Cognitive Debt: What the MIT "Your Brain on ChatGPT" Study Really Found

Cognitive debt is the term the MIT Media Lab’s 2025 EEG study “Your Brain on ChatGPT” used for the cost of letting an AI write in your place: participants who drafted essays with ChatGPT showed weaker connectivity between brain regions than people who wrote unaided, and most could not quote a sentence from their own essay minutes after submitting it. The study is small and was released as a preprint, so it does not prove that AI “makes you dumber”. What it does show is narrower and more useful: when the tool generates the language and you merely supervise, your brain does less of the work that writing normally forces — and skips the learning that comes with that work. The problem is not AI assistance. It is full ghostwriting.

That distinction matters for everyone, but it matters double if you write in English as a second language, which is the situation I write from myself. I will come back to why. First, the study — what it actually did, what it found, and where the headlines outran the data.

What the MIT study actually did

The study, run by a team at the MIT Media Lab, took 54 university students from the Boston area and split them into three groups to write short essays on the kind of open prompts you would meet in a standardised test. One group wrote with ChatGPT, one with a search engine but no AI, and one with nothing but their own head. While they wrote, an EEG headset recorded the electrical activity across their brains. The sessions repeated over several months, so the researchers could watch patterns build up rather than snapshot a single essay.

Three findings stand out.

First, neural connectivity scaled down with the amount of external support. The brain-only group showed the strongest and most widespread connectivity between regions while writing; the search group sat in the middle; the ChatGPT group showed the weakest. The more the tool did, the less the brain did — measurably.

Second, memory of one’s own text collapsed in the AI group. Asked to quote from the essay they had just finished, most ChatGPT users could not produce a single accurate line. Writers in the other groups mostly could. The AI group also reported a weaker sense of ownership over the text. That is honest of them, because in a real sense it was not their text.

Third, and most interesting, the debt lingered. In a later session, some participants swapped conditions. Those who had leaned on ChatGPT for the earlier essays and then had to write unaided still showed weaker engagement than people who had trained on their own effort all along. The phrase “cognitive debt” works like technical debt in software: the shortcut is not free, you repay it later, and interest accrues quietly in the meantime.

The essays themselves told a story too. Texts from the ChatGPT group came out more similar to one another — the statistical sameness that makes AI-assisted prose feel weightless and that human readers pick up on more often than we expect.

What the study did not show

If you searched “ai writing making us dumber” and landed here, this is the honest part. The study does not support that headline, and the researchers themselves publicly pushed back when the coverage ran that way.

The limitations are real and worth naming plainly:

• It is a preprint with a small sample. Fifty-four participants from one metropolitan area, and a smaller subset in the crossover session. That is enough to generate a hypothesis worth taking seriously, not enough to settle it.
• EEG connectivity is not intelligence. Weaker connectivity during a task means less coordinated engagement during that task. It is not a reading of anyone’s IQ, and nothing in the study measures lasting damage to general ability.
• One task, one mode of use. Participants used ChatGPT to produce essay text under time pressure. The study says little about using AI to check grammar, find a word, critique a draft you wrote, or explain a concept — modes where you remain the one generating the language.
• Effects were context-specific. The crossover result is suggestive, not proof of permanent change. Nobody’s brain was shown to be broken; a habit was shown to have a residue.

So the fair summary is not “ChatGPT damages your brain”. It is: on the specific task of producing text, outsourcing the production to a model produced measurably shallower engagement and much worse memory of the result, and the pattern persisted after the tool was taken away. That is a smaller claim than the headlines made. It is still a claim that should change how you work.

Why ghostwriting, specifically, is the problem

The MIT result surprised journalists more than it surprised memory researchers, because it lines up with one of the oldest findings in the field: the generation effect. Information you generate yourself — a word you retrieved, a sentence you assembled, an argument you constructed — is remembered far better than the same information read from a page. Psychologists have documented this since the late 1970s. Producing is encoding. Reading is barely encoding at all.

