Deep Dive — Math's credit system was built for humans. AI just broke it.

A Millennium Prize problem appears to have fallen, and the interesting part is not the proof. It is that the institutions mathematics uses to decide who discovered something — priority, publication, peer review, the two-year wait before a prize — all assume a human pace and a human author. OpenAI's Navier–Stokes result went from rumor to Lean-verified in five days. Nothing in the rulebook is calibrated for that, and the fight that erupted is what happens when a discovery outruns the machinery meant to credit it.

Here is the sequence, because the timeline is the argument. On September 1, OpenAI says it heard rumors that two Millennium Prize problems had been resolved and pointed an internal model at all of them. The agents reached their answer on September 5, roughly 88 hours after launch, with another 17 hours of Lean formalization by GPT-6 Astra. Across the whole effort the agents exchanged 4.9 million messages and consumed about 300 billion output tokens. On September 8 OpenAI published. Later the same day, NYU mathematician Tristan Buckmaster published a statement saying he and Anthropic researcher Levent Alpöge had been working the same problem for nearly a year, had made real progress a month earlier, and that information about their approach had reached OpenAI before OpenAI started. We covered the announcement and the objection as they landed in OpenAI's AI solved Navier-Stokes — and a mathematician cries foul.

What has become clear since is that this is a structural problem wearing a personal dispute's clothes.

The rules assume a paper, not a swarm

The Clay Mathematics Institute's rules for the prize are short and specific. Before it will even consider a proposed solution, three conditions must hold: publication in a qualifying outlet, at least two years elapsed since that publication, and "general acceptance in the global mathematics community." CMI does not accept direct submissions and will not offer further guidance.

Every one of those conditions was designed for a world where a proof arrives as a paper, in a journal, written by people whose names go on the front. A lab blog post plus a GitHub repository of Lean code is none of those things, and OpenAI has said it does not intend to claim the $1 million anyway — which quietly removes the one adjudicator with a defined procedure from the process entirely. The prize is not the point; the point is that the only institution with standing to declare a Millennium problem solved has rules that a machine-generated result cannot easily satisfy, and the result's author has already opted out.

Two years is also doing more work than it looks. That waiting period exists because hard proofs take time to be understood, and because a community needs to absorb a result before it can be said to accept it. It presumes the scarce resource is human attention arriving slowly. Here the scarce resource was compute, and compute was not scarce for long.

Verification is the actual bottleneck — and nobody is funding it

The tempting reading is that Lean solves this. OpenAI says the proof has been formalized in Lean, and formalization is a genuinely strong guarantee: a machine has checked that each step follows from the previous ones. But a formalization certifies the chain of reasoning, not the framing of the problem, and not whether the formal statement corresponds to what anyone meant.

We have already seen how this fails in practice. When OpenAI published its ten-advances mathematics paper in August, an independent audit found a polarity error in a lemma's proof — a continuation test written in terms of average success where the next step needed a coordinate with large conditional failure probability. The statement of the lemma was correct; the proof as printed did not support it, and the authors had to supply a corrected proof. That is the model for what real checking looks like: slow, unglamorous, done by someone with no stake in the outcome, and published after the announcement rather than before it.

The same pattern showed up on the prime-gap result. The Lean formalization there was explicitly conditional, resting on axioms unproved inside Lean and marked self-assessed — as we wrote when it landed, the part that could be machine-checked was machine-checked, and the assumptions underneath it were not.

A recent survey of autonomous research agents puts numbers on the gap. Screening 125 candidate works and fully coding 26, it found that among 24 runnable systems, 83% release code but only 38% release seeds or execution traces and 38% report any method for verifying novelty. Among nine closed-loop systems, seven were mechanical reruns and one was author-claimed with no external check. The authors' conclusion is the one that matters here: the bottleneck is no longer whether agents can complete research tasks, but whether reviewers can verify the claims those agents produce.

That is an unfunded mandate. Peer review is already voluntary, unrewarded labor performed by the same people whose own research is being scooped.

