A new front has opened in the credit war between mathematicians and OpenAI, and this time it predates the Navier-Stokes blowup by months. Andreas Thom, a group theorist at TU Dresden whose own research underpins one of OpenAI’s most celebrated math results, has gone public with a seeming accusation that a senior OpenAI researcher gave him a misleading — in his words, dishonest — answer about whether his private conversations with ChatGPT ever touched the company’s training pipeline.
The claim, laid out in a three-part post on Mathstodon, reaches back to August, when OpenAI announced that its Astra model had constructed the first-ever non-sofic group, resolving a 27-year-old open problem in geometric group theory first posed by Mikhail Gromov. The proof’s central technical step leaned directly on a 2019 paper by Gábor Kun and Thom himself, along with a related 2016 result by Kun.
The email that started it
Shortly after Astra’s announcement, Thom says he emailed OpenAI’s Mark Sellke and Sébastien Bubeck — the researcher who leads the company’s math efforts — noting that he and a colleague in Dresden had spent months discussing the expander matching problem and extensions of the Kun-Thom work with ChatGPT itself. He wanted to know two separate things: whether those conversations had entered the model’s training data, and separately, whether they had been accessible to the system while it was working on the proof.
Sellke’s reply, as Thom quotes it, was short: “that did not happen.” Thom argues that answer only addresses the second question — direct access during the solving process — while staying silent on the first, and more consequential, question of whether his conversations shaped the training data itself. He’s now calling that a materially misleading response rather than an honest one, particularly since OpenAI has since told a very different story in a separate, higher-profile dispute.
The Buckmaster-Alpöge shadow
That separate dispute is the one that’s dominated math and AI headlines this week: NYU’s Tristan Buckmaster and Anthropic researcher Levent Alpöge accused OpenAI of learning about their unpublished Navier-Stokes work and racing to publish a competing proof, with Bubeck allegedly pushing to strip Alpöge’s name from a joint paper because he works at a rival lab. OpenAI’s eventual written response said no researcher or agent saw the pair’s work before publication and that no specific user data was accessed — but the company added a carve-out it had never offered Thom: it “cannot rule out that de-identified data derived from their usage of our products helped improve” its models.
That’s the distinction Thom says his original question already drew, and that Sellke’s flat denial erased. He isn’t asking anyone to reverse-engineer OpenAI’s training pipeline — he’s arguing that only OpenAI holds the relevant data, and that a categorical denial needs to come with a disclosed basis: account settings, training checkpoints, and a plain-English definition of what “de-identified data derived from usage” actually covers. He also points out that opting out of training, which he says he did on June 29, is a forward-looking control with no way for users to audit it, and it says nothing about what happened to conversations from before that date.
Thom’s mathematical argument sharpens the suspicion rather than proving it: he notes that the Kun-Thom approach was never the leading candidate for resolving non-soficity — other lines of attack involving quantum games looked more promising — which is part of why he found it notable that Astra homed in on his specific technique. He also argues that de-identifying a conversation strips out a name, not the underlying mathematical idea, so privacy safeguards built for personal data don’t obviously do the job when the “data” in question is unpublished research.
Part of a wider pattern
This isn’t the first time the non-sofic group result has drawn scrutiny. Days after the original announcement, Cambridge group theorist Francesco Fournier-Facio and Yeshiva University’s Steven Miller separately argued that two of Astra’s ten flagship results, including the non-sofic group construction, leaned on existing 2016 and 2019 papers more heavily than OpenAI’s initial announcement let on, with Miller telling Scientific American that the pattern of unattributed prior work “points to research misconduct.” OpenAI subsequently edited its blog post to soften language claiming the problems had seen “no progress… for at least a decade.”
Thom’s post ties that episode to the current one directly: he argues that Bubeck’s conduct in the Buckmaster-Alpöge affair, combined with Sellke’s narrow answer to him months earlier, “deepen[s] the concern that there is a loss of moral compass” among the people setting OpenAI’s norms for how it treats mathematicians’ unpublished work. He now reads Sellke’s original answer as, at minimum, unjustifiably broad, and at worst, deliberately misleading.
OpenAI has not yet issued a specific response to Thom’s posts. Given how the company eventually walked back parts of its framing in the Buckmaster-Alpöge case, and given that this exchange with Thom predates that controversy by weeks, the pressure is likely to mount for OpenAI to explain, in concrete terms, what “de-identified data derived from usage” actually means — and whether it applies here too.