When headlines declared that an MIT study had proven ChatGPT “makes you stupid,” the researchers behind it published an FAQ asking journalists to stop using words like “dumb” and “brain rot,” because their paper, based on 54 students writing essays, never said anything of the kind.
The study came out of the MIT Media Lab in June 2025, and it examined what happens in the brain when people write essays with the help of ChatGPT. Within days it had been compressed, across headlines and social media, into that single claim, and the study became shorthand for a fear a lot of people were already carrying.
The gap between the study and its reception is the more interesting subject, so it is worth starting there.
This is one preprint, not a settled body of evidence. The paper, led by Nataliya Kosmyna and titled Your Brain on ChatGPT, was posted without peer review, and the authors said plainly that its conclusions should be treated as preliminary. Kosmyna has said she released it early because she was worried about schools rushing to put the technology in front of young children before anyone understood the effects, not because she believed the case was closed.
What the study actually did
Fifty-four people, aged eighteen to thirty-nine and drawn mostly from the Boston area, were split into three groups and asked to write short essays of the kind used in standardised tests. One group used ChatGPT. One used a conventional search engine. The third, called brain-only, worked with no tools at all. Each did this across three sessions. A smaller group returned for a fourth session in which the conditions were swapped.
Throughout, the researchers recorded the writers’ brain activity using electroencephalography, or EEG, which measures electrical signals at the scalp. The measure they focused on was connectivity, a rough indication of how much different brain regions were communicating during the task.
Two findings drove the coverage. The first was that connectivity scaled with how much help the writers had. The brain-only group showed the most widely distributed networks, the search engine group less, and the ChatGPT group least, with the paper reporting up to 55 per cent lower connectivity in the AI group compared with the brain-only writers. The second was about memory. When participants were asked, straight after writing, to quote a single sentence from the essay they had just produced, 83 per cent of the ChatGPT group could not, against around 11 per cent in each of the other two groups.
Why “thought the least” is the wrong gloss
The phrase that attached itself to the connectivity result was that the AI group “thought the least.” That is a loose translation of what EEG connectivity shows.
Lower connectivity during a task you have largely handed to a machine is close to what you would expect, and it does not measure intelligence. Kosmyna has been direct about this, pointing out that the study did not measure IQ and did not find anything she would describe as brain rot. EEG records the level of engagement during a particular activity. A person who asks a tool to draft a paragraph is, unsurprisingly, engaging less in the drafting of that paragraph. That is not the same as a person whose capacity to think has been diminished, and the study does not contain the evidence that would be needed to make the second, much larger claim.
The distinction matters because the two readings point to different worries. One is about a habit. The other is about damage. The data here support a statement about the habit.
The memory finding, read carefully
The quoting result is the most striking number in the paper, and also the one most easily overstated. The seed of truth is real: a large majority of the ChatGPT group, shortly after finishing, could not accurately reproduce a line of their own essay.
But the word “their own” is doing unstable work. Some of those essays had been generated by the model and, in the later sessions, increasingly pasted in with little change. You do not remember a sentence you never composed, in the way you do not remember a phone number you never dialled. The finding is better understood as a measure of how shallowly the material was encoded when the writer was not the one assembling it. The authors themselves read the poor recall as a sign that the essays were never integrated into the writer’s own memory in the first place, because the cognitive work had been handed off.
That is a real effect, and a sensible one to be curious about. It is not the same as the claim that using AI erases your memory.
An old pattern wearing new clothes
What the study describes is, in its general shape, familiar. Psychologists have long studied cognitive offloading, the way people stop retaining information they know a tool will hold for them. A frequently cited 2011 paper in the journal Science by Betsy Sparrow and colleagues found that people remember less of the information itself when they expect to be able to look it up later, while remembering where to find it. Drivers who follow satellite navigation form weaker mental maps of the routes they drive. The calculator did something similar to mental arithmetic, generations ago.
Offloading a task means you hold less of it in your head afterwards. This is not new, and it is not, in itself, a catastrophe. We have always traded effort for convenience and accepted the loss of the skill we stopped practising.
What gives the essay case a different weight is that the writing was the point. When the task you hand off is arithmetic, the answer was the goal and the tool delivers it. When the task you hand off is composing an argument in your own words, the effort was the goal, because that effort is where the thinking and the learning were supposed to happen. Skip it and you have the document but not the understanding it was meant to produce. That is the genuine question the study raises, and it is narrower and more useful than “AI makes you stupid.”
What the study does not show
A few limits are worth stating before anyone builds a policy on this.
The sample was small and narrow: fifty-four people, a single age band, one city. The task was one specific kind of timed essay. Four sessions across a few months is not a measure of long-term change, whatever the word “accumulation” in the title suggests. And because it is a preprint, it has not yet been through the scrutiny that would test whether the effects hold.
There is also a result in the paper that cuts against the simple reading and received far less attention. When people who had started in the brain-only group later used ChatGPT, their brain activity did not collapse. It rose, and they used more considered prompting. The order mattered. Coming to the tool with your own thinking already underway produced a different pattern than leaning on it from the start. AI use, on this evidence, is not uniformly an off-switch for engagement. What you bring to it shapes what it does to you.
The honest takeaway is smaller than the headline and harder to act on. A tool that does a task for you tends to leave you with less of that task in your head, and when the task was the learning, that trade has a cost worth noticing. The MIT group has asked for the obvious next step, which is larger and longer studies, ideally peer-reviewed, before anyone decides what this means for a classroom. That request has been much quieter than the headline it was buried under.
