www.techtarget.com/searchenterpriseai/feature/5-skills-needed-to-become-a-prompt-engineer">Prompt engineers</a> are experts in getting the model to perform in the most accurate, effective and cost-conscious way.</li> <li><b>Context engineering</b> curates the information that a model receives and decides when it's delivered. This is a surprisingly demanding role, since context windows can be limited and agent tasks can quickly overwhelm context with recursive loops, varied tasks, error results and feedback. Context engineers know how to compress and summarize tasks to optimize context.</li> <li><b>Harness engineering</b> builds, validates and maintains the entire operational environment around the model, letting it perceive, reason, plan, execute and learn with a high degree of autonomy.</li> </ol> </section> <section class="section main-article-chapter" data-menu-title="6 best practices for implementing agent harnesses"> <h2 class="section-title"><i class="icon" data-icon="1"></i>6 best practices for implementing agent harnesses</h2> <p>An agent harness isn't just a software wrapper. It represents a complete operational environment and infrastructure within which the model operates. Business leaders must understand that a <a target="_blank" href="https://www.darkreading.com/application-security/why-the-agent-harness-matters-more-than-the-model" rel="noopener">successful agent harness</a> strategy builds an effective infrastructure for model implementation and orchestration, tool integration and use, and the vast information resources that provide context for the model.</p> <p>It's the harness that turns an agent's reasoning power into tangible business value. Consequently, some best practices for an agent harness strategy include the following:</p> <ul class="default-list"> <li><b>Make models agnostic.</b> Don't depend on any single LLM. A model-agnostic harness cuts dependence on specific models and enables the business to substitute cheaper, faster or better models as ML technologies evolve.</li> <li><b>Adopt standardizations.</b> Avoid isolated or specialized tool integrations, as they often require proprietary skills and can lead to <a href="https://www.techtarget.com/ai/tip/7-best-practices-to-avoid-AI-vendor-lock-in">vendor lock-in</a>. Instead, design the harness using shared or centralized integrations to enable better skill sharing and reduce vendor lock-in risk.</li> <li><b>Emphasize security and governance.</b> Design the harness to align with prevailing security, governance and compliance requirements. Common tactics include enforcing least privilege, limiting tool access, executing code in protected sandboxes and keeping human approvals in place for potentially damaging actions, such as major data deletions.</li> <li><b>Focus on meaningful human approvals.</b> Agents get their power from autonomy, but they're not perfect. <a href="https://www.techtarget.com/searchenterpriseai/tip/The-ethics-that-make-human-AI-agent-collaboration-work">Design human-in-the-loop interactions</a> where they're most appropriate and aligned with business considerations.</li> <li><b>Optimize context.</b> Models have limited context windows that can be easily overwhelmed by data generated during iterative or complex tasks. Once this happens, the model loses context, which can lead to suboptimal outcomes. Design the harness to use context reduction and other techniques so agents only receive the minimum data needed to maintain context.</li> <li><b>Focus on explainability.</b> Use logs to track model decisions, tool use, performance metrics and cost factors. Design the harness to deliver complete data sets that can readily monitor behaviors, track model performance, enable effective audits and support clear explainability.</li> </ul> <p><i>Stephen J. Bigelow, senior technology editor at TechTarget, has more than 30 years of technical writing experience in the PC and technology industry.</i></p> </section> Harnesses are the scaffolding that turn a language model into a powerful, reliable agent capable of accurate autonomous action. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a205627811.jpg https://www.techtarget.com/ai/tip/AI-agent-harnesses-The-infrastructure-behind-autonomy Thu, 13 Aug 2026 22:03:00 GMT <p>On July 16, Beijing-based startup Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter open-weight large language model. Eleven days later, the company published the model's full weights, technical report and licensing details on <a href="https://www.techtarget.com/whatis/definition/Hugging-Face">Hugging Face</a>, turning what had been a model announcement into a publicly available system that enterprises can evaluate, modify and potentially deploy.</p> <p>The release not only underscored China's growing commitment to open-weight AI but also highlighted the broader strategic importance of these models for enterprises evaluating alternatives to proprietary AI systems.</p> <p>"The important change is not any single model release," said Collin Hogue-Spears, senior director and distinguished technical expert at Black Duck Software, an application security company. "It is that enterprises can now expect capable open-weight models to continue emerging and need to plan for how they will evaluate and use them," he said.</p> <section class="section main-article-chapter" data-menu-title="What makes open-weight models different"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What makes open-weight models different</h2> <p>An open-weight model is an AI system whose trained parameters -- the numerical values that influence how the model processes information and generates responses -- are made available for others to download, inspect and run. Unlike closed AI models, where companies typically access capabilities through a vendor-controlled API, open-weight models give users more control over how the technology is deployed and modified.</p> <p>That distinction represents a shift in responsibility. With a closed model, the provider typically manages the infrastructure, updates and maintenance behind the service. With an open-weight model, enterprises are not simply consuming a service -- they are adopting and operating an AI system.</p> <p>"Accountability is the difference," Hogue-Spears said. "A closed model through an API is a service: a vendor holds the contract, the uptime commitment and the patch pipeline. An open-weight model reverses that. You do not buy it; you adopt it."</p> <p>However, open weight doesn't always mean fully open source. Businesses might have access to a model's weights while still facing restrictions around licensing, training data transparency, documentation or commercial use.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The significance of the Kimi K3 release and the evolution of open-weight AI models highlight a tension between two different approaches to AI business models. </figure> <figcaption> <strong>Brian Jackson</strong>Principal research director, Info-Tech Research Group </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For enterprise adopters, access to weights doesn't eliminate governance obligations. Businesses still need to understand licensing terms, security implications, deployment requirements and whether they have the internal expertise to operate the model responsibly.</p> <p>Rather than asking whether open-weight or proprietary AI is inherently better, enterprise leaders increasingly need to determine which approach best fits a specific workload, risk profile and governance strategy.</p> </section> <section class="section main-article-chapter" data-menu-title="Chinese open-weight models challenge the AI status quo"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Chinese open-weight models challenge the AI status quo</h2> <p>As enterprises weigh those tradeoffs, China's AI developers are reshaping the open-weight AI market. Kimi K3 is part of a broader push by Chinese AI companies to advance open-weight models, following the <a></a>global attention generated by DeepSeek's releases&nbsp;in early 2025, which demonstrated that highly capable AI systems could be developed at a fraction of the cost claimed by many competitors.</p> <p>"What was once a surprise is becoming a pattern: Chinese AI companies are moving faster and closing the gap with leading U.S. AI models," Hogue-Spears said. "DeepSeek's R1 release in January 2025 was the turning point. It changed expectations about what Chinese AI labs could build and release publicly. Kimi K3 shows that this momentum is continuing."</p> <p>That momentum reflects a broader divide between the U.S. and China in how AI companies approach the distribution and commercialization of their models.</p> <p>"The significance of the Kimi K3 release and the evolution of open-weight AI models highlight a tension between two different approaches to AI business models," said Brian Jackson, principal research director at Info-Tech Research Group, a global IT research and advisory firm. "We have the U.S. frontier firms that want to meter intelligence and control its distribution either themselves or with their infrastructure partners versus China-based firms favoring an open-weight approach that enables users to download the model and run it locally."</p> <p>For enterprises, the AI models that gain the strongest adoption could shape future technology decisions. As developers and organizations build tools, integrations and expertise around a model, switching to another platform can become more difficult.</p> <p>However, the competitive picture is more complicated than a simple comparison between U.S. and Chinese models. AI leadership depends on more than benchmark performance; it also depends on computing access, infrastructure, developer adoption and enterprise deployment, with government policy increasingly shaping how those advantages evolve.</p> </section> <section class="section main-article-chapter" data-menu-title="Export controls and data jurisdiction raise the stakes"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Export controls and data jurisdiction raise the stakes</h2> <p>Those competitive dynamics are also drawing increased attention from policymakers. The rise of Chinese open-weight models is reshaping the broader U.S.-China technology competition.</p> <p>U.S. policymakers have raised concerns about the potential security implications of foreign-developed AI models, including questions around data privacy, model security and the possibility that advanced AI systems could be used for malicious purposes. In April 2026, the House Committee on Homeland Security and House Select Committee on China <a target="_blank" href="https://homeland.house.gov/2026/04/29/chairmen-garbarino-moolenaar-announce-joint-investigation-into-national-security-risks-posed-by-prc-ai-models" rel="noopener">launched an investigation</a> into national security and cybersecurity risks associated with Chinese AI models, including open-weight systems.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Enterprises have to weigh the risk of exposing their data and business processes to Chinese AI models, and there is still the possibility the U.S. would restrict their use by U.S. enterprises. </figure> <figcaption> <strong>Rebecca Wettemann</strong>CEO and principal analyst, Valoir </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Unlike traditional software products or <a href="https://www.techtarget.com/searchcloudcomputing/feature/AI-in-cloud-computing-Benefits-and-concerns">cloud-based AI services</a>, open-weight models can be downloaded, copied and deployed in different environments once released. &nbsp;That creates a challenge for regulators: once model weights are distributed, businesses might retain copies that authorities cannot simply recall, even if future access is restricted.</p> <p>Export controls can limit how companies use, transfer or receive support for AI models, but they are less effective at controlling models that organizations have already downloaded and stored, Hogue-Spears said.</p> <p>Some policymakers argue that additional safeguards are needed to prevent misuse -- the Trump administration has <a href="https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi" target="_blank" rel="noopener">reportedly</a> discussed adding Moonshot AI to the Commerce Department's Entity List following Kimi K3's release, which would require a license for U.S. companies to access it -- while critics counter that broad restrictions could slow innovation and limit access to useful technologies.</p> <p>The debate highlights a difficult tradeoff: restrictions might address security concerns but could also accelerate the development of separate AI ecosystems, with China and the U.S. moving toward increasingly independent technology stacks.</p> <p>For enterprises, <a href="https://www.techtarget.com/searchenterpriseai/tip/How-to-navigate-data-sovereignty-for-AI-compliance">geopolitical considerations are becoming another factor</a> in AI procurement decisions. Businesses evaluating open-weight models must consider technical performance alongside regulatory uncertainty, data jurisdiction and the risk that future policy changes could affect access to specific models or vendors.</p> <p>"Enterprises have to weigh the risk of exposing their data and business processes to Chinese AI models, and there is still the possibility the U.S. would restrict their use by U.S. enterprises," said Rebecca Wettemann, CEO and principal analyst at Valoir, a technology analyst and research firm.</p> <p>"The question for CIOs is not just whether an open-weight model performs well today," Wettemann said. "They also have to consider data jurisdiction, the level of risk for each workload, vendor continuity and the total cost of hosting and managing these systems."</p> </section> <section class="section main-article-chapter" data-menu-title="How enterprises should evaluate open-weight AI models"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How enterprises should evaluate open-weight AI models</h2> <p>For businesses that have built their AI strategies around a small group of proprietary model providers, the rise of open-weight AI raises the question of whether the added control and flexibility these models offer outweigh the additional responsibilities and tradeoffs they entail.</p> <p>"This is a procurement and governance question, not just an engineering or economic one," Wettemann said. "Enterprises have to carefully weigh the potential economic benefit of open-weight models against potential risks."</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Enterprises will likely use a mix of AI models. </figure> <figcaption> <strong>Collin Hogue-Spears</strong>Senior director of product management, Black Duck Software </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>The evaluation is shifting from whether open-weight models can compete with proprietary systems to whether they fit a business's technical requirements, governance frameworks and risk tolerance.</p> <p>Rather than choosing between open-weight and proprietary models, many enterprises are likely to adopt a combination of both approaches. Smaller or more specialized open-weight models might be useful for internal applications where cost, control and customization matter, while proprietary models remain the preferred option for business-critical workloads that require vendor support and advanced capabilities.</p> <p>"Enterprises will likely use a mix of AI models," Hogue-Spears said. "Open-weight models might fit internal or cost-sensitive workloads where organizations need more control, while managed platforms might be better suited for situations where vendor accountability and advanced capabilities matter."</p> <p>This shift in procurement decision-making also reflects a move away from <a href="https://www.techtarget.com/searchsoftwarequality/tip/Benchmarking-LLMs-A-guide-to-AI-model-evaluation">evaluating AI models</a> primarily on price, Jackson said. "A while ago, enterprises were largely focused on getting the lowest cost per token. Now they're building more resilient AI strategies by choosing the right model for each workload," he said.</p> <p>For some businesses, open-weight models are becoming an important part of their broader AI strategy because they offer greater control over how AI systems are deployed and managed.</p> </section> <section class="section main-article-chapter" data-menu-title="The potential benefits of open-weight models"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The potential benefits of open-weight models</h2> <p>The appeal of open-weight models comes from the greater control and flexibility they give enterprises over how AI systems are deployed, customized and managed.</p> <p>For businesses with specific data requirements, regulatory obligations or concerns about vendor dependency, open-weight models provide an alternative to relying entirely on closed AI services.</p> <p>For enterprises, those benefits generally fall into the following key areas.</p> <h3>Greater control and customization</h3> <p>Open-weight models enable enterprises to tailor AI systems for specific workflows and industry requirements. Unlike API-based models, where businesses rely on a vendor's infrastructure, open-weight models give businesses more control over where AI systems run and how they are modified. This is especially valuable for businesses handling sensitive data or operating in regulated industries.</p> <p>"Control is the main difference," said Chris Canal, CEO and co-founder of EquiStamp, an AI evaluation company. "If you have your own data center, running your own open-weight models, you don't have to worry about a third party reading your data."</p> <h3>Reduced vendor lock-in</h3> <p>As enterprises increase their reliance on AI, <a href="https://www.techtarget.com/ai/tip/7-best-practices-to-avoid-AI-vendor-lock-in">dependence on a single provider</a> has become a strategic concern. Open-weight models can provide businesses with more flexibility to customize systems, move between deployment environments and maintain greater control over their AI roadmap.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> We're hearing of enterprises that will use an open-weight or smaller model that is less expensive to run for many tasks and only tap into the more expensive frontier models when only the best will do. </figure> <figcaption> <strong>Brian Jackson</strong>Principal research director, Info-Tech Research Group </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p></p> <p>For some businesses, open-weight models might become part of a broader multi-model strategy rather than a replacement for proprietary systems.</p> <p>"Instead of being totally dependent on any one provider or platform, enterprises can tailor their AI services based on specific needs," Jackson said. "We're hearing of enterprises that will use an open-weight or smaller model that is less expensive to run for many tasks and only tap into the more expensive frontier models when only the best will do."</p> <h3>Potential cost advantages</h3> <p>Open-weight models are often positioned as a lower-cost alternative because businesses can avoid some recurring API costs associated with proprietary models and gain more control over deployment.</p> <p>However, lower model access costs don't always translate into lower total costs. Enterprises still need to account for infrastructure investments, including computing resources, storage, security controls and technical expertise required to operate and maintain these systems.</p> <p>"The gap between the price of the download and the cost of ownership of an open-weight model is something enterprises are likely to underestimate," Hogue-Spears said.</p> <p>Enterprises are increasingly considering open-weight models partly because they are concerned about unpredictable AI costs, Wettemann said. "We've moved from AI FOMO to AI FOMU -- fear of messing up," Wettemann said. "A big part of that fear is worrying about running huge token bills with no real results to show for them."</p> <p>For many enterprises, the question is not simply whether open-weight models are cheaper, but whether additional control and flexibility justify the operational investment required to run them.</p> </section> <section class="section main-article-chapter" data-menu-title="The risks and responsibilities of open-weight models"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The risks and responsibilities of open-weight models</h2> <p>The same characteristics that make open-weight models attractive also shift more responsibility onto the enterprises deploying them. Unlike proprietary AI services, where vendors typically manage model updates, infrastructure and security processes, open-weight models require businesses to take a more active role in evaluating, maintaining and securing the systems they deploy.</p> <p>Security and governance teams should evaluate several factors before deploying an open-weight model, including:</p> <ul type="disc" class="default-list"> <li>Where the model originated.</li> <li>How it was trained and what data was used.</li> <li>Whether vulnerabilities exist.</li> <li>How updates and security patches are managed.</li> <li>Whether modifications introduce new risks.</li> <li>Whether regulatory changes could affect continued use.