25 Fields Medalists Challenge AI Math Benchmarks

This is a sharp warning that benchmark wins can become regressions for the discipline they claim to advance.

• –Binary proof outcomes reward answers over explanations, context, and reusable mathematical ideas
• –Rapid AI-generated claims raise serious attribution, plagiarism, provenance, and verification concerns
• –Developers should prioritize transparent workflows, formal proof checking, citation tracking, and human-readable writeups
• –The declaration supports AI as a research copilot, but insists mathematicians—not model benchmarks—set the field’s priorities