OpenAI is facing renewed scrutiny from mathematicians who want clearer evidence that their private interactions with its products did not help produce the company’s recently announced mathematical advances.
Mathematician Andreas Thom said OpenAI’s work on non-sofic groups relied heavily on prior work by him and Gábor Kun, and raised concerns that conversations he and colleagues had with ChatGPT may have contributed indirectly to the company’s capabilities. The dispute follows a separate public challenge involving OpenAI’s claimed progress on the Navier-Stokes problem.
The immediate issue is not simply attribution. It is whether a model provider can distinguish between directly retrieving a user’s work and training or improving models on de-identified information derived from that user’s activity.
The distinction at the center of the dispute
OpenAI said in its Navier-Stokes announcement that its researchers and agents did not see other researchers’ work before it became public, including through “specific user data.” But it also said it could not rule out that de-identified data from product usage helped improve its models.
For Thom, that qualification is insufficient. Removing a user’s identity does not remove the intellectual content of an unpublished mathematical idea, he argued. He has called on OpenAI to disclose the relevant datasets and settings needed to substantiate a denial.
OpenAI did not immediately respond to *The Verge*’s request for comment.
The company had already amended its write-up on the non-sofic groups result after criticism that it did not adequately acknowledge Thom and Kun’s recent contributions. That change does not resolve the training-data question, but it raises the stakes around provenance and credit when AI systems operate near the research frontier.
Why this matters beyond mathematics
The controversy maps directly to a growing enterprise AI governance problem. Companies increasingly put valuable material into general-purpose AI tools: source code, product plans, customer research, legal analysis, technical designs and scientific hypotheses.
A provider’s statement that a particular conversation was not directly accessed may not answer the operational question users actually have: whether their inputs were retained, transformed, evaluated, or used to improve a future model. For research organizations, that ambiguity can affect publication strategy, intellectual-property controls and willingness to use external AI platforms at all.
The concern is especially acute in fields where a small insight can have outsized value. If researchers suspect private experimentation could help a well-resourced AI lab compete on the same problem, they may share less, delay collaboration, or keep work off commercial systems. Researchers told *The Verge* that the recent incidents could make mathematics more secretive.
What operators should ask vendors now
Organizations using AI services should treat data-use terms as an architecture and procurement issue, not just a legal checkbox. Key questions include:
- Are prompts, uploads and outputs used for training, evaluation, safety work, or product improvement?
- Is the policy different for consumer, team and enterprise products?
- Does “de-identified” data remain eligible for model improvement?
- What controls exist to opt out, and how are those controls verified?
- Can the vendor provide retention, access and audit details for sensitive workflows?
For high-value research or proprietary development, teams should establish explicit rules for which tools can receive nonpublic material. Where assurances are insufficient, a segregated environment, contractual data protections, or an internal model deployment may be more appropriate.
What to watch next
The near-term test is whether OpenAI provides a more concrete account of its data boundaries for research-related use. More broadly, the episode may increase pressure on AI providers to make training, product-improvement and data-retention policies more legible—and independently auditable.
As frontier models move from summarizing public knowledge toward generating plausible new results, transparency about where their capabilities came from will become a business issue as much as a scientific one.
