OPINIÃO
MIT’s EEG study of 54 people writing with and without ChatGPT found weaker connectivity and poorer recall of the text. It doesn’t show brain damage.
The headlines said "AI is shutting down your brain." The study behind them is an MIT Media Lab preprint that measures something more specific: 54 people writing essays, with and without ChatGPT, wearing EEG headsets. What it found is interesting. "Brain damage" isn’t among the findings.
What the study did
The paper is titled "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task". Participants, recruited from a few academic institutions close to each other, were split into three groups of 18:
- LLM: wrote using ChatGPT.
- Search engine: used Google.
- Brain-only: no tools at all.
Each person wrote 20-minute essays across three sessions. A fourth session was completed by only 18 people, with the groups swapped: those who used the LLM wrote without tools, and those who wrote alone started using the LLM. The study ran about four months and used a 32-channel EEG headset.
What it found
- Brain connectivity. According to the authors, the brain-only group had the strongest and widest networks, the search group was in the middle, and the LLM group had the weakest coupling. This reading is contested (see the limits).
- Memory of their own text. In the first session, none of the 18 people in the LLM group (0 of 18) could correctly quote a line from the essay they had just written. In the search group 3 of 18 got it wrong; in brain-only, 2 of 18. The difference between search and brain-only wasn’t significant, so the contrast is the LLM group against both. This improved later, without disappearing: in session 2, 4 of 18 in the LLM group still got the quote wrong, and in session 3, 6 of 18, according to the paper’s discussion.
- Sense of authorship. 9 of 18 in the LLM group said they fully owned their essay, versus 16 of 18 in the brain-only group.
- Quality. The abstract says the LLM group did worse than brain-only on scoring too (teachers and an AI judge), and that essays within each group were homogeneous (named entities, n-grams, topics).
- The swap. Those who went from LLM to writing alone showed weaker connectivity. Those who went the other way (brain-only to LLM) showed, per the abstract, higher memory recall and re-engagement of occipito-parietal and prefrontal regions.
The authors call the overall pattern "cognitive debt", a hypothesis: leaning heavily on the tool may reduce mental effort and the practice of thinking for yourself.
The limit: 54 people, one task, one model
- Preprint. When this post was written, the paper hadn’t been peer-reviewed. As far as I could verify, it still is a preprint.
- Small sample. 54 people, and only 18 in the swap session. The authors themselves say larger samples are needed.
- Narrow scope. Only ChatGPT, only text, only essay writing in an educational setting. You can’t extend it to other tools or tasks.
- EEG measures connectivity, not damage. Fewer significant connections in one group, in the study’s analysis, means stronger connections in another group, not necessarily less overall brain activity.
A critical commentary, by researchers at the University of Vienna and TU Dresden, goes further. According to them:
- Statistical power was lacking. By their calculation, about 159 participants would be needed. Some per-topic analyses used only 2 to 4 essays per group.
- The EEG pattern is less clean. In some frequency bands, the LLM group had more connectivity than the search group, and in beta more than brain-only. The statistical analyses and corrections for multiple comparisons are thinly described.
- There are reporting inconsistencies. The paper says 55 people completed the study but analyzes 54. Session 4 interview results appear in opposite directions in two parts of the text.
- Quoting a line doesn’t prove learning. In their view, “correct quoting” measures little, and session 4 may reflect a practice effect. Their conclusion: no strong evidence that LLM use harms performance itself.
The commentary’s authors acknowledge the study’s value. I haven’t read any reply from the original authors.
What this has to do with synthetic data
Little, in terms of method, but a lot in terms of how to read. A study with 18 people per group becomes a headline. Results about synthetic data also become headlines, and the rule for reading them is the same: what was the sample, what was measured, what’s the scope. That’s what we try to do in our paper analyses.
There’s a practical parallel too. What stood out in the study wasn’t the tool, but what happens when people stop checking what they produce. In a simulation pipeline, generation and labeling are automatic, but the decisions (what to generate, what to vary, how to validate against real data) stay human.
What to take away
- Read the method and the critiques. “MIT study” says nothing about sample, task or review, and there’s a public commentary with objections.
- Use the tool as a draft. The clearest finding is how little people recalled of text they were handed ready-made. Reviewing with judgment is what keeps the effort.
- Automate generation, not checking. Any automatic step needs a human who checks the result.
Dealing with an automatic step nobody checks anymore? Tell us about it in a diagnostic call. If synthetic data isn’t the answer, we’ll say so.
Sources
- Kosmyna, N. et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv 2506.08872, 2025 (preprint).
- Stanković, M., Hirche, E., Kollatzsch, S., Doetsch, J. N. Commentary on Kosmyna et al. (2025). arXiv 2601.00856, submitted Dec 2025.
Leave a Reply