Analysis
OpenAI's dispute over who deserves credit for its Astra model's claimed math breakthroughs widened this week, The Verge reported, after mathematician Andreas Thom raised concerns on Mastodon that his own prior interactions with ChatGPT -- in his specialty area of non-sofic groups -- may have contributed to one of the ten results OpenAI's internal model claimed to solve. Pulse first covered this dispute days earlier, when it centered on a single mathematician's Navier-Stokes-adjacent credit claim; Thom's is now the second named academic to publicly question the chronology.
What's changed: this is no longer a single researcher's grievance but a pattern, with a second mathematician in a different subfield making a structurally similar complaint -- that private conversations with ChatGPT prior to OpenAI's August announcement may have fed directly into results the company later presented as autonomous breakthroughs, without disclosure or acknowledgment.
“OpenAI has not issued a specific response to Thom's claim as of publication.”
OpenAI released a 249-page manuscript in early August claiming an internal, unreleased model called Astra had solved ten long-standing open problems in mathematics and theoretical computer science, backed by machine-checkable formal proofs in the Lean 4 proof assistant that anyone can independently verify. Six weeks later, none of the ten results has cleared traditional academic peer review, and OpenAI has not published a clear account of what training data or prior human interactions, if any, informed each specific result.
The dispute cuts at something more fundamental than credit: if models are being fine-tuned or prompted using researchers' own unpublished reasoning, captured through ordinary product usage, and then presented as independent machine achievement, that raises the same underlying question as the NYT copyright suit, just applied to individual academics instead of a media company. The difference is that mathematicians generally can't sue over an idea the way the Times can sue over published text, leaving public pressure and reputational risk as the only real leverage a working mathematician has against a company as large as OpenAI.
OpenAI has not issued a specific response to Thom's claim as of publication. The next real test isn't another statement -- it's whether any of the ten results clears formal peer review with a training-data provenance disclosure attached, which would settle the question data alone, not more mathematician testimony, can actually answer.
