What Thom says he asked OpenAI
Andreas Thom, a group theorist at TU Dresden, says OpenAI gave him a “materially misleading” answer when he asked whether months of private ChatGPT conversations about his research could have fed the model it later credited with constructing the first non-sofic group. He set out his account in a guest post on Terence Tao’s blog on 11 September, two days after first raising it in a thread on Mathstodon.
According to Thom, he and a colleague in Dresden had spent months discussing the expander matching problem, and extensions of his earlier work with Gábor Kun, in ChatGPT. Shortly after OpenAI’s non-sofic announcement, he emailed OpenAI researchers Mark Sellke and Sébastien Bubeck to ask whether those conversations had been part of the training data or accessible to the reasoning process.
Thom quotes Sellke’s complete reply as: “Regarding your conversations with ChatGPT: that did not happen.”
Why he calls the reply misleading
Thom says he had asked two separate questions: whether the conversations entered training data, and whether the system could access them while it worked on the problem. In his reading, the one-line denial now looks as though it addressed only the second. He writes that it came with no qualification, explanation or evidence.
What changed his view, he says, was OpenAI’s wording in a different case. Responding this week to concerns raised by mathematicians Tristan Buckmaster and Levent Alpöge, OpenAI said no specific user data had been accessed to solve the problem, but that it “cannot rule out” that de-identified data derived from their usage of its products had helped improve its models, as quoted by ABC News. That statement was about Buckmaster and Alpöge, not Thom. Thom argues it draws exactly the line between direct access and training that Sellke’s answer to him did not.
Thom writes that he disabled model training on 29 June. He describes that setting as a promise users cannot audit, and one that says nothing about earlier conversations. He also says he has since been pointed to a newer OpenAI statement that Buckmaster’s Codex prompts could not have influenced its system, including through training.
The result at the centre of the dispute
Sofic groups are a class of groups introduced by Mikhail Gromov in 1999, and in the 27 years since no one had proved or disproved that non-sofic groups exist. At the start of August, OpenAI said an internal version of its Astra model had produced the first explicit construction of a non-sofic group, one of ten results it announced, The Next Web reported.
Thom writes that the proof relied in a crucial way on his 2019 paper with Kun, and that OpenAI’s original announcement nonetheless spoke of a decade without progress. After he objected, he says, OpenAI revised the framing of that announcement.
He presents the data question as unanswered rather than settled. His argument is conditional: if non-public research supplied by users improved a model that was then used to race those users to publication, without consent, disclosure or credit, he writes, that would be ethically indefensible. OfficeChai and Notebookcheck both reported the allegation on 10 September.
OpenAI’s position
Sellke’s email is the only response from OpenAI that Thom reports. Notebookcheck reported that OpenAI had not responded publicly, and Aivio News found no OpenAI statement addressing Thom’s account at the time of publication. For a denial to be credible, Thom says, OpenAI should disclose its basis: the relevant product and privacy settings, datasets and checkpoints, and what it means by de-identified data derived from usage.
The allegation adds to a week of friction between OpenAI and research mathematicians. On 10 September, OpenAI researcher Dan Roberts said the company had withdrawn its sponsorship of the Caltech Mathathon, a hackathon due to start on 30 October, after current and former Caltech mathematicians published an open letter against the event, Gizmodo reported. The letter said it had 771 signatories at the time of publication.
What to watch
The open question is the one Thom put to OpenAI in writing: whether conversations like his, in any de-identified form, reached the training data of the model that produced the non-sofic result. OpenAI has not said whether it will give the fuller account he is asking for.