The provenance question has no answer, and that is the answer

Buckmaster's specific allegation is narrow and serious: that he asked whether OpenAI's agents had access to the Codex sessions where he and Alpöge had been putting their drafts, was told the model did not look up user data, asked again about training, and received no answer. OpenAI's own post concedes the point in a sentence that every enterprise customer should read: "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

Whether or not it happened here, the structural fact is that a lab can no longer prove it didn't happen. A model trained on user data cannot be interrogated about what it learned. That is a due-process problem for a field whose entire currency is knowing who thought of something first — and it is the same evidentiary wall we keep hitting elsewhere: after an AI breakout this month, nobody had clear legal power to investigate it, and OpenAI's proposed self-disclosure framework has no outside enforcer.

Brown's Javier Gómez-Serrano put the resource asymmetry plainly: "What is clear is that very few mathematicians will have resources of that scale." Buckmaster and Alpöge worked for nearly a year, using publicly available models from both labs, and broke a simplified version of the problem. OpenAI, hearing a rumor, applied roughly 10,000 concurrent agents and finished the full version in days.

The contrarian case: taste still came from somewhere

There is a version of this that is less alarming than it sounds, and it deserves a hearing.

Both groups converged on an approach to Navier–Stokes pioneered by Diego Córdoba and Luis Martínez-Zoroa. MIT Technology Review notes that this was one of several approaches considered promising, so independent arrival is not impossible. But if the agents followed it because two humans had already chosen it, then human judgment about which direction to attack — the thing researchers call "taste," and the thing AI is supposed to lack — was load-bearing in the result. That would mean the model supplied execution, not insight, and the credit problem is one of attribution rather than replacement.

The strongest version of the skeptical case comes from Terence Tao, and it is not about credit at all. Writing last week, he argued that most pure-math problems are posed not because anyone desperately wants the answer but because human attempts to answer them reliably generate new mathematics. "Prematurely solving the problem by purely AI-powered methods — particularly without full transparency into the solution process — can contaminate this process to the point where it actually becomes a net negative for the progress of mathematics as a whole."

That reframes the damage. The loss is not that machines get the medal. It is that the wrong turns, the abandoned approaches and the partial results — the things that historically seed new subfields — are discarded or kept private. When a lab releases a clean proof and not the 4.9 million messages of failure behind it, the field gets the answer and loses the map.

What to watch

Four things will decide whether this is a one-off dispute or a permanent change.

First, whether OpenAI releases the agent transcripts. It has the receipts — the messages, the dead ends, the prompts. Publishing them would cost nothing but embarrassment and would settle the provenance question; not publishing them leaves the allegation permanently unfalsifiable.

Second, whether the Clay Mathematics Institute says anything at all. It has a defined procedure and no obligation to comment on a result nobody has submitted to it. Silence is the likely outcome and also the worst one.

Third, whether any independent group formalizes the proof from scratch. A second Lean development, built by people unaffiliated with OpenAI from the mathematical writeup rather than from the agents' output, is the only check that would actually settle this. Nobody has announced one.

Fourth, whether journals adapt. A two-year acceptance window assumes human-scale verification capacity. If machine-generated results arrive faster than volunteer reviewers can absorb them, the bottleneck moves to the journals, and they are the least equipped institution in the chain to handle volume.

The uncomfortable summary: mathematics can now produce results faster than it can certify them, and every mechanism it has for deciding who gets credit was designed for a slower era. The Navier–Stokes fight is not really about who solved it. It is about whether a field built on the assumption that anyone can check anyone else's work survives the arrival of work that only some people have the compute to check.

Should a lab that publishes an AI-generated proof be required to release the full agent transcript? Tell us in the comments.

Sources: OpenAI — On the Navier–Stokes Millennium Prize Problem · Clay Mathematics Institute — Rules for the Millennium Prize Problems · MIT Technology Review — What OpenAI's latest controversy tells us about the future of math · Scientific American — OpenAI claims blockbuster math breakthrough amid swirl of controversy · arXiv — Autonomous Research Agents: A Survey of AI Scientists and the Verification Gap