Ethical AI surveillance balances oversight and transparency

How AI changes employee monitoring

Employee monitoring itself is nothing new. Companies use tools to track keystrokes, badge swipes, work hours and other activities. What changes with AI is what employers can do with that information.

Traditional monitoring is essentially a recorder. AI monitoring is an interpreter.

"Traditional monitoring is essentially a recorder," said Kirill Meshyk, head of AI data collection at Unidata, a data collection and annotation company. "AI monitoring is an interpreter."

An AI system could label an employee who is inactive for a certain period of time as disengaged or likely to leave the company, he said. "The problem is that organizations can mistake keystrokes, mouse movements and time online for measures of productivity or engagement," he added.

Workers whose behavior falls outside the patterns an AI model deems acceptable can therefore be penalized even when their performance is strong. While AI systems can influence decisions about discipline, scheduling, promotion or termination, the reasoning behind those decisions might be difficult for employees and the companies deploying the tools to understand or challenge.

The human might simply end up reviewing the AI's conclusion rather than forming an independent judgment.

That distinction becomes especially important when AI-generated assessments inform decisions about individual employees, said Emily Hartstone, founder of AI governance company Runtime Authority Control, which develops safeguards for AI-initiated actions. She compared a dashboard that provides a manager with data on an employee's activity and performance for interpretation to an AI system that automatically evaluates an employee and routes its assessment to HR. In the latter case, she said, the system has interpreted the data and reached a judgment about the employee.

"The human might simply end up reviewing the AI's conclusion rather than forming an independent judgment," she said.

The question is no longer what the company can monitor, but what it should monitor and how it should use its insights. These decisions can affect more than the technology itself, shaping employee trust and raising questions about privacy, fairness and regulatory compliance.

The trust and regulatory gap

Employee concerns about AI surveillance are already showing up in surveys. According to Owl Labs' "2025 State of Hybrid Work Report," of the 2,000 full-time knowledge workers surveyed, nearly half cited having their work activities monitored as a top workplace concern. This underscores the tension between employers' desire for greater visibility and employees' unease about being watched and evaluated.

Part of this friction involves how companies introduce monitoring. In nonprofit Jobs for the Future's "AI Is Getting Real, But the Real Work Is Still Ahead" survey, of the more than 3,000 American respondents, 56% said they weren't consulted about how AI tools are used in their work. Meanwhile, the 2025 American Job Quality Study, which surveyed over 18,000 workers, found that employees with some influence over how their workplace deploys technology are more than twice as likely to report high job satisfaction compared to those with none. These findings suggest that employee involvement and transparency regarding AI surveillance can play a role in how workers respond to new technology.

Secrecy is what transforms a tool into a form of surveillance.

Employees might accept certain forms of monitoring when there is a clear business justification, such as cybersecurity or compliance. Concerns become more complicated when AI evaluates productivity, predicts behavior or influences decisions that affect an employee's career.

"Secrecy is what transforms a tool into a form of surveillance," Meshyk said. "The tool is neutral by its nature -- deploying it secretly, however, is not."

Hartstone said meaningful transparency should go beyond a notice or privacy policy. Employees should know what is collected, what the system produces about them, who receives that information, whether it can affect performance, pay, promotion or layoffs, how they can see their own record and how they can challenge an assessment.

Organizations should collect only what's necessary for a defined purpose, use aggregate data where possible and limit how long employee information is retained.

Transparency, however, is only one part of the conversation. Organizations also need to consider how much employee data they collect. Srinivas Chippagiri, senior member of technical staff at Salesforce, said companies should pair transparency with data minimization. "Organizations should collect only what's necessary for a defined purpose, use aggregate data where possible and limit how long employee information is retained," he said. Keystroke or screen capture, he added, should require a specific, justified and disclosed need.

