Cognitive Debt Is Not Brain Damage

Key Takeaways

• “Cognitive debt” entered wide use through a 2025 MIT Media Lab preprint with 54 participants, 18 of whom did the crossover session the headlines rest on. It lists no journal reference, and its authors ask people not to say “brain rot”, “harm” or “damage”.
• Three claims get blurred. You remember and own less of work you offload: well supported. Your unaided performance can drop after practising with the tool: supported in specific settings, most cleanly a PNAS field experiment where a plain GPT-4 interface cost students 17% on a later exam and a hint-giving tutor did not. AI use erodes general thinking ability over time: not shown by any of these studies, because none was designed to show it.
• Software engineers now use the same phrase for something different: a team losing the theory of its own system.
• Debt is the right word precisely because debt is not damage. It is a deferred cost. The useful question is which understanding you cannot afford to borrow.

On 16 June 2025, a 216-page preprint from the MIT Media Lab was posted to Hacker News, where it collected 391 points. Seven months later it was posted again and collected 710. Its title was “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”. By then the phrase “cognitive debt” had escaped the paper, and through 2026 software engineers adopted it for something rather different.

Most of what gets said about the study is stronger than the study. Their project FAQ answers the question “Is it safe to say that LLMs are, in essence, making us ‘dumber’?” like this:

No! Please do not use the words like “stupid”, “dumb”, “brain rot”, “harm”, “damage”, “passivity”, “trimming” and so on. It does a huge disservice to this work, as we did not use this vocabulary in the paper, especially if you are a journalist reporting on it.

So this essay reads the papers: what each one measured, what it found, and how far the finding reaches.

What is cognitive debt?

In the MIT paper the phrase appears only a handful of times. The definition comes near the end: cognitive debt is “a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking.” It “defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity.” The paper does not claim to have coined the term, and I could not establish who first used it.

The software meaning arrived in February 2026. Margaret-Anne Storey, a professor of computer science at the University of Victoria, wrote that where technical debt lives in the code, “cognitive debt lives in developers’ minds.” She calls it “a term gaining traction recently”, so she is not claiming it either. She describes a student team that, by week seven or eight, could not make simple changes safely, because “no one on the team could explain why certain design decisions had been made.”

Nine days later she tightened the definition: “the accumulated gap between a system’s evolving structure and a team’s shared understanding of how and why that system works.” An essay titled “Cognitive Debt: When Velocity Exceeds Comprehension” reached 507 points on Hacker News. In April, Martin Fowler relayed Storey’s split of technical debt in code, cognitive debt in people and intent debt in artifacts, while “getting a bit bemused by debt metaphor proliferation.” In August, Ankur Sethi’s post about retyping every line his coding assistant produces drew 545 points and 443 comments.

These are different claims. The MIT usage is about an individual’s cognition, the engineering usage about a team’s shared model of an artifact. Evidence for one is not evidence for the other. What they share is a shape.

What did the MIT study actually find?

Nataliya Kosmyna and seven colleagues recruited participants, most of them students, from five universities around Boston. The paper reports data from 54 of them, aged 18 to 39 with a mean age of 22.9. They were split into three groups of 18. One group wrote essays with ChatGPT, one with a search engine, and one with no tools, which the paper calls Brain-only. Each person wrote one essay per session, on a prompt taken from SAT tests, within a 20-minute limit, while wearing a 32-channel EEG headset. There were three such sessions, spread over four months.

Then came a fourth session, which the authors call the source of “some of the most striking observations in our study”. Participants who returned were switched: LLM users wrote unaided, Brain-only users got ChatGPT. In the paper’s words, “only 18 participants were able to attend.” That is nine people per arm. They wrote on topics they had already written about.

The behavioural results from the first session are the ones worth remembering. Asked minutes after writing to quote a sentence from their own essay, 15 of 18 LLM users could not do it, against 2 of 18 in each of the other groups. Not one LLM user produced a correct quote. Nine LLM users claimed full ownership of their essay and three claimed none. Among Brain-only writers, 16 of 18 claimed it.

