An AI audit of 50,000 astrophysics papers found ~1,000 major flaws
A researcher says he used Opus 5.5 to index about 150,000 astrophysics papers, then audited 50,000 of them. The audit turned up nearly 1,000 major issues. These are his figures, and the post does not say how severe the problems were, how they were checked, or how many hold up. Even so, it is a striking test of a plausible job for machine readers: going through a field's literature at a scale no human referee could manage and flagging what looks wrong. The hard part, for science, will be the human follow-up.