OpenAI’s 722-paper maths dump has experts racing to work out what is real
OpenAI released 722 maths papers at once. Some were retracted or corrected almost immediately, and experts are now checking the biggest claims.
OpenAI has released 722 mathematics papers in a single batch, covering 372 open problems across subjects including algebra, geometry and theoretical computer science. The company says the work, produced largely by an unreleased AI model, is meant to help mathematics progress, but the rollout has already sparked confusion, excitement and criticism.
The release landed earlier this week and has already sent shockwaves through the maths community. Its sheer scale has been compared to everything from a giant data drop to a carpet bombing of the field, with researchers suddenly faced with hundreds of fresh claims to sort through.
That task has not been made easier by the way some of the papers are written. One mathematician who tried to read a paper tied to a favourite problem said it was so difficult to follow that, in a normal journal process, it would have been rejected immediately. Even ChatGPT, when asked about at least one of the headline-grabbing results, reportedly flagged it as something that should not be treated as a proof without expert checking.
The biggest change is simple: a vast amount of unvetted mathematical work is now out in the open at once, and mathematicians are the people most directly affected because they are the ones who now have to test whether any of it stands up.
Some of the claims are enormous
Among the papers are claimed advances on famous unsolved problems including the Riemann hypothesis and the Birch and Swinnerton-Dyer conjecture, both linked to US$1 million rewards as Millennium Prize Problems.
OpenAI says the aim is to enable further progress in mathematics. But the release has also revived questions after the company’s September publication of a claimed solution to a case of the Navier-Stokes problem. Mathematicians Tristan Buckmaster and Levent Alpöge alleged OpenAI had accessed their data and used their ideas, while also spending about US$15 million in computing power to reach a solution. OpenAI denied those allegations.
The wider backdrop is the race around artificial general intelligence. Research-level performance in pure mathematics could strengthen the case that AI systems are moving closer to human-level reasoning across difficult cognitive tasks. That matters commercially too, with OpenAI reportedly targeting a stock market launch and a valuation of up to US$1.4 trillion.
Now comes the slow part
In mathematics, new results are normally checked by specialist referees before they become part of the literature, and that process can take years. OpenAI’s papers are now entering that stage in public.
Three papers have already been retracted because of an elementary error, and several others were amended after mistakes were found that invalidated their results. OpenAI says 300 of the main results have been formalised using machine-checkable methods, part of a broader push that uses software such as Lean to verify logic step by step.
Even that is not a guarantee. AI-generated formalisations have had problems before, including in work connected to Navier-Stokes. For now, the release has left mathematicians in two camps at once, excited by the possibility of major breakthroughs and worried about what this volume and speed could mean for how mathematical research is judged in future.
