Does AI make us dumber? What the studies say

“AI makes you stupid”: the phrase comes back with every new study that does the rounds on social media. The problem is that these studies are not all worth the same, and their headlines rarely say so.

Our reading: the risk exists when you hand over to AI the task you are supposed to learn or master. It is far less established for supporting use, when you already know how to do the task and check what the tool produces. Here is how we reach this conclusion.

Four studies, four levels of strength

We present them from the strongest to the weakest: in this debate, the level of evidence matters more than the headline.

1. A controlled field experiment

In June 2025, PNAS published an experiment carried out with nearly a thousand high-school students in Turkey:

The tutor version largely limited this effect, without doing better than no AI at all.

It is the strongest of the four: an experiment with a comparison group, in real conditions. Its limit is its scope: maths, in a single high school in Turkey, over four sessions. We go into it in more detail in our article on AI in schools.

2. A self-reported survey

In April 2025, researchers from Microsoft Research and Carnegie Mellon University presented at the CHI conference a study on the impact of generative AI on critical thinking. They surveyed 319 knowledge workers about 936 examples of tasks. Result: the more people trust AI, the less they say they exercise critical thinking; the more they trust themselves, the more they exercise it.

Two limits: these are self-reports, not measurements, and it is a correlation, not a cause. Note also that Microsoft sells AI tools.

3. A correlational study, since corrected

A study published in January 2025 in the journal Societies, “AI Tools in Society”, surveyed 666 participants and found a negative correlation between frequent use of AI tools and critical thinking ability. It circulated widely. A correction was published in September 2025, pointing to an error in a table. We do not repeat any of its detailed figures.

Main limit: a correlation does not tell you the direction of the relationship. It is possible that people less inclined to critical thinking use AI more, rather than the other way round.

4. A preprint on a small sample

The study “Your Brain on ChatGPT”, circulated in June 2025, measured the brain activity of participants who wrote essays with or without an assistant:

The study concludes that the assistant’s users performed less well on several counts. As of October 2026, it is still a preprint, revised in December 2025 but not published in a peer-reviewed journal, on a very small sample. A critical comment, also a preprint, points in particular to the sample size, the analysis of the brain signals and inconsistencies in how the results are presented. It remains interesting as a line of research. It proves nothing on its own, and yet it is often the one cited first.

Position 1: by delegating thinking, we lose the habit of it

The first camp starts from an old idea: a skill you do not exercise weakens. If you hand writing, calculation, summarising or the search for arguments over to AI, you lose the habit of doing it yourself. The Turkish experiment supports this idea for learning, and the Microsoft and Carnegie Mellon survey suggests that greater trust in the tool goes hand in hand with less self-reported critical thinking.

A signal also comes from medicine. In four endoscopy centres in Poland, a study published in August 2025 compared colonoscopies carried out without AI, three months before and three months after the arrival of an AI polyp-detection tool:

The authors see this as a possible sign of a negative effect of the tool on doctors’ practice.

Position 2: the same was said of the calculator

The second camp points out that every cognitive tool has aroused this fear: writing, the calculator, the search engine, the GPS. We calculate less in our heads and remember fewer routes, but we free up attention for something else. AI would be one more tool, one that shifts the effort rather than removing it. And most of the alarming studies are too fragile to conclude anything.

What we think

The two camps describe different situations. Someone who already knows how to write and uses AI for a first draft that they rewrite is not in the same position as a beginner who hands it the whole job. The first keeps their judgement; the second never builds it.

It is also our conviction for automation in business: the most useful person is not the one who writes the best prompts, but the one who knows the job well enough to see when the machine is wrong. That skill is built without the tool, before it is exercised with it. We develop this argument in our analysis of human judgement in the face of AI.

What would change our mind: controlled experiments, on adults in work situations and over time, showing a measurable drop in ability among people who already mastered their task. As of October 2026, we know of none: the colonoscopy study points that way, but it is not controlled.

What this changes in practice for a company

• Train before deploying. The European AI Act requires, in its Article 4 as amended in July 2026, that companies using AI tools take “measures to support the development of AI literacy” among their staff, without requiring them to guarantee a specific level. It is an obligation, and it is also good protection against blind delegation. Our guide to integrating AI in business places this training among the rules to set before deploying.

• Protect juniors’ learning. Plan tasks that newcomers do without AI, then the same tasks with it, so that they can judge the result.

• Keep human validation on anything that commits the company: whatever goes to a customer, an employee or money. That is how our own expense-entry agent works: it extracts the data from a photo of a receipt, then a person validates before anything is written to the accounting system. We come back to these safeguards in our articles on business AI.

• Be wary of viral studies. Before changing an internal policy on the strength of a headline, check whether it is an experiment, a survey or a preprint.

• Choose what you hand over to AI, task by task. Some tasks can be delegated without risk; others are how people learn the job. That is the question settled by our digital and AI maturity audit, project by project, with a written GO or NO-GO for each.

This task-by-task sorting applies to the tool too: many tasks can be handled with a simple rule or a small model, without going through the most expensive one. That is the principle behind our AI at the right cost offer: each task goes to whatever is enough, with a budget and a ceiling for each use.

Frequently asked questions

Sources

1. Experiment carried out with nearly a thousand high-school students in Turkey (new tab)doi.org
2. The impact of generative AI on critical thinking (new tab)doi.org
3. “AI Tools in Society” (new tab)doi.org
4. Correction (new tab)doi.org
5. “Your Brain on ChatGPT” (new tab)arxiv.org
6. Critical comment (new tab)arxiv.org
7. A study published in August 2025 (new tab)doi.org
8. Article 4 as amended in July 2026 (new tab)ai-act-service-desk.ec.europa.eu