This topic generates a lot of alarmist headlines and very little nuance. We're going to look at what the research actually says, including its methodological limitations, which are significant. The conclusion is neither that AI is dumbing us down, nor that all is well: it's more useful than that, because it points to a way of doing things.
The question keeps coming back, often framed with a hint of guilt: by asking a machine to do everything, are we losing something? It deserves better than a knee-jerk answer in either direction. Here is the state of knowledge.
The concept: cognitive offloading
The mechanism at the heart of the debate has a name in cognitive psychology: cognitive offloading. It refers to externalising a mental task to an external support, and it has been studied long before AI. Writing a shopping list, working out a sum on paper, using a calculator: these are all cognitive offloading.
The phenomenon is therefore nothing new, and it is not bad in itself. What changes with generative AI is a matter of degree. Search engines had already displaced our memory: a classic 2011 study showed that heavy use of online search altered independent memory, with people retaining more where to find information than the information itself. With language models, a further threshold is crossed: AI no longer just stores data, it directly produces reasoning, texts and analyses. We no longer delegate storage, we delegate the intellectual act itself.
What the studies find
Several pieces of research converge, and they need to be taken seriously.
A study by Michael Gerlich, conducted in 2025 on 666 participants, found a significant negative relationship between frequency of AI use and critical thinking abilities, with cognitive offloading as an intermediate factor. The 17–25 age group, the heaviest users, showed the most reduced abilities.
A study conducted by Microsoft with 319 knowledge workers found a substantial negative correlation, around -0.49, between frequency of AI tool use and critical thinking score. And it identified a very telling aggravating factor: offloading intensifies when trust in the model exceeds trust in one's own abilities.
A study from MIT went further by measuring the brain activity of participants writing texts with or without ChatGPT. It observed weaker brain connectivity among users of the tool, and poorer recall of their own output. The researchers coined a phrase that sums up the idea well: cognitive debt. You get an immediate result, but you accumulate a hidden deficit, a skill that never gets built.
Almost all of these studies are correlational, not causal. They observe that heavy AI users have poorer critical thinking, but they do not demonstrate that AI is the cause. The link could just as easily run the other way: people less inclined to deep analysis may be more likely to delegate. Or a third factor explains both. Researchers currently working on randomised controlled trials say so explicitly: existing evidence relies largely on questionnaires and interviews. Some results, including those from MIT, had not yet been fully peer-reviewed at the time of their heavy media coverage.
The most useful finding, and it's encouraging
Here is the discovery worth remembering, because it is actionable. In the MIT study, the most revealing phase is the one where the roles were reversed.
Regular ChatGPT users, suddenly deprived of the tool, showed weaker brain activity than those who had never had assistance. As if prolonged use had left a trace, a form of deconditioning. But, conversely, those who had first worked without the tool for several sessions before moving to AI maintained good connectivity.
In other words: order seems to matter more than usage. Learning first without assistance, then relying on AI, does not produce the same effect as delegating from the start. It is not the tool that would be the problem, but skipping the stage of building the skill.
This is exactly the reasoning we developed about learning to code: you cannot supervise what you do not understand. Someone who first learned to write retains the ability to judge a generated text. Someone who has never written can only accept it.
The other risks named by researchers
Beyond critical thinking, two effects recur in the literature and deserve mention.
The homogenisation of thought. If everyone uses the same tools in the same way, intellectual output risks converging, losing diversity and originality. A risk that echoes, for humans this time, the logic of model collapse we described for machines.
Overconfidence. Accepting an answer without cross-checking or questioning it is particularly risky with systems that err with confidence, as we explained in our article on hallucinations.
So, dumber?
The honest answer is: it depends on how you use it, and that formula is not a cop-out.
There is a real difference between asking an AI to produce something in your place, and asking it to help you produce it yourself. Between asking it for a finished text and asking it to critique your draft. Between asking it for the answer and asking it to explain the reasoning so you can verify it. The first usage offloads you, the second exercises you.
The factor identified by the Microsoft study is probably the best personal warning signal: the day your trust in the machine exceeds trust in your own judgement, offloading becomes dependence. It is not AI that decides that shift. It is you, with every request. And of everything we have read on the subject, that is probably the most useful piece of information to keep in mind.