You spend twenty minutes working with ChatGPT on a short essay. It reads clean, the argument holds together, and you hit submit feeling good about it. Then someone hands it back and asks you to recite the third sentence from memory. In a 2025 study out of the MIT Media Lab, most people in exactly that position couldn't do it — 83% of participants who had just finished an AI-assisted essay were unable to accurately quote a single line of what they'd supposedly just written.

That isn't really a story about AI being unreliable. It's a story about what happens to your own mind while something else is doing the drafting — and it's the reason a fairly obscure idea from cognitive psychology, cognitive offloading, has suddenly become one of the more urgent questions in how people think.

Offloading your thinking is old news — this part isn't

Handing mental work to something outside your own head is not a new behavior, and it isn't inherently a bad one. Psychologists Evan Risko and Sam Gilbert described this formally in a widely cited 2016 paper in Trends in Cognitive Sciences: people offload cognition — onto notebooks, calculators, contact lists, GPS — whenever the expected benefit of using an external aid outweighs the mental effort of doing the task unassisted. Reaching for a calculator instead of doing long division in your head isn't a character flaw. It's usually just a sound trade of a limited resource, freeing up attention for whatever actually needs it.

What's different about offloading to a conversational AI is the shape of what gets handed over. An address book stores a phone number; it doesn't decide who you should call. A calculator computes exactly the operation you give it; it doesn't decide what the answer means. A large language model can generate the idea, the argument, the phrasing, and the conclusion — all before you've formed an opinion of your own — and it does this conversationally, in a register close enough to your own inner voice that the line between "my thought" and "the tool's output" gets genuinely hard to locate.

What the MIT brain-scan study actually measured

The 83% figure above comes from a 2025 MIT Media Lab study titled "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," led by researcher Nataliya Kosmyna. Fifty-four participants were split into three groups and asked to write SAT-style essays across multiple sessions: one group used ChatGPT, one used a standard search engine, and one wrote entirely unaided. Everyone wore an EEG cap so researchers could track brain activity — specifically, functional connectivity between brain regions — while they wrote.

The headline neural finding was a gradient, not a cliff: unaided writers showed the strongest, most widely distributed brain connectivity during the task; search-engine users showed moderate engagement; ChatGPT users showed the weakest. That doesn't mean using ChatGPT "shrinks your brain" — the study measured engagement during one specific task under one specific condition, not permanent structural change. What it captured is that the less of the cognitive work you're doing yourself, the less of your brain's networks light up doing it, which is a fairly unsurprising finding stated in a very measurable way.

The problem was never that the AI could write the essay. It was that outsourcing the thinking made the writing forgettable — even to the person who turned it in.

The behavioral results tracked the neural ones. Essays from the ChatGPT group converged on similar phrasing and ideas across different participants; two English teachers reviewing the set independently described them as strikingly uniform and, in their word, "soulless." Self-reported ownership — how much participants felt the essay was genuinely theirs — was lowest in the ChatGPT group and highest among unaided writers. In a fourth session, a subset of participants switched conditions, and the pattern moved with them: writers who'd relied on ChatGPT and then wrote unaided still showed under-engaged brain activity, as if the effect didn't reverse instantly; writers who switched from unaided to ChatGPT-assisted showed renewed activation in memory- and reasoning-related regions, closer to what the search-engine group had shown all along.

What skill atrophy actually looks like, day to day

"Atrophy" sounds dramatic, like something wasting away. In practice, it looks a lot more mundane — a handful of small, repeatable habits rather than one dramatic loss of ability:

• Reaching for AI before you've spent even a minute attempting the task yourself
• Weaker recall of material you technically produced, because you never fully processed it in the first place
• Passing along an AI's answer, summary, or draft without checking whether it's actually correct
• A noticeable drop in patience or performance the moment the tool isn't available
• The habit spreading past low-stakes tasks — emails, code, trivia — into higher-stakes ones: how you interpret a friend's text, whether to take a symptom seriously, how to read a conflict with a partner

That last pattern — creep from low-stakes to high-stakes decisions — is close to what a 2025 study by researcher Michael Gerlich, published in the journal Societies, actually measured. Gerlich surveyed 666 people and conducted in-depth interviews with 50 of them, looking at the relationship between how often someone uses AI tools and how they perform on measures of critical thinking. He found a negative correlation between frequent AI use and critical-thinking performance, and that the relationship ran largely through cognitive offloading — the more people habitually let AI do the reasoning, the weaker their independent critical thinking tended to be. The effect was strongest among 17-to-25-year-olds and smaller among people with more formal education.

It's worth being precise about what that study can and can't claim. It's correlational — people who already prefer to offload mental effort may simply be more drawn to heavy AI use in the first place, rather than AI use causing the decline outright. Gerlich frames his own conclusion carefully: AI tools aren't inherently harmful to thinking, and the outcome depends heavily on how — not just how much — they get used.

Where the line actually is

None of this adds up to a case against using AI. The same underlying logic that makes cognitive offloading potentially risky is also what makes it valuable: freeing up limited mental effort for what actually deserves it. A calculator doesn't erode your math skills if you already understand the math. A GPS doesn't erode your sense of direction if you occasionally navigate without it. The MIT study's own crossover data suggests these patterns aren't fixed, either — engagement shifted again within the same handful of sessions once the tool use changed.

The more useful question isn't "do you use AI," since nearly everyone reading this does. It's where, specifically, the handoff happens. Do you attempt a problem before asking, or ask first and see if you'd have gotten there yourself? Do you check what the tool gives you against your own judgment, or wave it through? If the tool vanished tomorrow, could you still do the version of this task you were doing five years ago — slower, maybe, but still do it?

None of that shows up in a single afternoon. It shows up as a pattern, across the small, low-stakes decisions of an ordinary week — and it tends to be more visible from the outside than it is from the inside.