The email arrived at 2:13 a.m., subject line: “We solved Navier‑Stokes.” I stared at the words and felt the floor shift under a room full of quiet proof-checkers. For many of us, that was the moment the theoretical began to behave like a practical threat.
I’m going to walk you through what happened, whom it hit, and why this matters to anyone who cares about how knowledge is made. You’ll see the lines where tech power meets academic practice, and why people like Tristan Buckmaster and Andreas Thom are no longer theorists in a vacuum but actors in a public fight over credit, control, and trust.
A midnight email landed in dozens of inboxes — and a prize problem suddenly had a corporate voice
OpenAI posted a solution to the Navier‑Stokes smoothness and existence problem on September 8. The announcement read like a headline meant to settle argument, not start one; that tone is what set off alarms. When an AI lab says it has cracked a near‑century puzzle, it touches nerves that go beyond mathematics: prestige, publication norms, and who gets named in history.
The immediate reaction from parts of the mathematics community was fury. NYU’s Tristan Buckmaster accused OpenAI of trying to force him to publish under the company’s terms after he and a coauthor had been close to releasing related work. OpenAI’s response—offering Buckmaster partial credit while allegedly trying to exclude his coauthor, Levent Alpöge of Anthropic—read as an attempt to rewrite authorship rules overnight.
A colleague noticed uncanny overlap — and asked, did the model see our work?
Andreas Thom of Dresden University of Technology spotted a solution from OpenAI that echoed recent work he and Gábor Kun had been testing with ChatGPT. He reached out. The question was simple: was our material used to train the model, or did the model see private drafts?
Did OpenAI use researchers’ private work to train models?
OpenAI’s public reply, via researcher Mark Sellke, was a flat “that did not happen,” and later a softer statement: no specific user data was seen by the human or agent teams, though de‑identified derivatives might have helped model improvement. Thom called the initial denial misleading. “You should say that you don’t know,” he told me. His point is procedural: in moments of high stakes, precision matters more than performance.
A coauthor felt steamrolled — and said the game had changed for mathematicians
Tristan Buckmaster told the Australian Broadcasting Corporation he felt defeated: “I think it’s pointless. Like, I think the game is up.” That sentiment rippled. When a researcher at NYU chooses resignation over revision, you can’t frame this as merely academic drama.
Will AI replace mathematicians?
Short answer: not like a factory machine replaces a human line worker. Long answer: AI models amplify certain capacities—pattern matching, brute-force search, heuristic testing—at speeds humans can’t match. That shifts the job, not always in ways that leave the human role intact. You and I both know that the value in mathematics lies in framing questions and interpreting meaning, but those tasks are under pressure when models produce publishable output.
Fields Medals weighed in — and the community demanded a say
Twenty‑seven Fields Medalists signed a statement arguing that AI companies using mathematical problems as benchmarks harms the science and raises “severe attribution and plagiarism questions.” That’s not a fringe protest; it’s the field’s highest honor speaking to the rules of scholarship.
What is the Navier‑Stokes problem?
Navier‑Stokes asks whether, for three‑dimensional fluid flow, smooth solutions always exist for all time or whether singularities can form. It’s a claymore of a problem: conceptually simple, devilishly resistant to proof. When a private lab claims a solution, the claim forces a public verification process that’s normally a slow, collegial grind.
OpenAI said it would form an advisory group with independent mathematicians to give the community a voice. That’s a step, but it’s also a bandage on a wound that cuts into reward structures, publishing norms, and the unpaid labor that feeds large models.
A hallway conversation at a conference made the risk personal — and academic culture is shifting
At a recent conference I overheard a graduate student ask a senior professor whether to keep certain notes private. The professor shrugged and said, “Once it’s on the web, it may be used against you.” That kind of caution changes how research gets shared and taught.
Andreas Thom told me we must “come together as a mathematical community to redefine our subject, to redefine what math students are supposed to learn.” He isn’t surrendering; he’s proposing a reappraisal. Mathematics now feels like an old library where books are being quietly re‑shelved by a stranger — and the librarians must decide new rules for access.
Here’s what I think is clear: AI companies—OpenAI, Anthropic, and others—are testing business models that intersect directly with the academic reward system. ChatGPT and similar agents have been used as tools by researchers, and that usage creates gray zones. When you train models on vast swaths of human creativity, attribution gets messy. When those models then publish “solutions,” we face an identity crisis about who deserves credit.
Some institutions will push for stricter data governance, journals will adjust authorship policies, and funders may demand clearer provenance for claims. The mathematics community is already moving—Fields Medalists signing statements; OpenAI forming advisory groups; individual researchers calling out opaque practices. You can see the momentum, but the rules will take time to settle.
I’ll leave you with one blunt question: if an AI model solves a problem because it mirrors centuries of human insight, should the victory be celebrated as a machine triumph, a community achievement, or a corporate score—and who gets the prize money, the paper credit, and the historical footnote?