Is AI making us stupid?

On 10 June 2025, 8 researchers from MIT and 2 neighbouring institutions posted online a study on writing essays with ChatGPT. 2 weeks later, a Belgian daily ran the headline "ChatGPT makes you less intelligent". Yet on the study's website, the authors answer "No!" to the question "can we say that language models make us dumber?", and ask that the words "stupid", "dumb" and "damage" not be used. The study covers 54 people and had not been published in a journal as of 13 November 2026, 17 months after it was posted.

The fear is widespread. In January 2025, 44% of French people aged 18 to 75 cited a decline in the ability to think among the main risks of generative AI.

We identified the studies that measure the abilities or effort of people using generative AI, excluding literature reviews. Among the most cited, 5 exceed 950 citations in OpenAlex, a public database of scientific publications. None measures intelligence.

2 of these studies measure critical thinking at a single point in time, through scores or self-reports. The other 3 randomly assign participants to groups and measure an effect on the same day. We set aside another much-cited study, on stories written with the help of AI, because it grades the texts without measuring the abilities of their authors.

The Haute Autorité de santé (France's National Authority for Health) classifies studies into 4 levels of evidence. Level 1 covers large randomised trials and meta-analyses of those trials; level 4, retrospective studies and surveys that describe a population at a given point in time. Our scale follows the same order, with the opinion poll, which measures a fear, at the bottom, and meta-analyses at the top, their level depending on the studies they pool.

The Swiss survey, the most cited with 2,120 citations, and the Microsoft survey are at level 4. The MIT study and the writing experiment randomly assign small numbers of participants, at level 2. The trial conducted in Turkey, in a single high school, is the only one of the 5 at level 1, and the least cited.

Meta-analyses do not yet settle the matter. In March 2026, one of them, covering 35 studies and 4,193 participants, found a "moderate" positive effect of ChatGPT on the results of pupils and students, while including non-randomised experiments with widely scattered results. Another, which concluded in May 2025 that the effect was "large", was retracted on 22 April 2026 because of inconsistencies in the analysis. 4 researchers audited some of the experiments pooled by a third: in their view, the gains observed cannot yet be attributed to ChatGPT with confidence.

In autumn 2023, about fifty classes in a Turkish high school revised mathematics in 4 sessions of 90 minutes, each ending with an exam taken without help. Each class was assigned to one of 3 groups: no AI, a ChatGPT-like AI, or an AI set up with the teachers to give hints rather than answers.

The authors read the chat logs. With the AI without safeguards, pupils often asked for the solution and copied it. They did not think they had done worse in the exam.

A larger trial, posted online in September 2026 without peer review, points in the same direction. At an American public university, randomly selected courses were given an AI tutor set by default to give the answer directly. The trial covers 2,379 students; between sections of the same course, the final grade fell by 0.37 standard deviations, a measure of how spread out grades are. The authors are cautious about inferring an effect on learning, since the grade is set by the instructor.

Smaller studies, on other tools and without random assignment, point the same way. At McGill University in Montreal, 50 drivers had poorer spatial memory the more they had driven with GPS; 13 of them, reassessed 3 years later, had declined more the more they had used it. In a Boeing 747 simulator, 16 airline pilots retained their manual flying skills but struggled to keep track of the aircraft's position and to decide on the next step, all the more so when their minds had wandered while the automation was flying. In 4 endoscopy centres in Poland, the detection rate of adenomas, polyps some of which become cancerous, fell from 28.4% to 22.4% in colonoscopies performed without AI in the 3 months after a detection AI was introduced; the authors write that AI "might" be the cause.

On the Haute Autorité de santé scale, these studies provide a scientific presumption or a low level of evidence. The Microsoft researchers point to the supposed mechanism, the "ironies of automation" described by Lisanne Bainbridge in 1983: handing routine work to the machine deprives the user of opportunities to exercise judgement.

The studies that come closest, the Swiss survey and the Microsoft survey, link AI use to critical thinking scores or to self-reports. A correlation of this kind can be read in 2 directions: AI may weaken critical thinking, and people who exercise it less may also use AI more.

The MIT study measures brain signals that are less connected to one another during writing with ChatGPT, and does not say what remains of this months later. 4 other researchers call for a more cautious reading of its results, pointing to the small sample and to results reported inconsistently.

2 older studies serve as precedents. In 2011, a study published in Science concluded that people remember information less well when they expect to be able to find it again on a computer; only one of its 4 experiments, which did not concern memory, has failed to replicate. In 1986, a meta-analysis of 79 studies concluded that calculators, used alongside conventional teaching, improved the pencil-and-paper arithmetic of average-ability pupils, except in the fourth year of primary school (CM1). The blog article on what the studies measure in the brain↗ finds that none of the studies it cites measures a lasting effect on the brain.

In the Turkish high school, the AI set up to give hints eliminated almost all of the exam loss, without producing a gain. In 18 Tennessee middle schools, a 2-year trial, posted online in August 2026 without peer review, gave randomly selected pupils a tutor set up to guide rather than answer: the result was a small gain in mathematics, close to that of the same platform's exercises without AI, and pupils rarely used the tutor. At Harvard, 194 physics students gained more than 2 times as much with a tutor that guided them step by step, following the teachers' instructions, as in an active-learning class, on a test taken immediately afterwards.

Among the studies we identified, none of those published in a journal follows the same people using generative AI for more than a few months.