OpenAI has released 719 AI-generated mathematical proofs covering 372 result families and numerous open research problems. The initial release on October 6 contained 722 manuscripts, but three were withdrawn on October 7 due to a symbol error.
The release has faced criticism for failing to meet standards set by the Advisory Group for Mathematics and Artificial Intelligence (AGMAI). OpenAI utilized proprietary models for these proofs, whereas the advisory group recommended using accessible models and disclosing specific data such as prompts, reasoning chains, and computational costs.
Only 10 of the published manuscripts include model reasoning chains. Furthermore, approximately 42% of the proofs lack formalization and machine-readable metadata, which the advisory group argues hinders the ability of mathematicians to review and learn from the AI output.
Mathematician Terence Tao has criticized the approach, describing it as a form of mathematical proof indigestion. He argues that the focus on speed and volume in solving famous problems undermines the traditional mathematical process of teaching, collaboration, and deep understanding.
Tao expressed concern that treating open problems as resources to be harvested by AI prevents the community from exploring different paths to solutions. He noted that once a problem is publicly solved, the exploration process is effectively contaminated, making it difficult to derive new insights.
