Is AI making us stupid?

Nobody has measured it. There is no study anywhere that takes a person's general intelligence, gives them an AI assistant for a year, and measures it again. What has been measured, repeatedly and by several methods, is something narrower and more useful: how well people do a task once the assistance is taken away. That answer is consistent, uncomfortable, and much more specific than the question as people ask it.

Three floors, and only two have anything under them#

The first floor is the oldest and the least contested. In 2011 Sparrow, Liu and Wegner ran four laboratory experiments on what people encode when they expect information to stay available. People who believed a fact would remain on the computer remembered where to find it rather than the fact. That is a trade in what gets stored, and the authors were careful to say it does not show total memory capability falling. Search engines did this before chat assistants existed.

The second floor is the one the last two years have filled in. Take the tool away and test the person, and a penalty appears. That describes a practice rather than a person, and it carries most of the evidence behind the popular question.

The third floor is empty. General cognitive ability, the thing an IQ test tries to estimate, has not been measured as a function of AI exposure by anybody. No cohort, no trial, no panel. A page claiming otherwise is reading one of the first two floors and labelling it as the third.

The population score everyone reaches for started falling in 1975#

Measured cognitive scores have fallen across a whole population in one well-documented case, and the date the fall began settles a lot. Bernt Bratsberg and Ole Rogeberg, working from Norwegian military conscription records, analysed birth cohorts from 1962 to 1991. Their sample was 817,611 men present in Norway on their eighteenth birthday, of whom 736,808 had a valid ability score. Mean IQ rose from 99.5 for the 1962 cohort to 102.3 for the 1975 cohort, then fell back to 99.4 by 1989.

What makes the paper matter is the design. They recovered the rise, the turning point and the fall from variation within families, comparing brothers. That rules out explanations depending on who was having children, which is where the debate had been stuck. Their selection-corrected estimate for the decline period is 0.33 IQ points per year within families, against 0.34 across families. The Flynn effect and its reversal are both environmental, and the environment doing it is shared by siblings.

The 1991 cohort sat that test around 2009. Every person in the falling half of that curve was assessed before generative AI existed. Whatever has been moving those scores for fifty years, it is not ChatGPT. The authors are equally clear about the limit of their own result: they cannot identify which environmental factor is responsible, and they list changing media exposure among the hypotheses their design leaves standing. So the paper does not exonerate screens. It dates the clock.

The experiments that withdraw the tool and then test you#

Four results, four populations, one shape.

Bastani and colleagues ran a field experiment with nearly 1,000 high-school mathematics students in three arms: unrestricted GPT-4, a hints-only tutor, and a control. With the tool present, grades rose 48 per cent under unrestricted access and 127 per cent with the tutor. With access removed, the unrestricted group scored 17 per cent lower than students who had never had it. The guardrailed tutor largely removed the harm. The interface decided the outcome; the presence of AI did not.

Liu and colleagues produced the effect causally and fast. Across randomised trials with 1,222 participants on mathematical reasoning and reading comprehension, assistance was available during practice and then withdrawn. People performed worse unassisted and were more likely to give up, after roughly ten minutes of exposure. Their reading moves the mechanism from knowledge to persistence, which travels further than a subject-specific result.

Shen and Tamkin tested 52 experienced Python developers learning an unfamiliar library, half with an AI sidebar, then quizzed them without it. The assisted arm averaged 50 per cent against 67 per cent, a pre-registered result at p = 0.010. The control arm hit more errors, with a median of 3.0 against 1.0, and got better at resolving them. Most of the gap sat in the debugging questions.

And outside a laboratory entirely, Budzyn and colleagues looked at 1,443 colonoscopies performed without AI assistance by 19 Polish endoscopists averaging 27.6 years of experience, before and after AI was introduced at their centres. Unassisted adenoma detection fell from 28.4 per cent to 22.4 per cent, a drop of 6.0 percentage points. It is observational rather than randomised, and the authors say so. It is also the only measurement here on people with decades of expertise doing paid clinical work.

None of these measures intelligence. All of them measure what a person can do alone after a period of not doing it alone.

The three studies that get quoted are the three weakest ones here#

Michael Gerlich surveyed and interviewed 666 people and found a negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading and strongest among the youngest. It is a correlation with a plausible mechanism and no causal claim, and it carries a published correction issued on 10 September 2025 which anybody citing it should read alongside.

Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of AI at work. Higher confidence in the tool went with less critical thinking, and the thinking that remained shifted from producing to verifying. It is self-report, and people who think differently may use AI differently.

Kosmyna and colleagues at the MIT Media Lab ran the EEG study that launched the phrase cognitive debt: 54 participants writing essays with an LLM, a search engine, or unaided, with the LLM group showing the weakest brain connectivity. It is a preprint. Stankovic and three colleagues published a methodological critique arguing it is underpowered, that a repeated-measures design at those parameters would need around 159 participants, and that the search-engine group relied on an external tool and showed no impairment, with that comparison returning p = 1. The critique is itself unreviewed and presents no competing data. What it establishes is that the most-quoted study in this area rests on a contested pilot.

