OpenAI recently announced that an internal AI model had solved part of the Navier-Stokes equations, notoriously difficult maths problems that have gone unsolved for decades. The breakthrough was immediately overshadowed by a dispute. NYU professor Tristan Buckmaster alleged that OpenAI had access to unpublished, related work by him and Levent Alpöge (an Anthropic employee researching in a personal capacity) which had been fed into OpenAI’s tools.
OpenAI’s Deus ex machina moment
Buckmaster and Alpöge had reportedly been working on the Navier-Stokes and Euler equations for many months and by late August had obtained significant results. The pair used AI, including OpenAI’s Codex tool and Astra model, heavily in their research.
Around early September, OpenAI deployed a new model on the maths problems as a way of testing the capability of the unreleased tool and after hearing rumours that others had made significant progress towards solving the equations.
Reportedly after just 88 hours the bots had succeeded. All 10,000 of them having exchanged around 3 million messages and 300 billion output tokens (the lines of text and code an AI model produces in answers) at a cost of around $22.5m.
With rumours about solving the unsolvable still circulating, discussions apparently took place between Buckmaster and OpenAI about the specifics of what maths problems had been solved, authorship and attribution. Here the two sides’ accounts differ.
OpenAI has said it didn’t see any of Buckmaster and Alpöge's work until they released it publicly and that the precise mathematical results differ. Equally the company has also said it can’t rule out that de-identified data derived from the pair’s use of OpenAI’s products helped them improve their models.
It will take a long time to unpick this dispute which happened in the US, but the issues it raises translate directly to the UK, and UK tech clients should not assume answers here would be any clearer.
Legal conundrums
AI models are trained on the accumulated work of many (often uncredited) humans, and it is often impossible to verify how much of a given output derives from any one source. So even if OpenAI’s model wasn’t trained or otherwise informed on Buckmaster and Alpöge's work, was it simply completing the final steps of a proof after borrowing ideas from a variety of human mathematicians’ work? That creates layered problems around attribution, provenance, and confidentiality whenever AI tools are used in R&D work, particularly where third-party AI tools have been used that may learn or retain data from what is typed into those models.
This example also raises practical and ethical questions for research programmes. In a world where being first to publish carries real consequences in academic credit, commercial advantage, and potentially patent priority, AI assisted work is likely to become more and more commonplace. It will be interesting to see if Buckmaster and Alpöge's experience (and others like it) will make researchers in academia and industry more cautious about what AI tools they use in their work.
For now, the UK government has decided not to legislate on issues like the transparency of training data, or the application of copyright or patent law to AI systems. That leaves clients relying on contract and internal process to manage who owns and gets credited for AI-assisted output.
Practical tips:
Keep a paper trail of human input in AI assisted R&D from day one, not just in a dispute - Document who defined the problem to be solved, made key judgment calls, and reviewed or validated AI output. This is the evidence you'd need to establish authorship, inventorship, or (if it comes to it) that your work was independently created rather than lifted from someone else's research.
Treat AI tool exposure as a confidentiality risk as well as a contract question - Unless you have an agreement with a third-party AI provider about their use of the data you input into the system, your protection against them exploiting your unpublished work rests on the law of confidentiality and trade secrets, which are notoriously difficult to prove. To avoid this, make sure you understand any terms between you and the provider about what the tool can retain, train on, or infer from your data. Shop around to find the tool and the terms that fit your needs. Spend time doing this before sharing anything novel or commercially valuable via your AI inputs.
Agree attribution and ownership upfront in genuine collaborations - Where you do have a formal partner, whether an external researcher, university, or another company, settle in advance how credit and IP ownership will be handled if AI-assisted work leads to a publication or patentable invention. This will not help against a stranger racing you to the same result, but it removes ambiguity between people who are meant to be on the same side.
Be aware of asymmetric competition - If your work has commercial or reputational value before it's published or filed, consider accelerating disclosure, filing any possible patent applications early, or using confidentiality measures precisely because a well-resourced competitor using AI can move from rumour to result far faster than you can.
Don't wait for the law to catch up - Government reform of AI and IP rules is unlikely in the near term. Businesses that plan around these risks now, rather than assuming IP law will protect them, will be far better placed than those hoping for legislative certainty later.
If you're affected by any of the issues in this article, please contact Felicity Lush or a member of Mills & Reeve’s IP team.
Our content explained
Every piece of content we create is correct on the date it’s published but please don’t rely on it as legal advice. If you’d like to speak to us about your own legal requirements, please contact one of our expert lawyers.