OpenAI’s 88-Hour Navier–Stokes Proof Sparks Fight Over Credit

OpenAI says roughly 10,000 AI agents, working in coordinated groups, solved one of maths’ seven Millennium Prize Problems in 88 hours. Seven weeks later, the row over who deserves credit has become the BBC’s lead example of a quieter worry. AI could wear down trust in how science gets checked.

What happened

OpenAI’s September 8 announcement says an internal model, which it calls significantly more capable than GPT‑6 Astra, produced a proof about the Navier–Stokes equations. The equations describe how fluids move. The claim is that a fluid starting smooth and at rest, with a smooth force pushing on it, can develop a singularity in finite time. OpenAI describes the shape as a vortex that spirals inward and stretches out like spaghetti.

The agents reached the result on September 5. Checking it in a machine-verifiable Lean proof took another 17 hours.

OpenAI says the push began on September 1, after it heard rumours that two Millennium problems had fallen. Those rumours traced back to NYU mathematician Tristan Buckmaster and Levent Alpöge, a mathematician employed by Anthropic. The pair had worked on this family of problems for about a year. They kept their drafts in OpenAI’s own coding tool, Codex.

That’s where it got messy.

Buckmaster described two Sunday-afternoon calls with OpenAI in a statement reported by ABC News. In his account, one option on the table had him writing up OpenAI’s result alone, leaving Alpöge off because he works for a rival lab. OpenAI’s Sebastien Bubeck says he never asked for Alpöge to be removed from his own work. Bubeck’s position is only that an Anthropic employee shouldn’t author OpenAI’s proof.

Buckmaster also asked whether his Codex sessions had fed OpenAI’s models. “I am not accusing anyone of anything,” he wrote.

OpenAI’s position has shifted along the way. Its original post said it couldn’t rule out that de-identified usage data had helped improve its models. A September 10 update replaced that line: an internal investigation had found that Buckmaster’s Codex prompts couldn’t have influenced the system. The teams also disagree on substance. ABC reports they used a very similar, niche method, while OpenAI says the proofs differ significantly.

Why it matters for anyone doing real work with AI

The maths is niche. The question underneath it isn’t. If you paste a thesis chapter, unpublished research or a client draft into an AI tool, do you actually know where it goes?

Monash University mathematician Melissa Lee told ABC that researchers need clear answers on what happens to unfinished work they put into AI systems. She expects AI companies’ promises about training data to face much closer scrutiny. The same concern applies to students submitting coursework and to freelancers handling a client’s confidential material.

The BBC widens the lens. Science journalist Dalmeet Singh Chawla told the broadcaster that generative AI has upended the research process. In his view, journal integrity teams aren’t growing fast enough to keep up. That’s the Navier–Stokes trust problem at scale, minus the million-dollar headline.

The integrity numbers behind the worry

The BBC points to several studies. A Stanford team led by Weixin Liang and James Zou found that up to 17% of peer-review text at four major computer science conferences had been substantially modified by a language model. A separate Cornell team found around 147,000 phantom citations last year, meaning references to research that doesn’t exist.

Publishers are reacting, though slowly. According to the BBC, Elsevier now uses about 100 signals to flag papers for closer human review. ArXiv, where scientists post early results, bans researchers for a year if they submit invented references.

Detection doesn’t settle anything on its own. AI-writing detectors have their own reliability problems, so a flag is a lead, not proof. Careful human checking stays the backstop, and it’s slow, expensive and already stretched.

Pushmeet Kohli, who leads science work at Google DeepMind, told the BBC that whoever runs an AI agent has to understand what it did and take responsibility for the output. That sounds obvious. The Navier–Stokes fight shows how hard it gets when the agent, the tool and the competing lab all belong to companies racing each other.

What to watch next

The Clay Mathematics Institute, which runs the $1 million prize, has said the problem appears to have been settled but that its review will take time. OpenAI says it won’t claim the prize. The next real test is whether independent mathematicians confirm that the Lean proof matches the official problem wording, not a close cousin of it. After that, watch whether AI labs rewrite their data promises for researchers before the next big proof lands.