Cognitive Debt: What AI Use Does to Your Brain

I ran a client deliverable through AI, start to finish. It was fast, genuinely fast, the kind of fast that used to take me a day and now took an hour. But the result was less precise than what I'd have produced working through it myself. And something was off in me too. I hadn't learned anything building it. I didn't own the reasoning behind a single choice in it. Explaining it back felt like reciting someone else's notes. The job got done. It just wasn't satisfying to do.

That gap, between output that looks fine and understanding that never actually formed, has a name now: cognitive debt. Researchers started using the term seriously in 2025, and the evidence behind it is more specific, and more useful, than a blanket "AI makes you lazy."

What cognitive debt actually looks like in the research

The setting matters more than the tool

A randomized controlled trial published in PNAS in 2025 put close to 1,000 high schoolers in Turkey through a semester of math practice under three setups: no AI at all, a standard ChatGPT-style assistant, and a modified tutor built to give hints instead of full answers. During the practice exercises, both AI groups looked great. The standard-assistant group scored 48% higher than the no-AI group. The hints-only tutor group scored a striking 127% higher. Handing students a tool that answers made a visible difference in the moment.

Then came the exam, with no AI allowed. The standard-assistant group scored 17% worse than the no-AI group. All that visible improvement had been borrowed, not earned, and the bill came due the moment the tool was taken away. The hints-only group, though, was statistically indistinguishable from the no-AI group. No measurable loss.

Same underlying technology, wildly different outcome for actual learning. The variable wasn't whether AI was present. It was whether the AI was configured to hand over the answer or to force the next step of reasoning. That's the finding I keep coming back to: cognitive debt isn't a property of the tool, it's a property of how you've set the tool up to interact with your thinking.

Correlation, not proof, but a pattern worth taking seriously

A 2025 study by Michael Gerlich, surveying 666 people in the UK with a validated critical-thinking scale plus interviews, found a negative correlation between frequent AI tool use and critical-thinking performance. It was strongest in the youngest bracket, ages 17 to 25, who also reported the heaviest reliance on AI. Gerlich's explanation is cognitive offloading: the more a mental task gets handed to a tool, the less practice your own reasoning gets.

Worth saying plainly: this is a correlational, self-reported, single-wave study. It cannot show that AI use causes weaker critical thinking, only that the two show up together, more so in the people who lean on AI the most. Treat it as a warning sign, not a verdict.

The 83%

The most viral number in this space comes from a preprint out of MIT Media Lab, still not peer-reviewed, so hold it loosely. Researchers had 54 people write essays: some using ChatGPT, some using a search engine, some using neither. Minutes after finishing, 83% of the ChatGPT group could not quote a single sentence from the essay they had just written. In the other two groups, about 11% couldn't. The EEG readings, qualitative rather than a clean percentage, told a matching story: the strongest, most distributed brain connectivity was in the no-tool group, moderate in the search-engine group, weakest in the ChatGPT group.

Again, this is a preprint, not yet peer-reviewed, on a small sample. But it rhymes with the PNAS exam result and with Gerlich's correlation. Three independent lines of evidence pointing the same direction is worth attention, even before peer review closes the loop.

Why doing the work sticks and reading the answer doesn't

Two older findings explain the mechanism underneath all of this. In 1978, Slamecka and Graf ran five experiments showing what's now called the generation effect: material you generate yourself, even filling in a single missing word, gets remembered better than the identical material simply read. Producing beats consuming, at the level of memory formation itself.

And in 2002, Rozenblit and Keil documented the illusion of explanatory depth: people are confident they understand how ordinary things work, a zipper, a toilet, a bicycle derailleur, right up until they're asked to explain the mechanism step by step, at which point the confidence collapses. A fluent AI explanation gives you the same false sense of understanding secondhand. The explanation is coherent, so you assume the coherence is now in your head. It isn't, until you've had to reconstruct it yourself.

It's not the tool. It's the missing friction.

This is the nuance I want to land, because "AI is bad for your brain" isn't what the PNAS study actually shows. The hints-only tutor group used AI just as much as the standard-assistant group, and lost nothing. What differed was friction: whether the tool produced the next step for you or made you produce it yourself.

David Brooks made a version of this argument in a 2026 essay for The Atlantic, sorting people into three profiles based on how much mental effort they tolerate. Productive Passengers let AI do the driving and don't mind. Reluctant Optimizers know AI might hollow them out, resolve not to let it happen, and get pulled in anyway once life gets busy and stressful. Mental Marathoners have a high tolerance for effort and actively resist over-relying on AI, using it to extend what they can do rather than replace the doing.

Brooks's read: your outcome in the AI era depends less on how smart you are than on your relationship to mental effort, and that relationship isn't fixed. It's not a personality you're stuck with. It's closer to a habit.

I think of AI the way I think of fire: genuinely useful, genuinely dangerous if you never learn to handle it, something you tame through practice rather than something you either fully trust or fully avoid.

My three rules

I've published these before, and I still run them daily.

