A field-wide audit uses AI to map theoretical astrophysics and test whether its literature corrects itself. Starting with 4,093 papers from Caltech’s TAPIR group, the project assembled a citation network of 158,059 papers and 554,670 citations, then screened 50,269 accessible full texts for signs of problems. It found little evidence of fraud, but examined influential work more deeply: AI reviewers re-derived equations and checked calculations in 659 papers, with a second reviewer required to confirm an error. The project argues that ordinary literature review cannot keep pace with a field where researchers cite only a small fraction of relevant work and known corrections often remain buried or unpublished.
The audit verified 872 errors in 433 papers; 159 errors were copied into later work, while only 47 were ever criticized in print. Among the 56 errors found in the 200 most-cited TAPIR papers, 52 remain in the latest preprint versions. Peer review removed 13 errors before publication but let 56 pass, and copied errors typically persisted for years. The estimate of roughly 34,000 errors in highly cited papers is an extrapolation, not a count: detailed checks covered only about 4% of such papers, and the project acknowledges limits in access and method. Its broader claim is that science more often leaves errors behind as a field moves on than actively corrects them, while AI makes systematic checking comparatively inexpensive.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.