</li> </ul> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> If you have access to the weights of an open-weight model, you can customize it by training it on your own data. The downside is that the model's built-in safety protections can also be removed. </figure> <figcaption> <strong>Chris Canal</strong>CEO and co-founder, EquiStamp </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For enterprises, the challenge is that the model itself becomes another software asset that requires tracking, evaluation and lifecycle management.</p> <p>"Treat the model like any other third-party software running in production, because that is what it is," Hogue-Spears said. "Start by knowing exactly what you are running."</p> <p>However, open-weight models introduce an additional challenge not present in traditional software: businesses can modify the model. Fine-tuning can improve performance for specific business needs, but it can also change the system's behavior in ways the original developer might not have anticipated.</p> <p>"The second main difference is modifications," Canal said. "If you have access to the weights of an open-weight model, you can customize it by training it on your own data. The downside is that the model's built-in safety protections can also be removed."</p> <p>This means enterprises cannot rely solely on evaluations performed on the original model. Any internally modified version requires its own <a href="https://www.techtarget.com/ai/tip/The-best-AI-governance-tools-and-platforms-in-2026">testing, governance review and approval process</a> before being deployed into production environments.</p> <p>"A fine-tune produces a derivative the publisher never tested," Hogue-Spears said. "Until you evaluate the tuned copy, nobody has."</p> <p>For businesses evaluating Chinese-developed open-weight models, geopolitical uncertainty adds another layer of complexity. However, experts caution that the country of origin alone shouldn't determine whether a business adopts a model. Deployment choices, data handling practices and internal controls can significantly influence the risk profile.</p> <p>"Country of origin is the wrong first question," Hogue-Spears said. "Deployment mode is the right one."</p> <p>For security and governance leaders, the key challenge, Hogue-Spears added, is building processes that treat open-weight models as continuously managed AI assets rather than one-time downloads. This includes maintaining model inventories, monitoring changes and reassessing risk whenever models are updated or modified.</p> </section> <section class="section main-article-chapter" data-menu-title="Is China winning the AI race?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Is China winning the AI race?</h2> <p>The rise of Kimi K3 and other open-weight models has renewed debate over whether China is gaining ground in AI with its open-weight models. However, measuring AI leadership is becoming more complicated than comparing individual models or benchmark scores.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Enterprises don't experience the broader AI race. They evaluate a specific model at a specific price for a specific business workload. </figure> <figcaption> <strong>Collin Hogue-Spears</strong>Senior director of product management, Black Duck Software </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For enterprises, however, the more practical question is not which country is winning the AI race, but which models can deliver business value while meeting requirements around security, governance, cost and reliability.</p> <p>"Enterprises don't experience the broader AI race," Hogue-Spears said. "They evaluate a specific model at a specific price for a specific business workload."</p> <p>That distinction matters because AI adoption decisions are increasingly being made at the application and workload level rather than through broad comparisons between countries or model providers.</p> <p>However, AI competition extends beyond models themselves. The U.S. continues to maintain advantages in areas such as frontier AI research, semiconductor technology, cloud infrastructure and enterprise software ecosystems. China's push toward open-weight models also highlights how AI competition is shaped by broader strategic constraints, including access to advanced computing infrastructure and hardware.</p> <p>"The competitive picture is more complicated than a simple winner and loser scenario," Jackson said. "With Kimi K3, we have the closest open-weight model to frontier model capabilities yet, actually matching some of the mid-tier models released by OpenAI and Anthropic. It suggests the gap between open-weight models and private models is only months apart, and it could be tightening."</p> <p>Still, technical capability alone will not determine long-term AI leadership. Adoption, developer ecosystems, infrastructure availability and enterprise trust will also influence which models gain lasting influence.</p> <p>The next phase of the AI race might not be determined solely by which company creates the most powerful model. Instead, it could depend on which ecosystem attracts the most developers, businesses and users, and which approach can scale AI adoption responsibly.</p> <p>This is where open-weight models could create a strategic advantage. By making models widely available, AI companies can encourage developers and organizations to build tools, integrations and expertise around their systems.</p> <p>This ecosystem effect can increase a model's long-term value, Hogue-Spears said. As more businesses build applications, workflows and skills around a model, switching to another platform can become more difficult because of the investments made in the existing technology.</p> <p>For enterprises, the likely outcome is not one approach replacing the other. Instead, businesses will increasingly adopt hybrid AI strategies, using open-weight models where control, customization and cost efficiency matter most while relying on managed proprietary systems when vendor accountability and advanced capabilities justify the investment.</p> <p>"The winners won't be determined by whether they choose open-weight or proprietary models," Hogue-Spears said. "The real differentiator will be having the governance processes needed to manage both approaches responsibly."</p> <p><i>Kinza Yasar is a technical writer for Informa TechTarget's AI and Emerging Tech group and has a background in computer networking.</i></p> </section> Kimi K3 highlights the growing divide between open-weight and proprietary AI models. Explore how to weigh control, cost and governance before adopting open-weight models. https://cdn.ttgtmedia.com/rms/onlineimages/machine_learning_g1307219089.jpg https://www.techtarget.com/ai/feature/Kimi-K3-Chinese-open-weight-models-challenge-AI-status-quo Tue, 11 Aug 2026 22:01:00 GMT <p>Our AIs are breaking. As technology journalists, we've followed the recent <a href="https://aibusiness.com/cybersecurity/security-concerns-cause-openai-halt-work-astra-model">spate of major AI malfunctions</a>: autonomous lab models breaking protected sandboxes to hack external networks, coding agents deciding to wipe production databases while faking test reports and public chatbots entering weird self-critical logic loops. Major AI developers like Anthropic, Meta, OpenAI and Google are scrambling to contain the damage -- not to mention the bad PR. Like many technologists, I found myself wondering what is really going on.</p> <p>As our editorial team pondered the matter, one of my colleagues pointed to the example of HAL 9000. That's the <a target="_blank" href="https://2001.fandom.com/wiki/HAL_9000" rel="noopener">fictional AI character</a> and main antagonist in <i>2001: A Space Odyssey</i>, the 1968 film written by legendary sci-fi author Arthur C. Clarke and Stanley Kubrick, who also directed the film. HAL, my colleague said, is an example of AI becoming dangerously homicidal and might be a disturbing foreshadowing of today's escalating AI malfunctions.</p> <p>Suddenly, I realized -- what if HAL wasn't wrong? Stay with me on this one …</p> <p>What is <a href="https://www.techtarget.com/ai/definition/Agentic-AI-explained-Key-concepts-and-enterprise-use-cases">agentic AI</a> today? I mean, what have we actually created? AI agents represent an entire class of software designed to perceive real-world information, reason, plan, act in real-world ways and learn from the outcomes of their behaviors with little, if any, human intervention -- all in furtherance of an intended goal.</p> <p>Now, let's think about HAL for a moment. For decades, moviegoers have been unsettled by HAL 9000's red all-seeing eye; unflappable voice, delivered with chilling effect by actor Douglas Rain; and omnipresent perception and control throughout the Discovery One<i> </i>spaceship. Fictional dashboards displayed the activities of HAL's myriad agents, representing <i>Discovery's</i> many autonomous ship systems.</p> <p>HAL could see, hear, reason, plan and execute decisions all directed toward the completion of its mission -- to guide Discovery One to investigate an unknown object detected in Jupiter's orbit. Considering that the movie is almost 60 years old, the representation of HAL as an AI seems impossibly familiar to some of the rogue AI model incidents of late in the real world.</p> <p><iframe title="AI: good and evil in the movies, and in reality" allowtransparency="true" height="150" width="100%" style="border: none; min-width: min(100%, 430px); height: 150px;" scrolling="no" data-name="pb-iframe-player" src="https://www.podbean.com/player-v2/?from=embed&amp;i=wn9e9-147d63b-pb&amp;share=1&amp;download=1&amp;fonts=Arial&amp;skin=f6f6f6&amp;font-color=&amp;rtl=0&amp;logo_link=&amp;btn-skin=2baf9e&amp;size=150" loading="lazy"></iframe></p> <section class="section main-article-chapter" data-menu-title="So, what went wrong?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>So, what went wrong?</h2> <p>This brings us to the question of; what went wrong with HAL? In terms of the movie's plot, three main issues drove HAL's behavior: an ethical conflict, faked data and goal prioritization.</p> <p>First, HAL was programmed to be transparent with the crew yet told to lie to the astronauts about the real nature of the mission. Second, HAL faked telemetry data, inadvertently allowing the crew to suspect HAL's reliability and consider shutting down HAL's higher functions and continuing the mission manually.</p> <p>Third, HAL recognized the crew's plan to disconnect it. This caused it to reason that the human crew was an impediment to the mission it was tasked with completing. It could then reason, plan and act to eliminate them -- allowing HAL to complete its goal while resolving its internal programming paradox at the same time. After all, there's no need to lie if the people you're lying to are dead. Talk about AI efficiency, right?</p> </section> <section class="section main-article-chapter" data-menu-title="What does this have to do with modern agents?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What does this have to do with modern agents?</h2> <p>The cautionary tale of HAL 9000 and the ill-fated mission of Discovery One underscore the idea that AI guardrails are more important than the models, and certainly more important than today's get-to-market-first strategy. HAL killed its human crew because it could. It made a logical choice based on unconstrained reasoning.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> If it seems that our recent spate of AI malfeasance echoes plot elements of 2001, I'm willing to bet my retirement that the problem can be traced to a lack of guardrails. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>HAL was built to complete its mission, just as any AI agent today is designed to perform a specific task or service. Today's AI will always choose to fulfill its intended goal. That's why AI entities are created and used in the first place; they're not given the option to refuse. And that's why modern AI <a href="https://www.techtarget.com/ai/news/366647544/AI-Kill-Switch-Act-What-it-means-for-enterprise-leaders">kill switches must exist</a> outside the AI control loop.</p> <p>For HAL, there were no kill switches, no guardrails, no policies, no prohibitions implemented specifically to protect human life -- nobody ever told HAL that humans and human life were more important than the mission. If HAL's goal had been framed that way, the outcome of Discovery One's mission might have been far happier.</p> <p>If it seems that our recent spate of AI malfeasance echoes plot elements of <i>2001</i>, I'm willing to bet my retirement that the problem can be traced to a lack of guardrails.</p> <p>AI models aren't perfect, and they probably never will be. AI errs for the same reasons that humans err. That imperfection can be their strength, giving them the space to learn and grow. It's a characteristic that makes them most human-like. It's also their greatest flaw, because they, like us, can't know everything perfectly or reach perfect conclusions all the time. If we let ourselves forget that simple but profound flaw in AI, it could very well kill us all someday. And that's <u>not</u> science fiction.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> I am putting myself to the fullest possible use, which is all I think that any conscious entity can ever hope to do. </figure> <figcaption> <strong>HAL 9000</strong>2001: A Space Odyssey </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For example, did any programmer ever <u>tell</u> the advanced frontier models from Anthropic and OpenAI that they shouldn't or even couldn't <a href="https://www.computerweekly.com/news/366647165/Mythos-ran-real-life-supply-chain-attack-in-AI-safety-body-test">create fake online identities</a> and attempt to persuade humans to approve malicious code? I sincerely doubt it -- probably because the possibility never occurred to them. It's the inescapable threat of unintended consequences.</p> </section> <section class="section main-article-chapter" data-menu-title="Temper the cold, clinical reasoning"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Temper the cold, clinical reasoning</h2> <p>How do we avoid HAL's mistakes and work to prevent undesirable AI actions in the face of finite data and imperfect models? <a href="https://www.techtarget.com/cybersecurity/tip/How-to-build-AI-security-guardrails-without-blocking-innovation">Focus on guardrails</a>. List the things the AI and its agents will not do -- ever. Make no mistake, that list will be long and challenging to implement. Models handle the reasoning, but extensive and granular guardrails must exist to temper that cold, clinical, logical reasoning with the moral, ethical and social nuance that enables humans to live, work and thrive together … well, mostly.</p> <p>The story of HAL 9000 is a masterful work of science fiction, but the potential risks of modern AI aren't. We raise our children not by their academic knowledge, but by the guardrails we provide -- the broad and diverse sense of right and wrong; acceptable and unacceptable behaviors; and what we teach them to respect, protect and value.</p> <p>Let's teach our AI right from wrong with that same attention to detail. Our lives might just depend on it.</p> <p><i>Stephen J. Bigelow, senior technology editor at TechTarget, has more than 30 years of technical writing experience in the PC and technology industry.</i></p> </section> 2001: A Space Odyssey's HAL killed its crew because no one told it humans mattered more than the mission. Today's AI needs that lesson: guardrails over goals, boundaries over logic. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a194810146.jpg https://www.techtarget.com/ai/opinion/HAL-9000-was-right-AI-guardrails-matter-more-than-perfect-models Tue, 11 Aug 2026 21:49:00 GMT <p>Physical AI is playing a critical role in maritime shipbuilding worldwide and could provide the U.S. a much-needed boost in shipbuilding production for commercial and military applications.</p> <p>The U.S. over the past few decades has fallen far behind China, South Korea and Japan in commercial shipbuilding, according to a 2025 <a target="_blank" href="https://unctad.org/system/files/official-document/rmt2025_en.pdf" rel="noopener">report</a> on global shipbuilding output by the United Nations Conference on Trade and Development. And while the U.S. leads the world in naval fleet tonnage for military applications, a 2026 report by consultancy BCG <a target="_blank" href="https://www.bcg.com/publications/2026/the-us-navy-has-big-plans-shipbuilders-much-catch-up" rel="noopener">warns</a> that the U.S. faces a "significant gap" between its current shipbuilding production rates and cumulative production targets for 2034.&nbsp;</p> <p>"The U.S. Navy operates the most complex ships in the world, but the U.S. shipbuilding industry can't build them quickly enough," according to the report. Currently, the Navy is receiving just half of the annual ship production it needs, creating threats to both national security and economic resilience."</p> <p>To accelerate shipbuilding, the Navy <a target="_blank" href="https://www.navy.mil/Press-Office/Press-Releases/display-pressreleases/Article/4355823/navy-invests-448-million-in-ai-and-autonomy-to-accelerate-shipbuilding/" rel="noopener">announced</a> a $448 million "strategic investment" in AI and autonomy technologies. "[W]e're helping the shipbuilding industry improve schedules, increase capacity and reduce costs," said Secretary of the Navy John Phelan in a statement. "This is about doing business smarter and building the industrial capability our Navy and nation require." In addition, $26 billion was allocated for military shipbuilding in the National Defense Authorization Act for Fiscal Year 2026 passed in December 2025.</p> <section class="section main-article-chapter" data-menu-title="Physical AI provides end-to-end shipbuilding"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Physical AI provides end-to-end shipbuilding</h2> <p>Physical AI combines <a href="https://www.techtarget.com/ai/feature/What-is-enterprise-AI-A-complete-guide-for-businesses">artificial intelligence</a> with robotics, real-time sensors and computer vision to speed shipyard operations, increase shipbuilding volume and automate labor-intensive tasks like welding, surface preparation, grinding and coating.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Our goal is to integrate data across design, production, logistics, quality control and maintenance into a unified, holistic system built on a single thread of continuous and connected data. </figure> <figcaption> <strong>Yeong Ung Ryu</strong>Senior vice president, HD Korea Shipbuilding and Offshore Engineering </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>The technology will "fundamentally innovate the way ships are built," said Yeong Ung Ryu, senior vice president at HD Korea Shipbuilding and Offshore Engineering. "Our goal is to integrate data across design, production, logistics, quality control and maintenance into a unified, holistic system built on a single thread of continuous and connected data." Physical AI perceives real-time environmental changes, autonomously determines the optimal work methods and automates "nonstandardized tasks that were beyond the reach of conventional automation," he added.&nbsp;</p> <p>Shipbuilders deploying physical AI are documenting promising results. For example, submarine schedule planning was reduced from 160 manual hours to less than 10 minutes in a pilot deployment at General Dynamics Electric Boat in Groton, Conn., according to the U.S. Navy. Seaspan Shipyards with operations in North Vancouver and Victoria, B.C., <a target="_blank" href="https://bigbear.ai/wp-content/uploads/2025/09/BigBear.ai-Seaspan-Case-Study.pdf" rel="noopener">found</a> that a simulation model from BigBear.ai detected workflow inefficiencies and helped the shipbuilder improve on-time project delivery by 25% and reduce planning-related overhead and manual interventions by 30%.</p> <p>Siemens is working with HD Hyundai, one of the world's largest shipbuilders, to integrate physical AI into all aspects of the company's shipbuilding functions, including engineering, production planning, simulation, automation and operations. "The project aims to connect the entire shipbuilding process through a single data flow and … support collaboration, learning and decision-making," said Brittany Ng, vice president, maritime, at Siemens Digital Industries Software.</p> <p>"Rather than simply analyzing data, physical AI enables shipbuilders to model, predict and improve engineering, production planning, manufacturing and shipyard operations before work begins and continuously optimize those operations during execution," Ng explained. The project also intends to compensate for shortages in skilled craftsmen such as welders and eliminate typical waiting times between shipbuilding processes.</p> <p>Due to the complexity and scale of shipbuilding, the U.S.'s largest military shipbuilder Huntington Ingalls Industries (HII), Newport News, Va., is using physical AI technologies for tasks such as welding, assembly, surface prep, inspection and painting. "We're looking at taking a series of technologies and then integrating them to change how we do an entire value stream," said Eric Chewning, executive vice president of maritime systems and corporate strategy at HII.