TrueSeeker AI · Verified claim report Case 3f14b5622b · 2026-08-25

§ Claim under review · Research

"MIT Media Lab's study 'Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task' (led by Nataliya Kosmyna) found that students who used ChatGPT to write essays showed weaker EEG-measured brain connectivity and struggled to accurately recall/quote what they had written, compared to those who used only their own brains or Google Search; researchers termed this potential long-term effect 'cognitive debt.'"

Circulating claim, as submitted.

Verdict

Mostly accurate

Confidence

High
§

Summary

This one largely checks out. The MIT Media Lab study exists with exactly that title, led by Nataliya Kosmyna, and its own abstract says the ChatGPT group showed the weakest EEG brain connectivity, the search engine group was in the middle, and the no-tools group was strongest, and that ChatGPT users struggled to quote accurately from essays they had just written. "Cognitive debt" is the authors' own phrase. The post is also unusually careful for viral coverage of this study, because it says outright that the researchers do not claim permanent brain damage, which matches a public FAQ where the authors ask people to stop using words like "damage" and "brain rot" about their work. What the post leaves out is that the paper is a preprint that has not been peer reviewed, that it involved 54 people from five Boston-area universities with only 18 completing the crossover session the post highlights, and that other researchers have published a formal critique raising concerns about sample size, EEG methods and reproducibility. The EEG measure captures how engaged the brain was during a short writing task, not learning or lasting change. One detail in the post, that the switchers were better at quoting their own essays, I could not confirm in the paper itself.

§

The readings

key figures from the evidence
18 participants

completed Session 4 crossover subgroups

§

Why this verdict

Every substantive element of the claim maps onto sentences in the paper's own abstract, retrieved from arXiv and mirrored on the MIT Media Lab publication page: the three conditions, the connectivity ordering, the quoting difficulty, and the authors' own coinage of "cognitive debt". The claim also preserves the paper's hedge by saying "potential", and the accompanying caption explicitly disclaims the brain-damage reading that the authors ask people to avoid. What is missing is context rather than accuracy: the preprint status, the small sample, and the fact that a published methodological critique exists. I considered "Accurate" and rejected it because omitting unrefereed status and n=54 (n=18 for the crossover) is a real gap for a claim framed as "MIT found". I considered "Source exists but framing is misleading" and rejected it because the framing here tracks the source closely and the post actively adds hedges rather than stripping them, which is unusual for this study's coverage. Confidence is High because the primary artifact was retrieved and the comparison is direct; the residual uncertainty concerns one secondary carousel detail, not the central claim. As of 2026-08.
§

Evidence

The paper exists, with the exact title quoted, by Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes of the MIT Media Lab. The abstract states that participants were divided into LLM, Search Engine and Brain-only groups, that 54 participants took part in Sessions 1 to 3 with 18 completing a fourth session, and that EEG was used to assess cognitive load. On connectivity, the abstract states that Brain-only participants showed the strongest and most distributed networks, Search Engine users moderate engagement, and LLM users the weakest connectivity, with cognitive activity scaling down in relation to external tool use. The abstract also states that self-reported ownership of essays was lowest in the LLM group and highest in the Brain-only group, and that LLM users "also struggled to accurately quote their own work." On Session 4, the abstract states that LLM-to-Brain participants showed reduced alpha and beta connectivity indicating under-engagement, while Brain-to-LLM users showed higher memory recall and activation of occipito-parietal and prefrontal areas similar to Search Engine users. The paper closes on the framing that while LLMs offer immediate convenience, the findings highlight potential cognitive costs.

The paper is a preprint. The authors' own project page states plainly that it had not been peer reviewed and that conclusions should be treated as preliminary, and that they released it early rather than wait months for review. As of this investigation I found no peer-reviewed journal version; the MIT Media Lab publications page still cites it as an arXiv preprint, and the latest arXiv version is dated 31 December 2025.

The authors have published a public FAQ explicitly asking readers and journalists not to use words such as "stupid", "dumb", "brain rot", "harm", "damage", "brain damage", "passivity" and "collapse", and separately to avoid "brain scans", "LLMs make you stop thinking", "impact negatively" and "terrifying findings". They also state the study had no funding.

A four-author methodological comment was posted to arXiv (2601.00856) raising concerns about limited sample size, reproducibility of the analyses, EEG methodology, inconsistencies in reporting, and transparency, while stating that some results could be interpreted more conservatively.

