Last week's issue traced the rising interpretive debt of AI-assisted discovery: the widening gap between what a system outputs and what any human can actually check. This week, the debt came due. On 8 September, OpenAI announced that it had solved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. A million-dollar question about whether the equations governing fluid motion can break down.
Within days, the celebration curdled into a priority dispute that says more about verification than about mathematics.
The controversy, as Matthew Sparkes reported in New Scientist, was not about whether the proof was wrong. It was about who got there first. Shortly before OpenAI's announcement, two researchers — Tristan Buckmaster at NYU and Levent Alpöge at Anthropic — rushed out their own solution to the Euler equations, a close cousin of Navier-Stokes that experts suggested could open a door to the full problem. They had heard rumours of OpenAI's imminent announcement and contacted the company for clarification, knowing their work was on the way. Buckmaster says he was met with opaque responses.
The controversy was not about whether the proof was wrong. It was about who got there first.
He then issued a statement that stopped short of accusation but noted, pointedly, that his own work was stored on OpenAI's servers — as a customer, not a research partner.
OpenAI denied any wrongdoing, stating that no person or AI from the company had seen Buckmaster and Alpöge's work, and that its solution was different in any case. Both claims may be true. That is precisely the problem. The dispute is not resolvable by the evidence anyone has made public. It is resolvable only by trusting one party's account of what its model ingested and when. This is what interpretive debt looks like when it stops being a metaphor: a priority claim in mathematics, the oldest credit economy in science, adjudicated by assertion rather than by record.
Note what OpenAI actually claimed. According to Nature, the company said it had generated a solution showing that the Navier-Stokes equations can break down. Not that they are smooth. That a singularity can form. The preprint was posted on OpenAI's website, not submitted to a journal. Martin Bridson, president of the Clay Mathematics Institute, called it "certainly an exciting day, as we contemplate the announcement of major advances in the human understanding of mathematics." Contemplate is the operative verb. The Clay Institute has not awarded anything. No referees have signed off. The mathematical community has a preprint and a press release.
This is the first time, Nature notes, that a truly major open problem in mathematics has been solved by a computer, according to OpenAI. The qualifier matters. "According to OpenAI" is doing more work in that sentence than any of the mathematics.
The verification problem compounds when the solver is not a person but a system whose reasoning cannot be reconstructed. A human mathematician who claims a proof can be interrogated: show the steps, defend the lemma, explain the trick. A model that outputs a proof can be checked only by checking the proof itself — which is fine if the proof is legible, and catastrophic if it is not. The oracle problem, as one analysis puts it, describes exactly this: we trust the output, yet we cannot explain the process. The speed of computational inference outstrips our ability to verify the underlying logic.
Philip Ball, writing in Nature, describes the test that DeepMind co-founder Demis Hassabis proposed at the India AI Summit in New Delhi this February. Train a large language model on everything known before a cut-off date — he suggested 1911 — and see whether it can reproduce general relativity. This is a clean benchmark because the answer is known. We can check the machine against history. But it is also a backward-looking test. It measures whether a model can rediscover what we already understand. It says nothing about whether we can verify what we do not.
That asymmetry is the whole game now. In a PNAS opinion piece published 9 September, Jacob Teeny, Andrew Luttrell, and Dongchan Lee name the "synthetic validity" problem in AI-generated science. The phrase captures something the Millennium dispute makes vivid: a result can be internally coherent, mathematically sound, and still fail to mean what it claims to mean in the world. Validity is not the same as consistency. A proof that no human can follow is a proof that no human can validate. The machine has answered a question we posed. We cannot confirm the answer is to the question we meant.
The infrastructure is already straining. Search engines, as one op-ed details, have stopped being conduits and become answer engines. Query fan-out breaks a user's question into parallel sub-queries, pulls material from disparate documents, and weaves a single narrative on the search interface. Organic clicks fall from roughly 15% to 8% when an AI summary appears. Citations embedded in those summaries receive about 1% of user engagement. Organic referrals dropped by nearly a third between late 2024 and late 2025. Smaller outlets and regional newsrooms have seen drops of up to 60%.
The knowledge is still being produced. It is simply no longer being traced to its source. The synthesis has replaced the citation. The answer has replaced the path to the answer. When a model generates a scientific claim, the same dynamic applies: the output circulates, the derivation does not.
So we are left with a strange inversion. The most celebrated AI result of the year is a claim about fluid dynamics that no one outside OpenAI has verified, shadowed by a priority dispute that no one outside the parties can adjudicate, announced as a breakthrough while the institutional machinery of mathematics — journals, referees, prize committees — has done nothing but contemplate. The Clay Institute has not spoken. The preprint sits on a company website.
The oracle problem is not a technical hurdle to be solved. It is a permanent feature of the landscape. The value of a scientist, as one analysis puts it, will no longer be determined by the ability to calculate but by the ability to ask the right questions of the machine. That is a lovely thought. But it assumes the machine's answers can be checked. The Millennium Problem affair suggests otherwise. The question is no longer whether AI can solve problems we cannot. It is whether we can recognise a solution when it arrives.