Breaking

Mathematicians Publish Leiden Declaration Cautioning AI Use in Math

July 5, 2026 at 18:10 EDT

Policy & Regulation
Research & Papers

On June 2, 2026, a group of 16 mathematicians and researchers, centered in Europe and the United States, published the "Leiden Declaration on Artificial Intelligence and Mathematics" on its official site, cautioning against the careless use of AI in mathematical research. The declaration calls for preserving certainty in proofs, maintaining peer review and safeguarding research autonomy, and the International Mathematical Union (IMU) voiced its support the same day.

June 2, 2026 · Leiden Declaration

Mathematicians Set Rules for AI in Proofs

The "Leiden Declaration on Artificial Intelligence and Mathematics" calls for full disclosure of AI use and human authorial responsibility — endorsed by the International Mathematical Union.

Drafters, across 15 universities

150+

Signatories immediately, later several hundred

IMU

International Mathematical Union endorsement

The trigger: an 80-year-old conjecture

In May 2026, an AI was announced to have disproved Erdős's unit distance problem — unsolved for nearly 80 years. The press release preceded peer review, raising concern.

~80 yrs

unsolved

1 day

AI claim

Speed impressed some — but undisclosed verification and weak citation alarmed others.

Not anti-AI — pro-transparency

The declaration aims to leverage AI while safeguarding mathematics' core values. Its main recommendations:

Full disclosure

Add a "Tool and computational resource disclosure" section to papers.

Human responsibility

Authors remain accountable for results, not the AI tools.

Attribution & openness

Covers LLMs, ML systems, proof assistants like Lean, symbolic & neural methods.

Welcomed as

"Well-thought-out" and "not anti-AI." Backed by Peter Scholze, Terence Tao, Kevin Buzzard and IMU vice president Ulrike Tillmann. Scholze: "mathematics is a human endeavor."

Open concerns

Corporate hype can bypass peer review; research may skew toward AI-friendly problems; verifying "plausible but incorrect" AI proofs is costly.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.