New Research Links AI Use at Work to Less Critical Thinking

A study from Microsoft Research and Carnegie Mellon, presented at the CHI 2025 conference, found that workers who reported the most confidence in generative AI also reported the least critical thinking when using it on the job. The finding landed inside a sample of 319 knowledge workers, who contributed 936 first-hand examples of how they actually used tools like ChatGPT and Copilot at work.

Researchers called the result a confidence paradox. Self-confidence went the other way: workers who trusted their own skills reviewed AI output more carefully, not less. That pairing matters because the better generative AI gets, the more natural it is to stop checking it. And the most checked-out users, in this data, are the ones with the highest trust in the system producing the answers.

The Confidence Trap at Work

The lead finding from the Microsoft and Carnegie Mellon team is sharp enough to repeat: higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking. The relationship held up after the researchers controlled for task type and the worker’s own experience.

The study surveyed people who use AI in their jobs, from software developers to social workers, and asked them to walk through specific tasks they had done with AI. The pattern held across occupations, and the authors saw the same shape in the qualitative examples: when a worker trusts the model, the verification steps get shorter. The authors describe this as a tradeoff that may look like efficiency, until the moment the AI is wrong.

Inside the 319-Worker Study

The work came from a collaboration between Microsoft Research and Carnegie Mellon University, published in the Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. The researchers asked participants to define what critical thinking meant to them when AI was in the loop.

Workers described how critical thinking shows up at work in three familiar moves: setting clear goals before prompting, refining the prompt, and verifying AI outputs against external sources and their own expertise. Those are the practices the rest of the study measured.

The data also showed where those practices break down. Why critical thinking drops with AI came down to three recurring barriers the workers themselves named.

• Awareness: not questioning whether the AI was competent for a simple task, so the verification step never started.
• Motivation: running out of time, or treating critical review as somebody else’s job.
• Ability: not knowing how to verify an AI output in a domain the worker did not already understand.

The researchers were careful not to claim causation. The study is a self-report survey, not an experiment. What it does show is a stable association across hundreds of workers and hundreds of task examples, which is what makes the corroboration from a different kind of study worth taking seriously.

Brain Scans Catch Up to the Survey

A separate team at the MIT Media Lab set out to find the same signal in a different place: the brain itself. Their paper, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” put EEG headsets on participants and watched what happened while they wrote.

Three groups wrote essays across multiple sessions: Brain-only, with no tools; Search Engine, with Google; and LLM, with an AI assistant. MIT’s EEG study on ChatGPT writing found that Brain-only participants showed the strongest, most distributed neural networks, Search Engine users were moderate, and LLM users had the weakest connectivity. The result held across the four-month observation window.

A fourth session swapped the conditions. Participants who moved from LLM back to Brain-only did not bounce back inside the study window; they still showed reduced alpha and beta connectivity. Workers who moved from Brain-only to LLM ramped up quickly. The asymmetry is part of the finding.

Self-reported ownership of the essays tracked the brain data. LLM users reported the lowest sense that the essay was theirs and, when asked to quote from their own text, they struggled to do it accurately.

A Familiar Pattern in New Clothes

The American Psychological Association’s Monitor pulled the threads together in its July-August 2026 issue, framing the Microsoft, MIT, and broader findings as an instance of a much older pattern. Psychologists call it cognitive offloading, using an external tool to reduce the mental effort a task takes.

“There’s a general consensus that people are ‘cognitive misers.’ If a task can be done with less cognitive labor, we tend to take that option,” Evan Risko, a psychology professor at the University of Waterloo, told the APA Monitor. The same issue quotes Brooke Macnamara at Purdue, who studies how skills degrade: “We essentially become passive passengers following directions without processing and synthesizing information.” That is the GPS analogy she returns to, and cognitive offloading and AI use at work reads, in her telling, like GPS for higher-order thinking.

Generative AI extends the analogy past navigation. “One of the things that’s seductive but also troubling about generative AI is the ability to offload much more executive functioning than we have in the past,” Nita Farahany, a professor of law and philosophy at Duke University, told the Monitor. “What does it mean for society if humans are passively receiving information but no longer able to critically interrogate it?”

By the numbers:

• 319 knowledge workers surveyed in the Microsoft and Carnegie Mellon study
• 936 first-hand GenAI use examples collected in that study
• 54 participants across the first three sessions of the MIT EEG study
• 4-month observation window in the MIT study
• 31 countries in Microsoft’s 2025 Work Trend Index

The Scale of What’s Already in Use

The Microsoft and Carnegie Mellon paper covered one slice of knowledge work. The wider rollout is visible in Microsoft’s 2025 Work Trend Index, which surveyed 31,000 full-time workers in 31 markets between February and March 2025.

