The research ยท last checked 13 November 2026
Does ChatGPT make you dumber? What 23 studies actually found
Short answer: nobody has shown that. No study has measured IQ before and after people started using chatbots, and none shows AI lowers IQ. What studies do find is narrower and more useful: when AI hands you the answer, you tend to learn and remember less from that task. When you try first, or the AI gives hints instead of answers, much of that loss goes away.
The one-line version. How you use AI matters more than whether you use it. Answer first, think later: weaker learning. Think first, ask second: much better.
What the strongest studies found
- Answers hurt exam scores; hints did not. In a field trial with about 1,000 high school students (Bastani and others, PNAS, 2025, peer reviewed), students who practised maths with plain GPT-4 did 48% better in practice, then scored 17% lower on the exam once access ended than students who never had it. A tutor version that gave hints, not answers, largely avoided the harm.
- Direct answers reduced persistence. In three experiments with 1,222 people (Liu, Christian and others, COLM 2026, peer reviewed), 10 to 15 minutes of AI help left people solving fewer problems on their own and giving up more. People who used the AI for hints did not show this.
- Writing alone first looked different from AI first. A small preprint from the MIT Media Lab (Kosmyna and others, 2025, 54 people) found lower brain engagement while writing with ChatGPT. In one session, 15 of 18 ChatGPT users could not correctly quote their own essay. People who wrote alone first, then used AI, showed higher connectivity than in AI-only sessions.
- What you ask matters. In a trial with 52 developers (Shen and Tamkin, Anthropic, 2026, a preprint), the AI group scored about 50% on a skills quiz against 67%. Those who asked conceptual questions kept learning. Those who handed the task over did not.
- Thinking first lowered blind trust. In a classroom study of 42 students (Islam and others, AIED 2026), "reflection first" gave the biggest drop in over-trusting the AI.
The try-first lesson
Put those together and one practical rule falls out. Before you ask, make your own attempt: a guess, a rough draft, what you already know. Then ask the AI to check your thinking or to coach you with hints. That is the difference between the groups that lost ground and the groups that did not.
It is also fair to say what cuts the other way. A carefully designed AI tutor at Harvard more than doubled learning gains (Kestin and others, Scientific Reports, 2025). A 2026 trial at the University of Maryland found a course AI tutor lowered grades. And in one study, answers with explanations beat layered hints. The evidence does not say AI is bad. It says design and order matter.
What no study shows
- That AI lowers IQ. No study measured IQ before and after chatbot use. The nearest thing is a 30-person preprint that gave four subtests of an adult intelligence scale before and after 4 weeks of ChatGPT use and found no change. That is too small to prove anything either way.
- Brain damage or permanent change. Brain activity recorded during a task is not damage. The authors of the MIT preprint ask the media not to use words like "dumb" or "brain rot" about their work.
- Long-term effects. Nearly every study runs for hours or weeks. One working paper follows school records for 30 months, and it is not a randomised experiment.
- That AI is bad for everyone. Several trials show gains when the AI makes people do the thinking.
Longer reads on single studies: does AI lower IQ, the MIT essay study, and the 1,000-student exam trial.
All 23 studies
Each row gives the year, how many people took part, what was measured, what was found, and whether the work has been peer reviewed. A preprint has not been checked by other scientists yet, so treat it as early evidence.
| Study | Year | People | What it measured | What it found | Status |
|---|---|---|---|---|---|
| Bastani et al., PNAS | 2025 | About 1,000 high school students | Maths practice scores with AI, then exam scores once the AI was taken away | Plain GPT-4 lifted practice scores 48%, but exam scores were 17% lower than for students who never had it. A tutor version that gave hints, not answers, largely avoided the harm. | Peer reviewed |
| Liu, Christian et al., "AI Assistance Reduces Persistence and Hurts Independent Performance" | 2026 | 1,222 adults, 3 experiments | Problem solving alone after 10 to 15 minutes of AI help | People who got direct answers solved fewer problems alone afterwards and gave up more. People who used the AI for hints did not. | Peer reviewed |
| Kosmyna et al., "Your Brain on ChatGPT" (MIT Media Lab) | 2025 | 54 (18 in the last session) | Brain activity (EEG) and recall while writing essays | Lower brain engagement while writing with ChatGPT. In one session, 15 of 18 ChatGPT users could not correctly quote their own essay. People who wrote alone first and then used AI showed higher connectivity than AI-only sessions. | Preprint |
| Shen and Tamkin (Anthropic), "How AI Impacts Skill Formation" | 2026 | 52 developers | A skills quiz after a coding task done with or without AI | The AI group scored about 50% against 67%. Those who asked the AI conceptual questions kept learning; those who handed the task over did not. | Preprint |
| Islam, Akgun, Billah, AIED 2026 | 2026 | 42 students | How much students over-trusted AI answers | "Reflection first" (think, then consult the AI) gave the biggest drop in over-trust. | Conference paper |
| Stromberg, Lei, Wu, CEPR discussion paper | 2026 | 26,811 students over 30 months | Homework and monthly exam scores after AI arrived (school records, not an experiment) | Homework scores rose 18% while monthly exam scores fell 20%. Students who kept spending a normal amount of time on homework had only small losses. | Working paper |
| Lee et al. (Microsoft Research and CMU), CHI 2025 | 2025 | 319 knowledge workers | Self-reported critical thinking across 936 real work tasks (survey) | More confidence in the AI went with less critical thinking. More confidence in yourself went with more. | Peer reviewed |
