OpenAI’s nearly 400 AI-generated math results may take years to assess

OpenAI released nearly 400 AI-generated mathematical results across more than 700 manuscripts, spanning fields from number theory and geometry to theoretical computer science and mathematical physics. Mathematicians who spoke with The Verge said understanding the collection could take years.

The release included a roughly 40-page table of contents and abstracts. Researchers said even reading that material took much of an hour, and the collection’s size made an initial assessment difficult.

Verification varies across the manuscripts. OpenAI said that 300 top-line results out of 719 manuscripts had been formalized, or around 42 percent, and that many—but not all—manuscripts had formalizations in Lean, a programming language and proof assistant. OpenAI said it plans to add more formalizations as it obtains them.

Researchers cautioned that Lean code does not settle whether a paper’s claims are correct: they must check that the formalized proof matches the result described in the manuscript. Several mathematicians told The Verge that the quality of proofs and their fit with the papers’ claims varied.

Some researchers said their first impressions were better than expected, but noted that this did not necessarily mean the work met academic standards. The report says the large volume and uneven verification make it difficult to judge the collection as a whole.