Midnight. The report is due at nine. You open ChatGPT, paste the task, and a minute later you copy a finished draft. The work is done. The real question is what happened inside your head during that minute.
Short answer: AI does not make people dumber on its own. It amplifies what you were already doing. Think alongside it and you learn faster. Hand over the thinking itself and the skill fades, quietly, without you noticing. Experiments show both outcomes in the same people.
Below: what researchers actually measured, where the line between the two runs, and why one study unsettled doctors with twenty years of practice. The doctors come later.
What happens in your brain when ChatGPT thinks for you
It gets noticeably quieter, and that shows up on instruments. At MIT in the United States, 54 people spent four months writing essays in one of three modes: with ChatGPT, with a search engine, or with nothing at all. Throughout, researchers recorded their EEG (electroencephalogram - a record of the brain's electrical activity through sensors placed on the scalp).
The pattern was clean. The densest connections between brain regions belonged to people writing unaided. Search engine users landed in the middle. The AI group came last.
Then it gets stranger. Minutes after submitting, people in the AI group could not quote their own essays. They barely felt the text was theirs. The brain does not store what never passed through it.
The authors called this cognitive debt: you save effort today and pay for it in thinking quality tomorrow. Picture a muscle that let the machine do the whole workout. The weight moved, the report shipped, the strength never came.
The most useful part came in session four, when 18 people swapped conditions. Those who had written unaided and then got AI showed strong brain activity and solid recall. Those used to AI and left without it dipped: the brain seemed to have forgotten how to fully switch on.
Which gives us the first rule, and we will come back to it: the order matters more than the tool. Your head first, AI second, and the result flips.
One caveat matters. This is a preprint (research posted before other scientists have reviewed it), the sample is 54 people, and the task was one short essay. The authors themselves asked people not to turn their findings into the headline "AI makes you stupid." One study is not enough. Luckily a bigger and stricter one exists.
Why AI makes some people sharper and others duller
The difference is not the AI. It is whether the AI hands over an answer or makes you work for it. A 2025 experiment in Turkey, published in PNAS - one of the world's most respected journals - split nearly a thousand high school students at random into three groups and set them loose on maths.
- Control group - textbooks and notebooks only.
- GPT Base - plain ChatGPT, no restrictions.
- GPT Tutor - the same GPT-4, but with guardrails: it gives teacher-written hints instead of answers.
While the AI was available, both groups soared: grades rose 48% for GPT Base and 127% for GPT Tutor. It looked like a clean win for technology.
Then the AI was taken away and everyone sat an exam. GPT Base scored 17% worse than students who never touched AI at all. Not worse than with a cheat sheet - worse than with nothing. GPT Tutor, meanwhile, matched the control group. Hints instead of answers cancelled the damage entirely.
Here is the detail that makes this personal. The students never noticed. While the AI was there, they felt they had genuinely mastered the material. The feeling of understanding and actual understanding are different things, and AI hands out the first far more generously than the second.
So that is the fork in the road. One tool, two modes. "Do it for me" costs you the skill. "Show me where I went wrong" builds it. Which raises the question of why adults keep picking the first mode even when they know about the second.
The more you trust AI, the less you check
Trust in the machine and your own checking behave like a seesaw. In 2025 researchers at Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers and examined 936 real situations where they used AI at work.
The results mirrored each other. People with more confidence in the AI verified less and applied less critical thinking (the habit of doubting a claim and checking it against something else). People with more confidence in themselves argued with the model and checked it more often, even though that costs effort.
Participants described exactly what checking involves: set a clear goal, rewrite the prompt, compare the answer against an outside source and your own experience. That is work. It is easy to skip when the text on screen reads smooth and certain.
Let us name the problem plainly. AI writes with confidence always - when it is right and when it is inventing. The tone tells you nothing about accuracy. If you have no view of your own on a topic, you have nothing to catch the error with.
Cognitive offloading, and why it hits young users hardest
The mechanism has a name: cognitive offloading, handing mental work to something outside your head. Writing a shopping list so you need not hold it in memory. Not learning a route because the map app knows it. None of that is harmful in itself, and it frees your head for harder problems.
Trouble starts when what gets offloaded is the reasoning rather than the trivia. A 2025 study in the journal Societies surveyed and interviewed 666 people of different ages and education levels. The finding: the more often someone used AI tools, the lower their critical thinking scores, with offloading as the link between the two.
Two details stand out.
- Age. Younger participants leaned on AI harder and scored lower on critical thinking than older ones.
- Education. The more of it people had, the stronger their thinking, regardless of how much AI they used. The habit of thinking works like insurance.
And the limit you must keep in mind: this is a correlation, not a proven cause. Maybe AI weakens thinking. Maybe people who dislike digging into things reach for ready answers more often. The data allows both readings.
Fine, you might say - students and office workers. What about a genuine expert with thousands of hours behind them? Now the doctors.
Even doctors with thousands of procedures lose the touch
They do, and this is the first such finding in real patients. In August 2025 The Lancet Gastroenterology & Hepatology published work from four endoscopy centres in Poland. Nineteen experienced doctors, each with more than 2,000 colonoscopies behind them, began routinely working with an AI assistant that highlights suspicious areas on screen.