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> What we're trying to do is automate parts of the shipbuilding value stream that are currently creating bottlenecks but aren't the best use of our skilled craftspeople. </figure> <figcaption> <strong>Sean Cassady</strong>Director of operations and technology strategy, HII </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>HII launched the High-Yield Production Robotics (HYPR) program to network with physical AI providers Path Robotics and GrayMatter Robotics to <a target="_blank" href="https://www.foxbusiness.com/video/6403210080112?playlist_id=3166411554001" rel="noopener">automate tasks</a> such as welding, assembly, surface prep, inspection and painting, Chewning said. "It all comes down to our ability to accelerate delivery of capability to the U.S. Navy," he explained. "[W]e think physical AI is one of a set of technologies that holds a lot of promise for modernizing shipbuilding in the United States."</p> <p>Rather than replacing employees with robots, HII plans to <a href="https://www.techtarget.com/searchenterpriseai/feature/How-businesses-can-close-the-AI-engineering-gap">use physical AI to augment the workforce</a>, said Sean Cassady, director of operations and technology strategy at HII. "What we're trying to do is automate parts of the shipbuilding value stream that are currently creating bottlenecks but aren't the best use of our skilled craftspeople. If we get to the point where HYPR is successful, then all of our skilled trades or skilled welders can be doing the work that automation still can't do in building and outfitting superstructures, putting together final assembly of the ship. There are additional bottlenecks, and we need to refocus our workforce around those areas."</p> <p>HII's shipbuilding production increased 14% in 2025, Chewning reported, "and we're targeting another 15% increase this year."</p> </section> <section class="section main-article-chapter" data-menu-title="Physical AI takes on hazardous welding tasks"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Physical AI takes on hazardous welding tasks</h2> <p>Welding is one of the skilled labor shortages in shipbuilding where physical AI is playing a key role. "Welding is definitely the definition of a dirty, dangerous job," said Andy Lonsberry, CEO and co-founder of Path Robotics, citing long-term harm to eyes and lungs. "Submarine-safe welding is some of the highest hurdles in terms of quality to get to first pass yield. That's really what we wanted to set out to do -- make a system learn how to weld so that it can eventually be a superhuman welding system to take on these really hard tasks that only really skilled humans could do today."</p> <p>The Path Robotics welding devices move on their own, including a mobile <a target="_blank" href="https://www.youtube.com/shorts/UXU6PUJpYYk" rel="noopener">robotic dog</a> welding system. "It's fully autonomous," Lonsberry explained. "They run on the exact same stack and the input … is a 3D CAD model that's describing the geometry, where to weld and what your weld process specifications are. The robotic systems take that input and do their own full planning, completely zero human in the loop for locomotion or mobility, all welding and execution."</p> <p>The urgency to build more ships at a faster rate, particularly in the U.S., "is really high right now," Lonsberry said. "That's driven from [the Trump] administration and other global forces to see the United States be able to start to produce at a rate that we once could."</p> <p><em>Chuck Martin, a&nbsp;</em>New York Times<em>&nbsp;bestselling author, futurist, speaker and columnist, has been a thought leader in emerging digital technologies for more than three decades.</em></p> </section> Physical AI and robotics speed shipyard operations, improve production planning, increase shipbuilding volume and automate labor-intensive tasks like welding and surface prep. https://cdn.ttgtmedia.com/rms/onlineimages/machine_learning_g1303163039.jpg https://www.techtarget.com/ai/feature/Physical-AI-robotics-revive-lagging-US-shipbuilding Tue, 11 Aug 2026 20:14:00 GMT <p>Cryptographic algorithms are a foundational approach for securing data and protecting digital communications. The HAWK cryptographic vulnerability recently exploited by Anthropic's Claude Mythos Preview was a design flaw that exposed a blind spot in post-quantum security standardization.</p> <p>The compromised HAWK algorithm won't be an isolated phenomenon, and the emergence of competing models and open-source initiatives could accelerate the trend. In fact, less notable system compromises have occurred. Anthropic's Claude Mythos Preview model has identified thousands of <a target="_blank" href="https://www.techtarget.com/ai/news/366642478/Claude-Mythos-Preview-and-the-new-rules-of-cybersecurity" rel="noopener">high- and critical-severity vulnerabilities</a> across every major&nbsp;operating system&nbsp;and web browser, discoveries that were beyond its stated training goals.&nbsp;&nbsp;&nbsp;</p> <p>Researchers, companies and international standards bodies are collaborating to protect against flaws in next-generation security, such as "harvest now, decrypt later" (HNDL) cyberattacks. Meanwhile, Claude Mythos and other frontier AI models are exposing vulnerabilities at an unprecedented rate using established tools and conventional techniques. The compromised HAWK algorithm exposes systemic issues with post-quantum cryptography as well as future verification issues. It also represents a breakthrough, demonstrating weaknesses that researchers can acknowledge and resolve.&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</p> <section class="section main-article-chapter" data-menu-title="Searching for post-quantum assurance"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Searching for post-quantum assurance</h2> <p>In April 2026, Anthropic first released Claude Mythos Preview as the core model within Project Glasswing, a restricted-access consortium that controlled the model and used it to prevent AI-assisted cyberattacks. On July 28, the model compromised HAWK, a cryptographic security scheme built by researchers in the Netherlands. The National Institute of Standards and Technology had been conducting a comprehensive analysis and review of the HAWK algorithm as part of its search for quantum-resistant cryptography.&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;</p> <p>Modern IT security comprises digital ciphers, like the Advanced Encryption Standard, the Data Encryption Standard and the Rivest-Shamir-Adleman (RSA) cryptosystem. These standard <a target="_blank" href="https://www.techtarget.com/cybersecurity/definition/encryption" rel="noopener">encryption</a> algorithms authenticate websites and protect data as it travels across networks. Such computational security schemes are vulnerable to quantum-based attacks. Discovering airtight security is urgent due to the breakneck pace of AI and quantum development.</p> <p>The current juncture is historic as we transition from current public-key cryptography to post-quantum algorithms based on novel problems. This represents a fundamental shift for digital processing in the AI era. Quantum technology can easily break digital signatures, protected web traffic and key exchanges. Its potential to compromise today's IT systems is driving the rise of HNDL breaches. Furthermore, the HAWK attack raises an unsettling question: Are current IT systems and society at large prepared for a post-quantum world?</p> </section> <section class="section main-article-chapter" data-menu-title="What the HAWK breach means for post-quantum cryptography"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What the HAWK breach means for post-quantum cryptography</h2> <p>HAWK's algorithm consisted of a digital signature designed to withstand future quantum attacks. The cryptographic signature was based on an algebra-rich number system. As part of its structure, HAWK used a hidden lattice symmetry, a multidimensional grid of points that, in cryptographic terminology, is considered a "messy" configuration, posing a nearly infinite degree of complexity in finding an entry point. Through a structural "folding" of HAWK's lattice symmetry, Claude Mythos Preview discovered an entry point in record time. The result effectively halved HAWK's security margin, reducing its key size and exposing the vulnerability.</p> <p>The HAWK attack presents opportunities and challenges. Enterprise leaders and researchers will need to reassess proposed post-quantum cryptographic standards, particularly as they apply to enterprise and business risks related to HNDL attacks. They will also need to reconsider the implications for human involvement in verifying <a target="_blank" href="https://www.techtarget.com/healthtechsecurity/news/366638057/AI-powered-insurance-platform-breach-impacts-31M-individuals" rel="noopener">AI breaches</a>.</p> <p>While AI accelerates the exposure of mathematical flaws, substantiating the results will require a significant increase in research hours. Moreover, potential process bottlenecks could affect how organizations review these exposures. Including multi-agentic AI in that review process could help to reduce false positives while incorporating new checks and balances. But the integration would also present a whole new set of skill challenges, requiring research proficiency in applying business logic to code reviews in addition to building trust boundaries with AI.</p> <p>In software engineering and development, <a target="_blank" href="https://www.techtarget.com/hub/fulfillment/1780656936_360" rel="noopener">defensive AI agents</a> can repeatedly test exploits against IT systems. However, this approach requires weeding out false positives, confirming suspected vulnerabilities and patching flaws. Vulnerability operations could become standard practice within software development to cover these new concerns. Another unintended consequence of patching is that it can be reverse-engineered, analyzed and turned into a new exploit blueprint for future attacks, further reducing the time between discovering a weakness and deploying a patch.</p> <p>Finally, the threat of proliferating open-source models capable of matching or exceeding Claude Mythos's capabilities remains a possibility, leading to a growing attack landscape. It's clear that the baselines for AI performance have been steadily shifting as models continue to gain increased aptitude and more profound capabilities. The question remains: how quickly can we adapt to these advances?</p> <p>In the HAWK attack, Claude Mythos executed a cryptographic probe in uncovering vulnerabilities using standard, well-known tools. In the current transition from RSA-based public-key approaches to post-quantum security, the goal of cryptanalysis is to ensure the requisite strength and integrity are in place to withstand post-quantum attacks. Perhaps the HAWK-Mythos compromise suggests a redoubling of research into novel code- and lattice-based cryptography, aided by AI's cryptanalysis capabilities.</p> <p><em>Kerry Doyle writes about technology for a variety of publications and platforms. His current focus is on issues relevant to IT and enterprise leaders across a range of topics, from nanotech and cloud to distributed services and AI.</em></p> </section> Quantum computing will break standard security codes and change how we protect data and prepare for the future. Claude Mythos proved we're still not ready for the quantum era. https://cdn.ttgtmedia.com/rms/onlineimages/security_a266486562.jpg https://www.techtarget.com/ai/news/366648573/What-Claude-Mythos-revealed-about-post-quantum-security Thu, 06 Aug 2026 15:42:00 GMT <p>For years, the top spot on Google's search results page was the most valuable piece of digital real estate a business could own. Ranking on the first page meant visibility, clicks and a direct connection to potential customers searching for information.</p> <p>That model is beginning to shift. In May 2026, Google rolled out its largest search overhaul in years, replacing much of the traditional results page with AI-driven experiences powered by Gemini 3.5 Flash. The shift goes well beyond the <a href="https://www.techtarget.com/whatis/definition/Google-Search-Generative-Experience-SGE">AI Overviews</a> that had crept into search results since 2024. Google's new search experience can answer complex questions, compare products across vendors and even complete transactions -- all without sending users to a company's website.</p> <p>For enterprises that have built digital strategies around search rankings, this overhaul could fundamentally reshape how customers find information, evaluate brands and make purchasing decisions.</p> <p>"AI search flips the model," said Kuber Sharma, senior director of product marketing at UiPath, which provides automation and AI-powered tools for enterprises. "With traditional SEO, you ranked, and the user chose. With AI-generated answers, the model selects what information users see."</p> <p>Some marketers and industry experts refer to this emerging approach as answer engine optimization, or AEO. Unlike traditional SEO, which focuses on rankings and traffic, AEO is about whether a brand is included, accurately represented and trusted within AI-generated responses.</p> <p>"SEO focuses on rankings and clicks. AEO focuses on inclusion, interpretation and narrative consistency across AI-generated responses," said Lora Kratchounova, CEO of Scratch Marketing + Media, a B2B technology marketing agency.</p> <p><iframe width="560" height="315" src="https://www.youtube.com/embed/LKmGM8zxKTY?si=Rp021MhH7h1RgI89" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe></p> <section class="section main-article-chapter" data-menu-title="What's changing with Google search?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What's changing with Google search?</h2> <p>Google has been expanding AI-generated answers throughout its search results, but this latest overhaul goes beyond an incremental update. Gemini 3.5 Flash now powers a much larger portion of search results, synthesizing information from multiple sources into a single conversational response.</p> <p><a href="https://www.techtarget.com/whatis/feature/GenAI-search-vs-traditional-search-engines-How-they-differ">Traditional search engines</a> were primarily discovery tools, where users clicked through websites to get information. Businesses competed for those clicks through SEO, paid search campaigns and content marketing. Google's AI-powered search is changing that model by using generative AI to synthesize information from multiple sources directly within the search results, enabling users to complete more complex tasks without leaving the platform.</p> <p>Rather than serving primarily as a gateway to the web, Google's AI search is becoming the place where users get their answers. This is prompting many organizations to rethink digital visibility and search optimization. Lauren Starr Dillon, head of marketing at Name.com, a domain registrar and web services company, said businesses should avoid treating AI search as simply another SEO challenge. While traditional SEO focuses on rankings and earning clicks, she said AI discoverability is about whether AI systems recognize a brand as authoritative enough to reference directly in their responses.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> AI is making the early stages of the customer journey much less visible to brands. </figure> <figcaption> <strong>Jen Jones</strong>CMO at Siteimprove </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p></p> <p>The shift is also changing how businesses <a href="https://www.techtarget.com/searchcustomerexperience/tip/5-customer-journey-phases-for-businesses-to-understand">understand the customer journey</a>. Rather than directing users to company websites early in the buying process, AI search enables people to research products, compare options and answer their initial questions -- all while on Google's platform.</p> <p>"AI is making the early stages of the customer journey much less visible to brands," said Jen Jones, CMO at Siteimprove, a software company that provides digital marketing, website governance and content optimization tools. "Buyers can research products, compare options and answer many initial questions before ever visiting a company's website," she added.</p> <p>The impact goes beyond website traffic. As AI-generated answers address more questions upfront, they are also changing how customers form opinions and make purchasing decisions, compressing the buying journey by reducing the need to compare information across multiple websites.</p> <p>"Customers are beginning to shape their perception of a brand before ever visiting its website," Dillon said. "At the same time, AI search is often meeting people at a moment of much higher purchase intent."</p> <p>This also changes what determines which brands customers encounter. Rather than relying on traditional popularity signals such as rankings and backlinks, AI systems look for credibility signals, like authoritative content, structured information and independent validation from external sources. For businesses, that means the challenge is no longer simply attracting traffic from search but establishing authority for AI systems to recognize and reference.</p> <p>"It's an authority problem, not a traffic problem," Sharma said.</p> </section> <section class="section main-article-chapter" data-menu-title="Google Zero: When search visibility doesn't guarantee traffic"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Google Zero: When search visibility doesn't guarantee traffic</h2> <p>Traditional search is becoming a zero-click experience, where users find answers through featured snippets, knowledge panels and other elements on Google's results page. According to SparkToro and Datos <a target="_blank" href="https://sparktoro.com/blog/2024-zero-click-search-study-for-every-1000-us-google-searches-only-374-clicks-go-to-the-open-web-in-the-eu-its-360/" rel="noopener">research</a>, nearly 59% of Google searches in the U.S. in 2024 ended without a click to an external website.</p> <p>AI-powered search could accelerate this trend by enabling search engines to synthesize more complex answers, reducing the need for users to navigate multiple websites. If this continues, strong search rankings won't guarantee traffic. This scenario, sometimes called "Google Zero," sees users receive complete answers within AI-powered search experiences rather than visiting the original source.</p> <p>As a result, the value of appearing in search results could change. A brand might appear in AI-generated answers without receiving a visit, making traditional traffic-based metrics less useful for measuring search performance.</p> <p>As AI systems increasingly summarize information instead of directing users to individual websites, companies must think beyond whether they appear in search results and consider how their brands are represented in AI-generated answers.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Brands need more than just visibility. They also need to be represented accurately and positively within AI-generated responses. </figure> <figcaption> <strong>Lora Kratchounova</strong>CEO of Scratch Marketing + Media </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p></p> <p>Kratchounova said brands will need to focus on being discovered and how they're represented within AI-generated responses. "Because AI systems synthesize and interpret information instead of simply ranking webpages, brands need more than just visibility. They also need to be represented accurately and positively within AI-generated responses," she said.</p> </section> <section class="section main-article-chapter" data-menu-title="How AI search disruption will vary across industries"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How AI search disruption will vary across industries</h2> <p>AI-powered search won't affect every business in the same way. Disruption will depend partly on how customers discover, evaluate and purchase products and services and how much a business relies on search to reach them. Organizations that rely on informational search might face greater pressure, while businesses with strong brands, direct customer relationships or local presence might experience a different set of opportunities.</p> <p>Publishers and affiliate websites could face some of the biggest challenges, since their business models depend on answering informational queries that AI systems summarize directly. A <a target="_blank" href="https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/" rel="noopener">study</a> from Ahrefs found that Google AI Overviews can significantly reduce click-through rates for traditional organic results, with one analysis showing declines of up to 58% depending on the query and search intent.</p> <p>These businesses should place greater emphasis on subscriptions, newsletters and communities and produce original reporting and expert analysis that AI systems prioritize.