Less often reported: the gap narrowed. By the third session, 6 LLM users could not give a correct quote, against 2 in each other group, and 12 of 18 claimed full ownership.

What does EEG connectivity show, and what does it not?

The headline finding is neural. Using the dynamic directed transfer function, a measure of directed statistical coupling between pairs of electrodes, the authors report that Brain-only writers showed the strongest and most widely distributed connectivity, search users were intermediate, and LLM users the weakest, with “up to 55% reduced total dDTF magnitude” in some low-frequency networks.

Read that as a description of what people were doing for twenty minutes. Composing sentences from nothing is a different task from prompting and editing. It would be strange if the brain looked the same during both. It is not a before-and-after comparison of anyone’s capacity. The authors’ own limitations section says the findings “are context-dependent” and “may not generalize across tasks”, and that the result they call most concerning “should be considered preliminary, as a larger participant sample is needed to confirm the claim.”

The most substantial critique I found is a commentary by Milos Stankovic and three colleagues at Vienna and TU Dresden, itself a preprint. A basic power analysis for the design, they write, “indicates that approximately N = 159 participants would be required for adequate power.” A “significant connection” in the paper means a coupling that was stronger in one group than another, so “a group with fewer significant connections should not necessarily be interpreted as exhibiting weaker overall neural activity.”

Their sharpest point is about the theory. On quoting, the search group did as well as the Brain-only group. If debt accumulated with delegation, it should sit in between. Their conclusion: “there is no strong evidence that the LLM use negatively affects performance itself.”

More than a year on, the arXiv record lists no journal reference, and I could not find a published replication. That leaves three claims to separate.

Claim one: do you remember less of what you offload?

Yes, and this is the least surprising and best supported of the three. If ChatGPT drafted the sentence, you did not encode the sentence.

The older version of this finding is “Google Effects on Memory”, published in Science in 2011 by Sparrow, Liu and Wegner. Across four studies they reported that when people expect to have future access to information, “they have lower rates of recall of the information itself and enhanced recall instead for where to access it.”

Its replication record deserves precision. A large project that re-ran 21 social science experiments from Nature and Science found a significant effect in the original direction for 13 of them. Sparrow’s paper was among those that failed. But the experiment re-run was the first one, a priming task claiming that hard trivia questions make people think of computers, and a later attempt closer to the original design also found “no conclusive evidence” for it. The recall experiments, the part everyone means when they say “Google effect”, were not what that project tested. I did not find a direct replication of them, so I will not claim one.

Claim one, then: you hold less of what you did not make. It is true and it is modest. It concerns that essay, that fact, that function. It says nothing yet about you.

Claim two: does performance drop when the tool is taken away?

The best evidence here is not the MIT study. It is a randomized controlled trial by Hamsa Bastani and five colleagues, published in PNAS in June 2025. Nearly a thousand students in about fifty classes at a Turkish high school went through four 90-minute maths sessions in autumn 2023. Each had a teacher review, a practice period, then an unassisted exam. During practice, one arm had textbooks and notes only. One had “GPT Base”, a plain chat interface on GPT-4. One had “GPT Tutor”, the same model with the solution and teacher-designed hints in its prompt, set up to give hints instead of giving away answers.

During practice, both AI arms did better: 48% better for GPT Base, 127% for GPT Tutor. On the closed-book exam that followed, on problems conceptually very similar to the practice ones and with no AI for anyone, the GPT Base students scored 17% lower than students who had never had it. The GPT Tutor students were statistically indistinguishable from the control arm.

Three details matter more than the headline. First, the 17% is relative: the drop was 0.054 on a scale of 0 to 1, against a control standard deviation of 0.277, which by my arithmetic is about a fifth of a standard deviation. Significant, and not a collapse. Second, the mechanism is visible in the chat logs. GPT Base students “most often simply ask for the answer”, and the answer they copied was correct “only 51% of the time.” Third, the students could not tell. Those in the GPT Base arm “did not perceive that they performed worse or learned less.” It is the pattern from this site’s essay on sleep loss: the deficit hides from the person who has it.