Researchers then compared how those same doctors performed without AI, before the clinic adopted it and after: 1,443 unassisted procedures in total, 795 before and 648 after.
The share of adenomas found (pre-cancerous polyps - growths that can eventually turn into a tumour) fell from 28.4% to 22.4%. That is a fifth of the previous rate, gone. Same hands, same experience, same equipment. But an eye used to the machine doing the spotting started spotting less.
The caveats here are serious and worth stating. The study is observational: doctors were not randomly assigned, only compared before and after, so it cannot prove cause. The project coordinator warned openly that this patient group differs from the main trial population and the finding should not be generalised yet. It is a hypothesis, not a verdict.
A sobering hypothesis all the same. If a skill built on thousands of repetitions can slip within months of AI arriving, a skill you are still building will slip faster.
So is there evidence that AI helps?
There is, and it is decent - it simply describes the other mode of use. In 2026 Humanities and Social Sciences Communications published a meta-analysis (a study that pools the results of many separate studies): 35 experiments, 4,193 participants. The overall effect of ChatGPT on learning outcomes was moderately positive, g = 0.670.
In plain words, g measures how much an intervention shifted the result. Around 0.2 is weak, 0.5 is medium, 0.8 is strong. So 0.67 sits comfortably above average. Only three of the 35 experiments produced a negative result.
The effect also depended on subject, duration and teaching format. What works is not "AI in general" but a specific way of building it into learning.
And here is a note you rarely see in articles on this topic. The most quoted ChatGPT meta-analysis, the one behind the headline figure of g = 0.867, was retracted by its journal in April 2026 over discrepancies in the analysis. It still circulates in summaries and slide decks. If you see that number, you now know not to lean on it.
Put it all together. The studies showing harm and the studies showing benefit are not in conflict. They describe two different ways of handling the same program.
Who this concerns: not just students
Almost anyone who goes online. Economists at NBER analysed 1.1 million anonymised conversations and estimated that by July 2025 roughly one in ten adults on the planet was using ChatGPT, with about 70% of messages unrelated to work.
One more figure from the same paper shows the scale: user numbers grew fastest in lower-income countries, more than four times the rate seen in wealthy ones. For millions of people AI is becoming their first and main "smart interlocutor" - with no school lesson on how to handle one.
Which leads somewhere simple. Working with AI is now a household skill, like reading a food label. It is worth learning deliberately rather than by instinct.
Six rules for using AI without going dull
The core rule: your thought first, the machine's second. Everything else follows from it.
- Draft it yourself. Write your own answer in two sentences before opening the chat. That exact order produced the highest brain activity and the best recall in the MIT experiment.
- Ask for hints, not answers. "Don't solve it - ask me three questions that lead me there." That is precisely the GPT Tutor mode that erased the damage in the Turkish trial.
- Argue with the model. "Find the weak points in my reasoning." "Give me the strongest counterargument." Models default to agreeing with you, and agreement teaches nothing.
- Verify numbers and links by hand. At least one key figure per conversation. The retracted meta-analysis above is a live example of how a pretty number outlives its own truth.
- Split your tasks. Routine where no skill is at stake - formatting, cleanup, template translation - hand it over freely. Anything you actually want to understand, keep in dialogue mode.
- Keep AI-free zones. One task a week done entirely in your own head. Not nostalgia - training. A muscle that occasionally lifts without the machine remembers how.
A word on children and teenagers. The link between heavy AI use and weak critical thinking was strongest in younger users. While a skill is still forming, a ready answer does not save time, it removes the learning.
If you want to understand how these tools actually work and how they differ, we have a full guide to ChatGPT, Claude and Gemini. And on which jobs are changing first under AI pressure, we wrote a separate breakdown.
Sharper or duller
Back to that midnight report. Whether you end up sharper or duller is not decided by opening the chat. It is decided by what you did in the thirty seconds before. Write your own version and AI extends it. Skip that and AI fills the emptiness, and an hour later you will not recall a single line - exactly like the MIT participants who could not quote essays they had just submitted.
Right now, in two minutes: take the task you were about to paste into a chat and write three lines in your notes first - what you think about it and what you do not understand. Then send it. You will feel the difference on the first try.
Sources
- MIT Media Lab. Your Brain on ChatGPT: Accumulation of Cognitive Debt (2025)
- Bastani et al. Generative AI without guardrails can harm learning. PNAS (2025)
- Lee et al. The Impact of Generative AI on Critical Thinking. CHI 2025
- Gerlich M. AI Tools in Society: Impacts on Cognitive Offloading. Societies (2025)
- Endoscopist deskilling risk after exposure to AI in colonoscopy. The Lancet Gastroenterology & Hepatology (2025)
- Wu et al. ChatGPT's impact on student learning outcomes: a meta-analysis of 35 experimental studies (2026)
- Chatterji et al. How People Use ChatGPT. NBER Working Paper 34255 (2025)
