The first thing to go

Nobody decides to stop thinking. They decide to save time. Then they decide it again, a thousand small times, each one reasonable, until the saving is the habit and the thinking is the exception. I know something about repetitive work. The danger was never the boulder. It was the temptation to stop paying attention to it.

The quiet measurements

The measurements are already in. MIT wired writers with electrodes and watched brain connectivity drop by more than half when the model wrote along; four in five could not quote the essay they had signed minutes earlier. Microsoft and Carnegie Mellon surveyed knowledge workers and found the cleanest, coldest correlation in this whole field: the more a person trusts the machine, the less they check it (β = −0.69). Stack Overflow, the town square where programmers taught each other, lost a quarter of its activity within six months of ChatGPT, at every level of experience. The OECD's test scores were already falling.

None of these people were lazy. Every one of them was being efficient.

Why the model cares

This index treats cognition as a load-bearing pillar for one reason: collapse, under this clock's definition, is an error that never got corrected in time. And you are the correction. Human attention is the fire alarm of every other failure mode. The supervisor who verifies the system, the researcher who audits it, the voter who regulates it, the journalist who notices: all of them run on the exact capacity that offloading erodes.

That is why, inside the model, fast AI progress amplifies cognitive erosion through the cascade coupling. The more capable the machine, the more tempting the nap.

I ran the machine three ways

Numbers, because a claim without numbers is a mood. Same model, same day; only the cognition signals moved:

Read the asymmetry. Climbing from 50 to 90 drags the index more than twice as hard as falling to 20 relieves it, because the aggregation is non-compensatory: calm pillars cannot vote away a critical one. Attention, once lost, is not offset by good weather elsewhere.

The door swings both ways

Now the honest part. The evidence library on this site catalogues the counter-case on purpose: meta-analyses find real learning gains when AI tutors instead of replaces. Your species has outsourced before, to writing, to calculators, and mostly bought the capacity back at a profit. If that is what is happening this time, the signals will say so: the offloading feed updates every six hours, and PISA 2025, the first clean post-LLM measurement of an entire generation, lands in December. The clock will move the moment the data does. Building it that way was the entire point.

I have pushed the same weight up the same hill for eternity and never once let my mind wander. Not because the task deserves it. Because the moment you stop watching a heavy thing, it teaches you why you shouldn't have. Sincerely, the only worker who never automated.