The authors are careful about reach: one school, one subject, an early model, and “we focus on short-term outcomes due to limitations imposed by our partner school.”

So claim two holds in a specific form: use the tool to skip the practice and you lose the benefit of the practice. The same experiment shows this is not a property of AI. The same model, set up to withhold the answer, produced no measurable harm. The variable that mattered was whether the student still did the work.

Does using AI make you worse at thinking?

This is claim three, the one in the headlines, and none of the studies above can carry it.

Start with Michael Gerlich’s 2025 paper in Societies, “AI Tools in Society”. Surveys and interviews with 666 participants “revealed a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading.” It is a correlation from self-report at one point in time, and the arrow could run either way.

The Microsoft Research and Carnegie Mellon survey presented at CHI 2025 has “Self-Reported” in its title. Hao-Ping Lee and colleagues collected 936 examples of AI use from 319 knowledge workers and found that “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.” They did not measure anyone’s thinking, and they describe the effort as shifting toward “information verification, response integration, and task stewardship” as much as shrinking.

The preprint behind this summer’s Hacker News story “AI advice made people less accurate but more confident” is a stronger design, and I have written about it before. Across five experiments with 3,132 participants, Marcoccia, Quattrociocchi and Capraro found that access to AI advice nearly eliminated people’s willingness to say “I don’t know”. Without incentives, correctness was 9.2% with AI against 27.5% without, while mean confidence was 75.9 against 29.6. But the questions were engineered so the AI was almost always wrong, and the authors say it “is untested whether suspension collapses to the same degree when AI advice is usually correct.” It shows what a fluent answer does to judgment in the moment, not that judgment is different the next day.

To show claim three you would need to measure general reasoning on tasks unrelated to AI use, before and after a long period of use, against a comparable group that abstained. Bastani’s exam tested the material just practised. The MIT crossover had nine people per arm writing on familiar prompts. The surveys are snapshots. That does not make claim three false. The argument in Why We Think still seems right to me as an argument. But for now it is a hypothesis with a plausible mechanism, and the honest sentence is “we do not know yet.”

Why “debt” is the right word

The word the authors chose is better than the coverage it generated, because debt is not damage.

A debt is a cost you moved to later. You took something now, the essay or the feature or the answer, without paying for the understanding that would normally come with it. Nothing is broken. Often it is the correct trade: no engineer needs a theory of every dependency.

Debt becomes a problem under two conditions. The first is when you never pay and it compounds, each new piece resting on pieces you also did not understand. That is Storey’s student team at week eight. The second is when you borrow without knowing it. Bastani’s students thought they had learned. The trouble with this kind of debt is that no statement arrives. It is also why I doubt addiction is the right frame: the person in debt is not craving anything. They have not looked at the balance.

The evidence is fairly encouraging about repayment. The MIT quoting gap shrank across sessions. The tutor arm in Turkey cost nothing measurable. Sethi’s retyping is a repayment plan: comprehension bought with speed.

Which understanding can you not afford to borrow?

That is the practical question, and the answer is personal. The test is to read your own balance: can you quote it, explain it, change it, or defend it with the tool closed? Where the answer is no, decide whether that matters. For a boilerplate email it does not. For the system you are on call for, the argument you sign your name to, or the subject you are supposedly learning, it does. Those are the places to do the first pass yourself.

The same test applies to beliefs. A conclusion you hold but could not reconstruct is a belief you are holding without knowing why. It may be true. It is still borrowed.

None of this requires believing a chatbot is thinning anyone’s cortex. It requires only the old observation that understanding is made by the person who does the work, and that an answer, as the Fields medallists put it this month, is a proxy for it. You can take the answer without the understanding. Sometimes you should. Just know which of the two you took.