“Cognitive debt”: what the MIT “Your Brain on ChatGPT” study actually found

The MIT study did find something: people who wrote essays with ChatGPT showed weaker EEG connectivity, felt less ownership of what they wrote, and could not quote their own essays as well as people who wrote unaided. It did not find that AI rots your brain. It is a preprint that has not passed peer review, it followed 54 people in total and only 18 into its key final session, and it tested one narrow task under time pressure. The honest summary is that it is a suggestive early result, consistent with what the wider literature on cognitive offloading predicts, and that the phrase “cognitive debt” has travelled much further than the evidence for it.

What the study did

The preprint, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, was posted to arXiv in June 2025 by a group at the MIT Media Lab and revised in December 2025. It has not been peer-reviewed as of 13 September 2026.

The design: “Participants were divided into three groups: LLM, Search Engine, and Brain-only (no tools). Each completed three sessions under the same condition. In a fourth session, LLM users were reassigned to Brain-only group (LLM-to-Brain), and Brain-only users were reassigned to LLM condition (Brain-to-LLM). A total of 54 participants took part in Sessions 1-3, with 18 completing session 4.” Participants wore EEG caps while writing timed essays; the researchers “used electroencephalography (EEG) to assess cognitive load during essay writing, and analyzed essays using NLP, as well as scoring essays with the help from human teachers and an AI judge”.

So: 18 people per group, one essay task, EEG as the main outcome, and a fourth crossover session that only a third of participants completed.

What it found

Three results carry the paper’s argument.

First, brain connectivity. “EEG revealed significant differences in brain connectivity: Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity.” The authors summarise this as “Cognitive activity scaled down in relation to external tool use.”

Second, the crossover session. “In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement.” The people who had used ChatGPT for three sessions and then wrote unaided did not bounce back to the Brain-only pattern. Meanwhile, “Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas, similar to Search Engine users.” This asymmetry is the basis for the “debt” metaphor: the authors argue that starting with the tool leaves a deficit that persists when the tool is removed.

Third, memory and ownership. “Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work.” This is the most intuitive finding: if a tool wrote the sentence, you do not remember the sentence.

The authors’ own framing is measured in places and expansive in others. They write that “While LLMs offer immediate convenience, our findings highlight potential cognitive costs”, and that “Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels.”

The limits, which are substantial

A formal commentary on the preprint, itself a preprint posted to arXiv in January 2026 by psychologists at the University of Vienna and TU Dresden, lays out the problems in detail. Its authors state that their “primary concerns focus on: (i) study design considerations, including the limited sample size; (ii) the reproducibility of the analyses; (iii) methodological issues related to the EEG analysis; (iv) inconsistencies in the reporting of results; and (v) limited transparency”. Those are the standard concerns a peer reviewer would raise, and the study has not yet been through that process.

The sample-size point is not cosmetic. The commentary runs a power calculation for the study’s design and reports that it “indicates that approximately N = 159 participants would be required” for adequate power. The study had 54, and its most-cited finding (the crossover) rests on 18. Underpowered studies do not just miss real effects; they inflate the effects that do reach significance, which is the kind of result that travels as a headline.

The commentary also questions what “weaker connectivity” means. A “significant connection” in the study is one that is stronger in one group than another, not an absolute measure, so “a group with fewer significant connections should not necessarily be interpreted as exhibiting weaker overall neural activity”. Lower connectivity could reflect less effort, but it could also reflect a different strategy or simply less time on task, because the essay was largely drafted by the tool.

And the behavioural finding cuts less cleanly than the debt metaphor implies. On the quoting task, the LLM group did worse, but the comparison between the search-engine group and the unaided group “was non-significant (p = 1), indicating a null effect”. If delegating to any external tool accrued debt, search users should have shown some. They did not. The finding is specific to having a model write your text, not to using tools in general.

Add the task itself: a timed essay under lab conditions, with a single product (ChatGPT). Nothing in this design speaks to code review, debugging, analysis, or any task where the person and the tool iterate. It is a study of what happens when you let a model draft your prose and then hand it in.

