Mathematicians accuse OpenAI's Astra of passing off stolen proofs as breakthroughs

Mathematician Andreas Thom has publicly alleged that OpenAI trained its Astra model on private conversations in which he and Gábor Kun were working through Gromov’s soficity conjecture — one of ten problems OpenAI later claimed Astra had solved on its own. In a detailed Mastodon post amplified by Valerio Capraro, Thom lays out evidence that his unpublished work, the product of roughly two decades on the problem, was absorbed into the model and then re-presented as an original machine discovery. Thom is a recognized authority on sofic and hyperlinear groups, which lends unusual weight to the accusation.

The claim does not stand alone. Capraro links it to similar allegations from Levent Alpöge and Tristan Buckmaster, arguing that a pattern of OpenAI showcasing human mathematical work as autonomous AI results is beginning to form. If substantiated, the disputes move beyond ordinary questions of citation and attribution into whether a company is laundering unpublished human research into marketing claims about machine reasoning.

All of this remains unverified allegation rather than established fact, and OpenAI has not responded in the thread. But the episode crystallizes a growing tension in AI evaluation: extraordinary claims about models “solving” hard open problems are difficult to audit when the training data is opaque, and the line between a model reasoning to a result and reproducing ingested human work is exactly what these researchers say is being blurred.