ICLR reviewer: Claude helped spot likely fraudulent numbers in a paper's results table
jacobcares · x · 2026-11-07
In a discussion on whether models should refuse to fully automate tasks needing human judgment like peer review, an ICLR reviewer shares a first-hand case: as the sole reviewer of a paper, he only caught likely fraudulent numbers in the main results table because Claude recognized damning patterns. The thread weighs a middle ground between full refusal and full automation, where AI outputs encourage rather than replace human thinking.