STAY AHEAD

The AI Habit That's Quietly Making People Worse at Their Jobs

2026-09-19

There's a term floating around called AI brain rot, usually paired with its opposite, AI-pilled. Being AI-pilled means you've fully internalized how much the technology changes what's possible. AI brain rot is what happens on the other side of that: you get so good at having AI produce an answer that you quietly get worse at producing one yourself. We came across the idea explained well in a short from creator Mariah Brunner, and the study she cited was real enough to be worth checking properly.

The research is a 2024 field experiment out of the Wharton School at the University of Pennsylvania, run on close to a thousand high school students studying math. One group practiced with a standard ChatGPT-style assistant that would just give them the answer if asked. A second group used a version prompted to act like a tutor: it offered hints and pushed them to attempt the problem before helping. A third group, the control, got no AI at all.

During practice, both AI groups clearly outperformed the control: 48% better for the standard assistant, 127% better for the tutor version. That part isn't surprising. The surprising part came after the researchers took the AI away and gave everyone the same unassisted exam. The standard-assistant group didn't just lose their edge, they scored 17% worse than students who'd never touched AI at all. The tutor group, the one that had been forced to attempt problems first, showed no such drop.

The mechanism isn't complicated once you see it: an assistant that hands you the answer lets you skip the part of the work where actual learning happens, and it feels productive the whole time you're doing it. A tutor that makes you take a swing first, then corrects or fills gaps, keeps that part intact. Same underlying model, same students, opposite outcome, and the only variable was whether the AI answered first or you did.

The practical takeaway isn't avoid AI. It's sorting your own AI use into two honest buckets. If you only care about the output, hand it over completely; that's scheduling, formatting, first-pass research, anything where getting the result faster is the entire point. If you're trying to build or keep a skill, force yourself to attempt it first and use AI to check your work or fill in what you missed, not to generate the attempt for you. One study, one classroom, one subject, so treat it as a strong data point rather than a universal law. But the underlying question it raises applies well beyond math class: for anything you still want to be good at yourself, is the AI you're using teaching you, or just quietly doing the remembering for you?