The Work Trend Index found that 64 percent of workers say they struggle with having the time and energy to do their job, and those workers are 3.5 times more likely to also struggle with innovation and strategic thinking. Sixty-eight percent say they do not have enough uninterrupted focus time during the workday. The same report found that 70 percent of workers would delegate as much work as possible to AI to lessen their workloads, even as 49 percent said they are worried AI will replace their jobs.

When the index asked leaders which skills employees will need for an AI-powered future, the top three were analytical judgment, flexibility, and emotional intelligence. Eighty-two percent of leaders said their employees will need new skills to be ready for the growth of AI, and the report identified analytical judgment as a new core competency, not just for technical roles. That is the skill the cognitive offloading literature says is most at risk of slipping.

The Irony of Automation

The Microsoft and Carnegie Mellon researchers flag a long-running concern inside their own field. Their paper warns that relying on AI for low-stakes tasks like proofreading might appear benign but can lead to significant negative outcomes in high-stakes contexts, because workers only apply critical thinking in those high-stakes moments, and cognitive abilities can deteriorate when routine practice disappears.

The same paper cites earlier findings that dependence on technology deprives humans of the chance to constantly hone their judgment skills, leaving them atrophied and unprepared when the exceptions do arise. That is the irony of automation: the more the system handles, the more important oversight becomes, and the less practice the human operator has for it.

Our survey-based study suggests that when people view a task as low-stakes, they may not review outputs as critically. However, when the stakes are higher, people naturally engage in more critical evaluation.

That is Lev Tankelevitch, a senior researcher at Microsoft Research and one of the study’s authors. The risk he names is not that people lack skill, it is that people stop using the skill on the routine tasks where it gets built. When a high-stakes moment arrives, the worker’s last real practice may be weeks or months old.

The same Microsoft Research team published a separate recommendation in the CHI 2025 blog: treat AI as a thought partner, one that can also act as a provocateur, rather than as a vending machine. The framing matters because the design of the tool shapes whether the worker practices verification or skips it.

What Could Bend the Curve

The papers do not leave the reader at the diagnosis. The Microsoft Research team at CHI 2025 lays out specific design moves that turn AI from a thought partner back into a thinking partner.

Their recommendations include proactive prompts that surface overlooked tasks, reasoning explanations that show how the model reached an answer, guided critiques that push back on the user, and cross-references that the worker can check. The broader pattern is to build friction back into the routine, so that critical engagement becomes the default rather than the exception.

• Proactive prompts that surface tasks the worker might otherwise skip
• Reasoning explanations and guided critiques instead of bare answers
• Cognitive forcing functions that require active engagement before proceeding
• Verification steps built into routine workflows, not only high-stakes ones
• Treat the AI as a thought partner, not a vending machine

The question isn’t whether AI stays at work. The workplace has already decided that. The question is whether the workplace builds in the friction that keeps thinking sharp, and whether the design of the tools makes verification the easy path rather than the hard one.

Frequently Asked Questions

Did the Microsoft study actually prove AI makes people worse thinkers?

The study found an association, not a cause. It surveyed 319 workers who use AI on the job and asked them to describe specific tasks. Workers who reported higher confidence in AI also reported less critical thinking in those examples. That is correlation across hundreds of self-reports, strong enough to warrant the follow-up studies now underway.

Is the MIT “Your Brain on ChatGPT” study peer-reviewed?

As of mid-2025, the MIT Media Lab paper was posted on arXiv as preprint 2506.08872. It had not gone through formal peer review at the time it was circulated, which the authors flag in their abstract. The findings align with the survey data from Microsoft and Carnegie Mellon, but the brain-scan result should be read as preliminary corroboration, not settled science.

What does “cognitive offloading” mean in plain terms?

It means using an outside tool to take over a thinking task you used to do yourself. Writing down a phone number instead of memorizing it is a small version. Letting a GPS pick the route is a bigger one. Letting an AI write the first draft and then accept the framing is the latest and largest version, and the one researchers say is most likely to reshape the underlying skill.

Does this mean workers should stop using AI?

None of the researchers surveyed recommend that. The Microsoft team frames AI as a thought partner that should challenge the user. The MIT team notes that switching from LLM to Brain-only does not immediately restore the neural patterns they measured, which is a caution about pacing, not a ban. The advice across the literature is to keep the verification step in the loop, even when the AI looks right.

Whose job is it to fix this, the worker, the employer, or the AI vendor?

The studies point at all three. Workers can keep verification in their workflow on routine tasks. Employers can design incentives that reward scrutiny instead of just output speed. AI vendors can ship tools that require a check before proceeding. The Microsoft recommendation treats it as a design problem first and a behavior problem second, which is where the most leverage sits.