Writing is where this bites hardest, because writing is not the transcription of finished thoughts. Anyone who writes seriously knows the thoughts are mostly not there until the sentence forces them into shape. The struggle — choosing this word over that one, noticing the argument does not follow, rebuilding the paragraph — is not friction around the thinking. It is the thinking.

Full ghostwriting removes exactly that. You type a prompt, text appears, you skim it, you accept it. You have become an editor of language you never generated, and the generation effect predicts precisely what MIT observed: you will not remember it, and you will not quite feel it is yours. There is a name for the extreme version of this workflow — vibe writing, where the human contribution shrinks to prompting and taste — and the cognitive debt framing explains why it feels efficient in the moment and hollow a week later.

For second-language writers, the debt compounds

English is my second language. When I read the MIT study, the finding that worried me was not the essay recall. It was what the mechanism implies for people still building the language itself.

A native speaker who ghostwrites an essay pays cognitive debt on that essay: weaker memory, weaker ownership, a skipped rehearsal of thinking. Their English, however, is finished growing. For an L2 writer the same shortcut costs twice, because every sentence you do not produce is also retrieval practice that never happened. Language ability grows through production. Forming a sentence forces you to notice the gap between what you want to say and what you can say, and that noticing is where acquisition happens. Read-only exposure does not do it; that is a large part of why so many learners get stuck on the intermediate plateau, understanding far more than they can produce.

Hand the production to a model and you freeze exactly there. The essays get better; you do not. Worse, you start to depend on the tool for a level of English you never actually acquired, which is a fragile place to stand in an exam hall, an interview, or a meeting where no tool is allowed. Cognitive debt for a native speaker is a study-skills problem. For a second-language writer it is a development problem: the interest compounds in the one asset you were trying to build.

The middle path: scaffolding, not ghostwriting

The useful question to ask of any AI writing feature is simple: who generated this sentence? If the answer is “the model, and I approved it”, you are borrowing. If the answer is “I did, and the tool helped me at the word and phrase level”, you are scaffolded: supported where support is cheap, still doing the work that pays.

This is the line Diglot is built along, and it comes directly from the ESL experience. When I draft in English and a word refuses to come, the old workflow was to stop, open a translator tab, lose the sentence, and come back. Weave keeps you generating instead: you write the failing word in your own language mid-sentence, a popup offers English options, you choose one and keep moving. The sentence is still yours — you assembled it, you selected the word, you noticed the gap. That noticing is the learning event ghostwriting deletes.

The same logic applies below and after the sentence. Contextual synonyms replace a word you chose with a word you choose: a decision, not a delegation. A grammar pass corrects text you produced, which is exactly the feedback-on-production loop language learning runs on. And diagnostic tools that score rather than rewrite — checking whether your draft reads at the level you intend or carries structures from your first language — tell you about your writing while leaving the fixing, and therefore the learning, to you.

How to keep the pen: five working rules

1. Draft before you prompt. Even a rough, mixed-language, ungrammatical draft means the generation happened in your head. Everything after that is assistance, not substitution.
2. Use AI below the sentence line. Words, phrases, prepositions, articles — borrow freely there. Sentence assembly and paragraph logic stay yours.
3. Ask for feedback, not text. “What is weak in this argument?” builds you. “Write this argument” bills you.
4. Close the loop on corrections. When a tool fixes your grammar, read what changed and why. A correction you never looked at is a lesson you paid for and did not attend.
5. Let your process leave a record. Writing this way produces a visible trail of human work — drafts, revisions, word-level decisions — which is also, incidentally, exactly the evidence that protects you if a detector ever flags your writing unfairly. Diglot can turn that trail into an Authorship Certificate, but the deeper point is that a real process existed to record.

The debt metaphor is the right one precisely because debt is not evil. It is a tool with an interest rate. Borrowing at the word level, where the interest is near zero, is what good tools are for. Borrowing the generation itself, the part where thinking and language grow, is the loan the MIT study caught people taking without reading the terms. Write the sentences yourself. Let the machine hand you words. That split is not a compromise between quality and learning. For a second-language writer it is the only arrangement that delivers both.