AI can do far more than simple counterexamples, but humans cannot sift the errors cheaply

Afinetheorem · x · 2026-07-21

AI can do more than simple counterexamples, but humans cannot sift the failures cheaply

The author argues that AI’s strength at finding trivial counterexamples is not because it can only do simple things. Rather, it can attempt far more ambitious tasks, but roughly 99 out of 100 attempts contain errors, and it becomes too expensive for humans to inspect all of them.

The post says this is also the mechanism behind the authors’ paper on "optimal AI": the system becomes valuable when the user already has enough knowledge and taste to know what to attempt, what to verify, and how to build on the output.

So the practical bottleneck is not raw capability alone, but the human cost of filtering imperfect attempts.

Related event: AI in Research: High Verification Costs and the Expert Sweet Spot(8 posts)→

Original post →

More from Research

Research channel →