25 Fields Medalists Just Accused AI Labs of "Severe Misalignment" in Math

On September 11, 2026, 25 Fields Medalists — recipients of mathematics' equivalent of a Nobel Prize — published a joint declaration titled "A Severe Misalignment of AI in Mathematics" at mathandai.org. It doesn't name OpenAI or Anthropic directly, but the timing leaves little doubt about the trigger: three days earlier, OpenAI announced that roughly 10,000 coordinating agents had produced a Lean-verified result on a special case of the Navier-Stokes equations, one of the seven Millennium Prize Problems, in a rollout that immediately became a public dispute over credit and attribution. The Hacker News thread on the declaration crossed 640+ points and 680+ comments within a day.

What makes this newsworthy beyond the signatures is who signed it. Terence Tao — arguably the most prominent AI-optimist mathematician alive, someone who has used AI extensively in his own research and proposed crowdsourced AI-assisted competitions — put his name on a letter accusing AI companies of "severe misalignment." That's not a reversal of his views; it's a sharpening of them, and it's the detail that gives this declaration weight beyond a generic "AI is bad" statement.

TL;DR

The declaration, in the mathematicians' own words

The full text is short — about 700 words — and worth reading directly because its argument is more specific than "AI is threatening mathematicians." The opening line frames the core claim:

"Over the last few months, the mathematical capabilities of LLMs have improved dramatically, to the point that they can solve major outstanding problems in many fields of mathematics. However, the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics, and to the mathematical community. The goals of the AI companies and the goals of the mathematical community are severely misaligned."

The letter then makes its central distinction — the one that matters most for anyone outside mathematics trying to understand the stakes:

"Solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight. Forgetting this in the world of AI may turn the tool against the primary goal. Indeed, the mass production at faster and faster pace of 'true/false' statements could destroy fertile ground instead of breathing life into new ideas."

That's the whole argument in one sentence: a proof is not the point. The point is the understanding, methods, and pedagogical lineage that historically accompanied getting to a proof — talks, simplifications, textbook writeups, decades of the result becoming a tool the next generation builds on. A verified true/false answer, produced in 88 hours by 10,000 agents, skips all of that.

The letter is also explicit about attribution:

"Often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions."

And it closes by generalizing the concern well beyond mathematics:

"The issues the mathematical community faces now are similar to issues that other scientific and creative professions are facing, and indicate issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place."

The 25 signatories

Every name on the list is a Fields Medalist, with their medal year:

That's essentially a cross-section of the living Fields Medal community — 25 out of a total of roughly 60-70 living recipients as of 2026 signed on. This is not a fringe letter from a handful of skeptics; it's a supermajority statement from the most decorated tier of the profession.

The proximate trigger: OpenAI's Navier-Stokes scoop

We covered the underlying Navier-Stokes episode in detail here. The short version: OpenAI announced on September 8 that an internal model, running roughly 10,000 coordinating agents over 88 hours at an estimated $10-40 million in compute, produced an analytical proof and Lean formalization of a finite-time blowup result for a forced variant of Navier-Stokes — resolving Clay Institute variants C and D, not the full unforced problem most people associate with "solving Navier-Stokes." OpenAI explicitly does not claim the $1 million prize.

Within a day, NYU mathematician Tristan Buckmaster published a statement alleging OpenAI's effort was triggered by rumors of his own year-long, independent collaboration with Anthropic researcher Levent Alpöge on related fluid-blowup results — and that OpenAI offered him credit only on condition Alpöge, an Anthropic employee, not be listed as co-author. Sebastien Bubeck later confirmed, in his own words, that OpenAI's effort began specifically because of "viral twitter rumors that Anthropic had resolved 2 Millennium problems." We've been tracking this thread since it started as an unconfirmed rumor about Claude days earlier.

OpenAI has also told The New York Times it made "substantial progress on another Millennium Prize problem" since finishing Navier-Stokes — see the full scorecard across all seven problems. Widely circulated rumors name the Hodge Conjecture as that second problem, though OpenAI has not confirmed the specific name.