§

Findings

✓ What's accurate 7

  • The study exists under exactly the title quoted, with Nataliya Kosmyna as first author and the MIT Media Lab as the institution.
  • Participants were assigned to ChatGPT, search engine, and no-tool ("Brain-only") conditions, exactly as described.
  • EEG was used, and the paper reports that connectivity was weakest in the LLM group, intermediate in the Search Engine group, and strongest in the Brain-only group.
  • The paper states that LLM users struggled to accurately quote their own work.
  • "Cognitive debt" is the authors' own term, in the paper's title, describing a hypothesized accumulating cost rather than a measured long-term outcome. The claim's use of the word "potential" matches the paper's own hedging.
  • The post's caption is accurate on two points that most viral coverage got wrong: the researchers do not claim permanent brain damage, and the paper calls for further work on long-term effects. This matches the authors' own FAQ and limitations statement.
  • The secondary Session 4 point in the carousel, that Brain-to-LLM participants showed higher memory recall and stronger connectivity re-engagement than the LLM group's earlier sessions, is supported by the abstract.

≈ What's misleading 4

  • Omitted qualifier: the claim presents the work as a finished MIT study without noting that it is an unrefereed arXiv preprint. The authors themselves state on their project site that conclusions should be treated as preliminary. That status is material to how much weight a reader gives the result, and it is absent from both the claim and the caption.
  • Omitted qualifier: the sample is not stated. The connectivity and quoting comparisons rest on 54 people split three ways, and the Session 4 crossover result the carousel highlights rests on 18 people split into two subgroups. A methodological comment on arXiv flags exactly this, along with EEG methodology and reproducibility concerns.
  • Subgroup generalization: the claim says "students". Participants were 54 adults aged 18 to 39 recruited from five Boston-area universities, mostly undergraduates but including postgraduates and some degree-holders working at their institutions, and the group was heavily female. The authors themselves list the narrow geographic and institutional recruitment as a limitation. Calling them "students" is close but broader than what was sampled, and it invites generalization to students everywhere.
  • Capability extrapolation, in mild form: "weaker EEG-measured brain connectivity" is a measure of neural engagement during a 20-minute essay task, not a measure of learning, intelligence, or lasting change. The claim does not assert otherwise, but the compressed phrasing lets a reader slide from "less engaged while using a tool" to "worse brain", which is the exact reading the authors' FAQ asks people not to make.

? What's uncertain 4

  • Whether the paper states that Brain-to-LLM participants were specifically better at quoting, as the third carousel slide asserts. The abstract supports higher memory recall for that subgroup but does not make the quoting comparison. I could not confirm that specific sentence in the full text from the excerpts I retrieved.
  • Whether a peer-reviewed version has appeared since. I found none, and the MIT Media Lab page still lists the arXiv preprint, but absence in search is weaker evidence here than a publisher record would be.
  • The widely circulated figures "83% unable to quote" and "up to 55% reduced connectivity" appear in refereed commentary and secondary coverage rather than in the abstract. They are not part of the claim under investigation, and I did not verify them against the paper's own results tables.
  • The status of the Stankovic comment, which is itself an unrefereed preprint. It documents that qualified criticism exists; it does not by itself establish that the criticisms are correct.
Distortion flags omitted qualifier subgroup generalization capability extrapolation
§

Sources

6 of 7 linked to records
[1]

arXiv:2506.08872 abstract page, Kosmyna et al., "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", latest version dated 31 December 2025

primary preprint, artifact of record
https://arxiv.org/abs/2506.08872 ↗
[2]

MIT Media Lab project and publication pages for the same study, including the authors' FAQ

primary institutional page of the authors
https://www.media.mit.edu/publications/your-brain-on-chatgpt/ ↗
[3]

Project site brainonllm.com, including the authors' limitations statement and FAQ

primary authors' own channel
https://www.brainonllm.com/ ↗
[4]

Full paper text excerpts (methods and participant demographics) as reproduced by ResearchGate, Fermat's Library and a hosted PDF copy

primary preprint text
https://fermatslibrary.com/s/your-brain-on-chatgpt-accumulation-of-cognitive-debt-when-using-an-ai-assistant-for-essay-writing-task ↗
[5]

Stankovic, Hirche, Kollatzsch and Doetsch, "Comment on: Your Brain on ChatGPT", arXiv:2601.00856, listed 29 December 2025, indexed by the University of Vienna research portal

primary preprint critique from named academics
https://arxiv.org/abs/2601.00856 ↗
[6]

Armitage, "Your brain on ChatGPT", British Journal of General Practice 2025;75(758):410

secondary refereed journal commentary
https://bjgp.org/content/75/758/410 ↗
[7]

Assorted secondary coverage (Medium explainers, Substack, Mad In America, aggregator sites)

secondary commentary of varying quality
This citation could not be independently verified.
How links are chosen. A source is linked only when the address comes from the investigation's own retrieval or from a registry lookup (PubMed, Crossref) that matches the citation's title and year. Author lists shown as registry-verified come from the registry record, not from the report text. Citations that cannot be matched are labeled, never guessed.
This is one case on the record See the full case, browse the archive, and search every checked claim on TrueSeeker AI Open on ai.trueseeker.com →