Does AI rot your brain? What 8 studies found

Updated 30 September 2026.

None of the studies below shows that AI damages the brain, and the MIT team behind the best known study asks people not to call its result brain rot. Controlled trials in PNAS and from Anthropic show something smaller. People who hand a whole task to AI tend to learn and remember less of it, and how you use the tool changes that.

Last verified 30 September 2026. Every study below links to the paper or the page it was read from.

What did the eight studies find?

MIT Media Lab, 2025. 54 university participants wrote essays with ChatGPT, with a search engine or with nothing. The ChatGPT group showed the weakest measured brain connectivity on EEG and had the most trouble quoting essays they had just written. A preprint.arXiv
Microsoft Research and Carnegie Mellon, CHI 2025. A survey of 319 knowledge workers and 936 real tasks. The more people trusted the AI with a task, the less critical thinking they reported doing.paper
Bastani et al., PNAS 2025. Nearly 1,000 high school maths students. Plain GPT-4 raised practice grades 48%, then exam grades fell 17% below students who never had it once it was taken away. A tutor version built to protect learning largely avoided the drop.PNAS
Anthropic, 2026. A randomised trial with 52 developers learning a new Python library. The AI group scored 17% lower on a quiz afterwards, 50% against 67%, with the biggest gap on debugging.Anthropic
Melumad and Yun, PNAS Nexus 2025. Seven experiments, 10,462 people. Learning a topic from an AI summary left people with shallower knowledge than learning it from web links, and the advice they then wrote was sparser and less original.PubMed
Stadler, Bannert and Sailer, 2024. 91 students researched one question with ChatGPT or Google. The ChatGPT group felt less mental effort and produced weaker reasoning.paper
Shaw and Nave, Wharton, 2026. As reported by Ars Technica and Wharton: across 1,372 people, participants who consulted a chatbot that was wrong about half the time accepted its wrong answers 73.2% of the time. The paper itself was not read for this page.Ars Technica
Budzyń et al., Lancet Gastroenterology and Hepatology, 2025. After four Polish clinics brought in AI polyp detection, doctors working without the AI found adenomas in 22.4% of colonoscopies, down from 28.4% before.Lancet

How much weight can each study carry?

The strongest designs here are the PNAS maths trial and the Anthropic coding trial: people were assigned a condition and then tested without the AI. Microsoft's work is a survey of how people feel about their own effort, and the colonoscopy paper is observational, which its authors say leaves it open to confounding.

The MIT study had 54 people, and only 18 took part in the fourth session, where the groups swapped tools. A published comment on it lists concerns about sample size, the EEG method and inconsistencies in the reported results.comment The authors list their own limits too, starting with a small sample from one area and a single writing task done with a single model.MIT

Every study here measured people over minutes, weeks or a few months. None followed anyone for years.

Which claims does the evidence not back?

Brain scans prove AI shrinks your brain. The MIT paper measured connectivity during a writing task, not brain structure. Its FAQ asks readers to avoid the words brain rot, brain damage and brain scans altogether.MIT FAQ

Heavy AI users think worse. The best known version comes from a UK survey of 666 people. It found a negative correlation between AI use and critical thinking scores, which cannot show which causes which, and the paper reports two different sizes for that correlation in two tables.paper

The Google effect already proved this happens with search. The famous 2011 result did not hold up in a large replication project or in a closer replication in 2020.replication

A big meta-analysis showed ChatGPT improves thinking. That paper, covering 51 studies, now carries a retraction notice on its publisher's page.Nature

What pattern do the studies share?

It is about how people used the AI. In the maths trial, the same model with different rules produced a different result. In the coding trial, developers who only asked the AI conceptual questions, or asked it to explain the code it wrote, scored between 65% and 86%. Those who handed the task over entirely scored between 24% and 39%.arXiv In the Wharton experiments, small rewards for accuracy plus feedback on each answer made people overrule the wrong AI answers more often, and a 30 second timer made them do it less.Wharton

A September 2026 Nature feature sums up the field the same way: the studies conflict, few are rigorous or long term, and AI seems to help learning when it gives feedback or tasks that make people think.Nature

Where does Mover OS fit?

Mover OS is a set of workflows that runs inside Claude Code, Codex or Gemini CLI. Several of them are built around the habits the stronger studies link with better results. The evening review asks you to score your own day before it shows you any analysis, then points out where your guess and the data differ. Your daily log and journal are yours to write. When your work drifts from your plan it asks for a reason.

Nobody has studied Mover. There is no evidence it protects anyone's thinking, and this page makes no such claim. It also does not stop you handing a task to your agent, and you can overrule every question it asks. Mover OS is not for anyone who wants a phone app or a chat window: it needs one of those three terminal agents.

Common questions

Is the MIT Your Brain on ChatGPT study peer reviewed?

As of 30 September 2026 the arXiv listing shows it as a preprint, version 2 posted 31 December 2025, with no journal named. A comment paper posted on arXiv asks for a more cautious reading of its results.

Does ChatGPT make you dumber?

No study has measured a lasting drop in intelligence from using it. Several found that people who let it do a whole task remembered or understood less of that task afterwards, and one maths trial found students did worse on an exam once their AI access was removed.

Is it safe to use AI for work, then?

The research points at how you use it. Doing a first pass yourself, asking for explanations, and checking answers are the patterns linked with better results in the trials above. The Microsoft survey found people reported less critical thinking when they trusted the AI more.

What is cognitive offloading?

Using an action or a tool to reduce the mental work a task needs, like writing a shopping list. The 2016 review that defined it notes that offloading often improves performance, and that people are not always good at judging when to do it.

Sources

Mover OS: $49 once. Runs inside Claude Code, Codex or Gemini CLI. Asks for your read of the day before it shows its own. Does not run while the agent is closed.

Get Mover, $49 once