These principles are increasingly reflected in laws and regulations covering employee privacy, AI-driven employment decisions and workplace monitoring. For enterprises, however, there is no single set of rules governing AI-powered employee monitoring. Requirements vary by state and by how an AI system is used, creating a patchwork that can be difficult to navigate.

Colorado's AI law, for example, covers certain high-risk AI systems used in consequential decisions, including those for employment. But the law has been delayed and rewritten. In May 2026, Gov. Jared Polis signed a narrower version focused on disclosure and transparency, with a new effective date of January 2027.

California is taking a more targeted approach to AI in employment. Its automated-decision-making rules are set to take effect January 1, 2027, with requirements for pre-use notices, risk assessments and employee opt-out rights for certain significant employment decisions.

At the federal level, there's no comprehensive law governing AI-powered employee surveillance. Nevertheless, existing labor and employment protections can apply in certain circumstances. The National Labor Relations Board's general counsel has warned that AI-enabled monitoring of union-organizing activity can violate workers' rights, while the EEOC has made clear that AI-based employment selection tools remain subject to federal antidiscrimination laws, including Title VII of the Civil Rights Act, which prohibits discrimination in employment based on protected characteristics such as race, color, religion, sex and national origin.

For enterprise leaders, the takeaway is that compliance depends on how their organizations use technology, collect data and make decisions. A monitoring system might be capable of collecting and analyzing vast amounts of employee data, but that doesn't mean an enterprise should -- or legally can -- use all of it.

What public-sector surveillance can teach enterprises

The risks of surveillance aren't limited to the workplace. Public-sector deployments show what can happen when monitoring expands without clear limits on how organizations collect and use data.

Recently, Flock Safety's automated license plate readers, used by thousands of U.S. law enforcement agencies, have faced growing backlash as communities question how Flock Safety deploys the technology and uses the data it collects. Some communities have removed or vandalized cameras, while cities including Santa Cruz and Mountain View have ended their contracts with the company.

In Massachusetts, the Framingham Police Department let its Flock Safety contract expire after months of residents citing privacy, surveillance and data sharing concerns. Residents specifically raised concerns about Flock's data sharing network and the potential for access by federal immigration authorities.

Organizations risk losing trust when they introduce surveillance without clear notice, explanations or limits on how they can use the resulting data. This leads to increased pressure to roll back technology after it's already deployed, which can be costly.

The scope of workplace monitoring can expand far beyond its original purpose without organizations revisiting how the data is being used.

The same dynamic can play out inside enterprises when employees discover that information collected for one purpose is later being used to evaluate them in another context. Hartstone said many companies rely on acceptable use policies written years earlier to justify newer forms of AI monitoring. A policy that once covered security monitoring of company devices, she said, might now be interpreted as permission to analyze collaboration patterns or feed behavioral scores into performance reviews.

"The scope of workplace monitoring can expand far beyond its original purpose without organizations revisiting how the data is being used," Hartstone said.

How ethical deployment works in practice

Some organizations are taking a more deliberate approach to workplace monitoring. 5app, a workplace learning and AI skills company, uses its Helix AI system to analyze meeting conversations for soft skills such as communication, accountability and growth mindset. Rather than making the system mandatory, 5app made Helix opt-in at the meeting level.

Philip Huthwaite, CEO of 5app, said the decision gives employees control over their data and recognizes that AI analysis isn't appropriate for every meeting. Employees can remove Helix from a discussion when it becomes confidential, while individual skills scores and feedback remain visible to the employee.

HR and administrators can access aggregate skills data to identify broader development needs, but not an individual's detailed feedback. Role-based permissions and data minimization measures provide additional safeguards.

"For us, the line comes down to purpose, access and consequences," Huthwaite said. A development tool, he said, should give the employee something useful in return, while the line is crossed when AI analysis becomes a hidden mechanism for surveillance or is used to make high-stakes employment decisions.

The approach followed concerns from employees about Helix becoming "Big Brother-esque," Huthwaite said. The company responded by explicitly positioning Helix as a personal development tool rather than an employment-decision system, with individual scores kept out of performance reviews and disciplinary actions.