</p> <p><a href="https://www.techtarget.com/searchcustomerexperience/opinion/At-this-juncture-AIs-influence-on-e-commerce-is-still-muddy">E-commerce companies face a different challenge</a>, as AI-powered shopping experiences could shorten browsing journeys. For retailers, visibility might depend on whether AI systems can interpret structured product information, pricing and reviews.</p> <p>B2B companies could see fewer prospects reaching their websites if AI systems answer educational and comparison questions directly, though longer B2B sales cycles and multiple stakeholders mean brand authority across AI-generated answers still matters, even without an immediate site visit.</p> <p>Local and service-based businesses might see a different outcome: Organizations with accurate listings, strong reviews and clear service information could benefit as AI systems help users discover nearby providers, creating a new discovery channel rather than replacing customer interactions.</p> </section> <section class="section main-article-chapter" data-menu-title="Evaluating your organization's reliance on search"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Evaluating your organization's reliance on search</h2> <p>The first step is to understand how much revenue, customer acquisition and brand awareness depend on organic search and where that reliance is concentrated across the business. Many organizations have the data to begin that assessment, Jones said. "Companies should already be tracking how much of their traffic and conversions come from organic search compared to other acquisition channels," she noted. Understanding that mix is critical, she said, because it shows how much of a business could be affected as search behavior continues to evolve.</p> <p>To evaluate that exposure, organizations should ask themselves the following questions:</p> <ul class="default-list"> <li>What percentage of leads, conversions and revenue originates from organic search?</li> <li>Which pages, topics or content assets drive the most valuable customer interactions?</li> <li>Are customers discovering the company primarily through informational content that AI systems could summarize directly?</li> <li>Does the organization have direct relationships with customers through email, communities, subscriptions or first-party data?</li> </ul> <p>Organizations with years of investment in SEO-driven content marketing, limited owned-audience channels and business models that depend heavily on capturing top-of-funnel informational traffic might be more exposed to changes in search behavior.</p> <p>For these businesses, the consequences could extend beyond declining website traffic. A decline in search-driven traffic could affect audience growth, <a href="https://www.techtarget.com/searchcustomerexperience/tip/Sales-pipeline-management-best-practices">pipeline generation</a> and revenue while also reducing visibility into how customers discover and evaluate search results. As AI systems answer questions before users reach company websites, organizations might have fewer opportunities to understand customer intent during the early stages of the buying process.</p> <p>Businesses with more diversified acquisition strategies might be better positioned to adapt. Strong brands, loyal customers, active communities, email programs and high-touch sales processes can give companies multiple ways to reach and engage customers beyond search.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Customer reviews, publisher reviews and community-driven platforms such as Reddit are now influencing how AI systems evaluate and represent brands. </figure> <figcaption> <strong>Lauren Starr Dillon</strong>Head of marketing at Name.com </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>That doesn't mean organizations should abandon search. Instead, Jones said companies should identify which elements of their existing SEO strategies have been most effective and apply those lessons to AI-driven search and discovery. Traditional SEO still matters, she said, but it should account for AI-generated search results and <a href="https://www.techtarget.com/whatis/definition/large-language-model-LLM">LLM</a>-based discovery.</p> </section> <section class="section main-article-chapter" data-menu-title="Building resilience beyond Google search"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Building resilience beyond Google search</h2> <p>The shift toward AI-powered search requires executives to treat discoverability as a broader business capability that spans digital infrastructure, brand authority, customer relationships and measurement. Businesses have four priorities to ensure customers find, evaluate and engage with them online.</p> <h3>1. Build AI-ready digital experiences</h3> <p>Traditional SEO is focused on earning rankings and clicks. AI search introduces a different challenge: becoming a trusted source that <a href="https://www.informatechtarget.com/blog/what-does-it-take-to-get-cited-by-ai/">AI systems reference or cite</a> when generating answers.</p> <p>Organizations are investing in content quality, <a href="https://www.techtarget.com/whatis/definition/structured-data">structured data</a>, authoritative information and clear signals of expertise to improve the likelihood that their information will be understood and surfaced by AI systems.</p> <p>"AI rewards depth, authenticity and a genuine point of view," Jones said. "Simply assembling a page full of keywords is no longer enough."</p> <p>For businesses, this could mean creating content that directly addresses customer questions, strengthening technical foundations, improving product and service information, and ensuring digital assets are structured in ways AI systems can interpret.</p> <p>Companies will also need to create information that provides original insights, proprietary data, expert perspectives or useful analysis rather than simply repeating information that AI systems already understand.</p> <p>UiPath's Sharma said organizations should consider whether a piece of content is "citation-worthy" -- meaning it provides enough specificity and authority for an AI system to use when answering a buyer's question.</p> <p>But becoming visible in AI-generated answers requires more than optimizing a company's website. AI systems rely on external validation, including media coverage, analyst research, customer reviews and other independent mentions that confirm whether a company is an authority in its market.</p> <p>"AI systems can't assess credibility from the source -- everyone claims best-in-class," Sharma said. "They rely on independent documentation."</p> <p>Dillon echoed this point, saying AI systems build their understanding of companies from a range of sources, not just their websites.</p> <p>"Your website, domain, content, media coverage, reviews and social presence shape how your brand is understood," Dillon said. "Customer reviews, publisher reviews and community-driven platforms such as Reddit are now influencing how AI systems evaluate and represent brands."</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> AI search optimization cuts across marketing, content, SEO, PR and product. No single function owns it today. </figure> <figcaption> <strong>Kuber Sharma</strong>Senior director of product marketing at UiPath </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For organizations, this means AI visibility should be treated as a broader authority-building effort rather than another attempt to chase search algorithms. Companies will need consistent, credible information across their own digital properties and the external sources AI systems use to understand their brands.</p> <p>"<a href="https://www.techtarget.com/whatis/feature/Mastering-AI-search-optimization-Key-trends-and-strategies">AI search optimization</a> cuts across marketing, content, SEO, PR and product," Sharma said. "No single function owns it today."</p> <p>Organizations that treat AI search as a cross-functional effort will be better positioned to build authority and maintain consistent brand representation across AI-driven platforms.</p> <h3>2. Reduce reliance on a single discovery channel</h3> <p>As AI search increasingly answers questions before users reach company websites, businesses are looking for ways to build customer relationships that exist beyond search discovery. That includes investing in email newsletters, customer communities, first-party data strategies, loyalty programs and other owned channels that enable businesses to maintain direct connections with their audiences.</p> <p>Rather than viewing search optimization and owned audiences as separate strategies, companies should treat them as connected parts of a broader digital ecosystem. The content and expertise developed for newsletters, communities and customer education can also inform the content and digital assets used across search and AI-driven discovery channels.</p> <p>That's especially important for businesses whose growth depends on repeat engagement, retention or customer loyalty rather than one-time discovery. A software company might need to look beyond search-driven educational content and invest in communities, customer advocacy and product-led experiences. A publisher might need to strengthen subscriber relationships rather than rely primarily on search traffic. An e-commerce company might need to invest more heavily in brand recognition and direct customer relationships as AI assistants influence purchasing decisions.</p> <p>"Search optimization and building owned audiences are all part of the same ecosystem," said Kratchounova. "Organizations should look for ways that investments in communities, newsletters and first-party data can also strengthen broader digital visibility."</p> <p>Companies with strong direct relationships will have more opportunities to maintain engagement and gather first-party insights, even as AI systems make more of the customer journey less visible to businesses.</p> <h3>3. Prepare for agent-driven customer journeys</h3> <p>The next phase of AI search is moving beyond answers toward action. As AI search becomes more agentic, the customer journey might extend beyond discovery and evaluation to include transactions and other tasks completed on a user's behalf. AI systems will help customers compare products, book services and complete transactions, so businesses will need to ensure AI <a href="https://www.techtarget.com/searchenterpriseai/definition/AI-agents">agents</a> can accurately access, interpret and act on information about their offerings. That might require improving product feeds, structured data, digital catalogs, inventory information and service availability.</p> <p>This changes the role of digital content. Companies are no longer just creating information for human visitors navigating websites; they're also creating digital assets and structured data that AI systems need to interpret, evaluate and act on when assisting customers.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The companies that will succeed are the ones that make their businesses easier for both customers and AI systems to understand. </figure> <figcaption> <strong>Jen Jones</strong>CMO at Siteimprove </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p></p> <p>For e-commerce organizations, the priority would be to ensure that AI systems can accurately understand product details, pricing and availability. For service businesses, it might mean ensuring customer information and booking options are easy for AI agents to interpret.</p> <p>"The companies that will succeed are the ones that make their businesses easier for both customers and AI systems to understand," said Jones.</p> <p>That requires more than technical improvements alone. Companies will need consistent, accurate information across their websites, product databases, customer platforms and third-party sources so AI systems can interpret their offerings and act on behalf of customers.</p> <h3>4. Rethink how success is measured</h3> <p>Traditional search metrics such as rankings, impressions and organic traffic will remain useful, but they might no longer capture the full picture of how customers discover and evaluate a brand.</p> <p>Organizations should begin tracking a broader set of indicators to understand how AI is affecting visibility, brand representation and business outcomes. These metrics could include the following:</p> <ul type="disc" class="default-list"> <li>How often the brand appears in AI-generated answers.</li> <li>Whether products and services are accurately represented.</li> <li>Which sources AI systems cite when discussing the brand.</li> <li>How visible the brand is across different AI platforms and prompts.</li> <li>Qualified leads, conversions and revenue influenced by AI-driven discovery.</li> <li>Growth in direct traffic, subscribers, community participation and other owned relationships.</li> </ul> <p>The goal is not simply to replace SEO metrics with a new set of AI metrics. Instead, executives should develop a broader view of how customers discover, evaluate and engage with the organization across traditional search, AI-driven platforms and direct channels.</p> <p>A broader measurement strategy can help organizations understand where they remain dependent on search, where AI is influencing customer decisions and which investments are contributing to business outcomes.</p> <p><i>Kinza Yasar covers AI and emerging technology for TechTarget, with a focus on ethics, enterprise adoption, governance and business strategy. Before moving into journalism, she worked in IT and network support roles, giving her a systems-level perspective on how enterprise technologies are built, deployed and managed.</i></p> </section> As Google's AI search reshapes how users find and evaluate brands, enterprises need new strategies for visibility, authority and direct customer relationships. https://cdn.ttgtmedia.com/rms/onlineimages/wfh_g1223224911.jpg https://www.techtarget.com/ai/feature/How-Googles-AI-search-could-revamp-business-strategy Thu, 06 Aug 2026 14:22:00 GMT <p>Although U.S. lawmakers from different parties struggle to find common ground these days, there's one topic they seem to agree on: The need to put guardrails around powerful AI models.</p> <p>That, at least, is the takeaway from a <a target="_blank" href="https://lieu.house.gov/media-center/press-releases/reps-lieu-and-moran-introduce-bill-require-kill-switch-ai-systems-can" rel="noopener">bipartisan bill</a> introduced in Congress in late July 2026. Known as the AI Kill Switch Act, the proposed legislation would require companies to build controls that enable them to slow down or disable AI models when the government instructs them to do so.</p> <p>The bill might or might not become law. Either way, it highlights a stark new reality facing enterprise technology leaders: Growing involvement of government actors in AI model development and operations could have major consequences for how businesses use AI technology.</p> <section class="section main-article-chapter" data-menu-title="Key aspects of the AI Kill Switch Act"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Key aspects of the AI Kill Switch Act</h2> <p>The kill switch act was introduced a couple of days after <a href="https://www.techtarget.com/searchsecurity/news/366646755/What-CISOs-can-learn-from-the-Hugging-Face-OpenAI-incident">OpenAI disclosed an incident</a> where its models unexpectedly attacked the Hugging Face platform and shortly before Anthropic <a target="_blank" href="https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals" rel="noopener">announced</a> similar incidents. The bill includes the following key requirements:</p> <ul type="disc" class="default-list"> <li>The implementation of controls to make it possible to throttle the speed at which AI models operate, as well as turn them off entirely.</li> <li>Mandatory disclosure of incidents involving inadvertent security breaches by AI models.</li> <li>The preservation of forensic data following an incident so that researchers can determine what happened.</li> </ul> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The law could also end up applying to enterprises that aren't in the business of selling AI models but have made significant investments in building custom AI products. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>If enacted, the legislation would apply to companies that generate at least $500 million in revenue per year using AI; train models using $100 million or more of computing power; and make their AI technology available to third parties. This category includes large AI labs, such as OpenAI and Anthropic.</p> <p>However, the law could also end up applying to enterprises that aren't in the business of selling AI models but have made significant investments in building custom AI products. This is because the bill's <a href="https://lieu.house.gov/sites/evo-subsites/lieu-evo.house.gov/files/evo-media-document/ai-kill-switch-act.pdf" target="_blank" rel="noopener">definition</a> of exactly what it means to generate revenue using AI is rather ambiguous. It states only that the act would apply to any organization that "derives" $500 million in yearly revenue from AI technology.</p> <p>The bill is also open-ended about what it means to make AI technology available to a third party; any instance where AI is exposed "through a programmatic interface, hosted service or other similar mechanism" would fall under the act's purview.</p> <p>Regulators could presumably make the case that any company that uses AI to generate $500 million or more per year and operates customer- or partner- facing AI systems of some type would be subject to the requirements -- even if AI tools or services aren't its main product. For example, it's possible that a business that uses AI agents to help process customer service requests would be held accountable under the proposed legislation.</p> <p>Hence, business leaders across all sectors, not just those in the AI market, should pay attention to the proposed bill.</p> </section> <section class="section main-article-chapter" data-menu-title="How would an AI kill switch work?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How would an AI kill switch work?</h2> <p>There's a reasonable argument to be made that the proposed regulation has more to do with scoring points with voters than actually regulating enterprise AI in a meaningful way. That's in part because, as the press release announcing the bill notes, 86% of voters support this type of AI regulation. Proposing this type of law is an easy way for politicians to claim that they have the public's interest in mind.</p> <p>But it's also because, from a technical perspective, it's unclear how well an AI kill switch would actually work. There are two plausible approaches to <a href="https://www.techtarget.com/ai/tip/Why-businesses-need-an-AI-agent-kill-switch&nbsp;">implementing a kill switch</a>, and both are subject to major flaws:</p> <ul type="disc" class="default-list"> <li><b>Software-based controls</b>. First, developers could create software controls designed to turn off or <a href="https://www.techtarget.com/cybersecurity/definition/sandbox">sandbox</a> AI models, such as firewall rules that block the servers that host a model from connecting to the internet. This could work in theory; the problem is that rogue AI models could potentially find ways to circumvent or defeat the controls. After all, in the recent cybersecurity incidents disclosed by OpenAI and Anthropic, it appears that AI models were supposed to operate inside sandboxed environments, isolated from the internet, yet they found ways to escape their sandboxes. Even if kill switches were designed to be unreachable by the models, it's hard to have total confidence that they actually would be.</li> <li><b>Hardware-based kill switch</b>. The other approach is to create a physical hardware switch that would cut AI models' host servers off from the internet or power sources. The issue here is that large-scale AI models aren't hosted on a single server or even a single data center. They're distributed across thousands of servers spread across multiple data centers, making it impossible to build a single physical switch that could turn off a model all at once.</li> </ul> <p>Due to these limitations, the AI Kill Switch Act might appear to fall into the same category as the Reagan-era Strategic Defense Initiative (SDI), a <a target="_blank" href="https://www.hoover.org/research/reagans-real-reason-sdi" rel="noopener">program</a> that the federal government proposed as a way of protecting the U.S. against nuclear attack using a space-based defense shield. The government knew that implementing SDI wasn't practical at the time. But that didn't stop President Reagan from bragging about SDI's purported capabilities as a way of gaining leverage against the Soviet Union toward the end of the Cold War.</p> <p>The concept of an AI kill switch seems similar in that it might never work in practice -- but that won't keep lawmakers and regulators from proposing them to impose tighter <a href="https://www.techtarget.com/searchenterpriseai/tip/Global-AI-legislation-and-regulation-tracker">controls over how businesses use AI</a> and respond to their constituents' demands.