1. Think first. Before I open a chat window, I write down, badly, in a few lines, what I already know, what I'm hypothesizing, and what's actually unclear to me. That forces the reflection to happen in my head before it happens in the tool's.
2. AI as coach, not flatterer. I ask it to challenge what I've written, not confirm it. "What's wrong with this" produces a different, more useful conversation than "does this look good."
3. Design the friction on purpose. Before I accept an explanation, I try to re-explain the concept myself, the way I would to a student who's never seen it. If I can't, I don't understand it yet. I've only read something that sounded like understanding.

None of this means using AI less. It means keeping the loop closed by me, not by the tool.

Wiring the friction into your pomodoros

The think-first pomodoro

Run one full pomodoro before you open any AI tool on a new problem. No prompting allowed, just you and the problem: what you know, what you'd try, where you're stuck. It's the PNAS lesson applied to your own schedule. The exam is the moment you're forced to think without the assistant, so create that moment on purpose, before you prompt, not after.

Alternate AI and no-AI blocks

Robert Bjork's research on desirable difficulties shows that certain forms of friction, spacing, self-testing, reduced hints, slow you down in the moment but strengthen what you retain. Applied here: don't run AI-assisted pomodoros back to back all day. Alternate a block where AI is off-limits with a block where it's fully available. The friction of the AI-off block is the desirable kind, deliberately slower and deliberately more durable.

Let your breaks rebuild what the work borrowed

What you do in the minutes between pomodoros matters too. A break spent scrolling stacks more passive input on top of a session that may already have been AI-heavy. A break that asks you to reconstruct, recall, or briefly reason without a tool does the opposite: it's a small, regular repayment on the debt. That's the whole premise behind Never Dumb breaks, and it connects to the bigger idea of cognitive sovereignty: staying the author of your own thinking even as the tools around you get more capable.

The week I tried to quit AI

In March 2025 I went a full week without touching any AI tool. Not a detox for its own sake, an experiment to see what I'd notice.

What I noticed first was depth. Thoughts that would normally get outsourced to a prompt sat with me longer, and went further than I expected. Second, I noticed the reflex itself: how often, mid-thought, some part of me reached for "let me just check what ChatGPT thinks" before I'd even finished forming my own opinion. That reflex was the real discovery of the week, not the depth.

I made it to Wednesday. At 5:23pm I cracked, in one sitting: a transcription I didn't want to do by hand, a client proposal against a deadline, a training summary. All three were genuinely infrastructure tasks, not thinking tasks, and that's the distinction the week taught me. AI is infrastructure. The vigilance isn't about whether you use the tool. It's about whether you're still thinking.

I'm not a doomer about any of this, and I'm not an evangelist either. My position, after the deliverable, the week, and the research above: it's not the LLM that makes you dumber. It's no longer thinking. This piece started as a note for my French newsletter, La Zone Sapiens, before I expanded it here.

Key Takeaways

• Cognitive debt is the gap between output that looks fine because AI produced it and understanding that never actually formed in your head.
• A PNAS randomized trial found the real variable isn't whether AI is used but how it's configured: a hints-only tutor scored 127% higher during practice and lost nothing on an unassisted exam, while a standard assistant scored 48% higher during practice and 17% worse on the exam.
• A 2025 study of 666 people found frequent AI use correlated with weaker critical thinking, strongest in the 17-25 age bracket, mediated by cognitive offloading. Correlation, not causation.
• An MIT preprint, not yet peer-reviewed, found 83% of people writing essays with ChatGPT couldn't quote a sentence of their own essay minutes later, versus about 11% without AI.
• Three rules that hold up in practice: think before you prompt, use AI as a coach that challenges you rather than a flatterer that agrees, and design friction on purpose, like re-explaining a concept yourself before trusting that you understood it.

Frequently Asked Questions

What is cognitive debt?

Cognitive debt is the gap between output that looks fine because AI produced it and understanding that never actually formed in your head. You get the deliverable fast, but you haven't practiced the reasoning behind it, so the knowledge isn't really yours, and it isn't available to you next time without the tool.

Does AI use actually cause weaker critical thinking?

The strongest single dataset on this, a 2025 study of 666 people, found a correlation between frequent AI use and lower critical-thinking scores, strongest in people aged 17 to 25. It's correlational, self-reported, and single-wave, so it shows association, not proof of cause. Treat it as a signal worth acting on, not a verdict.

Is the MIT "Your Brain on ChatGPT" study reliable?

It's a preprint, not yet peer-reviewed, with a small sample of 54 participants, so hold its findings loosely. That said, its headline result, 83% of the ChatGPT group unable to quote a sentence of the essay they'd just written versus about 11% in the other groups, lines up with the PNAS exam results and the Gerlich correlation. Three different studies pointing the same direction is worth attention even before peer review closes the loop.

Can you use AI without losing your thinking skills?

Yes, and the PNAS randomized trial shows how: a tutor configured to give hints instead of full answers produced 127% better practice performance and no measurable loss on an unassisted exam. The variable that matters isn't whether you use AI, it's whether the way you use it still forces you to do the reasoning.

How do I build this friction into my workday?

Run one pomodoro without any AI tool before you start prompting on a new problem, alternate blocks where AI is off-limits with blocks where it's fully available, and use your breaks to reconstruct or recall rather than scroll. Pomodorian's Never Dumb breaks are built around exactly that last piece.