Andreas Thom, a prominent mathematician, is accusing OpenAI of unethical behavior regarding the data used to train its AI models. This allegation follows closely on the heels of a similar dispute involving NYU professor Tristan Buckmaster. Thom’s concern centers on OpenAI’s recent announcement of breakthroughs in non-sofic groups, a niche area where he and colleague Gábor Kun have made significant prior contributions. He suspects that his own private interactions with ChatGPT may have inadvertently fed into the model’s training data, giving the AI an unfair advantage.
Thom reached out directly to OpenAI researchers Sébastien Bubeck and Mark Sellke, seeking clarity on whether their conversations were incorporated into the model’s learning process. The response he received focused narrowly on direct access to his chats but failed to address whether those interactions influenced the broader training datasets. Thom found this distinction insufficient and misleading. “No such qualification, explanation, or evidence was given,” he stated on Mastodon. “I take this as dishonesty to say the least.” He argues that without full transparency from OpenAI, researchers cannot verify if their unpublished ideas are being absorbed and repurposed by the company’s systems.
The core of the disagreement lies in OpenAI’s handling of “de-identified” user data. While the company denies accessing specific user work for its Millennium Prize solution, it admits that de-identified usage data might indirectly improve model performance. Thom rejects this defense, noting that stripping a name from data does not erase the intellectual substance of a mathematical proof. If private research inputs are refining the very tools competing against their creators, it creates an ethically fraught loop. “It would be ethically indefensible,” Thom argued, if nonpublic research helped improve models that then raced users to publication without consent or credit.
This growing unrest threatens to overshadow OpenAI’s technical achievements. The company’s solution to the Navier-Stokes equations is a monumental feat, yet the surrounding controversy casts a long shadow. Mathematicians are left questioning the integrity of the collaboration between human insight and machine learning. Until OpenAI provides concrete evidence of its data hygiene practices, trust within the academic community will likely continue to erode. The burden of proof now rests squarely on the tech giant to demonstrate that it isn’t cannibalizing the intellectual property of the experts it aims to assist.