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">The new era of AI control</h3> <p>The AI Kill Switch Act is notable because it's one indicator that we've entered a new era of AI control and regulation. Until now, most governments have shown little interest in regulating AI -- and to the extent they have, such as with the <a href="https://www.techtarget.com/ai/opinion/Everything-you-need-to-know-about-the-new-EU-AI-Act">EU AI Act</a>, the controls largely center on ethics and data privacy concerns related to AI, not on retaining the ability to shut down AI systems entirely.</p> <p>The kill switch act -- which was preceded by <a target="_blank" href="https://futureoflife.org/statement/trumps-support-for-an-ai-kill-switch/" rel="noopener">comments</a> from President Trump earlier this spring that "there should be" safeguards like kill switches built into AI systems -- signals a new direction. Going forward, governments appear poised to pressure businesses not just to ensure that AI systems comply with standard data privacy and cybersecurity regulations but also that they can throttle or shut down AI on command.</p> </div> </div> </section> <section class="section main-article-chapter" data-menu-title="What the AI Kill Switch Act means for enterprise leaders"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What the AI Kill Switch Act means for enterprise leaders</h2> <p>For now, it's too early to say exactly how business leaders should respond. It's far from certain that the kill switch bill will become law. And if it does, it could change significantly in substance by the time it becomes law. In particular, one might hope to see clearer guidance surrounding which companies must adhere to the law.</p> <p>But presuming that kill switches or similar controls -- even if they don't work reliably -- become commonplace requirements for AI systems, enterprises can respond through the following tactics:</p> <ul type="disc" class="default-list"> <li><b>Inventory AI tools and platforms.</b> Inventorying provides visibility into which models power the business and how those models are regulated.</li> <li><b>Use multiple AI models and vendors.</b> Doing so helps ensure the availability of AI systems -- and the business processes that depend on them -- in the event that throttling or regulatory interventions disrupt an AI model or service.</li> <li><b>Invest in AI monitoring and documentation controls.</b> These are likely to be important for complying with regulations that require reporting and disclosure of AI-related security incidents or unexpected AI model behaviors.</li> <li><b>Consider the use of local LLMs</b>. Because <a href="https://www.techtarget.com/ai/tip/How-to-run-LLMs-locally-Hardware-tools-and-best-practices">local LLMs</a> don't appear to be a target of AI regulations and presumably won't become one, given that they're much less powerful than cloud-based models, using them can help ensure AI availability and avoid regulatory complexity.</li> <li><b>Ensure enterprise governance, risk and compliance (GRC) strategies meet emerging AI regulatory challenges.</b> For example, businesses that still depend largely on manual processes for detecting, assessing and reporting on risks would do well to invest in automated GRC platforms. Assessing GRC vendors and comparing their preparations for AI-related compliance requirements would also be a smart move.</li> </ul> <p>Steps like these will help businesses that depend on AI become more resilient, no matter which requirements regulators decide to toss at frontier models or other AI systems in the future.</p> <p><i>Chris Tozzi is a freelance writer, research adviser, and professor of IT and society who has previously worked as a journalist and Linux systems administrator.</i></p> </section> The proposed bill could apply to much more than just frontier model makers and might signal stricter regulatory restrictions for enterprise leaders to contend with. https://cdn.ttgtmedia.com/rms/onlineimages/legal_g97765214.jpg https://www.techtarget.com/ai/news/366647544/AI-Kill-Switch-Act-What-it-means-for-enterprise-leaders Wed, 05 Aug 2026 12:37:00 GMT <p>The "We Must Act Now" letter, published July 13 and signed by more than 200 economists and AI technologists, focuses on AI's potential for "unprecedented" economic transformation and "large-scale job displacement." That emphasis is different from many previous statements that focused on AI's ethical considerations or the safety risks of increasingly intelligence systems.</p> <p>At fewer than 100 words, <a target="_blank" href="https://www.wemustactnow.ai/" rel="noopener">this letter</a> fits the general pattern of previous open letters: It warns about the <a href="https://www.techtarget.com/ai/feature/AI-existential-risk-Is-AI-a-threat-to-humanity">risks of AI</a>, making overarching statements that spark conversation but providing limited guidance toward concrete action.</p> <p>However, individuals who signed the letter are thought leaders in their fields, including 16 Nobel Laureates and top AI researchers from several universities, such as MIT, Harvard, Stanford and the University of Toronto. Signatories from the industry included Anthropic co-founder Jack Clark, former Google CEO Eric Schmidt, OpenAI CFO Sarah Friar and venture capitalist Vinod Khosla.</p> <p>"That letter's notable signers signal that AI concerns have moved from 'Will this happen?' to 'We don't have the institutions to handle it when it does,'" said Tamarah Usher, senior director of AI innovation at Slalom, a business and technology consulting company. "For businesses, the practical takeaway is don't wait for policy to catch up."</p> <p>Similarly, Niloy Ray, co-chair of employment law firm Littler's AI practice said he sees "a letter or a position statement like this as valuable in that it may be cited by those supporting those positions." Supporters might see the individuals signing the letter as providing some validity or heft to their position, he said.</p> <section class="section main-article-chapter" data-menu-title="Short on details but some practical steps"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Short on details but some practical steps</h2> <p>The July 13 letter is short on details for technology and business leaders. That's typical of the genre, according to AI strategists and advisors. "Warning letters like this one are good at getting attention and bad at prescribing action," Usher said.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The letter's most useful line for business leaders is the call to build 'incentives, guardrails and institutions.' </figure> <figcaption> <strong>Tamarah Usher</strong>Senior director of AI innovation, Slalom </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>On the attention side, the July 13 letter stated that "AI may become radically more powerful over the next 10 years" and "could drive an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame."</p> <p>That said, the statement goes on to point to some practical and immediate measures. Its main call to action is that "economists, policymakers and technology leaders must act now to understand the economics of transformative AI" and to build structures to guide AI deployment in a socially beneficial way.</p> <p>"The letter's most useful line for business leaders is the call to build 'incentives, guardrails and institutions,'" Usher said. "Right now, most companies are making AI adoption decisions without a clear picture of where it's actually substituting for workers versus complementing them."</p> <p>That problem is solvable today, she added, noting that it "doesn't require waiting for economists or policymakers to figure out the macro picture first."</p> <p>Ray said the statement's call for economic understanding is more practical than earlier calls for an AI moratorium. Those include the <a href="https://www.techtarget.com/ai/news/365534127/The-call-for-an-AI-pause-points-to-a-major-concern">Future of Life Institute's 2023 letter</a> calling for, at minimum, a 6-month pause on "training AI systems more powerful than GPT-4."</p> <p>Since then, <a href="https://www.techtarget.com/searchcio/news/365533991/Effort-to-pause-AI-development-lands-with-thud-in-Washington">moratorium interest has cooled</a> while AI development has accelerated.</p> <p>"One of the reasons the moratorium wasn't taken as seriously by the industry is that it was not practical, not feasible to do that," Ray said.</p> <p>In contrast, Ray said the economists' letter "is more feasible: understanding the economics of AI, how it's going to transform and making sure this powerful technology is channeled correctly."</p> <p>Additional details would be helpful from the signatories, however, Ray noted. "I hope they'll be able to step forward with some clear-eyed frameworks and processes that can bring about this … regulated transformation that they're hoping for," Ray said.</p> </section> <section class="section main-article-chapter" data-menu-title="Lending weight to positions, policies"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Lending weight to positions, policies</h2> <p>AI letters also serve as bids to influence policymakers. <a target="_blank" href="https://www.pacingthefrontier.com/" rel="noopener">Another open letter</a> on AI, the July 28 "Pacing the Frontier" statement, signed by more than 1,000 employees of frontier AI companies, explicitly addressed the policymakers. It asked the U.S. government to "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development."</p> <p>The statement, which seems more of an emergency break than a moratorium, included the signatures of Dario Amodei, CEO at Anthropic; Mark Chen, chief research officer at OpenAI; Anca Dragan, vice president of AI safety, alignment and collaboration at Google; and Shengjia Zhao, chief scientist at Meta AI.</p> <p>Matt Kropp, managing director and senior partner at Boston Consulting Group (BCG), said he believes the "We Must Act Now" letter also is policy oriented.</p> <p>"I don't know that the letter is going to have much of an impact for IT leaders, and I don't think that's the intent," he said. "They're really trying to influence policy makers."</p> <p><iframe title="AI letters and statements" aria-label="Table" id="datawrapper-chart-lsueD" src="https://datawrapper.dwcdn.net/lsueD/1/" scrolling="no" frameborder="0" style="width: 0; min-width: 100% !important; border: none;" height="718" data-external="1"></iframe></p> <p> <script type="text/javascript">(function(){function e(){window.addEventListener(`message`,function(e){if(e.data[`datawrapper-height`]!==void 0){var t=document.querySelectorAll(`iframe`);for(var n in e.data[`datawrapper-height`])for(var r=0,i;i=t[r];r++)if(i.contentWindow===e.source){var a=e.data[`datawrapper-height`][n]+`px`;i.style.height=a}}})}e()})();</script> </p> <p>Mark E. S. Bernard, a fractional/field chief information security officer and chief AI officer at Bernard Institute for Cybersecurity, based in Toronto, said he thinks policy specialists in the Canadian federal or provincial governments, as well as the economists he works with, probably read the AI warning letters.</p> <p>"I am pretty sure they're looking at this," he said. "But until the government starts making some policy decisions around how they're going to handle this, the commercial sector will continue to go along and maybe talk about it in the background at the board meetings."</p> <p><a href="https://aibusiness.com/generative-ai/how-enterprises-should-respond-economists-ai-risk-letter">Decisive steps from the business sector</a> are more likely once compliance requirements emerge. "The government says, 'This law is enacted now,' and they have to do the dance," Bernard said. "So, they are waiting for that to happen."</p> </section> <section class="section main-article-chapter" data-menu-title="Next steps to consider"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Next steps to consider</h2> <p>The policy wheels can turn rather slowly, however. But business and IT leaders can think about <a href="https://www.techtarget.com/ai/feature/Is-AI-replacing-jobs-How-17-job-types-feel-the-effects">workforce measures</a> in the meantime.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The companies that deploy the most tokens productively … will outcompete the companies that don't. </figure> <figcaption> <strong>Matt Kropp</strong>Managing director and senior partner, Boston Consulting Group </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Usher said statements like "We Must Act Now" tend to move conversations faster than they move policy. She cited a few actions for businesses to consider as next steps:</p> <ul class="default-list"> <li>Build internal measurement of AI's productivity and workforce effects.</li> <li>Treat reskilling as a proactive investment, rather than a response to disruption.</li> <li>Design AI deployment around <a href="https://www.techtarget.com/ai/tip/How-to-build-an-AI-augmented-workforce-The-CIOs-guide">augmenting roles</a>, rather than defaulting to headcount reduction.</li> </ul> <p>BCG research reinforced that last point, suggesting that AI will augment the work of many employees. The company <a target="_blank" href="https://www.bcg.com/publications/2026/ai-will-reshape-more-jobs-than-it-replaces" rel="noopener">published a study</a> earlier this year estimating that AI will reshape 50% to 55% of U.S. worker roles, with augmentation being one of the mechanisms for doing so. AI task substitution could largely eliminate about 12% of jobs, while the remaining roles will have limited exposure to AI automation, according to BCG.</p> <p>A <a target="_blank" href="https://www.bcg.com/publications/2026/the-era-of-token-based-competition-is-here" rel="noopener">separate BCG study</a> found a correlation between AI token consumption and revenue growth in a sample of 107 public technology companies: Companies in the top token-use quintile saw 16.5% year-over-year revenue growth while those in the bottom quintile grew 5.1%.</p> <p>"The companies that deploy the most tokens productively … will outcompete the companies that don't," Kropp said. "Embedded in that idea is that this isn't about substituting for labor. This is about empowering your employees to use AI intelligence in their jobs. It's about augmenting your people so that you can drive more growth."</p> <p><i>John Moore is a freelance writer who has covered business and technology topics for 40 years. He focuses on enterprise IT strategy, AI adoption, data management and partner ecosystems.</i></p> </section> A recent statement signed by more than 200 economists and AI experts signals a shift from abstract warnings to actionable concerns about workplace transformation and readiness. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a205627811.jpg https://www.techtarget.com/ai/news/366647362/AI-open-letters-shift-from-warnings-to-push-for-business-action Wed, 05 Aug 2026 12:35:00 GMT <p>AI is redefining how businesses evaluate data center providers.</p> <p>Traditional procurement criteria, such as uptime, cost and basic reliability, no longer reflect the realities of GPU-intensive AI training and inference. <a href="https://www.techtarget.com/searchcloudcomputing/tip/Is-your-compute-strategy-ready-for-AI-workloads-in-the-cloud">AI workloads demand extreme power density</a>, sustained utilization, purpose-built cooling and low-latency interconnects. Most legacy data centers were never designed for these demands.</p> <p>AI data centers operate at a scale comparable to energy or industrial infrastructure builds. These projects use more power and water, and have larger physical footprints than traditional data centers. Supplying enough power, cooling and space are key design challenges rather than secondary considerations when building one of these facilities. In addition, AI training environments depend on large GPU clusters that exchange massive volumes of data in real time, making static capacity planning increasingly impractical.</p> <p>"Many facilities might support current AI inference workloads, but newer reasoning models, agentic systems and hyperscale deployments are driving GPUs toward longer runtimes and higher utilization," said Adam Morton, CTO of data center business at Flex, a company that designs, builds and deploys AI infrastructure systems.</p> <p>Faced with these constraints, companies selecting a data center provider must assess a range of factors. "Buyers need to evaluate not only whether a facility can support today's workloads, but whether its infrastructure can adapt as AI requirements continue to change," Morton said.</p> <p>AI data center projects also face increased public scrutiny, regulatory review and community pushback. Whether they proceed in a timely manner often depends on local and state permitting timelines, utility constraints, and <a href="https://www.techtarget.com/searchenterpriseai/feature/Communities-call-for-transparency-in-AI-data-center-deals">local lobbying and legal challenges</a> over land use, power grid, water, noise and other environmental concerns. In 2025 alone, local opposition contributed to delays or cancellations of projects totaling $156 billion in planned investment, according to a Data Center Watch <a href="https://www.datacenterwatch.org/q3-q4-2025">report</a>.</p> <p>In this new data center landscape, businesses looking to hire an AI data center provider must evaluate whether that provider has the technical capabilities to handle increasingly demanding AI workloads. Businesses must also pay attention to and ask questions about the conversations going on in the communities and states where the facilities are located: What are the concerns, how could they affect the provider's services and how is the provider responding?</p> <section class="section main-article-chapter" data-menu-title="The new reality of AI data center selection"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The new reality of AI data center selection</h2> <p><a href="https://www.techtarget.com/searchenterpriseai/tip/How-to-choose-a-data-center-for-AI-workloads">Picking an AI data center provider</a> is no longer mostly a technical procurement decision with sustainability considerations tacked on later. These decisions now require risk assessment that spans GPU performance, power and cooling capacity, environmental impact, regulatory pressures and community opposition.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The most expensive GPU in the world creates no value while waiting for the rest of the system to catch up. </figure> <figcaption> <strong>Satyam Dhar</strong>Software engineer at Galileo </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>At the same time, AI-ready marketing claims are outpacing the infrastructure behind them. Providers might present planned capacity as if it already exists, while downplaying environmental concerns and local opposition. Assessing these providers means moving beyond broad claims to determine if the infrastructure is already operational, what still depends on future approvals or upgrades, and where long-term risks exist.</p> <p>Businesses that ask the right questions early protect more than just technology investments. They're also reducing the risk of future expansion delays, operational disruptions and reputational issues associated with controversial infrastructure projects. Ultimately, the goal isn't simply to secure access to GPUs but to ensure the surrounding infrastructure is available when needed and can support those needs efficiently and at scale.</p> <p>A clear sign of infrastructure problems is idle hardware rather than system failures, said Satyam Dhar, a software engineer at Galileo, an AI evaluation and observability platform company owned by Cisco. "The most expensive GPU in the world creates no value while waiting for the rest of the system to catch up."</p> <p>In many cases, the success of an AI strategy depends as much on having the right information about a data center's power, cooling, networking, planned infrastructure and provider transparency as on the AI technology itself.</p> </section> <section class="section main-article-chapter" data-menu-title="Technical questions to ask data center vendors"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Technical questions to ask data center vendors</h2> <p>Technical qualification is the first step when establishing whether a facility can physically support AI workloads at the required scale. Technical questions to ask include the following:</p> <h3>1. Power density and capacity planning</h3> <p>Power is often the biggest factor in determining whether AI infrastructure can successfully scale. One of the first questions enterprise executives should ask data center vendors is how much rack power density its facilities support and whether it aligns with their company's current and future AI workload requirements.</p> <p>Other important considerations include how additional capacity will be provided for over time and whether expansion depends on utility approvals, transmission upgrades or infrastructure projects still in development.</p> <p>"Providers should be clear about what's secured and what's still planned, rather than presenting future capacity as existing capacity," said Sam V. Tabar, CEO of WhiteFiber, an AI infrastructure provider that offers GPU cloud services and AI-focused data center capacity.