AI-first company

Study

Students writing with ChatGPT showed weaker brain connectivity and quoted their essays worst

The no-tool group showed the strongest brain connectivity and the ChatGPT group the weakest. After the first session, 15 of the 18 ChatGPT users could not quote from the essay they had just written, against 2 of 18 in each of the other groups. None of the 18 gave a correct quote.

By
Last reviewed
Type of study
Randomised experiment
Published
2025
Corrections
Report an error

At a glance

Source
Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes (MIT Media Lab) (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv 2506.08872, preprint (version 1, 10 June 2025; version 2, 31 December 2025). arxiv.org · Not peer-reviewed: other researchers have not yet checked it for a journal.
How it was done
Students were assigned by chance to write with ChatGPT, with a search engine or with no tool, in three sessions. EEG headsets recorded their brain activity while they wrote, and each student was interviewed after each session. In a fourth session, 18 of them swapped: the ChatGPT group wrote without tools and the no-tool group used ChatGPT. Randomised experiment. People were assigned by chance to work with or without the tool, and the two groups were compared.
Who took part
54 adults aged 18 to 39 from five universities in the Boston area, 18 in each group; 18 of the 54 came back for the fourth session.
Where
Timed essay writing (20 minutes per essay, on SAT essay prompts) in sessions run by MIT Media Lab researchers in the United States over four months, with OpenAI’s ChatGPT.

“Brain connectivity systematically scaled down with the amount of external support”

What it found

  • Brain connectivity. EEG headsets measured how strongly signals at one point on the scalp predicted signals at another, for every pair of the 32 electrodes. The no-tool group showed the strongest and widest links, the search-engine group came next, and the ChatGPT group showed the weakest.
  • The 79 and the 42. In one frequency band (alpha), over the first three sessions, the researchers counted 79 pairs of electrodes where the link was clearly stronger in the no-tool group, and 42 where it was clearly stronger in the ChatGPT group (appendix F of the paper). These are counts of pairs, not a share of the brain. The paper gives no percentage drop.
  • Quoting their own essay. After the first session, 15 of 18 people in the ChatGPT group (83.3%) could not give a quote from the essay they had written minutes earlier. In each of the other two groups, 2 of 18 (11.1%) could not. None of the 18 ChatGPT users gave a correct quote. In the search-engine group 3 of 18 did not, and in the no-tool group 2 of 18.
  • Later sessions. By the second and third sessions, most people in every group could quote from their essay. In the third session, 6 of the 18 ChatGPT users still gave no correct quote, against 2 of 18 in each other group.
  • Owning the essay. After the first session, half of the ChatGPT group (9 of 18) said the essay was fully theirs. In the no-tool group, 16 of 18 said so.
  • Switching in the fourth session. People who had used ChatGPT and then wrote without it showed stronger connectivity than the no-tool group in its first session, but weaker than in its second and third. People who had written without tools and then used ChatGPT showed stronger connectivity than the ChatGPT group had in its first three sessions.

Both versions of the paper on arXiv, from June and December 2025, give these figures.

When the result holds

The task was a short essay on a set topic, written under time pressure by university students and staff. The groups differed in how much of the thinking the tool did for them. The result says nothing direct about office work, where people use AI on tasks they already know well.

The quoting result matters most for managers. When a tool writes the text, the person may not remember what it says, even minutes later. That is a reason to check now and then what staff can still do without the tool.

Limits

  • Small groups. 18 people per group; 18 people in all in the fourth session, 9 in each switched group. The fourth session was optional, so the people who came back chose to.
  • One task, one tool, one place. Essay writing only, with ChatGPT only, by students and staff from universities in the Boston area. The authors write that their findings “may not generalize across tasks”.
  • The quoting test cannot separate thinking from authorship. People remember words they produced better than words they only read: the generation effect (Bertsch and others, 2007). The ChatGPT group wrote fewer of their essays’ words, so weaker quoting is what that effect predicts, with or without weaker thinking. The paper does not test the two explanations against each other.
  • EEG measures signals on the scalp. It shows how strongly brain regions worked together during the task. It does not show damage, and it does not measure intelligence. The authors ask reporters not to use words such as “harm” and “damage”.
  • Not peer-reviewed. Both versions are headed “Preprint, under review”. Other researchers have published a comment questioning the sample size, the EEG analysis and inconsistencies in how the results are reported (Stankovic and others, December 2025).
  • No funding. The authors say the study had no funding, so no vendor paid for it.

Practices that use this study

This study supports:

Claim checks that cite this study