Second Mathematician Accuses OpenAI of Using Unpublished Work to Train Its Models

A second mathematician has publicly accused OpenAI of training its models on unpublished research without permission, escalating a fight over data provenance that began only days earlier with a separate researcher's complaint. The new allegation, reported by The Verge, centers on whether OpenAI's math-focused models drew on private work that was never released or licensed. The mathematician described the company's conduct as unethical and dishonest, pointing to what they call a persistent lack of transparency about where training data comes from.

The dispute is not abstract. OpenAI has spent the past year promoting a string of mathematical results produced by its models, including competition-level problem solving and proofs that impressed working researchers. Those claims rest on the assumption that the models reasoned their way to answers rather than memorizing results from papers they had already ingested. If unpublished work was in the training set, the shine on those results dims considerably, because the models may be reciting rather than solving.

The first complaint surfaced just days before the second, suggesting a pattern rather than an isolated grievance. In that earlier case, a mathematician raised similar concerns about whether their unpublished findings had been fed into OpenAI's systems without credit or consent. Neither researcher has produced a full forensic audit, and OpenAI has not released detailed data documentation that would settle the question either way. That absence is itself the story: outside researchers have no reliable way to check the claim.

This is a familiar problem for large language model developers, but it lands harder in mathematics, where the boundary between public and private work is unusually clear. Papers are either posted to arXiv or they are not; conference submissions are confidential; grant proposals are confidential. A model that reproduces an unpublished lemma with suspicious precision is easier to flag than one that mimics a blog post. Mathematicians are also a small, tight-knit community, which means word travels fast and reputational stakes are high.

OpenAI has not issued a substantive response to the second accusation, and it is unclear whether it will publish training-data details or allow an independent review. What happens next likely depends on whether the mathematicians push for formal discovery, institutional complaints, or public documentation requests. The broader question — whether labs can keep claiming breakthrough reasoning when the provenance of their training data remains opaque — now has two named critics instead of one.