There is a pattern in that. The weaker a study's design, the closer its headline sits to the question people want answered, and the further it travels.

What the word smuggles in#

Stupid is a property of a person. Everything above is a property of a practice. That distinction changes who is responsible and what can be done.

Read as a claim about people, the evidence licenses nothing: there is no cohort getting dimmer. Read as a claim about practice, it licenses something quite sharp. A capability you stop exercising becomes a capability you cannot rely on, and the interval is shorter than anyone assumed. Liu's ten minutes is not skill decay, it is a within-session carry-over, and it still went in the direction nobody wanted. Budzyn's endoscopists had 27.6 years behind them. Experience did not insulate them.

This estate calls the accumulated version capability debt: the gap between what you can produce and what you could still do if the machine were switched off. It is invisible while the machine is on. The measurements above are all taken by switching it off, which is the only way anybody has found to see it.

The other thing the word hides is that the interface is doing most of the work. Bastani's three arms are the cleanest demonstration on record: the same technology, the same students, the same period, and the difference between a 17 per cent deficit and almost none was how the tool was built. That is a design question rather than a fact about human beings, which puts it in the hands of whoever chooses the tools.

Four things that follow for somebody using it daily#

• Test yourself with it switched off, on purpose. Every measurement above was produced this way, and there is no other way to see the gap. Pick something you used to do unaided and do it unaided, at intervals, and notice what has got harder.
• Prefer a tool that makes you produce the answer. The hints-only tutor beat unrestricted access on the withdrawal test and beat it on the assisted test too. Where a tool offers a mode that explains rather than completes, that mode is not the slower option.
• Watch for giving up rather than for forgetting. Liu's finding is that the first thing to go is persistence, not knowledge. The tell is reaching for the tool earlier in a problem than you used to, rather than failing to remember something.
• Keep the reps you would notice losing. The endoscopists did not stop being doctors. They stopped doing the looking. Whatever the equivalent looking is in your work, protect that. The missed reps covers how to tell which repetitions those are.

Where this page is guessing#

The link between the withdrawal experiments and anything cumulative is unestablished. Every result above measures a person shortly after a period of assistance, over minutes, hours or a few months. Nothing follows the same people for years, so whether the effect compounds, plateaus or reverses is unknown, and Shen and Tamkin say in their own paper that whether an immediate quiz score predicts longer-term skill is a question they do not resolve. Budzyn is the longest-running observation here and it is observational. That paper also carries a correction, at Lancet Gastroenterol Hepatol 2025; 10: e12, whose content could not be read from here, so the 6.0 percentage points above is the figure as published and not a figure anybody here has checked against the amended paper.

Nor does anything measure the other side of the ledger. If attention freed from a task is spent on a harder one, the trade could be positive and no study here would detect it. The estate's position is that the withdrawal penalty is real, specific and the only thing anyone has measured, and that calling it stupidity overstates it in one direction while ignoring it understates it in the other.

Key sources

• Bratsberg, B. and Rogeberg, O. (2018). Flynn effect and its reversal are both environmentally caused. Proceedings of the National Academy of Sciences, 115(26), 6674-6678. Published online 11 June 2018, in issue 26 June 2018. PMID 29891660. Full text read at source 23 September 2026. Graded entry.
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O. and Mariman, R. (2025). Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. Graded entry.
• Liu, G., Christian, B., Dumbalska, T., Bakker, M. A. and Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv 2604.04721. Preprint. Graded entry.
• Shen, J. H. and Tamkin, A. (2026). How AI Impacts Skill Formation. arXiv 2601.20245. Pre-registered, preprint. Graded entry.
• Budzyn, K., Romanczyk, M., Kitala, D. and 18 others (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology and Hepatology, 10(10), 896-903. A correction exists at 10(11), e12, DOI 10.1016/s2468-1253(25)00294-8, confirmed in the Crossref record on 23 September 2026 and not readable from here. Graded entry.
• Sparrow, B., Liu, J. and Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333(6043). Graded entry.
• Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. Correction at Societies 2025, 15(9), 252, issued 10 September 2025. Graded entry.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025, Microsoft Research and Carnegie Mellon. Graded entry.
• Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv 2506.08872. Preprint. Graded entry.
• Stankovic, M., Hirche, E., Kollatzsch, S. and Doetsch, J. N. (2025). Comment on: Your Brain on ChatGPT. arXiv 2601.00856. Preprint critique. Graded entry.

Related SuperSkills research#

The long version of the second floor is at does using AI stop you learning and does AI weaken critical thinking. On what the withdrawal penalty accumulates into, cognitive debt and capability debt and deskilling. On the mechanism, cognitive offloading, the Google effect and productive struggle. On whether it comes back, can you regain a skill you have lost and how fast skills decay. For parents, should children use AI.