OpenAI says an unreleased AI model helped produce more than 700 manuscripts on over 370 mathematical topics, with each manuscript taking an average of three hours. The work, published in October, has prompted debate among mathematicians about how AI-generated research should be evaluated and shared.

The topics include the Riemann hypothesis, a famous unsolved problem connected to prime numbers. OpenAI has not solved it, and earlier work by Anthropic also did not resolve the problem, though its results were described as progress in the field.

The Association for Human Mathematics, founded by mathematicians from around the world, criticized the scale of OpenAI’s release. Publishing more than 700 files at once, it argued, does not demonstrate scientific competence and may instead be a show of power.

A central concern is that mathematical research involves more than producing a correct answer. The German Mathematical Society (DMV) says AI systems tend to prioritize results, while researchers also value the methods and reasoning that lead to them. Demian Nahuel Goos of Humboldt University in Berlin said a proof can offer little insight on its own.

Researchers have also raised questions about verification: who can check such a large volume of work, how the results were reached, and whether they rely on false assumptions. The DMV warned that AI systems may draw on existing material that specialists have not yet reviewed, and said that using other people’s work without appropriate acknowledgment would violate scientific norms. In a separate dispute over a mathematical solution, two mathematicians have said they were overlooked.

The debate has also raised concerns about the effects on open exchange and early-career researchers. DMV president Jürg Kramer said some researchers had stopped putting their findings online, which he warned could slow progress and restrict discussion within the mathematical community. Stuttgart mathematics professor Christian Hesse told the German Press Agency that doctoral research may need to change if AI can solve in minutes problems that students have worked on for years. He said future work could place greater emphasis on developing questions, understanding mathematics and critically assessing results.

Some mathematicians have welcomed the advances. Hesse called OpenAI’s work in mathematics sensational, while stressing that its proofs still need checks for correctness and originality. He argued that AI’s shortcomings do not justify shielding the field from change. Robert Lange, an AI researcher and founding member of Japan-based Sakana AI, noted that the debate concerns pure science, where the path to making findings a public resource remains relatively long.

The DMV suggested that academic publishers could have a stronger role by publishing work that has passed peer review. Hesse objected to relying on paywalls, saying science depends on exchange rather than exclusion. OpenAI said it wants to expand the boundaries of human knowledge and evaluate its internal models in mathematics and other sciences to develop tools that advance those fields.

For Hesse, the deeper challenge is deciding what mathematical understanding means: a correct proof is not necessarily the same as understanding why a statement holds, what connections it reveals or what questions it raises. He believes AI will not eliminate mathematics, but may take over some of the work and challenge assumptions about how indispensable human mathematicians are.

Sources: t3n.