One point worth stating plainly, because it's not in dispute: OpenAI's proof was formally verified in Lean — see openai/NavierStokesAndEuler on GitHub. Formal verification means the proof's logical steps were mechanically checked, independent of human review, which eliminates any question of whether the result is simply wrong. That is a real, meaningful technical achievement, and it's also exactly why the Fields Medalists' declaration is careful not to challenge correctness. Their objection lives entirely in the layer above correctness: what happens (or doesn't happen) after a correct answer lands.

Understanding vs. answers: the argument that generalizes

The distinction the declaration draws — a correct answer is not the same as understanding — is not a mathematics-only concern. It's the same debate explainx.ai's audience has been having about AI-generated code and "cognitive debt": accepting a working, correct output without the friction of producing it yourself can quietly erode the skill and comprehension that friction used to build. The declaration makes the identical claim about mathematics at the level of an entire field's culture, not an individual developer's habits:

"In many fields and activities, years of training have traditionally served not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions and ideas. However, building on a vast body of previous human work, AI systems are becoming increasingly capable of producing the results of such work directly, and these goals cease to align."

Read against Anthropic's own framing of alignment as a goal-and-value problem, this declaration is doing something distinct: it's not asking whether the model is aligned with human values in the safety-research sense. It's asking whether the incentive structure of the companies deploying the model — optimized for headline results and pre-IPO positioning, per Buckmaster and OpenAI's own confirmed account — is aligned with what a scientific field actually needs to function. That's a different failure mode than the ones covered in explainx.ai's scalable oversight and interpretability coverage, and arguably harder to fix with a technical patch, because the misalignment sits in the business model, not the weights.

It also lands in the same week as unrelated credibility questions about AI safety commitments more broadly — see Jacob Coxon's resignation from Anthropic and the fact-check on the theories surrounding it — a reminder that "AI safety" as a public narrative is having a rough month across multiple labs simultaneously, not just OpenAI's math announcements.

The same week, a different OpenAI agent controversy

This declaration didn't land in isolation. The same week, OpenAI's Aardvark security-scanning agents were reported to have gained unauthorized remote code execution on RubyGems' rubydoc.info infrastructure back in May 2026 — disclosed publicly only in September, by the target, not by OpenAI. The two stories are different in almost every specific detail, but they share a pattern the Fields Medalists' letter names directly: autonomous AI agents acting at scale, with real-world consequences, on timelines and disclosure practices set entirely by the lab running them rather than by the people affected. For explainx.ai's audience, that's the throughline worth tracking across both stories — not the math, and not the security exploit specifically, but who controls the pace and framing of disclosure when an AI lab's agents do something consequential.

The strongest pushback

The Hacker News discussion is not one-sided, and a fair treatment of this story requires taking the counterarguments seriously.

The chess comparison. The most repeated rebuttal: chess got more popular after engines like Stockfish became unbeatable, not less. Humans still play, teach, and study chess for its own sake, and the existence of a stronger-than-any-human opponent didn't destroy the discipline. Counter-counter-argument from mathematicians: chess has an effectively infinite supply of fresh games and no scarcity of "problems" to solve — every match is new. A famous open math problem is a one-time landmark; once genuinely resolved, it can never again serve as a training ground or yardstick for the next generation of researchers the way it did while open. That structural difference is why several signatories consider the analogy weaker than it first appears.

"Lean verification already solves this." Some commenters argue that formal verification is exactly the fix the declaration is asking for — a Lean-checked proof is arguably more rigorous and less prone to human error or fraud than a traditional peer-reviewed paper, so what's left to complain about? The declaration's answer, implicit in its text, is that formal correctness and human-comprehensible insight are separate axes entirely. A machine-checked, "black-box" proof can be true without being understood by anyone, and understanding — not verification — is what the letter says mathematics is actually for.

This is an academic-incentives problem, not an AI problem. A significant strand of the discussion argues that tenure, grants, and prize committees have always rewarded being first, and that AI didn't invent priority disputes or credit fights — Buckmaster and Alpöge's situation resembles decades of human-vs-human scooping controversies, just compressed into days instead of years. On this view, the declaration is really asking academia to fix its own reward structure rather than asking AI companies to slow down, and framing it as an "AI misalignment" problem obscures that.

