The Pulse
Mathematicians debate proof quality after AI “mathocalypse”
At an MIT panel, mathematicians questioned how AI-generated proofs should be assessed and what researchers should be rewarded for. The discussion followed OpenAI’s release of 722 manuscripts claiming AI-derived solutions to unsolved problems.

Mathematicians at an MIT panel questioned how to judge AI-generated proofs and the people who produce them, following a wave of claimed solutions to open problems. The discussion followed OpenAI’s release of 722 manuscripts claiming AI-derived solutions, an event widely dubbed the “mathocalypse,” MIT News reported.
The panel brought together Jacopo Borga of MIT, Michael Hopkins of Harvard University, Bryna Kra of Northwestern University and Tristan Buckmaster of New York University. Science writer Kevin Hartnett moderated the discussion, which covered proof quality, research credit and the future of mathematical careers.
A contested proof
In September, MIT News reported, OpenAI focused thousands of AI agents on a mathematical challenge and, after 88 hours, the agents produced a proof related to the Navier-Stokes equation. The equation is associated with one of the seven Millennium Prize Problems.
The company’s work coincided with research by Buckmaster and a collaborator at Anthropic, who were close to solving the problem after a year of progress, according to the report. MIT News reported that the race raised questions about credit, whether mathematicians’ data had leaked, how the agents reached their result and whether it was meaningful.
The manuscripts released a month later claimed to use AI entirely to solve many major unsolved mathematical problems, the outlet reported. Mathematicians examining AI proofs have questioned whether they can understand how the systems arrived at their conclusions.
At the panel, Hopkins called the papers “nonsense” and said AI-generated proofs contain “strange ideas.” Buckmaster argued that mathematical work needs human understanding to be important.
Rethinking priority
Kra urged mathematicians to reconsider the value placed on being first to prove an idea. “Speed should not be a priority,” she said, arguing that the field needed other ways to assess understanding.
Hopkins countered that people would still want to be pioneers even if the profession discarded priority as a standard. Kra suggested that pioneering work could take different forms, including explaining how a proof came about.
Kra helped draft the Leiden Declaration on Artificial Intelligence and Mathematics, a statement signed by mathematicians around the world that calls for action on risks AI poses to mathematical research. She said communicating mathematics had become more important.
Journals and careers
Kra said interviews were becoming more important when mathematicians are considered for positions or promotion. In her view, candidates need to demonstrate that they understand their ideas and can explain them, rather than relying on what they have written or published.
Borga questioned the reliance on journals to judge mathematicians and called for a new measure that could evolve. Academic journals have long been used to evaluate and validate mathematical advances, MIT News reported.
As a managing editor at a specialty journal, Kra has seen a large increase in submissions, mainly reporting AI-generated or AI-assisted results; the journal hired an additional managing editor to handle them. She said some publications were placing greater value on exposition, the part of a paper that explains the reasoning behind a conjecture or proof.
Kra also pointed to experimental journals that highlight human contributions, including explanations of older results for modern mathematicians. She mentioned Mathematical Discourse, a peer-reviewed video journal that will feature talks about mathematical research.
Students face conflicting messages about how they should learn mathematics and can feel overwhelmed, Borga said. He argued that they need space to learn the fundamentals.
Sources
- Original source · Press release MIT News news.mit.edu
Related stories
- Vals AI says GPT-6.2 Astra produced Lean proof of algebra gap theorem
- Anthropic mathematician claims Claude helped prove Riemann hypothesis
- Amazon releases tools to check AI-generated mathematical proofs
- OpenAI says its models improved AMD GPU designs by up to 14%
- OpenAI says agents completed 46% of delegated two-to-four-hour tasks