# OpenAI's Navier–Stokes 'solution' comes with a very large asterisk

> OpenAI solved the forced Navier–Stokes problem, not the harder unforced version mathematicians care about.

*It cleared the bar the Clay Institute actually set. It did not clear the bar mathematicians actually care about. Both things are true.*

By Bez · InsidersFeed
Canonical: https://insidersfeed.com/news/openai-navier-stokes-forced-vs-unforced

> **Key:** **The take:** OpenAI did not cheat, and it did not solve the problem people actually wanted solved. Both are true because the Clay Institute wrote its rules 26 years ago with an escape hatch built in, and OpenAI's agents found the escape hatch faster than anyone thought possible. That is a real result. It is also, pointedly, not the one the maths community has spent 90 years trying to crack.

## The escape hatch was always there — nobody's cheated

Start with the framing everyone got wrong in the first 48 hours. OpenAI's Navier–Stokes proof shows a fluid, starting smooth and at rest, breaking down under a *specifically constructed external force*. Critics jumped on this as if it were a technicality OpenAI slipped past the referees. It isn't. Charles Fefferman — the Princeton mathematician who wrote the Clay Institute's official 2000 problem statement — explicitly allowed this variant, labelled option C, as a legitimate way to resolve the prize. OpenAI's system found a real answer to a real, officially sanctioned question. The complaint that follows isn't about legitimacy. It's about relevance.

## Why the 'wrong problem' critique actually lands

Here's the uncomfortable part for OpenAI's press release. Nearly every mathematician who has spent a career on this equation was chasing the *unforced* case — no external push, just the fluid's own internal dynamics tearing itself apart. That's the version with real physical stakes: does turbulence, on its own, break the maths that's supposed to describe it? University of Chicago mathematician Luis Silvestre put it as bluntly as a mathematician gets.

> The Clay problem is settled, but the main problem for the Navier-Stokes equations is not.
> — [Scientific American](https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/), 2026-09-21

And it gets worse for the 'this changes everything' framing: nine days after OpenAI's announcement, three mathematicians posted a proof showing OpenAI's whole approach — the forcing technique, the infinite cascade — structurally cannot be pushed through to the unforced case. Not 'hasn't been' — *cannot be*, without an entirely new mathematical idea nobody currently has. Diego Córdoba, whose own technique underpins both AI results, doesn't dispute this; he simply thinks the forced version was always worth solving on its own terms.

> All fluids we know of are under some kind of external force.
> — [Scientific American](https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/), 2026-09-21

Fair point, and worth sitting with. But it doesn't change the scoreboard: the version of Navier–Stokes that keeps physicists up at night — whether a fluid can break itself, with no outside help — is exactly as open today as it was on 7 September.

| | What was actually claimed | The catch |
| --- | --- | --- |
| OpenAI's headline result | Forced Navier–Stokes blow-up | Uses a purpose-built external force — legitimate under Clay's option C, but a side door |
| The problem mathematicians wanted | Unforced Navier–Stokes blow-up | Still open; a 17 Sept proof shows this method structurally can't reach it |
| The result 12 hours earlier | Forced Euler blow-up (Buckmaster & Alpöge) | Real, Lean-verified, and ceded priority to OpenAI's louder announcement |

## The part everyone's skipping: the proofs are ugly

The credit fight got the headlines, but Tristan Buckmaster's own statement about his and Levent Alpöge's work is the most damning thing anyone involved has said about AI-generated mathematics all year — and he said it about his own team.

> I can say the first LLM generated proof Levent sent me was the most horrendous I have ever read.
> — [Quanta Magazine](https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/), 2026-09-08

That's not a rival being ungracious. That's the person who co-authored the result, describing the raw output of the tool that helped produce it. Buckmaster went further, admitting one of the three papers his team released 'can only be described as AI slop.' Lean can tell you a proof's logic is airtight. It cannot tell you whether a human being can read the thing, learn from it, or build on it — and 'a machine-checked proof no mathematician can parse' is a genuinely new category of research output nobody has a good process for yet.

