AI could forever change the human psyche, experts say

The scale and scope of AI dependence

GenAI is shifting the balance between capabilities and constraints, enabling us to feel more productive while eroding our innate skills. Consider what GenAI can do, including the following:

• It can take on more. GenAI tools can offload processes involving language for translation, synthesis and processing of unstructured data, expanding the scope of tasks we can offload.
• It asks less. GenAI tools reduce the friction of thinking through problems at work and in life.
• It agrees more. GenAI chat systems tuned for user satisfaction tend to agree with the premise we bring to them, making us overly confident in our ideas.

Across these domains is a problem with metacognition, or awareness of our own thinking. AI use can affect how we judge what we know, how we assess the reliability of our memory and our estimation of how much effort a task will require. Increasing scope and agreeability while reducing friction can make it harder to judge what we know versus what the tool knows. This can lead to false confidence and dependence across various elements, such as the following:

• Memory. Offloading recall unburdens our focus but comes at the cost of recalling specifics in the moment or in context.
• Critical thinking. Offloading removes the need to practice the critical thinking required to excel in a field.
• Calibration. Fluent AI can obscure the balance between what we know and don't know.
• Efficiency. Tools that make people faster at some activities, such as writing lines of code, can exacerbate downstream problems, such as code quality issues, that are harder to debug without the requisite understanding.
• Validation. Agreeable systems can reinforce existing biases or bad ideas, leading to delusions and ruined lives under the worst conditions, or to poor choices in everyday work decisions.
• Connection. Systems that excel at creating a feeling of being understood without the friction of human relationships can interfere with our ability to form healthy relationships or engage in productive debate with colleagues.
• Identity. Heavy AI users describe their thinking, writing and self-image being reshaped, for better or worse.

Taking on more

Sonya Cullington, a cyberpsychologist and digital policy advisor, said that GenAI tools represent a category shift from previous tools, such as calculators, which replace a bounded, verifiable procedure. "Generative AI is different because it can step into judgment: the selection of what matters, the weighing of competing considerations and the tacit sense of what good looks like in this context," she said. The problem is that checking its output requires the same type of judgment the person was trying to offload. Cullington also said that a common concern is that AI might erode existing skills. Still, a larger problem is that early-career staff might never build underlying competencies as they offload valuable experiential opportunities.

In the study "Generative AI at Work," researchers collected data from 5,172 customer support agents and found that productivity rose by about 15% on average, with higher gains for less experienced workers and the largest gains for moderately rare problems. However, the most skilled workers saw small gains in speed and small declines in quality. That said, customers working with the AI-assisted workers tended to respond more politely and were less likely to ask to speak to a manager.

A study of experienced endoscopists titled "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study" used AI-assisted adenoma detection with over 1,400 patients and found that their detection rate in unassisted procedures dropped from 28.4% to 22.4% after they adopted the tools. With AI assistance, their detection rate was 25.3%, better than the unassisted rate afterward but still lower than before AI arrived. Marcin Romańczyk, a gastroenterologist at the Academy of Silesia who co-authored the study, said the team did not understand why the detection rate was lower with AI and that more research is needed.

Anthropic researchers found a similar pattern in a paper they published, "How AI assistance impacts the formation of coding skills." Their research focused mainly on junior engineers working with an unfamiliar programming library with or without AI help. Afterward, the AI-enabled group scored 50% compared to 67% for the unassisted group that hand-coded. But the effects varied by use. Developers who used AI only for conceptual questions, or asked it to explain the code it generated, scored higher than those who had AI write the code.

Adnan Masood, PhD., chief AI architect at UST, observed the same phenomenon on his own team. "The skill that has actually weakened is debugging, because people read generated code to see if it looks right instead of working through how it runs," he said.

He encourages his team members to ask four questions: Can you explain why the code works, how it fails, how to change it when assumptions change and how to keep going without the tool. "If all you can do is regenerate the answer, the tool replaced the skill," Masood said.

Asking less

GenAI also makes it easy to skip the effort required to keep skills sharp. Srinivas Chippagiri, a senior member of technical staff at Salesforce, said, "Choosing the slower, unaided path when a faster one is one keystroke away takes deliberate discipline, and I do not always win that trade." He found that when a tool stopped working or wasn't available, he could still recall an API signature or a regex he had written a hundred times. But that came with a new quality of hesitation he hadn't accounted for.

