A dispute over an artificial-intelligence claim involving one of mathematics’ major unsolved problems has intensified concerns about whether rapidly developing AI systems could weaken trust in scientific research. The disagreement centres on the Navier–Stokes equations, a famous mathematical problem, and on questions about how research stored in online tools may be accessed or used.
OpenAI, the company behind ChatGPT, said last month that its AI agents had solved the problem in 88 hours. The organisation responsible for a prize linked to the challenge said it had “apparently been settled”, while stressing that its assessment would take time. That review is important because a claimed solution must be examined carefully before the mathematical community can establish whether it is correct.
“Apparently been settled.”
The announcement prompted concerns from Tristan Buckmaster, a mathematician at New York University. Buckmaster had been working on the same objective with Levent Alpöge, a mathematician who is employed by Anthropic, an AI company. Their independent team kept draft material in Codex, OpenAI’s coding tool.
Buckmaster said his colleague had received an indication that news of their progress had reached OpenAI. He did not explain how that might have happened, and he made clear that he was not accusing anyone of wrongdoing. His concern was whether work held in Codex could have contributed to OpenAI’s own effort.
OpenAI said that no person or AI system at the company had seen the other team’s work before it was published. The company initially said it could not completely exclude the possibility that anonymised information from users’ interaction with its products had helped improve its systems, although it regarded that possibility as unlikely. Anonymised data is information processed so that it is not directly linked to an identified individual, but questions can still arise over how such data is used and whether researchers understand those arrangements.
A wider question about scientific evidence
The dispute matters beyond one mathematical problem because scientific progress depends not only on producing an answer, but also on showing how that answer was reached. Researchers need to be able to examine the evidence, reproduce the work where possible and identify which contributions shaped the final result. When AI tools are involved, those tasks can become more difficult if the system has been trained on large collections of material or if drafts are stored on commercial platforms.
The issue is emerging as companies and researchers seek to apply AI to increasingly large scientific datasets. At a former watch factory in central London, Basecamp Research is storing glass tubes containing what Dr Oliver Vince, the firm’s co-founder, described as 100 billion genes. The company intends to gather as much DNA as possible to develop an AI system capable of working with the language of biology.
Vince said two of the facility’s boxes could contain 10 times more information than all the text on the internet. The comparison illustrates the scale of the biological information being assembled, although the amount of data alone does not determine whether an AI system will produce reliable scientific findings. Data must still be collected responsibly, interpreted correctly and assessed by people with appropriate expertise.
Speed must be matched by scrutiny
Supporters of AI in science believe the technology can accelerate research. The attraction is clear: an AI system can process large bodies of information and assist with difficult technical tasks. Yet the mathematics dispute shows why speed cannot replace independent checking. A result announced after less than four days may be important, but its value depends on the quality of the reasoning and on whether the work can withstand detailed examination.
The central concerns raised by the dispute include:
- Research access: whether material placed in online tools can influence later work without the authors’ knowledge;
- Attribution: how contributions from human researchers and AI systems should be identified and credited; and
- Verification: how mathematical or scientific claims should be reviewed before being accepted.
| Issue | What is known |
|---|---|
| Mathematical claim | OpenAI said its AI agents solved the Navier–Stokes problem in 88 hours. |
| Independent review | The body responsible for the associated prize said the claim appeared settled, but that its review would take time. |
| Data concern | Researchers questioned whether work stored in Codex could have influenced the result; OpenAI denied that its people or AI had seen it before publication. |
The disagreement remains a dispute about process and provenance rather than a final finding of misconduct. The available information does not establish that the independent team’s work was used. It does, however, demonstrate why clear rules are needed as AI becomes more deeply embedded in research. Scientists will have to distinguish between a system’s ability to generate a plausible answer and a demonstrated scientific result that has been independently checked.
