Second mathematician questions OpenAI's math claims

Andreas Thom, Professor of Geometry at TU Dresden, has accused OpenAI of being dishonest about its training data, writing on Mastodon that his own exchanges with ChatGPT may have contributed to the mathematical breakthroughs OpenAI announced. Quartz reported the accusation on September 10.

At a glance

• The complaint centers on how the material reached the model: Thom says conversations he had with ChatGPT could have supplied the mathematics that OpenAI later described as results produced by its system.
• According to The Verge, one of the 10 mathematical results OpenAI announced concerned non-sofic groups, Thom's own field, and leaned heavily on earlier work by Thom and Gábor Kun.
• The Verge counts Thom as the second mathematician to raise such objections publicly, days after Tristan Buckmaster disputed OpenAI's account of a proof involving the Navier-Stokes equations.

The dispute is less about a single proof than about provenance: an AI lab presenting mathematical results as machine-generated has to account for what the model was shown, and conversational logs are the least visible part of that chain. With two mathematicians now objecting in public within days, the reporting burden appears to shift from the announcement to the audit trail behind it. Training data provenance is the recurring fault line in AI disputes, and mathematics makes it unusually testable: a specialist can recognize his own argument.

Thom set out the complaint on Mastodon, where he wrote that material from his interactions with ChatGPT may have found its way into the results OpenAI presented. His central charge is that OpenAI has not been honest about what went into the training data behind those announcements.

According to The Verge, one of the 10 mathematical results OpenAI announced dealt with non-sofic groups, the area Thom works in, and built heavily on earlier work by Thom and Gábor Kun. The result therefore sits in the same field as Thom's published work.

Thom is the second mathematician to go public with objections of this kind, according to The Verge, which places his posts days after Tristan Buckmaster disputed OpenAI's account of a proof concerning the Navier-Stokes equations. Both cases concern results OpenAI presented as new mathematics.

What the Mastodon posts leave open

Thom's posts do not state what remedy he is seeking, and no timetable has been given for a fuller account of how the 10 results The Verge describes were produced. The accusation so far rests on his reading of the overlap between his work with Gábor Kun and the non-sofic groups result, rather than on documentation of what entered the training set.