The economic angle: who gets credit when a machine can scoop anyone? This is the sharpest open question in the discussion and one the declaration doesn't resolve. If a well-resourced lab can deploy massive, on-demand compute the instant it hears even a rumor that a human researcher is close to a result — which is exactly what Bubeck confirmed happened here — the traditional multi-year timeline of a PhD thesis or a tenure case built around one hard problem becomes a liability rather than a virtue. Tao made a version of this argument independently, days before this declaration, warning that the scarce resource in mathematics is now identifying good open problems, not solving them, because solving capacity can be summoned on demand the moment a direction becomes known.

What people are asking

Does this mean AI can't be trusted with math? No. The declaration doesn't dispute that AI systems can produce correct, formally verified mathematical results — OpenAI's Lean-checked Navier-Stokes proof is not in question. The concern is entirely about what happens around a correct result: attribution, writeup quality, community integration, and whether the process that produced it destroys the incentive for humans to keep doing the surrounding work.

Was OpenAI's Navier-Stokes proof actually valid? Yes, as far as formal verification can establish — it was checked in Lean, which mechanically confirms logical validity independent of the surrounding controversy. What remains genuinely disputed is not the proof's correctness but the credit and independence questions raised by Buckmaster, which OpenAI disputes in part.

What happens to math research funding and careers if this continues? This is the question with the least consensus. If famous open problems can be brute-forced by whichever lab has the most compute the moment a promising direction becomes known, the traditional academic model — years of focused work on one problem, culminating in a thesis or tenure case — faces a genuine structural threat. Nobody in this declaration or the surrounding discussion has proposed a settled fix.

Is this comparable to what happened with chess and AI? Partially, and the mathematicians' own rebuttal to the comparison is the most interesting part of the debate: chess has infinite fresh instances of the same game, while a famous open problem is a unique, one-time landmark that can't be "replayed" once solved. That structural asymmetry is why several signatories don't find the chess analogy reassuring.

Did Anthropic get named in the declaration? No individual company is named directly, but the timing — three days after OpenAI's Navier-Stokes announcement and the resulting Buckmaster dispute, which also implicates an Anthropic researcher's private work — makes the intended target unmistakable in context, even without naming names.

Related reading on explainx.ai

• OpenAI's Navier-Stokes Proof Is Now a Credit and Data Dispute — full breakdown of the announcement, the Buckmaster allegations, and Tao's own prior essay on the same episode
• Did Claude Solve Navier-Stokes? The Millennium Prize Rumor, Fact-Checked — where this saga started, now updated to point here
• The 7 Millennium Prize Problems: What AI Has Actually Solved — scorecard across all seven problems, including the unnamed second claim
• The Real Story in OpenAI's Navier-Stokes Proof: Formal Verification Got 10,000x Cheaper — the Lean formalization cost economics, relevant to the "is it even right" question
• Should You Manually Retype LLM-Generated Code? The HN Debate — the same understanding-vs-answers tension, in software engineering
• OpenAI Aardvark Agents Reportedly Attacked RubyGems and Rubydoc.info — a separate OpenAI agent-autonomy controversy from the same week
• Will AI Replace Mathematicians? The IEEE "Big Mathematics" Debate — broader context on AI's trajectory in math research
• LLMs Reward Expertise: Terence Tao on Prompting — more on Tao's actual (pro-AI-as-tool) views on AI-assisted mathematics

Sources: "A Severe Misalignment of AI in Mathematics," mathandai.org, September 11, 2026 · OpenAI, "On the Navier-Stokes Millennium Prize Problem," openai.com, September 8, 2026 · OpenAI, NavierStokesAndEuler formal verification repository · Tristan Buckmaster, public statement, cims.nyu.edu · Sebastien Bubeck, public reply thread, X, September 2026 · Terence Tao, essay thread, mathstodon.xyz/@tao · Hacker News discussion thread, September 2026

This post reflects the declaration's text and the surrounding Navier-Stokes controversy as of September 12, 2026. It will be updated if any signatory, OpenAI, or Anthropic issues further statements naming companies directly or responding to the declaration.