A letter appeared as a guest post on Terence Tao's blog with the title: "Why I didn't sign the Fields medallists' letter." Tao himself was one of the letter's 25 signatories.
Hosting a dissenting opinion on his own turf was itself the third act of September's drama.
Quick recap of the backstory. On September 8, OpenAI announced that its internal multi-agent system had produced a complete proof of the Navier–Stokes blow-up scenario—10,000 agents running for 88 hours, with Lean formalization attached. TechCrunch later tallied the compute bill for that week: 300 billion output tokens, roughly over $20 million. A priority dispute erupted the same day—NYU's Buckmaster alleged the approach overlapped heavily with his group's unpublished progress, and claimed OpenAI mathematician Bubeck had demanded the removal of a collaborator's name from authorship, even allegedly saying "do you want to ruin your career?"; Bubeck denied everything. On September 10, OpenAI withdrew sponsorship of a math event at Caltech. On September 11, 25 Fields Medalists issued a joint declaration whose core accusation was that AI companies were treating the solving of famous math problems as a benchmark competition—rushed announcements, inadequate citation—seriously misaligned with the goals of the mathematical community.
The holdout is Timothy Gowers, 1998 Fields Medalist, a major figure in combinatorics, author of the famous essay "The Two Cultures of Mathematics," and initiator of the Polymath mass-collaboration experiment. Another non-signer was Tsimerman. By contrast, the signatory list reads like living history of mathematics: Deligne, Scholze, Viazovska, June Huh, Maynard, Kontsevich, and this year's newest laureate Deng Yu.
Gowers opened by positioning himself carefully: I agree with most of the letter, the crisis is real, I simply want to offer an alternative analysis of the crisis. I do not want the mathematical community to fracture.
He gave five reasons, each more piercing than the last.
1. Temperament. The letter's logic demotes "solving famous problems" to a proxy metric for conceptual understanding—solving is merely a means; understanding is the end. Gowers cited his own essay from over two decades ago: the math world has always contained problem-solvers and theory-builders, and this is a spectrum with legitimate mathematics at both ends. Declaring one end of the spectrum "wrong" tells a large group of mathematicians that their lifelong temperament has been struck from the register.
2. The digestion crisis is not as scary as claimed. A thousand AI proofs a year works out to roughly thirty per subfield, and the math community's digestion is inherently massively parallel. He offered a streaming-media analogy: massive content floods in, yet both the undigested important mathematics and the properly digested mathematics are growing. His exact words: this seems like a pretty good deal.
3. The credit system will collapse anyway. Once AI drives down the cost of "finding a striking proof," the event will become like an amateur discovering a comet—congratulations-worthy, but no longer counted as an intellectual achievement. No one needs to push the credit system over; it will fall on its own.
4. His real worry is people, not proofs. Many young people who would have pursued math PhDs and become stewards of the discipline may no longer come. Policymakers, following the "AI is better than mathematicians" narrative, may cut funding. Once the chain of transmission breaks, the repair timescale is decades.
5. The most destabilizing point: "I'm not sure the letter asks for anything that isn't already happening." The letter makes no executable demands, and an open letter cannot stop any company from releasing a model—at most it forces a pre-written apology. Acknowledging that change is underway and finding ways to respond is more useful than posturing.
He also disclosed his interests plainly: he has ties to OpenAI's math group and has seen early models, but received no money. And there's a small easter egg at the end—the article notes that its first draft's format conversion was done by AI. The most gracious dissenting voice in this entire debate used AI to tidy up his own manuscript.
The signatures keep climbing: as of September 18, endorsements verified via ORCID or academic email on the declaration page had passed 7,400. And September's whole debate has unfolded almost entirely on Tao's blog—over twenty related posts, from benchmark-ification to fast math versus slow math, making that platform a de facto public hearing. The Hacker News discussion was also high quality. Some defended the letter, arguing the declaration targets less the technology than OpenAI's conduct. Some defended Gowers, citing Maynard's tightening of prime gaps—human breakthrough proofs likewise become course material overnight; methods and understanding don't only grow in slow labor. Others sided directly with Gowers's "pretty good deal," considering it the fulcrum of the whole piece.
Both sides carry Fields-level intellectual backing—that's what makes this genuinely worth watching. The disagreement makes neither side look foolish; instead it maps the true shape of the problem: the real point of contention has shifted. Whether AI can prove things is no longer the question; who pays the verification bill and who carries the transmission to the next generation—those are the questions. A declaration with 7,400 signatures, and the dissenting vote of one of the strongest minds hung alongside it—the mathematics community's negotiation has only just entered its main text.
And the first sentence of that main text belongs to Gowers: figure out exactly what we want before deciding whether to stand in the way.
--- Sources: Tao's blog guest post "Why I didn't sign the Fields medallists' letter" (Timothy Gowers, 2026-09-17); the declaration and signature page at mathandai.org; coverage by The Economist and Scientific American; HN discussion 49738091.