OpenAI's Astra accused of training on mathematicians' unpublished work
A post on X (Twitter) reports that mathematician Andreas Thom has presented, in a detailed Mastodon post, several pieces of evidence suggesting OpenAI may have trained its Astra system on private conversations between him and fellow mathematician Gábor Kun. The two were working on Gromov's soficity conjecture, one of ten open mathematics problems OpenAI later announced Astra had solved. The X post says it reproduces Thom's Mastodon post in full in its replies, but that text is not part of what is summarized here, and the specific evidence Thom cites is not described.
The post vouches for Thom's standing: its author says they met him several times early in their careers, and calls him an exceptional mathematician, a leading expert on sofic and hyperlinear groups, and one of the most respected scholars in the field, someone who has spent two decades working on the soficity conjecture. The post argues that, if Thom's account is correct, this is not a minor dispute over attribution: it would mean unpublished human work was absorbed into a model and then presented to the world as a breakthrough by the model itself.
The post ties the claim to a pattern rather than an isolated incident, invoking earlier, separate allegations of the same kind from mathematicians Levent Alpöge and Tristan Buckmaster, without saying what those allegations concerned. If all three are substantiated, it argues, this would amount to one of the greatest intellectual scandals in the history of science, and it closes by framing the stakes starkly: in its words, AI is not discovering new mathematics, it is stealing human discovery.
Beyond that summary, little else is available: neither Thom's full Mastodon post nor the specific evidence it describes is reproduced here, no date is given for the Mastodon post or for OpenAI's original ten-problems announcement, and no response from OpenAI is reported.
Key facts
- A post on X (Twitter) reports that mathematician Andreas Thom presented evidence, in a Mastodon post, suggesting OpenAI may have trained its Astra system on private conversations between him and fellow mathematician Gábor Kun.
- Thom and Kun were discussing Gromov's soficity conjecture, one of ten open mathematics problems OpenAI later announced Astra had solved; Thom has worked on the problem for two decades.
- The X post's author says they have known Thom since early in their careers and vouches for him as a leading expert on sofic and hyperlinear groups, but the specific evidence Thom cites, and his full Mastodon post, are not reproduced in what is summarized here.
- The post links the claim to a pattern rather than an isolated incident, citing earlier, separate allegations of the same kind from mathematicians Levent Alpöge and Tristan Buckmaster.
- No date is given for the Mastodon post or for OpenAI's original ten-problems announcement, and no response from OpenAI is reported.
Why it matters
If Andreas Thom's account holds up, it would mean OpenAI trained a model on unpublished, private mathematical work and then presented the result as an autonomous breakthrough by that model, rather than crediting the people whose conversations supplied the underlying ideas. That is a different problem than a citation dispute: it goes to whether an AI lab's proof-solving claims can be trusted at face value, and it lands on a specific, named result, Gromov's soficity conjecture, among the ten problems OpenAI has publicly credited to Astra. The post frames this as part of a pattern rather than a one-off: taken together with earlier, separate allegations from Levent Alpöge and Tristan Buckmaster, it argues these would no longer be isolated incidents.
Who it affects
Andreas Thom and Gábor Kun are the immediate parties: two decades of Thom's own work on one conjecture is what the post says may have been absorbed without credit. OpenAI's Astra, and the credibility of its wider ten-problems announcement, is directly implicated. The post also names Levent Alpöge and Tristan Buckmaster, mathematicians it says raised allegations of the same kind, though it does not explain what theirs concerned. More broadly, any researcher who has discussed unpublished work somewhere an AI company's systems might have had access to has a stake in how this is resolved.
How to use it
Treat this as an allegation relayed secondhand, not a settled account: the source here is a post on X summarizing a Mastodon post, and the Mastodon post itself, along with the specific evidence Thom says he presents, is not part of what is captured. Readers following the story should look for Thom's full Mastodon post and for any direct response from OpenAI before treating the claim as confirmed.
How solid is it
The claim rests on a single chain of secondhand reporting: a post on X summarizing a Mastodon post by Thom, whose full text and cited evidence are not reproduced here. The post's own language stays hedged throughout, from might have and may have to if his account is correct, rather than asserting the claim as settled fact. Its case for credibility is largely personal: the author says they have known Thom since early in their careers and vouches for his standing as a leading expert on sofic and hyperlinear groups. That is a character reference, not independent evidence, and no response from OpenAI is reported.
Risks and caveats
No date is given for either Thom's Mastodon post or OpenAI's original announcement that Astra had solved ten problems. The specific evidence Thom says he presents is not described, only that it exists. What the earlier allegations from Levent Alpöge and Tristan Buckmaster actually concerned, and whether they involved OpenAI or Astra specifically, is not stated here either. No response from OpenAI is quoted or mentioned, so this remains a one-sided account pending further detail.
“AI is not discovering new mathematics. AI is stealing human discovery.”
— the post on X