‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop

OpenAI released nearly 400 AI-generated mathematical results this week, spread across more than 700 manuscripts in fields from number theory and topology to mathematical physics, The Verge reports. More than three dozen mathematicians described the drop with words like overwhelming, surreal and unprecedented.

The work is not uniformly verified. OpenAI said 300 top-line results out of 719 manuscripts had been formalized in the Lean proof assistant, around 42 percent. By October 8 the public record already showed revisions to more than a dozen manuscripts and three papers removed over a sign error.

And yet several researchers said some results would have been career-making in a pre-AI world, with Stanford's Jared Duker Lichtman pointing to progress toward the Riemann hypothesis, a special case of the Hodge conjecture, and a solution to the four-dimensional Kakeya conjecture. If even part of this holds up, the speed of mathematical discovery just changed.

We think two things are true at once, and the story only makes sense if you hold both. This may be one of the most significant scientific outputs any AI system has produced. It was also released in a way that hands the hardest, least rewarded part of the work to a community that did not ask for it.

Start with the upside, because it is real. Mathematicians who use AI and are optimistic about it told The Verge the release contains genuinely impressive work, including attacks on problems people have tried for decades. Lichtman's point that the results come from existing techniques used in ingenious ways is encouraging, not deflating. It suggests there is far more reachable territory than experts thought, and that human mathematicians who explain, verify and connect these results will have plenty to do.

Now the cost. Verification at this scale is not instant. Even where Lean code exists, researchers have to confirm the formal statement matches what the paper claims, and several said that mapping was inconsistent. Kevin Buzzard of Imperial College London summed up the bind: read possibly incorrect work, wait for others to do it, or wait for formalization. Brendan Hassett at Brown said the write-up he knew best made little sense on a quick read. When a lab ships hundreds of papers in one go, peer review becomes unpaid labor at industrial volume.

The human toll is the part we cannot wave away. Colleagues' grant proposals wiped out. Research programs described as obliterated. PhD students and untenured researchers whose dissertation problems may vanish overnight. One mathematician warned of a possible collapse in academic culture and said AI labs seem to be ignoring it. That is not a fringe complaint from AI skeptics. It is coming from people who use the tools.

OpenAI did take some advice from the new Advisory Group on Mathematics and Artificial Intelligence. It disclosed more than before, including that its model attempted more than 4,000 problems and a typical result used about three hours of ChatGPT Pro thinking compute, promised to preserve the release history, and said it would fund workshops. But it did not name the model, publish the prompts, or list every problem attempted. Those omissions matter, because they are exactly what would let outsiders reproduce and assess the work, and they keep the whole exercise looking, as the advisory group warned, like a marketing vehicle.

Here is where we land. AI doing new mathematics is good news, and we want more of it. But a responsible release at this scale should look like a scientific collaboration, not a product drop: fewer results at a time, full formalization where feasible, disclosed methods, and real funding for the people doing the digesting. If every lab adopts the dump-and-move-on model, mathematics risks becoming a field where humans only grade machine homework. The advisory group made the same point: mathematicians cannot be reduced to people who decode what the labs produce.

Earlier OpenAI math releases drew criticism from researchers for sloppy write-ups and weak attribution, including a dispute over the Navier-Stokes problem. Lean is a programming language and proof assistant that lets mathematical proofs be checked by computer. The Advisory Group on Mathematics and Artificial Intelligence was formed to advise labs on responsible release of AI-generated mathematics.