the cognitive debt we don't talk about

in june of 2025 a small team at the mit³ media lab did something unusually unfashionable. they didn’t benchmark a new model or argue about agi timelines. they took 54 people, wired them up to EEGs, and asked them to write SAT essays. 3 groups: chatgpt-only, google-only, brain-only.

the chatgpt group showed the lowest neural engagement across the board. their writing scored lower on the linguistic and behavioral measures the team tracked. and by the end of the 4-session study, most of them had stopped writing entirely. they were copy-pasting the model’s output and submitting it.

the part of the paper that stuck out is what happened next. weeks later, when the chatgpt users were asked to write without the tool, their brain activity was still sluggish. the researchers used the phrase cognitive debt.

we help enterprise product teams get unstuck and ship. we use ai constantly, drafting, research, reformatting specs, generating variants. we’d be hypocrites to tell you to stop. but the study is making us think harder about how we use it.

there are 2 kinds of ai usage. one lifts your output past what you could do alone. you still do the thinking, the tool sharpens the artifact. the other skips the thinking entirely. it produces an artifact you couldn’t have produced and, more to the point, didn’t. the first compounds. the second is a loan.

cal newport¹, the georgetown computer-science professor who wrote deep work and slow productivity, calls this the augmentation/automation line. he frames writing as cognitive fitness: what running is to your heart, writing is to your mind. you don’t outsource the workout and expect to stay in shape.

clarai sells velocity arbitrage. our whole pitch is shipping in weeks what agencies bill for months, because we use ai well. so this is a strange essay for us to write. there’s no contradiction.

going fast on the work doesn’t require going fast on the thinking. different speeds. the deliverable can run on a tight cycle and the judgment behind it can still come from a slow, well-rested human brain. that’s the only way the deliverable stays good. if the human also speeds up, you get convergence. and microsoft²/cmu measured exactly that earlier in 2025: heavy ai users produce a less diverse set of outputs for the same task. the work flattens.

so we’re trying something this summer. one piece of work a week, no model. strategic memo, discovery question list, retro…doesn’t matter. the rule is that the muscles stay attached to the brain.

small thing. but the mit team’s word, debt, is the one we keep coming back to. you can run a credit card for a while and feel fine. the bill always comes.