OpenAI posts 722 AI-written maths papers, withdraws three after a sign error

OpenAI, the maker of ChatGPT, has published 722 mathematics manuscripts written by an artificial intelligence model it has not released to the public, setting off a sharp debate among mathematicians. The papers were posted in a public repository on the code-hosting site GitHub on Tuesday, 6 October 2026, and cover purported progress on 372 open problems in areas including geometry, algebra and computer science. OpenAI said the results came from an unreleased model more capable than its commercial ones. A day later the company withdrew three of the manuscripts after a sign error was found that broke the argument in one paper and in two others that depended on it.

Mathematics is one of the few fields where an AI claim can, in principle, be checked line by line. A proof is either correct or it is not. That makes this release an unusually public test of what today's most advanced AI systems can do in original research, rather than in exams or benchmarks built from known answers.

The scale is the problem as much as the promise. A single new proof in a specialised area can take experts weeks or months to referee. Releasing hundreds at once, written by a system no outside researcher can run, places a heavy checking burden on a small community of specialists. The quick withdrawal of three papers after one error showed both sides of this: errors were found fast, but they were there. Some researchers praised specific results as genuinely impressive; others questioned whether OpenAI could have checked so many results itself before publishing, and argued that the release skipped the norms of academic publishing, where work is reviewed before it is announced.

There is also a question of credit and process. Traditional mathematics relies on named authors who stand behind their arguments and answer questions about them. When the author is a model that cannot be inspected, the question of who is responsible for an error, and who gets credit for a correct idea, becomes harder to settle.

For India, which has a strong tradition in number theory and a large community of research mathematicians at institutes such as the Indian Statistical Institute, the Tata Institute of Fundamental Research and the IITs, the episode is directly relevant. Indian researchers may be among those asked to verify, extend or challenge these results, and Indian universities will face the same questions about how to treat AI-produced work in theses and journals.

Short-term: Specialists will spend the coming weeks checking individual results, and more corrections or withdrawals are possible. Debate over how AI labs should release research claims is likely to intensify.

Long-term: If a meaningful share of the results holds up, AI could become a regular source of new mathematics, changing how proofs are produced, checked and credited. Journals and funders may set rules for AI-generated submissions.

Who is affected: Research mathematicians and journal editors, AI companies making research claims, students considering mathematics careers, and universities in India and abroad setting rules on AI in research.

Questions and Answers