Cognitive Debt: What an MIT Study Says About AI and Your Notes

A 2025 MIT Media Lab study wired 54 people to EEG caps and asked them to write essays. The people who leaned on an AI assistant showed the weakest brain connectivity, reported the lowest ownership of their own essays, and struggled to quote what they had just written.¹ It is a preprint, not yet peer-reviewed, with a small sample. The direction still matters for anyone who takes notes.

The study is called "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task."¹ Its coined term, "cognitive debt," names something most of us have felt: the help that makes today easy can quietly leave a bill. The authors themselves urge caution — as of its June 2025 upload, the paper "has not yet been peer-reviewed, thus all the conclusions are to be treated with caution and as preliminary."² We will hold that caveat the whole way through.

This is not an argument against AI. It is an argument about who does the encoding. There is a real line between AI that helps you think about what you wrote and AI that writes so you never have to. For notes — the place you do your actual thinking — that line is the whole ballgame.

What the MIT study actually found

The MIT study split participants into three groups — LLM, search engine, and brain-only — and tracked them with EEG across multiple essay sessions. The headline result was a clean ranking: "Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity."¹ More tool, less brain.

The mechanism is stated plainly in the abstract: "Cognitive activity scaled down in relation to external tool use."¹ In other words, the brain spends what it must and no more. Hand the hard part to a tool and the circuits that would have done it stay quiet. The official project site puts the same finding in one sentence: "Brain connectivity systematically scaled down with the amount of external support."³

Two results land hardest for note-takers. First, "self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group."¹ Second, "LLM users also struggled to accurately quote their own work."¹ You can produce a document and still not feel it is yours — and still not remember what it said.

What "cognitive debt" means

Cognitive debt is the long-run cost of letting a tool do your thinking for you. The Decoder, summarizing the MIT paper, describes the term as the idea that "when relying on AI makes it harder to build and maintain your own thinking skills."⁴ The convenience is real and immediate. The interest accrues out of sight.

The trade has a shape. As The Decoder puts it: "In the short term, cognitive debt makes writing easier; in the long run, it may reduce critical thinking, increase susceptibility to manipulation, and limit creativity."⁵ Over the study's full run, the pattern held — "over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels."¹

The word "debt" is exact. A debt is not a punishment; it is a deferral. You get the essay now and pay in comprehension later. For a note — a thing whose entire purpose is to be remembered and reused — paying later defeats the point.

Why this matters for your notes

Notes are not output. They are the residue of thinking, and the value lives in the act of making them. The MIT ownership finding is the cleanest tie-in: if a model wrote it, the study found you are less likely to feel it is yours and less able to quote it back.¹ A note you cannot recall is a file, not a memory.

Independent work points the same way. Michael Gerlich's 2025 study in Societies surveyed and interviewed 666 people and reported that "the correlation between AI tool use and critical thinking was found to be strongly negative (r = −0.68)."⁶ That figure is correlational and self-reported, not causal, and the paper carries a published correction — treat it as a consistent direction from a separate, larger sample, not as proof. Two different methods, pointing the same way, on different people.

There is also an age wrinkle worth holding. Writing in Psychology Today, Timothy Cook notes that "what AI does to a 45-year-old's brain is categorically different from what it does to a 14-year-old's."⁷ An adult outsourcing a skill they already own is not the same as a student outsourcing one they have not built yet.

Assist versus author: where to draw the line

The useful distinction is who does the encoding. Use AI to react to thinking you already did — to summarize a note you wrote, to surface a contradiction, to ask what you missed — and the thinking stays yours. Use it to produce the note from a prompt, and you have skipped the part that builds memory. The test is simple: did the words arrive in your head, or only on your screen?

Five honest places to draw the line:

1. Write the first pass yourself. The friction is the feature. The MIT brain-only group did the most work and owned the most.¹
2. Let AI question, not generate. "What's weak here?" keeps you in the loop; "write this for me" takes you out.
3. Summarize what you already understand, not what you skipped. A summary of a thing you read is review; a summary of a thing you never read is a substitute.
4. Quote yourself. If you cannot restate a note without rereading it, you outsourced the encoding.¹
5. Keep the raw thinking somewhere durable. The drafts, the cross-outs, the half-formed lines — that is the part worth keeping.