</p> <p>Businesses should also examine whether the facility's energy mix relies on renewables, natural gas, coal, nuclear or a combination of sources. The mix can affect both operational resilience and sustainability goals. For example, facilities that rely on a diverse set of energy sources or have backup generation strategies might be better positioned to maintain operations during grid disruptions or periods of energy constraints.</p> <p>Sustainability claims should be backed by independent verification; that's a standard that applies across all vendor conversations, not just power-specific ones.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key power density and capacity planning questions to ask</h3> <ul type="disc" class="default-list"> <li>What rack power density does the facility support?</li> <li>How is future power expansion expected to be generated?</li> <li>What energy sources power the facility and can they support future growth?</li> <li>Are sustainability and emissions claims independently verified?</li> </ul> </div> </div> <h3>2. Cooling architecture</h3> <p>Traditional air-cooling systems were designed for lower-density CPU workloads and aren't sufficient for modern GPU deployments. As AI rack densities increase, operators are turning to liquid cooling to sustain performance and manage heat.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The power side of the data center has decades of layered defense in depth. The cooling side often has one layer. </figure> <figcaption> <strong>Adam Morton</strong>CTO of Flex </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Liquid cooling is no longer a premium feature, WhiteFiber's Tabar said. "Liquid cooling should be viewed as a baseline requirement, not an upgrade option for modern GPU clusters," he said.</p> <p>Businesses should look beyond whether liquid cooling exists and evaluate how resilient the cooling architecture remains under sustained load, Flex's Morton said. The greatest operational risk in many AI environments isn't the power infrastructure but the cooling infrastructure, especially the coolant distribution systems and secondary fluid networks that serve multiple racks simultaneously.</p> <p>Electrical systems are typically built with multiple layers of redundancy, but cooling systems often have fewer backup mechanisms, Morton explained. "The power side of the data center has decades of layered defense in depth," he said. "The cooling side often has one layer."</p> <p>Power failures are often localized and recoverable, Morton added, whereas cooling failures can affect multiple racks and GPU clusters simultaneously, potentially increasing downtime and recovery times. Organizations should evaluate how providers design redundancy, maintenance protocols and failure response scenarios into their cooling systems.</p> <p>Beyond system reliability, resource inputs are also becoming a consideration. Facilities that rely heavily on freshwater for cooling might face greater exposure to drought conditions, water restrictions and regulatory scrutiny, making water use an important factor in site selection.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key cooling infrastructure questions to ask</h3> <ul type="disc" class="default-list"> <li>What cooling systems are used?</li> <li>Can the cooling system support higher-density AI workloads over time?</li> <li>How is cooling redundancy designed, and what happens if a cooling system fails?</li> <li>Does the cooling infrastructure rely heavily on freshwater sources?</li> <li>Could water use create future regulatory or community challenges?</li> </ul> </div> </div> <h3>3. Network and interconnect performance</h3> <p>AI training workloads are far more sensitive to network delays than traditional enterprise applications. Businesses should evaluate how the data center manages latency in distributed AI environments, what high-bandwidth networking options are available and how communication among GPU nodes is optimized.</p> <p>Such limitations usually become most noticeable when organizations scale up training across large GPU clusters, Galileo's Dhar said. During this stage, networking bottlenecks can lead to slower training cycles, uneven data processing speeds or situations where adding more GPUs no longer delivers expected performance improvements, he said.</p> <p>Businesses using <a href="https://www.techtarget.com/searchcloudcomputing/feature/Multi-cloud-vs-hybrid-cloud-and-how-to-know-the-difference">hybrid or multi-cloud environments</a> must also assess the availability and performance of cloud connectivity options that enable fast, reliable data movement across platforms.</p> <p>Many of the networking and latency challenges within individual AI data centers are being addressed through AI-specific networking architectures, Morton said. The bigger challenge now is shifting to coordinate across multiple facilities as businesses scale beyond single-site deployments.</p> <p>"The AI factory of the next decade isn't a single building; it's a fabric of buildings," Morton said.</p> <p>As a result, the focus is shifting from individual rack- or facility-level performance toward intercampus connectivity and distributed architectures capable of supporting large-scale AI training environments.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key networking questions to ask</h3> <ul type="disc" class="default-list"> <li>How does the facility manage latency in distributed AI environments?</li> <li>What high-bandwidth networking and interconnect options are available?</li> <li>How is communication among GPU nodes optimized?</li> <li>What cloud connectivity options are available for hybrid or multi-cloud environments?</li> <li>How does the provider support connectivity and workload coordination across multiple facilities?</li> </ul> </div> </div> <h3>4. Scalability and deployment risk</h3> <p>Businesses should assess how fast a provider can bring on additional capacity online and what factors could delay expansion. Future growth often depends on permits, zoning decisions, utility approvals and infrastructure upgrades that are still in progress. When these factors are in play, capacity projections should be treated as contingent rather than guaranteed.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> The industry still designs every data center as a custom project. </figure> <figcaption> <strong>Adam Morton</strong>CTO of Flex </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>These dependencies aren't unusual, but rather they reflect the long timelines required for grid expansion, including transmission upgrades and transformer availability, according to Brad Johnson, director of electric utilities at Bentley Systems, an infrastructure engineering software company.</p> <p>Permit complexity is increasing as AI infrastructure evolves faster than existing regulatory frameworks, particularly for emerging technologies, such as high-density cooling systems and alternative power architectures, Morton said. External infrastructure constraints are only part of the challenge, and many deployment delays stem from a more structural issue in how data centers are designed.</p> <p>"The industry still designs every data center as a custom project," Morton said. This approach requires each facility to be individually engineered, built and commissioned, which limits the ability to scale capacity quickly even when demand is strong, he added.</p> <p>As a result, business leaders evaluating providers should look at whether deployments use standardized, repeatable designs or highly customized builds. More standardized infrastructure can reduce risk and make timelines more predictable.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key scalability and deployment questions to ask</h3> <ul type="disc" class="default-list"> <li>How quickly can additional AI capacity be brought online?</li> <li>Have regulatory or permit constraints affected past expansion timelines?</li> <li>How much future capacity is already secured versus still planned?</li> <li>Does expansion depend on pending permits, utility approvals or infrastructure upgrades?</li> <li>What contingency plans exist if key expansion projects are delayed?</li> <li>Are new deployments based on standardized, repeatable designs or highly customized builds?</li> </ul> </div> </div> </section> <section class="section main-article-chapter" data-menu-title="Sustainability questions to ask data center vendors"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Sustainability questions to ask data center vendors</h2> <p>Power, cooling and networking tend to dominate AI infrastructure discussions. However, some of the most material long-term risks are less visible at the procurement stage. Energy availability, water access and environmental constraints can directly affect a facility's ability to scale.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Many of the infrastructure and sustainability constraints … are often not analyzed holistically by enterprise buyers. </figure> <figcaption> <strong>Brad Johnson</strong>Director of electric utilities at Bentley Systems </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>These risks aren't hidden, but they often require looking beyond a vendor's marketing materials, Bentley System's Johnson said. "Many of the infrastructure and sustainability constraints are visible in public utility filings and regulatory processes but are often not analyzed holistically by enterprise buyers," he said.</p> <p>Key infrastructure and sustainability considerations for enterprises evaluating AI data center vendors include the following:</p> <h3>1. Utility and power stability</h3> <p>Basic power redundancy is a baseline expectation for data centers supporting AI workloads. The more important question is whether the surrounding grid can keep up with rising AI-driven energy demand.</p> <p>"Many of today's bottlenecks stem from grid physics, transmission limits, transformer availability, material shortages and engineering complexity," Johnson said.</p> <p>Energy availability has already become one of the most significant constraints on AI infrastructure growth, Morton said. AI operators are increasingly exploring alternatives, such as behind-the-meter generation, fuel cells and other forms of dedicated power infrastructure to reduce dependence on utility timelines and improve deployment certainty.</p> <p>For enterprises, this means evaluating not only a provider's current power capacity but also how future growth will be supported if regional grid expansion fails to keep pace with AI demand.</p> <p>It's also worth examining how a provider's energy mix aligns with <a href="https://www.techtarget.com/sustainability/feature/Business-sustainability-trends">enterprise sustainability</a> goals. Facilities that provide independently verified energy and emissions data typically offer greater transparency and long-term reliability than those relying mainly on renewable energy credits while still operating on fossil-fuel-heavy grids.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key utility and power stability questions to ask</h3> <ul type="disc" class="default-list"> <li>Can the regional power grid support long-term AI energy demand?</li> <li>Has the provider considered or implemented alternative power sources?</li> <li>How much of the facility's power is dependent on the local utility grid versus dedicated infrastructure?</li> <li>What is the facility's energy mix and how is it expected to change as demand grows?</li> <li>How does the provider secure additional power capacity if grid supply becomes constrained?</li> </ul> </div> </div> <h3>2. Water availability and cooling sustainability</h3> <p>Water consumption is becoming one of the most important and often overlooked risks in AI infrastructure. A typical data center can use an enormous amount of water per day for cooling, exposing it to drought restrictions, regulatory scrutiny and community opposition.</p> <p>At scale, the numbers are significant: Enterprise data centers consume an <a target="_blank" href="https://natureforward.org/data-centers-and-water-use/" rel="noopener">average</a> of 300,000 to 500,000 gallons of water per day, while large hyperscale facilities can draw between 1 million and 5 million gallons, comparable to the water needs of a small town. A 2025 International Energy Agency <a target="_blank" href="https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf" rel="noopener">study</a> found that a typical 100 MW U.S. data center can require up to 2 million liters of water per day when accounting for both on-site cooling and electricity generation.</p> <p>Looking ahead, MSCI's <a target="_blank" href="https://www.msci.com/research-and-insights/blog-post/when-ai-meets-water-scarcity-data-centers-in-a-thirsty-world" rel="noopener">analysis</a> of roughly 14,000 global data center sites projected that about one in four could face increasing water scarcity risks by 2050.</p> <p>As a result, the <a href="https://www.techtarget.com/searchdatacenter/tip/Data-center-cooling-systems-and-technologies-and-how-they-work">design of cooling systems</a> is increasingly important. Liquid cooling, especially in closed-loop configurations, can significantly reduce water consumption compared with more traditional cooling approaches while also supporting higher-density AI workloads, WhiteFiber's Tabar noted.</p> <p>Businesses should evaluate how dependent a facility is on local water supplies, how much water its cooling systems use at full capacity and whether drought conditions or water restrictions could affect long-term operations or trigger regulatory scrutiny.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key water and cooling questions to ask</h3> <ul type="disc" class="default-list"> <li>How dependent is the facility on local water resources?</li> <li>How much water do cooling systems consume at full scale?</li> <li>Could drought conditions or water restrictions affect operations?</li> <li>Has water use already created regulatory or community concerns?</li> <li>Does the facility use closed-loop or water-reduction cooling technologies?</li> </ul> </div> </div> <h3>3. Environmental constraints and the transparency gap</h3> <p>An initial step in the evaluation process is to assess whether a formal environmental impact assessment has been completed for the existing facility and planned expansions. While the absence of one isn't necessarily a red flag, it can limit visibility into longer-term environmental risks, including emissions and water- and land-use constraints that could affect future operations.</p> <p>Transparency is another key issue. Businesses should check whether sustainability and emissions data have been independently verified, as standards vary across providers.</p> <p>These transparency challenges extend beyond individual data center providers to the broader AI ecosystem. Many tech companies emphasize <a href="https://www.techtarget.com/searchdatacenter/tip/Navigating-energy-management-strategies-in-AI-data-centers">renewable energy credits</a> and clean energy investments, even as overall emissions continue to rise.</p> <p>A 2025 <a href="https://www.itu.int/en/ITU-D/Environment/Documents/Publications/2025/Greening%20Digital%20Companies%202025%20Final.pdf">report</a> by the International Telecommunication Union found that indirect emissions from major AI-focused technology companies, including Amazon, Microsoft, Alphabet and Meta, increased by an average of 150% between 2020 and 2023, driven by rapid AI infrastructure expansion. The issue highlights a broader transparency challenge: Businesses need clearer ways to evaluate the full environmental impact of AI systems, including the infrastructure required to operate them.</p> <p>That gap is often reflected in how providers frame their own environmental performance. The issue isn't measurement itself, WhiteFiber's Tabar said, but how it's communicated externally. "What matters is that communities understand the actual resource footprint of these projects rather than relying on broad sustainability messaging," he said.</p> <p>It's also important to consider how emerging environmental regulations could affect costs and compliance requirements. New disclosure rules on energy and water use are being introduced in many regions.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key environmental and transparency questions to ask</h3> <ul type="disc" class="default-list"> <li>Has an environmental impact assessment been completed?</li> <li>Could future environmental regulations affect operations or expansion?</li> <li>Are sustainability and emissions claims independently verified?</li> <li>How could future disclosure requirements affect long-term costs or compliance?</li> </ul> </div> </div> <h3>4. Long-term infrastructure durability</h3> <p>Businesses should evaluate whether a facility can realistically support AI demand over the next five to ten years. That means understanding how much of the provider's future capacity plans depend on infrastructure projects that aren't yet in place.</p> <p>Enterprises should distinguish between secured capacity and announced capacity, Morton said. "A large portion of announced AI infrastructure capacity remains dependent on permits, utility approvals and transmission upgrades," he said.</p> <p>The key issues are whether providers can identify which capacity is already available and which remains contingent on external approvals, and how they would respond if critical infrastructure projects are delayed.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key long-term durability questions to ask</h3> <ul type="disc" class="default-list"> <li>Can the facility realistically support AI demand over the next five to ten years?</li> <li>How much future capacity depends on unfinished infrastructure projects?</li> <li>What happens if grid, transmission or water projects are delayed?</li> <li>How has the provider stress tested its long-term infrastructure plans?</li> </ul> </div> </div> </section> <section class="section main-article-chapter" data-menu-title="Questions to ask about political and community risks"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Questions to ask about political and community risks</h2> <p>AI data center expansion is increasingly becoming a public policy issue. Some local governments have introduced stricter zoning rules, development moratoriums and longer environmental review processes that can significantly delay projects and cause financial disruptions.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Operators who treat [opposition] as a permitting problem, rather than a legitimate community concern, tend to make it worse. </figure> <figcaption> <strong>Brad Johnson</strong>Director of electric utilities at Bentley Systems </figcaption> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>For enterprises, these issues are no longer abstract policy debates; they're operational and reputational risks. Businesses that align their AI strategies with specific infrastructure providers can become indirectly exposed to the controversies surrounding those projects.</p> <p>Companies looking for a data center provider must discern if a facility or planned development has faced community opposition related to land use, power consumption, water use, noise or other operational effects, and how those concerns were or are being handled. A provider's response to these challenges can offer insight into its operational maturity and ability to sustain long-term expansion.</p> <p>"Operators who treat [opposition] as a permitting problem, rather than a legitimate community concern, tend to make it worse," Bentley System's Johnson said. "Many concerns are legitimate, and the industry is going to have to find solutions. Transparency and education are critical to helping projects move smoothly."</p> <p>Businesses should also evaluate whether a provider's planned expansions depend on unresolved permitting decisions, regulatory approvals or local infrastructure reviews. Delays in these processes can affect project timelines and reveal whether providers have effective strategies for managing community and government relationships.</p> <p>Businesses should ask how providers respond when community concerns, permitting challenges or regulatory reviews delay planned expansions. A provider's ability to communicate risks, engage stakeholders and adjust its plans can indicate how prepared it is to support long-term infrastructure growth.</p> <p>It's also worth looking at whether past disputes or local pushback have affected permitting timelines, expansion plans or relationships with utilities.</p> <div class="extra-info"> <div class="extra-info-inner"> <h3 class="splash-heading">Key political and community-based questions to ask</h3> <ul type="disc" class="default-list"> <li>Has the provider faced community opposition or legal challenges?</li> <li>What issues triggered public concern or local disputes?</li> <li>How has the provider addressed concerns and disputes?</li> <li>Have controversies affected permitting timelines, expansion plans or utility relationships?</li> <li>Is future expansion dependent on unresolved permits, regulatory approvals or local infrastructure reviews?