OpenAI Math Breakthrough Faces Credit Claims

A second mathematician has gone public with claims that OpenAI's models drew on his unpublished work without proper credit, days after a similar dispute erupted over a different researcher's findings. Andreas Thom, whose expertise centers on so-called non-sofic groups, laid out his concerns in a series of Mastodon posts, questioning whether conversations he'd had with ChatGPT fed into the company's later announcement of a mathematical breakthrough in his own field.

OpenAI touted the non-sofic groups result as one of ten major achievements unveiled last month, and the company itself acknowledged the work leaned heavily on research by Thom and mathematician Gábor Kun. That admission came only after mathematicians criticized the initial announcement for glossing over Thom and Kun's contributions, prompting OpenAI to revise its writeup without much fanfare.

Thom says he started scrutinizing his own exchanges with the chatbot after Tristan Buckmaster, a math professor at New York University, raised similar questions about whether his use of OpenAI's Codex tool had shaped the company's results. Non-sofic groups, for context, are infinite structures that resist approximation by finite ones, a niche corner of math now at the center of a very public argument about credit and data.