New Studies Find ChatGPT Is Quietly Reshaping Critical Thinking

Five years ago, asking an AI chatbot to summarize a plot you half-watched would have sounded absurd. Today it takes longer to type the question into Google than to ask ChatGPT to catch you up. You hand the bot a one-line prompt, it hands back a tidy recap, and the next episode starts without the cognitive inconvenience of actually thinking about what you saw.

That shift is now under the microscope. Three separate research groups, working independently, have landed in the same place: heavy reliance on generative AI is showing up as weaker neural engagement, less critical evaluation of outputs, and, in younger users, lower scores on standard critical thinking tests. The findings are not proof that AI is breaking our brains. They are the first measurable sign that something is bending.

What the New Studies Actually Measured

A central paper in the conversation is “Your Brain on ChatGPT,” from MIT Media Lab. It tracked brain activity in adults writing essays, with one group using ChatGPT, one using Google’s search engine, and one with no external tools. The full comparison of who did what sits in the table below.

The second paper comes from Microsoft Research and Carnegie Mellon. Presented at CHI 2025, the study surveyed knowledge workers on when they apply critical judgment to AI-assisted tasks and when they don’t. A third piece, from Michael Gerlich at SBS Swiss Business School, was published in the journal Societies on January 3, 2025 and tested adults across age and education bands on their AI habits and critical thinking.

All three reached overlapping conclusions through different doors. The table lays out who ran what, on whom, and what they found.

How the Brain Reacts to an AI Assistant

The MIT Lab’s EEG readings placed the LLM group at the bottom of the three. The brain-only group showed the strongest, most distributed connectivity, the search-engine group sat in the middle, and the LLM group registered the weakest. “Cognitive activity scaled down in relation to external tool use,” the team wrote. Two English teachers who graded the essays later called the LLM-assisted submissions largely “soulless.”

The result that landed harder came in session four. The researchers asked the LLM group to write one more essay, this time without ChatGPT. The participants remembered little of what they had previously submitted, and their alpha and beta waves, the bands tied to memory and attention, registered reduced engagement.

The task was executed, and you could say that it was efficient and convenient. But as we show in the paper, you basically didn’t integrate any of it into your memory networks.

That framing comes from MIT Media Lab researcher Nataliya Kosmyna, lead author of the study. Her team is now running the same protocol on software engineering tasks. Kosmyna says the early findings for code are “even worse” than for essays, a warning for companies hoping to swap entry-level developers for AI before there is data on what that swap costs in cognitive skill.

Trust in AI Tracks Less Thinking

The Microsoft and Carnegie Mellon study zeroes in on the condition that decides whether AI use sharpens or shrinks a worker’s own thinking: confidence. “Higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.” Workers who trusted ChatGPT’s answers were the ones most likely to skip the verification step.

That insight reframes the old question of whether to use AI into a sharper one: how you use it, and how much you believe it. The 936 task examples clustered into three recurring patterns, information verification, response integration, and task stewardship, each with a different effect on cognitive effort. Workers who outsourced the verification step outsourced the most thinking.

Mutlu Cukurova, a professor of learning and artificial intelligence at University College London, calls the dominant pattern the danger zone. “At the moment, most interactions between people and generative AI are transactional. You ask a question; you get an answer; you move on. Unfortunately, these types of interactions are more likely to lead to cognitive atrophy than cognitive development.”

The Younger the User, the Steeper the Curve

Gerlich’s study surfaces a finding that should worry parents and school administrators more than office workers. Participants between 17 and 25 reported the heaviest AI use and the largest amount of cognitive offloading. They also scored lowest on the critical thinking assessments. Participants aged 46 and up used AI less, offloaded less, and scored higher.

“One surprising finding was the extent to which younger participants exhibited higher dependence on AI tools and, correspondingly, lower critical thinking scores,” Gerlich said in an interview. “This suggests that digital natives, who have grown up with AI-integrated technologies, might be more prone to cognitive offloading than older generations.”

Kosmyna released her paper before peer review for one reason. “I am afraid in 6-8 months, there will be some policymaker who decides, ‘let’s do GPT kindergarten.’ I think that would be absolutely bad and detrimental at the most at-risk age. Developing brains are at the highest risk,” she told TIME. Dr. Zishan Khan, a child and adolescent psychiatrist quoted in the same piece, treats children who lean on AI for schoolwork. “From a psychiatric standpoint, I see that overreliance on these LLMs can have unintended psychological and cognitive consequences, especially for young people whose brains are still developing. These neural connections that help you in accessing information, the memory of facts, and the ability to be resilient: all that is going to weaken.”

