Product1 publisher3 min readPublished
OpenAI offered Tristan Buckmaster partial credit for its Navier-Stokes solution
Buckmaster was preparing to publish when OpenAI announced it had solved the problem and offered him a share of the credit. A second mathematician, Andreas Thom, then asked the company whether his own ChatGPT conversations had reached its model.
The Product Desk · Product desk
What happened
- OpenAI announced on September 8 that it had solved the Navier-Stokes smoothness and existence problem, a question that had vexed mathematicians for close to a century, according to Gizmodo.
- The company drew accusations of steamrolling two researchers who had made recent progress on a related problem and were preparing to publish it.
- Andreas Thom of Dresden University of Technology said last Wednesday that OpenAI's solution to a separate problem looked eerily similar to work he had recently done with Gabor Kun.
- Thom asked whether results produced by AI should be considered results of the general mathematical community, since the system's creativity is trained on human creativity.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Any researcher who has thought out loud in a chat window is in Thom's position: the company's assurance covers what its people and agents read on the project, and it leaves de-identified training data explicitly unresolved.
- decision Heads of research groups now have to decide whether unpublished work counts as disclosed once someone types it into a product owned by a competitor for the same publication.
- contradiction A researcher's flat denial and the company's later carve-out for de-identified data cannot both be the complete answer, and Thom's reading is that the careful version arrived only once the situation got tense.
- precedent Partial credit offered by the lab that published first, to one of two co-authors, is the only template so far, and no journal or university stands behind it.
Andreas Thom wrote to OpenAI after reading the company's solution to a problem he had been working on with Gabor Kun, work the two of them had been talking through with ChatGPT. He asked whether their material had been fed directly to the model that produced the solution, or used in training it [6]. Thom said the reply from OpenAI researcher Mark Sellke was "that did not happen" [7].
Then the company posted its own account on X. Neither the human researchers nor the AI agents on the Navier-Stokes work had seen "any specific user data", it said, while adding that OpenAI "cannot rule out that de-identified data derived from their usage of our products helped improve our models" [8]. Thom told Gizmodo that the second statement made the first look worse. Once the situation got tense, he said, the company became much more careful about what it was saying [9]. Sellke's flat denial was "not honest," he said. "You should say that you don't know" [10].
A research group can act on that gap, even if the credit fight is the part everyone will quote. OpenAI's position is that offering Buckmaster partial credit for the Navier-Stokes solution was good faith [3]. Buckmaster's position is that the company tried to strong-arm him into publishing his results on OpenAI's own terms [11].
Two separate disputes run through this, one over Navier-Stokes and one over the unrelated problem Thom and Kun were working on, with four named mathematicians between them [15]. Neither has produced a rule. Gizmodo reports that the tech industry and academia have yet to work out attribution in an era when human researchers compete with AI companies to publish first [17]. The partial-credit offer is what one company did once, under criticism, for one of two co-authors, and Gizmodo also reported that OpenAI tried to push the other one, Levent Alpoge of Anthropic, out of it [4].
For anyone writing a policy for a group of researchers next week, it comes down to whether the vendor publishes results in your field, and whether its written commitment covers model training or only what its staff and agents are allowed to read. A vendor that does not compete with you and commits on both counts is an ordinary tool. A vendor that competes with you and limits its promise to specific user data is a place for finished work only. OpenAI's own wording on the Navier-Stokes project puts it in the second group [8].
Thom is not arguing for abstention. "We have to come together as a mathematical community to redefine our subject, to redefine what math students are supposed to learn," he said [13]. Buckmaster, in an interview with the Australian Broadcasting Corporation, reportedly said: "I think it's pointless." He added: "Like, I think the game is up" [12].
What to watch
- Whether OpenAI answers Thom's training-data question with documentation rather than a denial.
- Whether Anthropic or Levent Alpoge says anything publicly; so far only OpenAI's attempt to push him out is reported.
- Whether any journal or mathematics department writes a crediting rule that cites the Buckmaster offer.