In 2025, a team at MIT's Media Lab ran an experiment most developers have felt the shape of without ever seeing the data: they hooked people up to EEG headsets and watched what happens to the brain when you let a chatbot do your thinking for you.
The paper is called "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." Researchers split 54 university students into three groups: one used ChatGPT exclusively, one used only Google search with no AI, and one wrote unaided, relying on their own thinking. Everyone wore EEG sensors while they worked.
"AI brain rot" vs. what the study actually says
You've probably seen the phrase "AI brain rot" somewhere — it's the term headlines reached for once this study went viral. Worth being precise about where it comes from: the researchers themselves never call it that. Their term is "cognitive debt." "Brain rot" is the media's shorthand for it, and it's punchier but also more alarmist than what's actually in the paper. Useful to know the difference before you repeat either one as if it were a diagnosis.
What "cognitive debt" actually means
The term comes directly from the study — when people fail to critically engage with a topic, their writing tends to get shallower and more generic, a pattern the researchers describe as cognitive debt: the accumulated cost of letting external systems repeatedly stand in for the effortful mental work that independent thinking requires. It's not a diagnosis. Nobody's brain is rotting in a literal sense. It's closer to what happens to a muscle you stop using — the capacity doesn't vanish, it just stops being exercised, and it shows up on a scan before it shows up in how you feel.
What the EEG actually showed
Brain network connectivity in the ChatGPT group came in as much as 55% lower than the unaided group, concentrated in the frontal and parietal regions — the areas tied to executive reasoning and contextual memory. The Google-search group landed in between: more engaged than the ChatGPT group, less than the unaided one.
Then came the part that should sound familiar to anyone who's shipped a PR they didn't fully read: 83% of the ChatGPT group couldn't accurately quote from the essay they'd just written. Not summarize badly — just couldn't produce their own sentences back. Recall of their own text after 24 hours sat around 17%, against roughly 46% for the group that wrote unaided.
The part that should worry you more than the low numbers: when the ChatGPT group was later asked to write without any AI assistance, their neural engagement stayed low — as if the brain had adapted to outsourcing the effort and didn't just snap back once the tool was gone.
The honest caveats
This isn't peer-reviewed. It's a preprint, based on 54 participants, with only 18 completing the full follow-up session, and a formal commentary from psychology researchers at the University of Vienna and TU Dresden raised concerns about the sample size, reproducibility, and how conservatively some of the results should really be interpreted. One critique worth taking seriously: the "cognitive debt" pattern in the final session might partly reflect a familiarization effect, since the brain-only group had already practiced the task three times unaided before the final round, while the AI group hadn't built that same repetition.
None of that makes the finding nothing. It makes it what it is: an early, provocative signal from a serious lab, not settled neuroscience. The study also only measured essay writing over a few months — not coding, and not years. Treat the specific numbers as directional, not gospel.
Why this matters even though it's about essays, not code
The mechanism the study points at isn't really about writing. It's about what happens when the effortful part of a task — the part where you actually reason, structure, and remember — gets handed to something else on a loop. Writing an essay and shipping a feature both run on the same muscle: holding a problem in your head long enough to actually work through it, not just approve someone else's answer to it.
That's the whole premise behind the test we built. Not "AI is bad" — the study's authors aren't saying that either. It's that the gap between using a tool and being quietly replaced by it is smaller than it feels, and you generally can't see it from the inside. You need something external to check.
This is a summary of ongoing, contested research — not medical or clinical advice. If you're noticing real difficulty concentrating, remembering, or thinking clearly day to day, that's worth raising with a doctor, not just a coding quiz.