Why the finding is still worth taking seriously

Having said all that, the direction of the result is not surprising, and that matters. There is a well-developed literature on cognitive offloading, the “use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand”, which is what using ChatGPT to draft an essay is. That literature’s core warning is that we decide when to offload based on a fallible sense of our own abilities, so we often lean on the tool more than we should. The MIT participants who felt less ownership and could not quote their work are a small demonstration of the same mechanism.

A second, independent data point arrived in January 2026: a randomised experiment in which developers learned a new programming library with or without an AI assistant. The authors “find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average”. That study is also a preprint, and it was written by researchers at Anthropic, which sells an AI assistant, so treat it as a vendor-authored result; but its finding runs against the vendor’s commercial interest, and it points the same way as the MIT data on a different task with a different population. Two small preprints agreeing is not proof. It is a pattern worth watching.

The key nuance in both studies is the same: the effect depends on how the tool is used. The Anthropic preprint identified “six distinct AI interaction patterns, three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance”. The MIT study’s Brain-to-LLM group, who thought first and used the tool second, showed the higher-recall pattern. Delegation and engagement are different behaviours with the same tool.

What “cognitive debt” should and should not mean

As a metaphor, cognitive debt is useful: if you let a tool do the part of a task that would have formed a memory or built a skill, you have deferred that cost, and you may pay it later when the tool is absent. That is compatible with decades of work on offloading and with the learning literature, and we cover the general version in cognitive offloading and the AI confidence gap.

As an empirical claim about brains, it rests on one underpowered, unreviewed EEG study of essay writing, with a crossover finding from 18 people and a definition of “connectivity” that the study’s critics say cannot support the interpretation placed on it. “AI makes you dumber” is not what the data show. “Letting a model write for you means you do not learn what it wrote” is closer, and much less exciting.

A disclosure: this site drafts its articles with AI assistance, which is a reason to be careful in both directions, neither flattering AI nor waving away findings that criticise it. The MIT study deserves neither the headlines it got nor dismissal. It deserves replication, at three times the sample size, with peer review. Until then, treat “cognitive debt” as a hypothesis with a good name. For how we weigh preprints against published work, see how we research.

Common questions

What did the MIT “Your Brain on ChatGPT” study find?

Across 54 participants writing timed essays with ChatGPT, a search engine, or no tools, EEG showed the weakest brain connectivity in the ChatGPT group, who also reported the lowest ownership of their essays and struggled to quote their own work. In a crossover session with 18 people, those who switched from ChatGPT to unaided writing showed reduced alpha and beta connectivity. It is a preprint, not peer-reviewed.

Is the cognitive debt study peer reviewed?

No. It was posted to arXiv in June 2025 and revised in December 2025, and as of September 2026 it has not been published in a peer-reviewed journal. A separate arXiv commentary from psychologists at the University of Vienna and TU Dresden raises concerns about sample size, EEG methodology, reproducibility and reporting inconsistencies. Treat the findings as preliminary.

Does using ChatGPT make you dumber?

The study does not show that. It shows that people who let a model draft an essay engaged less, remembered less of the text, and felt less ownership, on one narrow task in a small sample. The search-engine group showed no memory impairment relative to unaided writers, so the effect is specific to delegation, not tool use. Nothing in the data measures general intelligence.

What is cognitive debt in the context of AI?

The MIT authors use it to describe a deferred cost: cognitive work you skip by letting an AI do it, which you may pay for later when the tool is not there. As a metaphor it fits the cognitive-offloading literature, which warns that our judgements about when to offload are often wrong. As a measured phenomenon it currently rests on one small, unreviewed study.

How should a developer use AI without losing skills?

Both the MIT study and a January 2026 preprint from Anthropic researchers point to the same distinction: engaging with the problem before or alongside the tool preserved learning, while full delegation did not. Draft your own reasoning first, ask the tool to explain rather than only to produce, and keep some tasks tool-free. The Anthropic preprint found three of six interaction patterns preserved learning even with AI help.