OpenAI's 722 AI-generated maths papers leave mathematicians with years of checking

Formalisation reduces a proof to steps a computer can confirm. Without it, the confirming falls to people, and some of these papers run to hundreds of pages of dense computation ^[2]. New Scientist estimates that checking the release could take mathematicians years, and that understanding its significance could take far longer ^[4].

Three retractions in 722 papers is about 0.4 per cent ^[15]. That figure reflects what readers caught in the first days, while the batch has not yet been combed through, let alone checked ^[2]. It is an early detection count. New Scientist does not report how many of the papers have been formalised.

The batch mixes claims of very different weight. Papers claiming partial versions of the Birch and Swinnerton-Dyer and Hodge conjectures ^[10] mean that, with the quasi-Riemann paper, the release touches three of the six Millennium problems still unsolved ^[16]. Each carries a $1 million prize from the Clay Mathematics Institute ^[7]. The four-dimensional Kakeya paper builds on earlier work: in 2025 Nets Katz, a mathematician at Rice University, said a paper claiming the three-dimensional case was "perhaps the biggest breakthrough in mathematics of the current century" ^[12]. Further down the list sits a result that seems to rule out five colours for a plane-colouring problem already known to need between five and seven ^[14].

I think the mix matters more than any single claim. A referee has to read a paper before knowing whether it is a step toward a Millennium problem or a speed-up too small to notice, so sorting the batch is itself expert work. The model's output arrived as one release of 722 papers ^[1]. The reading has to be done by specialists, paper by paper.

For an operator, the case carries over with a limit. The evidence comes from one field and one release. Mathematics also has a machine route to checking, formalisation, even where many of these papers have not taken it ^[3]. I would expect much commercial output, such as a drafted contract or a forecast, to lack an equivalent, so its review falls to people. The trade-off is between commissioning output at the speed of a model and funding review at the speed of qualified staff. This quarter's decision is how much review capacity to pay for. Next quarter, that choice shows up either as a backlog of unchecked work or as a rollout held to the pace of review.

The backlog in mathematics may grow. New Scientist calls the assumption that this release is a one-time event far from safe, and says there is no way to predict whether AI mathematics at this pace will help or harm the field until the work is validated ^[5].