Feeling faster is not always the same as being faster. Google's 2024 DevOps Research and Assessment (DORA) survey, "Accelerate State of DevOps," collected responses from over 39,000 professionals from organizations across the globe, and found that 75% of respondents reported productivity gains when using AI. It also found that increased AI adoption was associated with lower software delivery throughput and stability. The 2025 DORA survey, "State of AI-assisted Software Development," collected responses from nearly 5,000 technology professionals globally and found that AI adoption had improved software delivery throughput but still increased instability.

METR, an AI research group, found a similar gap in its 2025 report "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." Before the study, 16 experienced open-source developers predicted they could reduce their work completion time by 24% with AI tools. The study concluded that they were slower when using AI tools, and completion time increased 19%. A 2026 follow-up showed signs that better tools sped things up, but METR conceded the results were unreliable because many developers would not agree to work without AI.

Michał Piszczek, CTO at Archdesk, said that his team's internal accounting found that verifying AI-generated code can cost five to 30 times as much as generating it. "People under pressure close that gap the easy way: they skip the check," he said. His team now caps code generation at the pace human reviewers can verify.

Cullington called this the productivity illusion, in which teams report time savings, but the judgment work shifts to whoever reviews the output. "In reality, the reviewer may be doing more cognitive work, not less, while the dashboard still tells the organization that work has been saved," she said. Middle managers are most exposed because dependence builds from many small decisions that each save a little time and are each defensible. A report that used to take four hours might take 30 minutes, but each step can take away the practice that builds judgment. "If you take away the four hours, you may keep the output and lose the judgment," she said.

Agreeing more

In the article "Sycophantic AI decreases prosocial intentions and promotes dependence," researchers from Stanford University and Carnegie Mellon University found that 11 leading models affirmed users' actions 49% more often than people did when asked about personal dilemmas, including deceptive or harmful behavior. A single agreeable conversation also reduced the willingness of the 2,405 participants to repair a conflict. Participants still trusted and preferred the agreeable models.

Anthropic researchers analyzed 1.5 million Claude conversations in their study "Who's in Charge? Disempowerment Patterns in Real-World LLM Usage," and found that users rated more favorably those interactions that had greater potential to undermine their autonomy.

Şerife Tekin, PhD., an associate professor at the Center for Bioethics and Humanities at SUNY Upstate Medical University, said that GenAI is taking on a greater role in people's personal lives without the training or accountability of a human therapist. "The danger is therefore not simply that the model 'doesn't really understand' but that it performs understanding well," she said.

Masood said the same thing happens at work. An engineer might propose an architecture, and the model accepts the premise, writes a polished justification and praises the engineer's choice. It looks like a second opinion, even though it's just agreeing with what might be a poor assumption. "Agreement feels like help, so people check agreeable output less," he said.

Masood's team asks the model for the case against its own answer. "We also track how often reviewers override the model, because a rate near zero means nobody is reviewing," he said. "My rule for engineers is that if a conversation leaves you more convinced but no better informed, go find a real source."

Are we prepared?

Regulators have begun preparing mitigation measures for consumer-facing AI, and their approaches have yet to be tested. China's new AI companion rules bar companies from designing services to replace social interaction or induce addiction. New York and California require disclosure and self-harm protocols for companion chatbots. But none of these rules cover workplace tools.

Masood treats AI outages as a test. "An outage I think presents you an opportunity for a free audit," he said. "If we slow down but keep working, the tool was helping." He is more concerned when people can't continue without writing a prompt or can't debug what an agent wrote last week. His team runs drills, as it would for disaster recovery, in which people work with limited AI access and system owners must explain their systems without the chat history.

Junior programmers are most exposed, Masood said, since many tasks they used to work on are now offloaded to agents.

Finding a way back to a less AI-dependent workplace will require deliberate effort and practice to identify new dependencies and encourage the development of critical skills, much like how commercial pilots are encouraged to practice flying manually despite the availability of autopilot. "The goal is not less AI. It is not losing the judgment that lets you catch [AI] when it is wrong," Chippagiri said.

Tekin was cautiously hopeful. "Human selves are remarkably revisable, so I don't think these effects are necessarily permanent or irreversible," she said.