68% of audited ML conference submissions contained fabricated citations, authors, or LLM-generated text

A July 30 audit of 22 submissions to NeurIPS, WACV, and TerraBytes found 68% contained fabricated citations, fabricated author lists, or were clearly LLM-generated. Of 11 papers checked for citations, 55% had fabricated ones; 82% of 11 checked for fabricated authors or LLM writing were flagged. The sample was small and non-random. The author released a bib-audit tool on GitHub.