Mathematician says OpenAI left unanswered whether ChatGPT-derived data informed unpublished math
- Mathematician Andreas Thom says OpenAI's earlier categorical reply that his and a colleague's ChatGPT discussions "did not happen" failed to distinguish direct access from possible use in model training or improvement.
- Thom says he and a Dresden colleague had discussed the expander matching problem and extensions of Gabor Kun's work with ChatGPT before OpenAI announced a non-sofic group result using related methods.
- OpenAI's reported response in the Buckmaster-Alpöge case says it found no access to specific user data, while saying it cannot rule out that de-identified data derived from product use improved its models.
- A commenter asked whether Thom had disabled ChatGPT's default "Improve the model for everyone" setting; Thom replied that he opted out on June 29.
- The post frames the unresolved issue as whether researchers can safely discuss unpublished mathematics with OpenAI products when user conversations may contribute de-identified data used to improve models.
Hacker News opinions
The evidence here is weak. Someone says they discussed a topic with an AI while working on it, but does not claim they had a proof or that the model saw a decisive result.
There is still a real distinction between independent discovery and a model making the final connection from relevant prior discussions. If the latter happened, OpenAI's public claims would be overstated.
These tools seem to solve the specific problems for which they had human chat data. If the system had independently developed the method, I would expect it to solve other nearby problems too.
For highly specialized mathematics, even an unpublished discussion of a technique can be highly informative because only a handful of people may know it. It is reasonable to question why the model reached for that tool.
Calling this human-AI collaboration ignores why mathematicians do research. A lab can spend far more on a capability demo than the result is worth to the field, while researchers need work that supports their careers.
The Navier-Stokes result looks more like a math search engine that started near the answer and spent about $10 million of compute testing combinations, like a chess engine for proofs.
If the allegation is true, researchers will not keep collaborating with systems that can front-run unpublished work. Getting scooped does not pay their bills.
The issue is not only credit. If chat-derived material makes models look more capable than they are, it feeds exaggerated AGI claims and AI investment hype.
OpenAI's wording leaves a loophole: it may not put chat transcripts directly into pretraining, but derived or transformed data can be difficult for any user to trace or prove.
Cloud LLMs are a liability for unpublished work. I would think carefully before making one part of my research or development workflow, because pricing and terms can change too.
I would timestamp research in public repositories or arXiv before discussing it with a model. That will not stop copying, but it makes a later priority dispute easier to establish.