A second mathematician has publicly accused OpenAI of unethical and "dishonest" behavior over the data behind its celebrated mathematics breakthroughs, saying the company has failed to prove that his private interactions with ChatGPT did not contribute to results announced with great fanfare last month.
In a series of posts on Mastodon reported by The Verge on Wednesday, mathematician Andreas Thom raised concerns that conversations he and his colleagues had with ChatGPT before OpenAI's announcement may have fed into the model's success. His intervention came just days after New York University mathematics professor Tristan Buckmaster publicly questioned whether OpenAI's models had benefited from his own use of the company's Codex tool — turning a moment of triumph for the lab into a running dispute over training-data provenance. For more context on this story, see our ongoing AI industry coverage.
From Breakthrough to Backlash
In August, OpenAI announced that its system had produced ten significant results in mathematics and theoretical computer science, each checked with machine-verified Lean proofs — a milestone this site covered at the time, alongside the company's claimed solution to a Millennium Prize problem. One of the ten results established the existence of so-called non-sofic groups, a long-open question in group theory.
That result sits at the center of the new controversy. According to The Verge, OpenAI acknowledged that the proof built heavily on previous work by Thom and fellow mathematician Gábor Kun — and after criticism from mathematical circles for initially failing to acknowledge their recent contributions, the company quietly amended its writeup.
For Thom, the acknowledgment raised an uncomfortable question: how did the model come to command his and Kun's techniques in such detail?
The Question Thom Says Went Unanswered
Thom said he was struck by "OpenAI's detailed command of our techniques," which he noted were neither the most obvious nor the most promising routes to a solution at the time. He wrote emails to OpenAI researchers Sébastien Bubeck and Mark Sellke — the latter also a statistician at Harvard — asking whether his interactions with ChatGPT had been "part of the training data or accessible to the reasoning process" and could therefore have contributed to the result.
The answer he received addressed only whether his conversations could be accessed directly, not whether they had entered the vast pools of data used to train and improve the company's models. "No such qualification, explanation, or evidence was given," Thom wrote. "I take this as dishonesty to say the least."
In a follow-up quoted by The Verge, he sharpened the accusation: "Sellke's categorical answer was, at minimum, unjustifiably broad and materially misleading; looking back it was plainly dishonest."
Thom's argument is less about what OpenAI did than about who bears the burden of proof. Researchers, he said, are not equipped to reverse-engineer the company's training pipeline. "Only OpenAI has the relevant data for that," he wrote — so if the company wants to deny that user data contributed to its results, the responsibility is on OpenAI to prove it by disclosing the relevant datasets and clarifying how it uses customer interactions.
OpenAI's Careful Denials
The dispute echoes the company's handling of its claimed Navier-Stokes breakthrough, work on which Buckmaster had been pursuing with Anthropic researcher Levent Alpöge in a personal capacity, according to The Verge. In the blog post announcing that result, OpenAI stated: "We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem."
But the same post conceded an indirect channel: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
For Thom, that distinction is the whole problem. "De-identification may remove a name; it does not remove the intellectual content of a mathematical idea," he wrote. OpenAI did not immediately respond to The Verge's request for comment.
He added that it "would be ethically indefensible" if nonpublic research supplied by users helped improve models that the company then used to race those very same users to publication — without consent, proper disclosure, or credit.
A Chill Over Mathematics
The circumstances of the Millennium Prize push have done little to calm nerves. OpenAI has said it pursued the problem after hearing rumors online that other researchers had made major progress — effectively racing academics to a result they had spent years approaching.
Numerous researchers told The Verge they worry that such behavior will push mathematics into a more secretive state, if mathematicians come to believe that even rumors of an imminent breakthrough could ignite a race against a well-resourced tech giant eager for glory. The Verge's related coverage — "OpenAI's sly mathematical breakthrough sends a chill through academia" — captured the mood in the field.
The episode lands at a delicate moment for AI companies and the research communities they increasingly brush up against. Whatever the merits of any individual result, the asymmetry Thom describes is structural: labs can ingest user interactions at scale, while the people generating that data have no way to audit what was taken. Until companies either open their training pipelines to scrutiny or impose hard, verifiable firewalls between customer conversations and model training, every "we did not use your data" statement will carry an implicit asterisk — and every breakthrough will come with mathematicians, and increasingly researchers in other fields, asking for the proof.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
