Andreas Thom Seeks Proof OpenAI Did Not Use His Mathematical Work

What changed

According to reporting by The Verge, mathematician Andreas Thom is demanding evidence that his conversations with ChatGPT did not enter OpenAI’s training data or influence its work on non-sofic groups. OpenAI acknowledged that the result relied heavily on work by Thom and Gábor Kun, then quietly amended its write-up after criticism; the dispute follows an earlier challenge from mathematician Tristan Buckmaster.

Why This Matters

The important question is not only whether OpenAI copied a proof. It is whether researchers can safely discuss unfinished work with an AI system when they cannot inspect what happens to those conversations afterward.

Thom says OpenAI answered a narrower question: whether researchers directly accessed his chats. That does not settle whether those interactions entered training data or affected the system’s reasoning. OpenAI has previously said that no specific user data was accessed before its Navier-Stokes result, but that statement does not resolve the broader concern Thom raised.

That creates a sharp imbalance. Researchers may hold the unpublished ideas, but OpenAI holds the logs, datasets and technical settings needed to establish what happened. If that uncertainty persists, mathematicians may stop putting active work into ChatGPT. Universities and research groups could also tighten rules around confidential research, trading some AI assistance for clearer control over unpublished ideas.

The effect reaches beyond this dispute. Future AI-generated mathematical results will need more than an impressive answer and a polished announcement. Attribution, reproducibility and data provenance become part of the result itself. That raises OpenAI’s burden of proof and gives an advantage to competitors able to offer clearer boundaries around research interactions.

Impact assessment

For independent mathematicians, the immediate risk is practical: sharing a promising technique with an AI system may carry an unclear chance of later reuse. If OpenAI cannot clarify the relevant data pathways, researchers could withhold unpublished methods, reducing the expert input available to its systems over the coming weeks and months.

For OpenAI, the cost is credibility. The second public challenge, combined with the amended non-sofic-groups write-up, makes vague assurances less useful. Future announcements may require fuller attribution and technical provenance details before researchers accept the claims.

Academic institutions face a trade-off over six to 12 months. They may formalize restrictions on entering confidential research into commercial AI tools, protecting unpublished work but limiting access to useful assistance. The chain could be interrupted if OpenAI provides independently checkable evidence that addresses both direct system access and broader training-data use.

Scenarios

Most likely

Our outlook (informed speculation): OpenAI faces sustained demands for provenance explanations, while mathematicians become more cautious about sharing unpublished work through ChatGPT over the next several months. This is most likely if OpenAI continues to address particular conversations without explaining broader training-data pathways and more researchers raise similar concerns.

The practical change would be a higher verification burden around every major mathematical announcement. Future write-ups would need clearer attribution and evidence about the limits of the systems involved.

Upside

If OpenAI publishes auditable documentation covering training-data use, system access and prior mathematical contributions, researchers could resume exploratory use of its tools with clearer boundaries. Consistently crediting work by mathematicians such as Thom and Kun would also make future results easier to evaluate.

This path depends on evidence that resolves the central uncertainty, not another general assurance about direct access to specific conversations.

Downside

If additional mathematicians report comparable experiences and OpenAI still cannot explain how unpublished interactions are handled, researchers and institutions could withdraw more unfinished mathematical work from its systems over six to 12 months.

That would reduce the pool of expert interactions available to OpenAI while increasing the scrutiny and reputational cost attached to its mathematical claims. Universities or research groups issuing formal restrictions would be a sign that the dispute had moved from an attribution fight into research policy.

What to watch next

• Whether OpenAI publishes a technical explanation of whether user interactions can enter training data or affect reasoning systems.
• Whether another mathematician identifies a specific interaction, technique or attribution issue tied to an OpenAI result.
• Whether OpenAI’s next major mathematical announcement includes detailed attribution and data-provenance disclosures.