The caveats, stated plainly

This is a preprint, and small. "A total of 54 participants took part in Sessions 1-3, with 18 completing session 4."¹ Eighteen people in the crossover phase is thin, and the authors say so. The sample was also narrow: the experiment "divided 54 mostly college students from five Boston-area universities into three groups."⁸ Five schools in one metro is not the world.

The design has named limits, too. As The Decoder records the authors' own list: "First, the sample size was small: 54 participants, with only 18 in the fourth (crossover) session. Larger samples are needed for stronger conclusions. Second, only ChatGPT was used as the LLM."⁹ One model, one task, one region. None of this makes the finding worthless; it makes it preliminary.

And the news is not all grim. Cook's reading of the literature offers a recoverable frame: "AI-driven cognitive atrophy is recoverable. Cognitive foreclosure may not be."⁷ Atrophy reverses when you start using the muscle again. The risk is never starting — letting a skill go unbuilt. The same trap shows up in code, where developers describe relearning to write by hand after a stretch of letting a model do it.¹⁰

How this shapes a note tool

A note app should be where you do the thinking, not where thinking is done for you. That points to a few design choices: plain Markdown you write yourself, kept locally on your own device, in your browser, offline, with no account in the way. The friction of writing is not a flaw to automate away. It is the point.

AI still has a seat — as an assist, not an author. Bring-your-own-key AI can react to what you already wrote: clarify a tangled paragraph, surface a contradiction, ask what you left out. That keeps you doing the encoding while the model does the chores. The ownership finding is the design constraint: a note you did not write is not really yours.¹

Frequently asked questions

Is it bad to let AI write my notes?

The MIT study suggests a real cost: people who let an LLM write reported the lowest ownership of their work and struggled to quote it back.¹ The honest answer is to use AI as an assist — to question and summarize what you already wrote — rather than as an author. Notes you did not write are harder to remember and harder to call your own.

Does using AI hurt memory and thinking?

In the MIT EEG study, brain connectivity ranked brain-only highest, search engine in the middle, and LLM lowest — "cognitive activity scaled down in relation to external tool use."¹ An independent 666-person survey reported a strong negative correlation between AI use and critical thinking.⁶ Both are preliminary, but the direction is consistent: outsourced effort means less of your own.

What is "cognitive debt"?

Cognitive debt is the long-term cost of offloading mental effort to a tool. The Decoder defines it as what happens "when relying on AI makes it harder to build and maintain your own thinking skills."⁴ You get the easy output now and pay later in weaker comprehension, recall, and critical thinking. The term was coined in the MIT Media Lab paper's title.¹

Is the MIT "Your Brain on ChatGPT" study peer-reviewed?

No. It is a preprint on arXiv. The authors state that as of June 2025 it "has not yet been peer-reviewed, thus all the conclusions are to be treated with caution and as preliminary."² The sample is small — 54 people, with only 18 in the crossover session — and only ChatGPT was tested.⁹ Treat it as a strong signal, not a verdict.

Will AI make me forget how to write?

Not permanently, if you keep writing. Psychology Today frames the effect as recoverable atrophy rather than permanent loss: skills return when you use them again.⁷ The deeper risk is for skills never built in the first place. Drawing the line at assist-versus-author — writing the first pass yourself — is how you keep the muscle.

Should I use AI to summarize my notes or write them myself?

Write them yourself; use AI to react afterward. A summary of a note you understood is useful review. A summary of material you never engaged with hands the encoding to the model — exactly the move the MIT study links to lower ownership and worse recall.¹ Let AI question your thinking, not replace it.

The convenience that writes your notes for you is also the convenience that keeps them from becoming yours. Write the first pass yourself; let AI react to thinking you already did.

If you want a place to do that thinking, mnmnote.com is a browser-based Markdown editor that keeps your notes on your own device, with optional bring-your-own-key AI as an assist rather than an author.