One in four math preprints now admits to using AI

Epoch AI went through every mathematics preprint posted to arXiv between January 2025 and August 2026 and counted how many acknowledged using AI. The share climbed fast. In April it was about 4 percent. By August it had reached 25 percent, with the steepest jump between July, at 15 percent, and August. Around 6 percent of August papers went further and said AI had made a substantial contribution to the research itself, not just to writing or editing.

The numbers are voluntary disclosures, so they undercount real use rather than overstate it. To build them, Epoch filtered for papers mentioning AI, then used GPT-5.6 Sol with Claude Fable 5 as a check to classify each acknowledgment; a human reviewer agreed with the labels 93 percent of the time on a sample of 200. The acknowledgments split across writing, code and computation, literature review, formalization, and genuine research help, and Epoch is careful that the line between a tool that assists and a tool that contributes is blurry.

Why it matters

Mathematics has been one of the last research cultures to treat AI as a serious collaborator, so a jump from 4 to 25 percent in four months is a real signal about adoption. If you referee papers or set journal policy, disclosure norms are already lagging behind practice, and the 6 percent claiming substantial help is the number worth watching next.