Short answer: no, and also the question is wrong.
No study has shown that AI use reduces measured intelligence. What the research does show is narrower and more interesting — that people engage less cognitively while working with an AI assistant, and remember less of what they produced. That is a collapse of practice, not of capacity. Whether practice lost for long enough becomes capacity lost is a separate question, and the honest answer is that nobody has measured it yet.
Here is what the evidence actually says.
What the MIT study found
The study everyone cites is Your Brain on ChatGPT, from MIT Media Lab in 2025. Fifty-four participants wrote essays under three conditions — with an LLM, with a search engine, or with nothing — while researchers recorded EEG.
Three results came out of it:
- The LLM group showed the weakest and least distributed neural connectivity of the three groups. The unaided group showed the strongest.
- Across four sessions, the LLM group performed worst at the neural, linguistic and scoring levels.
- Asked to quote from essays they had submitted minutes earlier, LLM-group participants frequently could not.
That last finding is the one worth sitting with. It is not about intelligence. It is about encoding. The essays existed; the learning did not. The participants had produced text without the process that normally makes text yours.
The researchers called the accumulated cost of this cognitive debt — the idea being that each delegation saves effort now and leaves a deficit that compounds over subsequent tasks.
What the MIT study does not establish
The paper is frequently reported as proof that AI damages your brain. It is not, and the authors do not claim it.
- Fifty-four participants, with only eighteen completing the fourth crossover session. That is a small study.
- One task type, one model. Essay writing with ChatGPT. Not coding, not analysis, not decision-making.
- It measures engagement during the task, not capability afterwards. Reduced neural connectivity while someone else does the work is close to tautological — it is what delegation looks like from the inside of a skull.
- It is a preprint, not a peer-reviewed publication.
What it demonstrates is that assisted writing is a cognitively lighter activity than unassisted writing, and that lighter activity produces weaker memory of the result. That is a real and useful finding. It is not the same as damage.
What the larger studies found
The survey work reaches further and grips less firmly.
Gerlich's 2025 study of 666 participants found that frequent AI tool use correlated negatively with critical thinking scores, with cognitive offloading mediating the relationship. Younger participants showed higher AI dependence and lower critical thinking scores. A comparative study of AI-assisted against manual task performance in young adults found a consistent pattern.
The problem is direction. These are correlational designs, and there is an obvious alternative reading: people who are already inclined to avoid effortful thinking reach for AI more often. The tool would then be a marker of the disposition rather than a cause of it. No survey design can separate those, and the honest studies say so.
The question that is actually answerable
"Is AI making us dumber" is unanswerable as posed, because it asks about a population, a decade, and a construct nobody agrees how to measure.
Here is a version you can answer: has your own capability changed since you started delegating?
That question is tractable because it has a fixed comparison point — you, at a known earlier time — and it does not require anyone to define intelligence. But it needs the comparison point to exist. Almost nobody has one, which is exactly why the change is invisible.
Try this now, before reading further. Open a blank document and write 300 words on something you know well. No AI, no notes, no search. Then notice how it felt.
Most people find that reaching for the assistant is now the reflex, and that sitting with the blank page is harder than it used to be. That feeling is not proof of anything. It is data you did not have five minutes ago.
What I think is actually happening
The strongest available frame is not intelligence at all. It is desirable difficulty — the learning-science finding that conditions making a task harder in the moment produce more durable learning than conditions making it easy.
Composing an argument from a blank page is difficult. That difficulty is not an inefficiency in the writing process; it is the mechanism by which the argument becomes yours. Remove it and you get the text without the encoding, which is precisely what the MIT participants demonstrated when they could not quote themselves.
This reframes the whole debate. The question is not how much AI you use. It is which difficulty you are removing. Offloading grammar checking so you can concentrate on the argument keeps the difficulty that develops you. Offloading the argument does not.
That distinction — between beneficial and detrimental offloading — is where the useful conversation is, and it is almost entirely absent from the "is AI making us dumber" coverage.
Where this leaves you
The evidence supports three claims and not a fourth:
- AI-assisted work involves less cognitive engagement than unassisted work. Well supported.
- People remember AI-assisted work worse. Well supported, small samples.
- Frequent AI use correlates with lower critical thinking scores. Supported, direction unknown.
- AI use causes lasting capability loss. Not demonstrated. Inferred from skill-decay research that predates AI entirely.
Anyone telling you the science is settled — in either direction — has not read it. The full set of studies, with what each one does and does not show, is indexed on the research page.
What you can do about it individually is smaller and more boring than the headlines: notice which difficulties you are removing, keep practising the ones that matter to you, and get a fixed measurement of your own work while you still recognise it. That last part is what IMPRINT is for, and how it measures is published in full, including where it falls short.