| Gerlich, Societies | 2025 | 666 UK adults | AI use and critical thinking scores (survey and interviews) | Heavier AI reliance went with lower critical thinking scores (r = -0.68). A correlation: it cannot show cause. | Peer reviewed |
| Kestin et al., Scientific Reports | 2025 | 194 Harvard physics students | Learning gains with a carefully designed AI tutor against active classroom teaching | The AI tutor, which made students work each step, more than doubled median learning gains in less time. | Peer reviewed |
| Liu et al. (University of Maryland), EdWorkingPaper 26-1598 | 2026 | 2,379 undergraduates | Final grades and participation with a course AI tutor | The course AI tutor lowered final grades by 0.37 standard deviations and participation by 0.90. | Working paper |
| Fan et al., British Journal of Educational Technology | 2025 | 117 students | Essay scores, knowledge gain and transfer | ChatGPT users wrote the best-scoring essays but showed no more knowledge gain or transfer. | Peer reviewed |
| Stadler, Bannert, Sailer, Computers in Human Behavior | 2024 | 91 students | Mental effort and quality of reasoning, ChatGPT against web search | ChatGPT lowered mental effort but students gave weaker reasoning than those who used a search engine. | Peer reviewed |
| Melumad and Yun, PNAS Nexus | 2025 | 10,462 across 7 experiments | Depth of knowledge after learning from AI summaries or web search | Learning from AI summaries gave shallower knowledge than web search, even with the same facts. | Peer reviewed |
| Kumar et al., CHI 2025 | 2025 | 1,100 across 2 experiments | Creative performance with AI help, then without it | Direct answers raised scores on one task while the AI was there. Afterwards, working alone, the groups that had coach-style AI scored lower than people who never had AI. The direct-answer groups did not differ significantly. | Peer reviewed |
| Doshi and Hauser, Science Advances | 2024 | 293 writers, 600 judges | Creativity and similarity of short stories written with AI ideas | AI ideas made individual stories more creative, and made the stories more alike. | Peer reviewed |
| Becker et al. (METR), developer trial | 2025 | 16 experienced developers, 246 tasks | Time to finish real coding tasks with early-2025 AI tools | In this 2025 trial, experienced developers were 19% slower with AI but believed they were 20% faster. METR's 2026 follow-up suggests a possible speedup and calls its own data unreliable. | Preprint |
| Budzyn et al., Lancet Gastroenterology and Hepatology | 2025 | 1,443 colonoscopies, 4 centres | Doctors' detection rate without AI, before and after AI was introduced (observational) | After AI was introduced, doctors' unaided detection rate fell from 28.4% to 22.4%. | Peer reviewed |
| Anthropic Education Report | 2025 | 574,740 student conversations | How university students use Claude (usage analysis) | About 47% of student conversations were direct requests for answers with little engagement. | Company report |
| Fang et al. (MIT and OpenAI) | 2025 | 981 people over 4 weeks | Loneliness and emotional dependence with daily chatbot use | The type of chatbot made no difference. The heaviest users reported more loneliness and dependence; the cause is unclear. | Preprint |
| Phang et al. (OpenAI and MIT) | 2025 | About 40 million conversations, 4,076 surveyed | Emotional use of ChatGPT | Emotional use is concentrated in a small group of heavy users. | Preprint |
| De Freitas et al., Journal of Consumer Research | 2025 | Several experiments | Loneliness after talking to an AI companion | AI companions reduced loneliness in the moment, over a day to a week. | Peer reviewed |
| Benitez et al., "Efficiency Without Cognitive Change" | 2025 to 2026 | 30 | Four subtests of an adult intelligence scale (WAIS-III) before and after 4 weeks of ChatGPT use | Faster task work and no change on the four subtests. Too small to prove anything either way. | Preprint |
| Stanford SCALE evidence review | November 2026 | 52 causal studies | What the causal evidence says about AI and learning | Effects depend on design. AI that explains tends to beat AI that does the work, but the gains are inconsistent. | Review |
Download studies.csvLast checked 13 November 2026. Free to reuse with a link back.
How to read headlines about this
- "MIT proved it." The MIT paper is a preprint with 54 people. It is interesting and early. It did not test IQ.
- "83% could not remember." That number is 15 of 18 people in one session who could not correctly quote a sentence from their own essay.
- "AI reduces critical thinking." The two best-known papers here are surveys (319 workers; 666 UK adults). They find that heavier AI reliance goes with lower critical thinking scores. A survey cannot show which causes which.
- "Developers are 19% slower with AI." In one 2025 trial, 16 experienced developers were 19% slower with early-2025 tools and felt faster. The same group's 2026 follow-up suggests a possible speedup and calls its own data unreliable.
What to do with this
You do not need to quit. Change the order.
Questions
Is ChatGPT bad for your brain?
No study shows brain damage or lasting change. A small MIT preprint found lower brain engagement while people wrote with ChatGPT, which is what you would expect when a tool does part of the work. The practical risk is learning less from tasks you hand over.
Does AI make us lazy?
Studies find people put in less mental effort when AI gives the answer, and that this goes with weaker reasoning and recall on that task. They also find the effect shrinks when people try first or get hints. Effort is a choice the tool can make easier or harder.
Does AI hurt students' learning?
It depends on the design. Answer-giving AI hurt exam scores in a large school trial; a hint-based tutor did not. A November 2026 Stanford review of 52 causal studies found AI that explains tends to beat AI that does the work, with inconsistent gains.
Where does this page get its numbers?
From the papers themselves, linked in the table. We note whether each is peer reviewed, a preprint or a working paper, and we re-check the list. If you spot an error, tell us and we will fix it.