A second mathematician is accusing OpenAI of being dishonest about its AI training data

Free daily briefing on global business news.

Andreas Thom says OpenAI's response to his questions about ChatGPT data use was "materially misleading" and "plainly dishonest"

Cheng Xin / Getty Images

Andreas Thom, a mathematician whose research overlaps with one of OpenAI's recent announced breakthroughs, has accused the company of "dishonesty" over its handling of questions about whether his interactions with ChatGPT contributed to its results, according to The Verge. The accusation comes days after a separate and bitter dispute erupted between OpenAI and New York University mathematics professor Tristan Buckmaster over the company's claimed solution to a Millennium Prize problem.

Thom said he began scrutinizing his own ChatGPT interactions after Buckmaster publicly questioned whether OpenAI's models had benefited from his use of the company's Codex tool. One of the 10 mathematical results OpenAI announced last month involved non-sofic groups, an area where Thom had done significant work, and OpenAI had already revised its writeup after facing criticism for failing to acknowledge contributions from Thom and fellow mathematician Gábor Kun, according to The Verge. Non-sofic groups are infinite mathematical structures that cannot be approximated by finite ones.

Thom said he found it notable that OpenAI had shown such a precise grasp of his techniques, which at the time he did not consider the most obvious path toward a solution. He contacted Sébastien Bubeck and Mark Sellke — OpenAI researchers, with Sellke also holding a position as a statistician at Harvard — to ask whether his ChatGPT conversations had been incorporated into the company's training data or made available to its reasoning process. Thom said the reply he received was narrowly confined to the question of direct access to his conversations, leaving entirely unanswered whether those conversations had ever become part of the company's training pipeline. "No such qualification, explanation, or evidence was given," he wrote, according to The Verge. "I take this as dishonesty to say the least."

Thom said that only OpenAI has the data needed to answer the question and that, if the company intends to deny any use of his work, the burden falls on it to disclose the relevant datasets and clarify its data-use terms. He also challenged OpenAI's practice of drawing a line between direct data access and de-identified training data, noting that the company had leaned on the same hedged framing in its dealings with Buckmaster. "De-identification may remove a name; it does not remove the intellectual content of a mathematical idea," he said, according to The Verge. Looking back at Sellke's response, Thom said it was "at minimum, unjustifiably broad and materially misleading; looking back it was plainly dishonest."

The dispute with Buckmaster centered on OpenAI's claimed proof of the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems carrying a $1 million award each. OpenAI maintained that neither its researchers nor its agents viewed Buckmaster's work prior to its public release, while stopping short of categorically excluding the possibility that anonymized data generated through his product use had shaped model development. Buckmaster claimed that an OpenAI researcher pressed him to drop his collaborator from the credit and cautioned him that making the dispute public would damage his career. OpenAI's head of mathematics research Sébastien Bubeck confirmed that certain remarks were "a bad choice of words."

OpenAI did not respond to The Verge's request for comment on Thom's accusations. The Clay Mathematics Institute, which administers the Millennium Prize, has said it will conduct a detailed review of the Navier-Stokes situation.

Join 500,000+ readers who start their day with Quartz.

By subscribing, you agree to our Terms of Service and Privacy Policy.

Global business news for a smarter world

Do Not Sell or Share My Personal Information

We won't sell or share your personal information to inform the ads you see. You may still see interest-based ads if your information is sold or shared by other companies or was sold or shared previously.

Dismiss

Opt out