Twenty-five Fields Medal-winning mathematicians have signed an open letter warning that competition among AI labs to solve famous problems could damage the norms that make mathematical knowledge usable: careful verification, clear attribution and open exchange.
The immediate flashpoint is OpenAI’s claimed solution to the Navier–Stokes problem, which remains unverified, alongside allegations from NYU professor Tristan Buckmaster that OpenAI pressured him not to credit a collaborator employed by Anthropic. Buckmaster also questioned whether work produced through Codex may have helped OpenAI generate its result. Those are allegations, not established findings.
OpenAI also withdrew sponsorship from a Caltech mathematics event after criticism from researchers there, according to a post cited by TechCrunch. Taken together, the events show a widening gap between the operating tempo of frontier AI labs and the validation practices of an academic field where priority and proof are inseparable.
The issue is more than who gets credit
The signatories’ central argument is that a mathematical result is not complete simply because a system produces a plausible proof. It must be written up, checked, connected to previous work and taught to other researchers. That process identifies which techniques are genuinely new, exposes errors and allows ideas to enter the broader mathematical canon.
A rush to announce results can undermine every part of that chain. The letter warns that insufficient time for documentation and citations creates attribution and plagiarism risks. It also argues that AI-generated ideas depend on mathematicians to develop and integrate them before they can become durable knowledge.
For operators building AI systems for research, this distinction matters. Benchmark performance or a headline result may demonstrate capability, but adoption in high-stakes technical domains depends on provenance, reproducibility and independent review. A breakthrough that cannot be audited may carry limited practical value—and substantial reputational risk.
Incentives could push research behind closed doors
The mathematicians also identify a more structural concern: if labs can spend tens of millions of dollars on inference to pursue a promising line of inquiry, individual researchers may stop sharing early ideas or using AI tools openly.
That would invert a long-standing research model. Instead of circulating partial results and collaborating toward a proof, researchers may treat prompts, model outputs and unpublished work as proprietary defensive assets. The result could be less openness precisely when human review is most needed.
The concern extends beyond mathematics. Similar questions are emerging across software, science, design and other knowledge work: when an AI vendor can observe user interactions, train on permitted data, or rapidly productize a workflow, what protections assure users that their contributions will not be used to compete with them?
A governance problem for AI companies
The June Leiden Declaration had already outlined recommendations for mathematicians, institutions and policymakers confronting AI-assisted proofs. The new letter turns the issue into a direct challenge for AI labs.
Companies working in research should expect demands for clearer boundaries around user data, better disclosure of model involvement, audit trails for outputs, and independent processes for validating prominent claims. They will also need publication and attribution policies that match the norms of the disciplines they seek to transform.
What to watch next
The near-term question is whether OpenAI’s Navier–Stokes work receives independent verification and a complete scholarly treatment. More broadly, watch whether major labs establish credible provenance standards before releasing research claims, and whether universities begin requiring them for collaborations, sponsorships or access to research communities.
The dispute is an early test of a wider proposition: AI can accelerate discovery, but its commercial deployment must preserve the human institutions that determine which discoveries are trustworthy and how they are shared.