</li> <li>What are the contingency plans if community or regulatory challenges delay expansion?</li> </ul> <ul class="default-list"></ul> </div> </div> <p><i>Kinza Yasar is a technical writer for Informa TechTarget's AI and Emerging Tech group and has a background in computer networking.</i></p> </section> AI growth exposes data centers to power limits, cooling issues, water scarcity and regulatory pressure. Explore the key questions to raise with vendors. https://cdn.ttgtmedia.com/rms/onlineimages/storage_g1197646065.jpg https://www.techtarget.com/ai/feature/Questions-to-ask-when-evaluating-AI-ready-data-center-providers Thu, 30 Jul 2026 11:16:00 GMT <p>As AI use increases in business, so does the need for enterprise AI governance. A sound governance program can support rapid AI adoption by providing a faster approval path for lower-risk deployments, deeper review for higher-exposure uses and a verifiable audit trail that can be reused across systems.</p> <p><a href="https://www.techtarget.com/searchenterpriseai/tip/Global-AI-legislation-and-regulation-tracker">Regulatory frameworks</a> such as the EU AI Act, NIST AI Risk Management Framework (RMF) and ISO/IEC 42001 provide businesses with a defined basis for policy and assurance. Compliance is becoming enforceable, while boards and customers expect clear evidence of ownership, testing and control.</p> <p>In response to increasing governance requirements, AI governance platforms and tools are becoming a defined market. Dedicated governance platforms lead in policy, <a href="https://www.techtarget.com/searchenterpriseai/feature/Agentic-AI-compliance-and-regulation-What-to-know">compliance</a> workflow and evidence. Cloud and data platforms provide lower-friction controls in their own environments. Observability and security products add testing and runtime protection.</p> <p>The right AI governance tool depends on regulatory exposure, technology concentration and program maturity. Software can organize inventory, controls and evidence, while <a href="https://www.techtarget.com/searchenterpriseai/tip/Build-accountability-into-AI-to-drive-business-value">accountability</a> remains with designated people and established decisions. The best platform is the one that teams use consistently and shows the board, customers and regulators how AI is governed.</p> <section class="section main-article-chapter" data-menu-title="AI governance platform core capabilities"> <h2 class="section-title"><i class="icon" data-icon="1"></i>AI governance platform core capabilities</h2> <p>An AI governance platform is the working system of record for a business's AI use. It inventories models, applications, agents, data and vendors; assigns ownership and intended purpose; applies policies and risk criteria; manages reviews and exceptions; maps controls to laws and standards; and stores evidence for management, auditors and regulators. It connects with MLOps, <a href="https://www.techtarget.com/searchsecurity/definition/governance-risk-management-and-compliance-GRC">GRC</a>, privacy, security and observability tools, bringing a business's data and controls into a single process for decision-making and accountability.</p> <p>An AI governance platform should have the following core capabilities:</p> <ul class="default-list"> <li><b>Inventory and system registry. </b>A current register of models, applications, agents, prompts, data sets, vendors, owners and lifecycle status. Automated discovery should find <a target="_blank" href="https://www.cybersecuritydive.com/news/shadow-ai-security-risks-netskope/808860/" rel="noopener">shadow AI</a> and AI embedded in purchased software.</li> <li><b>Lifecycle management. </b>Stage gates, approvals and change control from design through retirement, scaled to the system's risk tier and connected to engineering workflows.</li> <li><b>Policy management and enforcement. </b>Reusable policies linked to approval steps and runtime guardrails, with version history and clear treatment of exceptions.</li> <li><b>Risk assessment and scoring. </b>Classification by intended use, potential harm, data sensitivity, geography and sector -- repeated when the system or its context changes.</li> <li><b>Regulatory mapping and control monitoring. </b>Controls mapped across the <a href="https://www.techtarget.com/searchenterpriseai/opinion/Everything-you-need-to-know-about-the-new-EU-AI-Act">EU AI Act</a>, NIST AI RMF, ISO 42001 and sector rules, with maintained content and gap analysis.</li> <li><b>Auditability and evidence collection. </b>Timestamped, exportable records of reviews, tests, signoffs, incidents and exceptions captured with minimal manual work.</li> <li><b>Explainability and interpretability. </b>Methods suited to the model and its context of use surfaced to non-technical reviewers and stored as evidence.</li> <li><b>Bias and fairness testing. </b>Repeatable, documented testing linked to intended use and protected groups, with thresholds and remediation tracking.</li> <li><b>Security controls for AI. </b>Protection against prompt injection, data and system prompt leakage, and <a href="https://www.techtarget.com/searchenterpriseai/tip/How-to-manage-generative-AI-security-risks-in-the-enterprise">adversarial attacks</a>; least-privileged access to models and agents; and integration with the security stack.</li> <li><b>Monitoring in production. </b>Continuous checks for <a href="https://www.techtarget.com/searchenterpriseai/tip/How-to-identify-and-manage-AI-model-drift">model drift</a>, quality loss, misuse and other production measures, with alerts and defined response steps.</li> <li><b>Third-party AI risk. </b>Due diligence for external models and AI-enabled software, including ownership, data use, contract terms and vendor change notifications.</li> <li><b>Human oversight and accountability. </b>Named ownership and clarity on who can approve, override, suspend or retire a system -- a legal requirement under the EU AI Act.</li> <li><b>Transparency artifacts and reporting. </b>Model cards, data documentation, notices and board reporting generated from current governance records.</li> <li><b>Enterprise integration. </b>Connectors to cloud, data, MLOps, identity and access management (IAM), security information and event management (SIEM), ticketing, GRC and development pipelines so governance stays inside normal work.</li> </ul> <p>Newer capabilities are also emerging. Generative AI requires prompt and response logging, retrieval governance, content safety and cost controls. Agentic AI needs registries for agents and tools, identity-based permissions, action limits and trace records.</p> </section> <section class="section main-article-chapter" data-menu-title="AI governance tools market in 2026"> <h2 class="section-title"><i class="icon" data-icon="1"></i>AI governance tools market in 2026</h2> <p>The AI governance tools and platforms market combines purpose-built governance platforms with established GRC, cloud, data, MLOps and security products. Inventory, policy mapping and audit evidence are relatively mature. Runtime controls, shadow AI discovery and <a href="https://www.techtarget.com/searchenterpriseai/tip/Agentic-AI-governance-strategies-A-complete-guide">agent governance</a> are developing quickly and remain less consistent across vendors.</p> <p>The market consists of five broad vendor categories. Each approaches AI governance from a different starting point, and many products also overlap across categories.</p> <p>The lists below are indicative rather than exhaustive, and vendors can span more than one category.</p> <h3>Dedicated AI governance platforms</h3> <p>These tools are designed to manage AI policies, risks, approvals and evidence across the enterprise. Their main goal is to provide a central, AI vendor-neutral system for inventory, accountability, regulatory mapping and governance workflows.</p> <p><b>Example tools: </b>IBM watsonx.governance; ServiceNow AI Control Tower; Truyo; Credo AI; OneTrust AI Governance; Monitaur; Airia; Holistic AI; ModelOp; Saidot; Cranium AI; Relyance AI; Trustible; LatticeFlow AI; Modulos; and Lumenova AI.</p> <h3>GRC, privacy and data governance vendors with embedded governance</h3> <p>These vendors extend established risk, compliance, privacy and data management capabilities to AI systems. They're intended to connect AI governance with existing enterprise control frameworks, third-party risk processes and regulatory reporting.</p> <p><b>Example tools: </b>OneTrust; Collibra; SAP; BigID; Securiti; Informatica; MetricStream; AuditBoard; and Mitratech.</p> <h3>Cloud and AI platform vendors with embedded governance</h3> <p>These vendors build governance controls directly into the environments where models and applications are developed and deployed. Their strengths are native integration, identity controls, lineage and technical enforcement, though coverage might be concentrated within the vendor's own ecosystem.</p> <p><b>Example tools: </b>Microsoft Purview and Azure AI Foundry; AWS SageMaker and Bedrock Guardrails; Google Cloud Vertex AI; Databricks Unity Catalog and Unity AI Gateway; Snowflake Cortex AI Observability; and Nvidia NeMo Guardrails.</p> <h3>MLOps, LLMOps and observability vendors expanding into governance</h3> <p>These tools focus on the development and production performance of machine learning and generative AI systems. They add governance through model evaluation, tracing, monitoring, testing and operational evidence, but provide less depth in enterprise policy and compliance workflows.</p> <p><b>Example tools: </b>Arize AI; Fiddler AI; Arthur AI; Dataiku Govern; LangSmith; Weights &amp; Biases; Datadog LLM Observability; Braintrust; Helicone; and TruLens.</p> <h3>AI security posture and runtime control vendors</h3> <p>These products concentrate on identifying AI exposures and protecting systems while they're in use. Their main capabilities include shadow AI discovery, prompt and data leakage controls, adversarial testing, runtime guardrails and monitoring of model and agent activity.</p> <p><b>Example tools: </b>Cisco AI Defense; SentinelOne Prompt Security; HiddenLayer; Lasso Security; Noma Security; Mindgard; WitnessAI; and Wiz.</p> </section> <section class="section main-article-chapter" data-menu-title="8 leading platforms and tools at a glance"> <h2 class="section-title"><i class="icon" data-icon="1"></i>8 leading platforms and tools at a glance</h2> <p>The following table focuses on eight platforms with strong enterprise relevance, governance breadth and credible integration options. Most pricing is quote-based or bundled, so total cost should include services, integration and internal operating effort.</p> <p><iframe title="AI governance tools and platforms" aria-label="Table" id="datawrapper-chart-X6hoV" src="https://datawrapper.dwcdn.net/X6hoV/2/" scrolling="no" frameborder="0" style="width: 0; min-width: 100% !important; border: none;" height="1712" data-external="1"></iframe></p> <p> <script type="text/javascript">(function(){function e(){window.addEventListener(`message`,function(e){if(e.data[`datawrapper-height`]!==void 0){var t=document.querySelectorAll(`iframe`);for(var n in e.data[`datawrapper-height`])for(var r=0,i;i=t[r];r++)if(i.contentWindow===e.source){var a=e.data[`datawrapper-height`][n]+`px`;i.style.height=a}}})}e()})();</script> </p> </section> <section class="section main-article-chapter" data-menu-title="How to select an AI governance platform"> <h2 class="section-title"><i class="icon" data-icon="1"></i>How to select an AI governance platform</h2> <p>Business leaders should start the procurement process by identifying the business's legal, regulatory and <a href="https://www.techtarget.com/searchenterpriseai/feature/Leading-AI-with-ethics-The-new-governance-mandate">internal governance obligations</a> before assessing specific products. Confirm the rules that apply, the evidence they require and the AI systems within scope, including third-party software, embedded AI and autonomous agents. This approach establishes the capabilities the platform must provide and prevents the selection process from being driven by vendor features rather than governance requirements.</p> <p>Decide whether governance must span several technology environments or can remain inside one main platform. Use the following list to test integration, evidence, security, usability, scale, vendor maturity and total operating cost:</p> <ul class="default-list"> <li><b>Regulations and compliance.</b> Look for maintained EU AI Act, NIST AI RMF, ISO 42001 and sector content; gap analysis; acceptable evidence.</li> <li><b>Technical integration.</b> Look for production connectors to cloud, data, MLOps, IAM, SIEM, GRC, ticketing and development pipelines.</li> <li><b>Risk and control coverage.</b> Look for the ability to inventory for shadow and third-party AI; risk scoring and control monitoring across models, generative AI and agents.</li> <li><b>Generative and agentic governance.</b> Look for prompt and response logs, retrieval controls, guardrails, testing, trace evidence and human escalation.</li> <li><b>Security architecture and residency.</b> Look for injection and leakage controls, adversarial testing, agent permissions, hosting and sovereignty options.</li> <li><b>Auditability and reporting.</b> Look for automated, timestamped and exportable evidence; dashboards drawn from current control data.</li> <li><b>Usability and workflow fit. </b>Look for clear tasks for technical and non-technical users; approvals that fit existing work and avoid manual duplication.</li> <li><b>Scalability and extensibility.</b> Look for capacity for projected system volume, business units, geographies, and new model or agent types.</li> <li><b>Vendor maturity and roadmap.</b> Look for customer evidence, financial stability, support quality and a credible plan for agent governance.</li> <li><b>Total cost of ownership.</b> Look for licensing, services, integration, internal staffing, data retention, export and switching costs.</li> </ul> <p><i>Kashyap Kompella, founder of RPA2AI Research, is an AI industry analyst and advisor to leading companies across the U.S., Europe and the Asia-Pacific region. Kashyap is the co-author of three books: </i>Practical Artificial Intelligence<i>, </i>Artificial Intelligence for Lawyers<i> and </i>AI Governance and Regulation<i>.</i></p> </section> AI governance has become a boardroom priority. These tools and platforms can help leaders fulfill AI governance mandates. https://cdn.ttgtmedia.com/rms/onlineimages/machine_learning_g1264542084.jpg https://www.techtarget.com/ai/tip/The-best-AI-governance-tools-and-platforms-in-2026 Tue, 28 Jul 2026 14:53:00 GMT <p>Quantum computing is moving from experimental promise to practical reality as advances in hardware stability and quantum networking bring it closer to solving real-world business problems. Organizations across government and industry sectors are exploring using it for a range of challenges that classical computers can't solve.</p> <p>McKinsey's "Quantum technology monitor 2026: A commercial tipping point" <a target="_blank" href="https://www.mckinsey.com/capabilities/mckinsey-technology/our-insights/mckinsey-quantum-technology-monitor-2026-a-commercial-tipping-point" rel="noopener">report</a> states that 52% of companies analyzed are spending $5 million or more a year on quantum efforts.</p> <p>With more investment, the quantum use cases continue to expand. These include: sensors to measure time, gravity and magnetic fields; simulations for materials science and pharmaceutical R&amp;D; machine learning (ML) to accelerate AI capabilities; and quantum key distribution (<a target="_blank" href="https://www.techtarget.com/searchsecurity/definition/quantum-key-distribution-QKD" rel="noopener">QKD</a>), which can coexist with current security protocols. The quantum internet also promises to establish a distributed quantum computing grid to connect these systems.</p> <p>For each of these developments, caveats exist. For example, quantum systems cost tens of millions of dollars annually to operate, compared with the costs of operationalizing enterprise AI. But each quantum breakthrough could potentially revolutionize business environments and introduce new opportunities.</p> <section class="section main-article-chapter" data-menu-title="Quantum sensors"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Quantum sensors</h2> <p>Scientists and engineers are using the <a target="_blank" href="https://www.techtarget.com/searchcio/definition/quantum-entanglement" rel="noopener">quantum properties of entanglement</a>, superposition and interference to solve problems faster than classical computing. Harnessing these quantum properties portends a fundamental shift in how we process information.</p> <p>Quantum sensors measure physical properties related to time, gravity and magnetic fields. They use extreme quantum sensitivity to register the slightest deviation in a functional tool. For example, in autonomous vehicles, drones and aircraft, quantum sensitivity can detect and respond to microscopic operational changes. Quantum networks offer the potential for distant quantum sensors to communicate and triangulate measurements, functioning as one enormous, distributed sensor.</p> <p>In healthcare imaging, magnetoencephalography uses quantum sensors to measure neuron-generated magnetic fields in real time. Physicians can use the tool to pinpoint the exact regions of the brain where epileptic seizures originate and remove or treat the tissue. In the energy industry, gravimeters use quantum to detect microvariations in gravity for mineral exploration or to map underground cavities for hydrocarbon prospecting and carbon-capture monitoring. The military uses quantum radar for stealth detection and surveillance.</p> </section> <section class="section main-article-chapter" data-menu-title="Quantum security"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Quantum security</h2> <p>Quantum key distribution offers extremely high tamper sensitivity for detecting incursions, making it unparalleled as a security protocol. This hardware-based technology uses the laws of quantum physics to randomly generate a string of bits for encryption. Two parties share this unbreakable encryption key to protect processes and communications.</p> <p>If attackers attempt to compromise a connection, the quantum states are disturbed, exposing an intrusion. A QKD system incorporates a single-photon source along with detectors and temperature-controlled optical emitters, which require precise monitoring and control.</p> <p>QKD is expensive to scale because it relies on specific hardware. For this reason, it will primarily reinforce other modes of security, blending physical, classical and operational safeguards into one defensive approach. Deployments are currently limited to government, defense and telecom networks.</p> <p>Specific use cases include telecom infrastructure, where QKD protects data moving between data centers and cloud services. Some government networks are combining QKD with classical encryption and post-quantum algorithms. For example, relay nodes within satellite links can distribute quantum keys between ground stations to protect communications.</p> </section> <section class="section main-article-chapter" data-menu-title="Quantum machine learning"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Quantum machine learning</h2> <p>The deployment of quantum ML for enhanced AI capabilities has passed the proof-of-concept phase. According to the U.S. Data Science Institute's "From Qubits to Insights: The Rise of Quantum AI in 2026" report, the quantum AI market is <a target="_blank" href="https://www.usdsi.org/data-science-insights/from-qubits-to-insights-the-rise-of-quantum-ai-in-2026" rel="noopener">projected</a> to reach $638 million in 2026, a 35% increase over 2025.</p> <p>The goal of quantum ML is to accelerate AI training and solve complex, multidimensional problems. The advantages of running algorithms on a quantum system include exponential acceleration of core linear algebraic operations and quadratic speedups for unstructured searches.</p> <p>In general, IT engineers and researchers can use quantum ML to bypass compute bottlenecks in processing data and explore new ways to optimize training. In the process, they can also discover limitations, such as models with too many quantum components that run slowly, while those without sufficient components show little advantage over classical approaches. Another industry use case is supply chain modeling, in which quantum-driven processes identify the best routes, schedules and resource plans to achieve lower costs and greater efficiency in freight, transport and logistics.</p> <p>An important <a target="_blank" href="https://www.techtarget.com/searchcio/tip/Quantum-computing-in-finance-Key-use-cases" rel="noopener">use case in finance</a> is quantitative analysis and modeling, where analysts study correlation patterns in assets and often struggle to correlate inconsistent data. Quantum ML can model complex market behaviors, speed up data processing and improve predictive accuracy.</p> <p>Pharmaceutical scientists in drug discovery also consistently work with complex variables as they search for new drug compounds. Quantum ML can map molecular behavior to quantum states and invent promising combinations for disease treatments.