Research group Epoch AI analyzed math preprints on arXiv, finding that the share of papers acknowledging AI use increased from 4% in April 2026 to 25% by August 2026. The study utilized a multi-step pipeline involving the GPT-5.6 Sol model for categorization and Claude Fable 5 for validation.
The analysis revealed significant variance across subfields. In differential geometry, the proportion of papers crediting AI rose from approximately 8% to 57%. Across all analyzed preprints, 6% of authors in August 2026 attributed a substantive, co-author-level contribution to AI.
Epoch AI reported that human reviewers agreed with the automated classification results 93% of the time. The organization noted that disclosure rates remain consistent among both established authors and new researchers, regardless of their publication history since 2023.
The underlying dataset, which includes metadata for math preprints dating back to 1989, has been made public under a Creative Commons BY license. Epoch AI suggests that the findings reflect an evolving shift in academic norms regarding transparency in research workflows.