Unidata offers another example of these safeguards in practice. When the company rolled out an AI-based quality monitoring system across 25 platforms, employees were told before deployment what the system would collect, what it would ignore and who could access the data. A human could also challenge and review every automated flag. Meshyk said productivity increased by roughly 18% during the first two quarters, which he attributed largely to employees no longer having to work defensively against an unknown monitoring system.

But higher activity metrics don't necessarily translate into greater productivity, Chippagiri cautioned. Organizations should also measure actual business outcomes and employee trust, he said, because higher activity under surveillance can reflect fear of being watched rather than genuine productivity.

Automaker Nissan offers another example of workplace monitoring for a specific operational purpose. At its Canton, Miss., plant, cameras use computer vision to analyze how assembly workers move, identifying positions that could indicate excessive bending, prolonged awkward postures or deviations from the plant's preferred motion sequence. According to Nissan, it initially explored the system to improve manufacturing efficiency and train workers, and it later recognized the potential to identify ergonomic risks.

You can turn off the technology, but you can't undo the loss of trust that happens when employees discover they were being monitored.

A framework for the C-suite

Research on AI and algorithmic management identifies several safeguards that involve employees or their representatives in choosing what gets measured, publishing clear policies on AI use and maintaining meaningful human oversight for consequential decisions.

For CIOs and other enterprise leaders, those principles can be distilled into six practical considerations.

1. Start with a specific business problem

Businesses should identify what they are trying to solve and determine whether AI monitoring is necessary to address it. Don't collect data simply because a tool can capture it.

"Just because an organization can measure something doesn't mean it should," Huthwaite said.

Leaders should also be cautious about framing surveillance primarily to cut costs or reduce head count, which can quickly undermine employee trust.

2. Be transparent before deployment

Companies should inform employees about what data they collect, why they're collecting it, who can access it and how they will use it. Businesses should enable employees to ask questions or raise concerns before deployment and provide feedback after the system is in place.

Huthwaite suggested giving employees five straightforward answers to the following questions: What is being analyzed? Who can see it? Why is it being analyzed? What will it be used for? And what will it never be used for?

"You can turn off the technology, but you can't undo the loss of trust that happens when employees discover they were being monitored," Hartstone said.

3. Keep AI out of sole decision-making authority

AI-generated insights can inform decisions about performance or behavior, but they shouldn't be the sole basis for discipline, termination, compensation or promotion. Employees should also have a way to review and challenge inaccurate or misleading information.

Hartstone said organizations can have a human in the loop while still letting an AI-generated conclusion drive the outcome. A stronger safeguard is to require a named person to authorize consequential actions rather than reviewing an automated recommendation after the fact.

4. Give employees access and a way to challenge results

Employees should be able to see assessments made about them and understand the evidence behind those assessments. An appeals process can protect workers while also exposing blind spots in the model.

5. Test systems for bias, accuracy and model changes

Enterprises shouldn't rely solely on vendors' claims about fairness. They should evaluate how monitoring tools perform against their own workforce data and continue auditing them after deployment.

Meshyk recommended testing model outputs across work patterns such as role, tenure, shift and location, as well as legally permissible demographic characteristics. Employers should check whether the system flags certain groups of employees more often than others in ways that can't be explained by differences in the work itself. Testing should also continue after deployment, he said, because model updates can change system behavior.

6. Set clear data, access, retention and regulatory boundaries

Enterprises should determine what information is necessary, who can access it, how long it will be retained and which legal requirements apply to the specific use case.

Leaders should also scrutinize vendors. Meshyk recommended asking what the model was trained on, how bias is tested, whether individual features can be disabled, who owns the resulting data and whether the organization can export or delete it.

"Vendors that resist independent testing or market predictions about employee intent or emotions should receive particular scrutiny," Meshyk said.

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.