OpenAI's millennium proof dispute raises the question of whether researchers can trust AI labs
Researcher Tristan Buckmaster accuses OpenAI of using his and Levent Alpöge's drafts (uploaded to Codex) to train its model on the Navier-Stokes Millennium Problem, alleging plagiarism, pressure tactics, and the exclusion of Alpöge (an Anthropic employee) as a co-author OpenAI CEO Sam Altman and researcher Sébastien Bubeck deny the allegations, claiming the team "acted with integrity" and that the effort was triggered by rumors that Anthropic's models had solved the problem OpenAI acknowledges i
Analysis
TL;DR
- Researcher Tristan Buckmaster accuses OpenAI of using his and Levent Alpöge's drafts (uploaded to Codex) to train its model on the Navier-Stokes Millennium Problem, alleging plagiarism, pressure tactics, and the exclusion of Alpöge (an Anthropic employee) as a co-author
- OpenAI CEO Sam Altman and researcher Sébastien Bubeck deny the allegations, claiming the team "acted with integrity" and that the effort was triggered by rumors that Anthropic's models had solved the problem
- OpenAI acknowledges it "cannot rule out" that de-identified data from user inputs may have helped improve its models, despite employees downplaying the likelihood of training on the researchers' solutions
- Mathematician Terence Tao warns that the incident could reverse centuries of open science traditions, as researchers may stop sharing promising directions for fear that tech companies will mobilize massive resources based on rumors and overtake their work
Why It Matters
This dispute strikes at the heart of trust between the academic research community and AI labs, raising urgent questions about data provenance, consent, and the ethics of training on user-submitted research. It also highlights a structural power imbalance: a single rumor can trigger a well-funded lab to deploy massive compute resources and potentially publish first, undermining the incentive structure that has long sustained open scientific collaboration.
Technical Details
- The dispute centers on the Navier-Stokes equations, one of the seven Clay Mathematics Institute Millennium Problems carrying a $1 million prize, and involves claims that an AI system produced a significant mathematical result on this problem
- Buckmaster and Alpöge allege they had entered a similar approach into OpenAI's Codex system, and that this input likely ended up in the training data used to develop OpenAI's model
- OpenAI employee Boaz Barak disputed the necessity of external input, claiming the model independently proved a stronger result and did not need "hints" from the researchers
- OpenAI's official blog post concedes uncertainty, stating that while unlikely, it cannot rule out that de-identified data derived from user product usage helped improve its models
- The incident underscores the opacity of AI training pipelines: there is no public verification of whether the researchers' opt-out settings for data training were effective or even available
Industry Insight
- AI labs must establish transparent, verifiable data governance policies if they wish to maintain credibility with the research community; vague disclaimers about "de-identified data" are insufficient to prevent erosion of trust
- Researchers should treat any input into commercial AI systems as potentially contributing to competitive training data, and reconsider sharing preliminary results through such platforms until clearer legal and ethical safeguards exist
- The incident signals a broader trend where AI labs operate as rapid-response research entities, capable of pivoting massive resources based on rumors—a dynamic that could fundamentally reshape academic publishing, priority claims, and the economics of scientific discovery
Disclaimer: The above content is generated by AI and is for reference only.