Everyone's favourite AI study asked us not to say the thing everyone says
Somebody asked the MIT team behind Your Brain on ChatGPT whether their work shows that LLMs make people dumber. Their answer opens with an exclamation mark: "No! Please do not use the words like 'stupid', 'dumb', 'brain rot', 'harm', 'damage'".
That's the study everyone's citing to prove exactly that.
Research Brief, Edition 3. I teased this one last time and then spent longer on it than I meant to, which is what happens when the footnotes turn out to be better than the paper.
If you've sat in a curriculum meeting lately, somebody's waved this study at you. It's the MIT one with the EEG caps. I've seen it in more slide decks this year than I've seen my own street, always at the same moment, right when the room's ready to agree that AI is quietly hollowing out the students. Meanwhile there's a bigger, peer-reviewed, thousand-student study almost nobody's waving, where the harm turned out to be a setting somebody could change. Numbers on that further down, and they're the ones worth your Monday.
None of this is me defending AI, by the way. I've got no interest in that fight and the internet's already well staffed for it. I'd just like us to read a paper before we put it on a slide.
Key takeaways
- The MIT "Your Brain on ChatGPT" study is a small preprint: 54 people across three sessions, and only 18 in the fourth session its strongest claims lean on. It measured brain activity during writing, not what anyone remembered later.
- Its own authors asked people not to describe it with words like "stupid", "dumb" or "brain rot".
- The stronger evidence is Bastani's study of nearly a thousand students: unrestricted GPT-4 left them worse off once it was taken away, and a hints-only version largely removed that harm.
What the MIT study actually did
Straight from the preprint:
- 54 people across sessions one through three. Eighteen made it to session four, which is, awkwardly, the session the strongest claims lean on. Eighteen is a seminar. A small one.
- EEG measured brain connectivity while people wrote essays. The LLM group came out weakest, the search group in the middle, the unaided writers strongest.
- Two teachers read the LLM essays and called them "soulless," which is a fair review of most first drafts, honestly.
- It's a preprint. No peer review yet, the authors figure that'll take months, everyone was recruited from one geographic area, and the authors ask that the conclusions be read as preliminary.
Now, the measurement was brain activity during the writing. Not what anybody knew a week later. And I know that sounds like pedantry (it sounds like pedantry to me, and I'm the one typing it) right up until you notice it's the load-bearing wall. If you hand the essay to a model, your brain does less essay-writing while that's happening. That isn't a finding so much as a definition. Whether you retain less afterwards is a different study, and nobody's run that one at this scale.
Good work, careful authors, a much smaller claim than the one it's been conscripted into. Time's write-up shows the distance between what a paper says and what a headline does with it, and I don't blame the reporter one bit. Nobody's ever been promoted for the headline "Small Preliminary Study Finds Something Modest."
The finding that matters is older than ChatGPT
Six years before any of this, Deslauriers and colleagues ran the study that should've settled it and somehow didn't. Undergraduate physics, randomly assigned classrooms, identical content, identical handouts, same instructor, nobody selling either format.
The students in the active classrooms learned more. They also reported learning less.
The study everyone cites to prove the point contains a request from its own authors not to make it.
That gap is the whole story, sitting in PNAS since 2019 with random assignment behind it, waiting for somebody to notice. Effort feels lousy and works. Ease feels wonderful and doesn't. AI didn't invent that. It just made the easy option free, frictionless, and awake at two in the morning when you aren't.
Worth sitting with which way that points, too. The uncomfortable classroom was the good one. So if you want the MIT preprint as evidence that new technology rots the brain, you're also on the hook for the beautifully polished lecture everyone rated five stars and couldn't reconstruct by Thursday.
When somebody measured it across nearly a thousand students, the design was the problem
Bastani and colleagues took nearly a thousand Turkish high school students through real maths classes with the thing actually switched on.
- Unrestricted GPT-4 access: performance up 48%.
- A tutor version that gave hints instead of answers: up 127%.
- Take the tool away again and the unrestricted group scored 17% below students who'd never touched it.
- Guardrails largely removed that harm.
