"AI is making people lazy" is a comfortable story. It fits what we already suspect and it doesn't require us to look very closely at what's actually happening in the brain. A 2025 study out of the MIT Media Lab looked closely — with EEG electrodes, not vibes — and the picture it found is more specific, and more useful, than "lazy."

The setup

Researchers led by Nataliya Kosmyna and Pattie Maes split just over fifty participants into three groups and asked each to write SAT-style essays under one of three conditions: using only their own mind, using a search engine, or using ChatGPT. Everyone wore EEG caps so the team could watch brain connectivity — how strongly and how broadly different brain regions were talking to each other — while the writing happened, across several sessions.

This wasn't a one-off. Participants came back multiple times, and in a final session the researchers swapped conditions: essay-only writers got access to ChatGPT for the first time, and habitual ChatGPT users had it taken away.

What the EEG showed

The headline finding: the more of the cognitive work was handed to the LLM, the weaker and narrower the brain's connectivity patterns were during writing. The brain-only group showed the strongest, most widely distributed coupling between regions. The ChatGPT group showed the least — consistently, across sessions, not just once.

That's a description of engagement, not intelligence or effort. Nothing here says the ChatGPT group's brains were "damaged" or that anyone was being lazy in the sense of not trying. It says that when a tool absorbs a chunk of a cognitive task, the brain measurably does less of that task — which is, on its own, not surprising. A calculator does the same thing to arithmetic. What's notable is what came with it.

The part that isn't "laziness"

Two other results round out the picture, and they're the ones worth sitting with:

  • Ownership dropped. Shortly after finishing their essays, ChatGPT-group participants were much more likely to struggle to quote even a single sentence they had supposedly just written. The words had passed through them without quite becoming theirs.
  • The essays converged. Independent evaluators (and an AI judge) rated the ChatGPT group's essays as more homogeneous — more similar to one another — than essays from the other two groups. Outsourcing the drafting didn't just change effort; it narrowed the range of what got said.

The researchers frame this with a term they call cognitive debt: doing less of the underlying cognitive work now doesn't cost you anything in the moment, but it can leave you less equipped to do that work unassisted later — a debt that comes due when the tool isn't available. In the session where regular ChatGPT users had to write without it, their brain connectivity and recall lagged behind the group that had been writing unassisted all along.

Why this isn't new, either

If "the mind adapts to what you stop practicing" sounds familiar, it should. A 2020 study in Scientific Reports by Dahmani and Bohbot followed habitual GPS users over time and found that leaning on turn-by-turn navigation was associated with worse performance on tasks that depend on the hippocampus — the brain's own spatial mapping system. Same underlying mechanism, different tool: when the environment does the remembering, the brain does less of it, and the skill doesn't stay sharp on standby.

What this does and doesn't tell us

A few caveats the video (and this study) take seriously, because overclaiming here would be its own kind of laziness:

  • This is a single preprint, not yet through peer review, with a modest sample doing one narrow task — SAT-style essay writing — over a handful of sessions. It is evidence, not a verdict on "AI and the brain" as a whole.
  • Weaker EEG connectivity during a task is a measure of engagement, not proof of lasting harm. Whether "cognitive debt" compounds over months of real-world AI use, or whether it's fully recoverable once someone re-engages the skill, is an open question the authors themselves flag for follow-up.
  • None of this is an argument against using AI tools. It's an argument for noticing which parts of a task you're handing over, and whether that's a trade you'd make on purpose.

The useful takeaway isn't "AI bad." It's that cognitive offloading has a cost structure, and that cost structure is now something researchers can actually see on an EEG readout instead of just theorize about.

Watch the full breakdown — including how the study measured brain connectivity and what the fourth-session swap revealed — in the film below.