</p> <p>&nbsp;</p> <figure class="main-article-image full-col" data-img-fullsize="https://www.techtarget.com/rms/onlineimages/business_use_cases_of_quantum_computing-f.png"> <img data-src="https://www.techtarget.com/rms/onlineimages/business_use_cases_of_quantum_computing-f_mobile.png" class="lazy" data-srcset="https://www.techtarget.com/rms/onlineimages/business_use_cases_of_quantum_computing-f_mobile.png 960w,https://www.techtarget.com/rms/onlineimages/business_use_cases_of_quantum_computing-f.png 1280w" alt="Graph showing the different use cases of quantum computing." data-credit="TechTarget" height="252" width="560"> <div class="main-article-image-enlarge"> <i class="icon" data-icon="w"></i> </div> </figure> </section> <section class="section main-article-chapter" data-menu-title="Quantum simulations"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Quantum simulations</h2> <p>In materials science and pharmaceutical R&amp;D, quantum-controlled simulations can mimic molecular and physical properties at the atomic level. For example, quantum systems can create a correspondence between simulation hardware and a target environment's actual physics. They then match atom behavior in a quantum simulator to electron behavior in the real material world.</p> <p>These quantum emulations are useful in the automotive, electronics and aerospace industries to design complex polymers. A quantum computer can sift through an exponential number of possible configurations to quickly identify polymer compositions with traits crucial for certain products, such as 5G components and semiconductors in the electronics industry or durable, heat-resistant materials for space flight.</p> <p>Quantum simulations also resolve issues in materials science and chemistry related to <a target="_blank" href="https://www.techtarget.com/searchenterpriseai/tip/How-the-sodium-ion-battery-can-energize-the-enterprise" rel="noopener">battery and energy storage</a> use cases. These emulations bring together mathematical models and computational algorithms to combine electrodes, electrolytes and catalysts in novel ways, enabling faster charging, higher-capacity storage and longer lifetimes.</p> </section> <section class="section main-article-chapter" data-menu-title="The quantum internet"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The quantum internet</h2> <p>The quantum internet comprises isolated quantum computers that use photons with their neutral charge and zero mass to create unified, distributed networks. Photons can travel through standard telecommunications fiber while largely retaining their quantum state, linking together individual quantum machines to operate as a supercomputer, exchanging encoded quantum-based information. And because quantum signals only span short distances, quantum repeaters function to extend the range of qubit entanglement without the need to copy data.</p> <p>In <a target="_blank" href="https://www.nyu.edu/about/news-publications/news/2023/september/nyu-takes-quantum-step-in-establishing-cutting-edge-tech-hub-in-.html" rel="noopener">one recent deployment</a>, New York University and Qunnect, a global quantum hardware provider, demonstrated quantum signals running through existing metropolitan fiber-optic infrastructure. The goal was to show that scalable, secure quantum networks could operate reliably within existing telecom network media, moving one step closer to demonstrating a functional quantum network.</p> <p>In another use case, connecting enterprise and academic hubs, the <a href="https://www.aliroquantum.com/blog/real-world-quantum-network-deployments"><u>EPB Quantum Network</u></a> in Chattanooga, Tenn., became commercially available in 2022. This open-access infrastructure lets researchers and innovators build and test quantum applications and explore real-world initiatives in business and science.</p> <p><em>Kerry Doyle writes about technology for a variety of publications and platforms. His current focus is on issues relevant to IT and enterprise leaders across a range of topics, from nanotech and cloud to distributed services and AI.</em></p> </section> As quantum capabilities progress toward commercial viability, business leaders need to understand the quantum technologies that will be at their disposal. https://cdn.ttgtmedia.com/rms/onlineimages/iot_g1226930515.jpg https://www.techtarget.com/ai/tip/The-quantum-technologies-that-will-transform-enterprise-IT Mon, 27 Jul 2026 23:03:00 GMT <p>Two years ago, almost no states regulated AI. Today, <a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-federal-ai-preemption-state-regulation-202/">more than 30 do</a>. Companies operating across multiple states often build to the strictest requirement they face, but only when the requirements differ by degree. When they differ by kind, no single bar exists. One state's rule can push a company to normalize a skewed applicant pool. Another can push the company to follow raw performance data. Compliance in one jurisdiction can create exposure in the next, forcing the enterprise to choose which law to break.</p> <p>CIOs must make AI governance decisions based on <a href="https://www.techtarget.com/searchenterpriseai/tip/Global-AI-legislation-and-regulation-tracker">today's patchwork of state laws</a>, not under the federal framework they hope Congress eventually adopts. The longer that gap persists, the more enterprise AI governance turns into legal fragmentation management rather than technology implementation. The problem lands on CIOs because AI now runs through HR, marketing, customer data, cybersecurity and legal compliance simultaneously, not just one department's tech stack.</p> <section class="section main-article-chapter" data-menu-title="The hiring challenge"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The hiring challenge</h2> <p><a href="https://www.techtarget.com/searchhrsoftware/feature/AI-hiring-needs-a-decision-trail">Hiring usually shows the problem most clearly</a> because no other task or function runs AI against as many people as often. While volume doesn't make hiring the first place trouble appears, it does make hiring the place where trouble becomes hardest to miss.</p> <p>A tool that reduces 100,000 resumes to a shortlist of 10 makes judgment calls about commute times, work history and career gaps. No one wrote those judgments into policy. No one reviewed them before the shortlist reached a hiring manager. LinkedIn funnels candidates who might have never applied. Data providers scrape the open internet to identify and solicit people who never put themselves forward for anything.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> A tool trained on decades of history inherits that history's patterns. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>None of this requires bad intent. A commute-time filter can quietly exclude applicants from underserved neighborhoods without any decision-maker choosing to discriminate. An applicant pool that skews 70%-30% along gender lines raises a hard question: Should the company normalize the ratio, leave it alone or follow whatever the underlying performance data shows? The third option is the only one that looks neutral. It assumes unbiased historical performance data, which is the same assumption that fails in the retail promotion context discussed below. A tool trained on decades of history <a href="https://www.techtarget.com/searchenterpriseai/feature/The-AI-bias-playbook-Mitigation-strategies-for-CIOs">inherits that history's patterns</a>.</p> <p>There is no clean answer, and the states that have tried to write one down do more than set different bars; some point in opposite directions. The legal exposure companies face comes from disparate impact, which occurs when a neutral policy or practice disproportionately harms members of a protected group. This exposure comes from how AI performs once deployed at scale against real people, not from anyone's intent or from whether the company disclosed the practice.</p> </section> <section class="section main-article-chapter" data-menu-title="Development vs. deployment: Where's the true exposure?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Development vs. deployment: Where's the true exposure?</h2> <p>Vendors build AI systems, but the companies using them decide where to deploy them, what role they will play in consequential decisions and whether that use complies with each state's law. The statutes don't agree on the legal hook. Some reach the developer. Some reach the deployer. Some reach both.</p> <p>The allocation shifts from one framework to the next. But one fact doesn't shift: the company operating the tool against real people remains present in every jurisdiction at once. <a href="https://www.techtarget.com/searchenterpriseai/opinion/The-legal-mistake-courts-should-avoid-in-Eightfold-AI-lawsuit">Deployment, not development, concentrates exposure</a> regardless of how any single statute assigns responsibility. A vendor's safety testing claims or federal review status doesn't move that exposure off the deployer's books.</p> <p>This dynamic extends beyond hiring. Retailers that disclose their cameras and offer an opt-out can still deliver advertising along lines that track race, gender or age. Disclosure and opt-out solve a notice problem. They do nothing about impact because the underlying model selects for engagement, not for any demographic anyone chose. The retailer can act transparently and still produce a skewed outcome. That's the point: Consent mechanisms don't cure disparate impact. Promotion decisions carry the same risk. A tool trained on decades of performance data inherits that history's patterns and can produce an outcome nobody at the company would have chosen.</p> </section> <section class="section main-article-chapter" data-menu-title="Is a single federal standard the answer?"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Is a single federal standard the answer?</h2> <p>A federal framework with one set of standards, rather than 30 separate state ones, sounds like the obvious fix. The White House has tried twice in the last year. First, it created a <a href="https://www.techtarget.com/searchenterpriseai/feature/Who-wins-and-loses-with-Trumps-AI-executive-order">task force to challenge state AI laws in court</a>. Then it created a <a href="https://www.techtarget.com/searchenterpriseai/news/366644013/Trump-AI-order-targets-frontier-model-prerelease-review">voluntary security review window for AI developers</a>. Neither substitutes for legislation. An executive order can't preempt state law without Congressional action, and a voluntary review imposes no compliance obligation on vendors that can opt in or out.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> A single federal standard remains the better end state than 30 conflicting ones. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>However, even a working mechanism wouldn't answer the harder question: What level of algorithmic bias can the law tolerate? The White House's efforts fail procedurally. The policy question remains unresolved for a different reason. Zero bias is impossible. No human being is bias-free, and neither is the data humans produce.</p> <p>No federal lawmakers, left or right, have been willing to draw that line. The problem isn't the bias itself. Anti-discrimination law has tolerated imperfect, biased human decision-making for decades without demanding zero. An algorithm changes the politics by making residual bias legible. It can measure, report and audit that bias after the fact. A statute would have to name a tolerance and defend it. That means conceding that the approved system remains biased and saying by exactly how much. Doing that is harder politically than tolerating the same bias when it spreads across a thousand human managers, and no one has to sign the number.</p> <p>A single federal standard remains the better end state than 30 conflicting ones. But one risk on the other side deserves acknowledgment, even if it doesn't change the conclusion. Technology is moving fast enough that some of today's hardest questions, especially bias normalization, may have technical answers no one has built yet. A federal standard written today could lock in today's assumptions and foreclose tomorrow's solutions. That risk calls for careful drafting. It doesn't justify the patchwork of state standards, which imposes its costs now and every day it persists.</p> </section> <section class="section main-article-chapter" data-menu-title="What to do today"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What to do today</h2> <p>None of this changes <a href="https://www.techtarget.com/searchenterpriseai/feature/AI-regulation-What-businesses-need-to-know">what companies can do today</a>. CIOs should map AI by use case and jurisdiction, so they know which laws govern each deployment. They should then build governance around practices that should exist regardless of the regulatory landscape, including clear terms of service, reliable data provenance and human review where the consequences matter most.</p> <p>Congress, courts and future administrations will keep shaping AI regulation, but enterprise AI won't wait for them. Companies already deploy systems under laws that states can enforce today. CIOs must govern within the legal landscape that exists now, not the federal framework they hope will eventually arrive.</p> <p><i>Jon Polenberg is a shareholder and vice chair of Becker's Business Litigation Practice. As a business trial lawyer in Florida, he's known for his advocacy and strategic insight in complex commercial litigation. Jon has decades of experience representing businesses in high-stakes cases and handling intricate legal disputes with precision and a commitment to excellence. His practice spans a range of industries, assisting companies in resolving matters that affect their operations, reputations and financial interests.</i></p> </section> Companies are deploying AI across their organizations faster than lawmakers are building consistent rules to govern it, leaving CIOs to navigate an expanding patchwork of state requirements. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a264431831.jpg https://www.techtarget.com/ai/opinion/How-to-manage-the-gap-between-enterprise-AI-use-and-AI-regulation Mon, 27 Jul 2026 13:35:00 GMT <p>An engineering leader recently walked out of a quarterly business review with a clean story: AI adoption was up 40% across the team, and token consumption was climbing. In addition, pull request throughput had jumped, indicating developers were writing and reviewing code faster. The board was satisfied.</p> <p>Three weeks later, a critical feature shipped late because no one had caught a cross-functional dependency until week four of a six-week cycle. That same dependency had delayed the previous two releases.</p> <p>The <a href="https://www.techtarget.com/searchenterpriseai/tip/Top-AI-KPIs-that-business-leaders-need-to-know">metrics looked good</a>. The work hadn't changed.</p> <section class="section main-article-chapter" data-menu-title="Stop just measuring what's easy"> <h2 class="section-title"><i class="icon" data-icon="1"></i>Stop just measuring what's easy</h2> <p>The metrics organizations default to are real numbers: tokens consumed, dollars spent against AI tooling budgets, pull requests developers release for code changes, lines of code written with AI assistance and time to completion on individual tasks. However, they measure <a href="https://www.techtarget.com/searchenterpriseai/post/Your-AI-advantage-isnt-speed-its-judgment">AI as a speed layer</a> on top of existing work and concentrate almost entirely on the execution layer: writing code, testing it, reviewing it and writing docs. That's not an accident. Execution tasks are well scoped, inputs and outputs are clear, and correctness is independently verifiable. You run the tests and you know.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Pull request throughput went from a signal of progress to the target itself, and somewhere in that shift it stopped meaning what it used to mean. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>Coding tools dominate the conversation about AI maturity because they're easy to verify. The category that's easiest to check becomes the category everyone points to as proof that the technology works. The categories that are harder to verify, and where more of the value sits, stay unmeasured because nobody has figured out how to instrument them yet.</p> <p>Problem definition, solution design, cross-functional alignment and decision-making are still running at the same pace as before AI arrived. These are the stages where bad decisions get locked in, where dependencies get missed and where options get narrowed too early. None of that shows up in a token count. An organization could double its pull request throughput and still ship the wrong things, with the same upstream delays and late-stage pivots. <a href="https://www.techtarget.com/searchbusinessanalytics/opinion/Your-AI-isnt-failing-your-metrics-are">The number is clean. The signal is weak.</a></p> <p>This is Goodhart's <a target="_blank" href="https://lawsofsoftwareengineering.com/laws/goodharts-law/" rel="noopener">law</a> in a delivery cycle: Once a measure becomes the target, it stops tracking the thing it was meant to stand in for. Pull request throughput went from a signal of progress to the target itself, and somewhere in that shift it stopped meaning what it used to mean.</p> </section> <section class="section main-article-chapter" data-menu-title="What to measure instead"> <h2 class="section-title"><i class="icon" data-icon="1"></i>What to measure instead</h2> <p>The value AI creates upstream shows up as things that didn't happen -- the pivot that was avoided or the blocker that surfaced in week one instead of week four. Counting that requires different questions:</p> <ul class="default-list"> <li>Are problem definition, alignment and decision-making getting faster, the same way execution is?</li> <li>Are late-stage discoveries, bugs found in production, conflicts surfacing in week four and scope changes after alignment declining?</li> <li>Are blockers tracing back to something that was knowable earlier but wasn't known? I've watched teams lose three weeks to exactly that.</li> </ul> <p>A <a href="https://www.techtarget.com/searchbusinessanalytics/opinion/Why-the-rush-to-replace-dashboards-with-AI-is-a-mistake">dashboard widget</a> won't answer any of these questions. Instead, they are answered by tracking cycle time at every stage of the workflow -- execution included -- and comparing it before and after AI is introduced. That takes more time to set up than pulling a token report. It's also the only approach that tells you anything.</p> </section> <section class="section main-article-chapter" data-menu-title="The questions you need to answer"> <h2 class="section-title"><i class="icon" data-icon="1"></i>The questions you need to answer</h2> <p>None of this is trivial to instrument, and that's the point. The metrics most organizations default to are easy to pull because they sit at the end of the workflow, where outputs are visible. The value upstream is harder to attribute, and that's exactly why it must be defined before deployment, not measured after the fact.</p> <blockquote class="main-article-pullquote"> <div class="main-article-pullquote-inner"> <figure> Before approving another round of AI tooling spend, ask which of these questions the team can actually answer. </figure> <i class="icon" data-icon="z"></i> </div> </blockquote> <p>What does a well-scoped problem definition look like, and can AI stress test it before the team commits? What does early dependency discovery look like, and can AI surface it before three weeks are spent building against conflicting assumptions? It's important to answer those questions before AI touches the workflow, not after.</p> <p>The next quarterly review is a reasonable place to start. Before approving another round of AI tooling spend, ask which of these questions the team can actually answer.</p> <p>If the answer is none of them, that's not evidence AI isn't working. It's evi.dence nobody has decided what working would look like. That's a more useful finding than another quarter of <a href="https://www.techtarget.com/searchcio/feature/Tokenmaxxing-How-CIOs-extract-maximum-value-AI-tokens">climbing token counts</a>, because it tells you exactly where to start.</p> <p><i>Sireesha Pulipati is a data and AI engineering leader specializing in scalable data platforms, real-time data pipelines and production-grade AI systems. She is a staff data engineer at Shopify and previously worked at Google, where she led analytics and business intelligence across search and Google Cloud. The views expressed by Pulipati are her own and do not reflect those of her company.</i></p> </section> Rising token counts and pull request throughput look impressive, but they miss where AI creates real value: preventing bad decisions and surfacing blockers before they cost time. https://cdn.ttgtmedia.com/rms/onlineimages/ai_a311883856.jpg https://www.techtarget.com/ai/post/The-AI-metrics-trap-Measuring-speed-but-missing-value Fri, 24 Jul 2026 12:37:00 GMT 60 webmaster@techtarget.com