Productivity Gains Are Real, and So Are the Costs

The new research does not sit alone. The productivity numbers it lives beside are not imaginary. A Harvard Business School study published in May 2025 found that generative AI made workers more productive and, in the same breath, less motivated. A Wharton Budget Model estimate projects AI lifting productivity growth by 1.5% by 2035, almost 3% by 2055, and 3.7% by 2075.

Those numbers do not contradict the cognitive findings. They are two measurements of the same shift: one in output, the other in the thinking underneath. A worker finishes a task faster while the mental work that used to come with the task is shrinking. Companies paying for AI subscriptions see the productivity line move up; the workers doing the work see the thinking part move out.

The Microsoft paper put the migration in plain terms: critical thinking shifts from generating ideas toward verifying AI’s output, integrating it into a final piece, and stewarding the result. That cocktail still requires thought. It just requires a narrower kind, and the band of what counts as “thinking” gets thinner each year the defaults hold.

Use AI Deliberately, or Use It Less

Nita Farahany, a professor of law and philosophy at Duke University, framed the cognitive stakes of generative AI in a single question. Her research focuses on the legal and ethical questions raised by emerging technologies.

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. What does it mean for society if humans are passively receiving information but no longer able to critically interrogate it?

Brooke Macnamara, a psychologist at Purdue who studies skill acquisition, drew the lesson from GPS. “We essentially become passive passengers following directions without processing and synthesizing information. But if a technology is very good and almost always available, maybe that’s a skill we’re willing to lose.” The point is the one the research keeps circling back to: the cost of AI depends on whether the user is steering or riding.

None of the studies argue for stopping. ChatGPT has reached 900 million weekly active users as of February 2026, and the productivity gains are real enough to justify its place at most desks. The choice in front of every user is whether to treat AI as a transaction, prompt in and answer out, or to deliberately keep the verification, integration, and judgment steps inside their own head.

The data on related work points the same direction. BairesDev’s Q2 2026 Dev Barometer found that only 16% of senior engineers trust junior AI-assisted code in a survey of 1,569 developers, because the verification step is the one being skipped at every level of the stack.

The cognitive offloading framework that underpins the new research has a longer history. A short ledger:

• The “cognitive offloading” concept was formalized by Evan Risko and Sam Gilbert in their 2016 review of how people delegate thinking to external tools.
• The phrase “cognitive misers,” used by researchers in the new studies, dates to Susan Fiske and Shelley Taylor’s 1984 social psychology work.
• Earlier MIT Media Lab research found that more time spent talking to ChatGPT correlated with higher reported loneliness.

Frequently Asked Questions

Does ChatGPT actually hurt your brain?

The strongest evidence comes from the MIT Media Lab’s EEG study of 54 adults, which found lower brain connectivity during essays written with ChatGPT than during essays written without any tool. Researchers called the effect “cognitive debt,” because some participants showed reduced activity even after the AI was removed. The paper has not been peer reviewed and the sample size is small, so it is a signal, not proof.

How much AI use is too much?

None of the three studies set a numeric limit. The pattern is conditional: the higher a user’s confidence in AI, the less critical thinking they apply to its output. Use becomes risky when the verification step is the one being outsourced.

Are these results just correlational?

Yes for the survey studies, no for the EEG study. Gerlich’s 666-person poll finds correlations between AI use and lower critical thinking scores, but cannot rule out reverse causation (people with weaker thinking might choose AI more often). The Microsoft study is also self-reported. The MIT Lab paper measures brain activity directly, though it is a preprint with a small sample.

What about kids and teenagers?

Both Gerlich’s data and Kosmyna’s commentary point to the 17-to-25 age window as the most exposed. Kosmyna released her paper before peer review specifically to warn against AI-led early education. The 18-to-39 cohort in the MIT study already shows reduced neural engagement after a few months of LLM-assisted writing.

Should I stop using AI at work?

The researchers do not argue for stopping. They argue for keeping the verification, integration, and judgment steps. Workers who treat AI as a tool they double-check tend to retain thinking capacity; workers who treat it as an answer they trust tend to offload it.

Have the findings been peer reviewed?

The MIT Media Lab paper is a preprint released before peer review. The Microsoft and Carnegie Mellon paper was peer reviewed and presented at CHI 2025. Gerlich’s Societies paper was peer reviewed and published on January 3, 2025. The MIT result is the most striking and the least formally vetted of the three.