Mathematician demands OpenAI prove it didn't use his research to train AI models

A prominent mathematician has publicly accused OpenAI of "dishonesty" and opacity regarding the sources of its training data, intensifying scrutiny around how the company's models achieved recent high-profile results in mathematics.

The challenge#

Andreas Thom, a mathematician whose research OpenAI cited while announcing a breakthrough in non-sofic groups last month, says the company failed to adequately address whether his interactions with ChatGPT could have shaped the AI system's capabilities. In posts on Mastodon, Thom explained that when he inquired whether his own ChatGPT conversations had entered OpenAI's training pipeline, the responses he received were evasive and technically misleading.

OpenAI researchers told him his direct conversations with the chatbot were not accessed for the specific result—but they stopped short of ruling out whether de-identified versions of his interactions influenced the underlying models. For Thom, this linguistic hedge was the problem: "De-identification may remove a name; it does not remove the intellectual content of a mathematical idea."

The larger pattern#

Thom's concerns echo those raised by mathematician Tristan Buckmaster, who recently questioned whether OpenAI's claimed solution to the Navier-Stokes problem—a Millennium Prize challenge—may have benefited from his use of OpenAI's Codex tool. In OpenAI's public statement about that result, the company made a similar distinction: it denied using "specific user data" while admitting it "cannot rule out that de-identified data derived from their usage of our products helped improve our models."

Thom argues this formulation is fundamentally inadequate. He points out that only OpenAI possesses the technical capability to verify whether unpublished research has been incorporated into its systems. "The responsibility is on them to prove that by disclosing all necessary datasets and clarifying various settings and terms setting out how it uses data," he wrote.

Why it matters#

The row reflects deepening unease in academic mathematics circles. Thom warns it would be "ethically indefensible" for OpenAI to use researchers' unpublished work to train models and then race those same researchers to publication for glory—all without consent or proper credit. Multiple mathematicians have told journalists they fear the incentives now favor secrecy: if rumors of a major breakthrough can trigger a corporate push to solve it first, researchers may stop sharing progress openly.