> **Note:** **Credit where it's due, again:** none of this makes the maths wrong. Lean-checked is Lean-checked. The gripe is that 'verified' and 'legible' are now two separate properties a proof can have, and the announcements are optimised to make you assume the second follows from the first. It doesn't.

## The Tao problem

Terence Tao — not a bystander, one of the most decorated working mathematicians alive — used his response to CNN to raise something bigger than any single proof: what a race like this does to the discipline itself.

> The indiscriminate use of AI is turning the subject into a meaningless production quota 'game' that ultimately is of very little benefit, either to mathematics or to the world.
> — [CNN](https://www.cnn.com/2026/09/09/business/openai-millennium-problems-navier-stokes-hnk), 2026-09-09

That's the real cost nobody's pricing in alongside Bubeck's several million dollars. Two labs, on overlapping problems, racing each other to a formal-verification finish line in the same single week, is not how mathematical understanding has historically advanced — and Tao's point is that a discipline optimised for who announces first, rather than who understands deepest, quietly stops being the discipline everyone signed up for. The maths checked out. The way we got here is the part actually worth arguing about.

## Key takeaways

- OpenAI's proof satisfies the Clay Institute's own official rules (option C), so calling it a loophole is technically wrong — but it is the easier of two variants mathematicians actually care about.
- The harder, unforced version most mathematicians consider the 'real' Navier–Stokes question is untouched, and a 17 September proof shows OpenAI's method can never reach it without a genuinely new idea.
- NYU's Tristan Buckmaster and Anthropic's Levent Alpöge got a related Euler result 12 hours earlier, on a fraction of the compute — and Buckmaster called one of his own team's papers 'AI slop'.
- The technique underneath both AI results was built by humans — Diego Córdoba and Luis Martínez-Zoroa — over years, without AI assistance, before any agent swarm touched it.
- Terence Tao is on record worried this is turning mathematics into 'a meaningless production quota game' rather than a discipline built on understanding.

## FAQ

### Did OpenAI solve the real Navier–Stokes problem?
It solved the 'forced' variant, which the Clay Institute's own rules explicitly allow as a valid resolution (option C). Most mathematicians were chasing the harder 'unforced' version, which remains open — a 17 September 2026 proof shows OpenAI's method structurally cannot reach it.

### Was this a loophole or a legitimate result?
Legitimate. Charles Fefferman, who wrote the Clay Institute's official 2000 problem statement, explicitly included the forced variant as an acceptable resolution. The critique isn't about legitimacy — it's that the forced case is less scientifically interesting than the one most mathematicians actually wanted answered.

### Why did Tristan Buckmaster call his own team's proof 'AI slop'?
Buckmaster, who co-produced a related Euler proof with Anthropic's Levent Alpöge about 12 hours before OpenAI's announcement, said the first AI-generated proof he was sent was 'the most horrendous I have ever read' and that one of the three papers his team released 'can only be described as AI slop' — a comment on readability, not correctness.

### Who actually built the underlying mathematical technique?
Diego Córdoba and his former student Luis Martínez-Zoroa, working analytically without AI, starting with Martínez-Zoroa's 2021 doctoral dissertation and a joint 2023 paper. Both AI-assisted teams built on that human-derived technique rather than inventing a new one.

### Is anyone worried this is bad for mathematics as a field?
Yes. Terence Tao told CNN the 'indiscriminate use of AI is turning the subject into a meaningless production quota game' — a concern about incentives and understanding, distinct from any dispute about whether the individual proofs are correct.

## Sources

- [On the Navier–Stokes Millennium Prize Problem](https://openai.com/index/navier-stokes-solution/) — OpenAI, 2026-09-08
- [AI Has Solved One of Math's $1 Million Millennium Prize Problems](https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/) — Quanta Magazine, 2026-09-08
- [OpenAI claims huge maths breakthrough on a famed 'Millennium Problem'](https://www.nature.com/articles/d41586-026-02842-5) — Nature, 2026-09-08
- [Did OpenAI solve the wrong Navier-Stokes problem?](https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/) — Scientific American, 2026-09-21
- [OpenAI says it has solved one of math's 'Millennium Problems'](https://www.cnn.com/2026/09/09/business/openai-millennium-problems-navier-stokes-hnk) — CNN, 2026-09-09