So it wasn't the model. It was the model handing over finished answers. Somebody chooses that setting (a vendor, a curriculum lead, whoever signs the purchase order at the end of a long Thursday) and somebody can unchoose it. That's why I rate this paper over the famous one. Wharton's plain-language version runs about four minutes, which makes it the link you forward to a dean.
| MIT preprint (2025) | Deslauriers et al. (2019) | Bastani et al. (2025) | |
|---|---|---|---|
| Sample | 54 people; 18 reached session four | Undergraduate physics classrooms, randomly assigned | Nearly 1,000 Turkish high school students |
| What was measured | Brain activity (EEG) while writing essays | Actual learning versus felt learning | Maths performance with the tool, and after it was removed |
| Peer review | Preprint, not peer reviewed | PNAS, peer reviewed | PNAS, peer reviewed |
| Result | LLM group showed weakest connectivity during writing; authors ask for a preliminary reading | Active classrooms learned more and reported learning less | Unrestricted access: +48% in practice, then 17% below controls. Hint version: +127%, harm largely removed |
Three papers came up while I was reading that I haven't cited here.
The Microsoft Research and Carnegie Mellon work on knowledge workers and cognitive offloading is the obvious omission, and it's almost certainly relevant. A 2025 paper by Gerlich in Societies is the second, widely referenced. A third preprint on productive struggle and the illusion of learning is directly on subject and turned up in search.
I couldn't open a canonical URL for any of the three, so they're out. Not because they're wrong, but because I don't cite things I haven't read at the source, and neither should the people quoting me. If you have verified links to any of them, send them and I'll read them properly for a future edition.
Three checked peer-reviewed sources beat six half-checked ones, particularly on a subject where half-checked citation is the entire problem.
Here's the question I'd put in place of "does AI make us dumber," which isn't answerable by anybody right now, very much including the study being used to answer it.
Is this tool doing work I needed to do myself?
That's the substitution test, and it's answerable about a specific tool, in a specific course, on a specific Tuesday, by the person standing in the room. No EEG cap required, which I'm told is a relief. They aren't comfortable.
Five things worth doing Monday
- If you've been citing the MIT preprint as proof of harm, cite it for what it is: a small unreviewed study of brain activity during writing, whose authors asked us not to read it that way.
- Ask any vendor whether the tool answers or prompts, then ask to see the setting that controls it. If there's no setting, that's the answer. Paste this straight into your next RFP if it saves you a draft:
The vendor shall state whether the system returns direct answers or scaffolded hints, confirm whether that behaviour is configurable, identify the default mode for student accounts, and describe how an instructor changes it.
- Keep "learners liked it" and "learners learned" in separate columns. In 2019 they pointed opposite ways and nothing since has swung them back.
- Test retention afterwards instead of watching process during.
- In clinical teaching, the struggle is the curriculum. Working through a differential, or a first pass at a discharge summary, is the training. A tool that tidies that up early has quietly removed the part you were paying for.
That's the edition. If there's a study you keep seeing quoted and you'd like a second pair of eyes on it, send it my way. I read this stuff anyway and it'd be rude not to pass along what I find.
Standing policy, as always: I disclose every use of AI in my own work.
How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.
These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.
Common questions
What did the MIT "Your Brain on ChatGPT" study find? People who wrote essays with ChatGPT showed the weakest brain connectivity on EEG, search-engine users came in the middle, and unaided writers were strongest. It is a small preprint, and it measured activity during the writing, not how much anyone retained afterwards.
Does ChatGPT make you stupid? The study most often cited to say so doesn't show that, and its authors asked people not to use words like "stupid" or "dumb". The better question is whether a tool is doing work you needed to do yourself, which you can answer for a specific tool in a specific course. The older evidence points the same way: in a 2019 physics study, students in the more demanding classes learned more and felt they had learned less.
How many people were in the MIT ChatGPT brain study? 54 people took part in sessions one to three. Only 18 completed session four, the session the strongest claims rely on. Eighteen is a seminar, which is part of why the authors ask that their conclusions be read as preliminary, and why the paper has not yet been through peer review.
References
Peer-review status noted, because it changes how hard a claim can lean.
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv:2506.08872. Preprint, not peer-reviewed. https://arxiv.org/abs/2506.08872
- MIT Media Lab. Your Brain on ChatGPT, project page and FAQ. Source of the authors' caveats and the quoted line. https://www.media.mit.edu/projects/your-brain-on-chatgpt/
- Time (2025). Coverage of the MIT study. https://time.com/7295195/ai-chatgpt-google-learning-school/
- Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom. PNAS, 116(39), 19251–19257. Peer-reviewed. https://www.pnas.org/doi/10.1073/pnas.1821936116
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS. Peer-reviewed. https://www.pnas.org/doi/10.1073/pnas.2422633122
- Knowledge at Wharton. Without guardrails, generative AI can harm education. https://knowledge.wharton.upenn.edu/article/without-guardrails-generative-ai-can-harm-education/