25 Fields Medalists Told AI Companies to Stop Using Math as a Benchmark and Hacker News Spent 1,000 Comments Picking Sides

On September 11, twenty-five Fields Medal winners published a joint declaration at mathandai.org arguing that AI companies are causing "a severe misalignment" in mathematics. The signatories span nearly five decades of the award, from Pierre Deligne (1978) to Deng Yu (2026), and include Terence Tao, the most publicly visible mathematician alive. Tao posted the full text on his blog the same day. Within 24 hours, the Hacker News thread hit 1,051 points and 1,005 comments. X picked it up as a trending topic. We read the declaration, Tao's commentary, and the threads. The argument is more specific than "AI bad" and deserves to be engaged with on its own terms.

What the declaration actually says

The core claim is precise: "The push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics." This is not about whether AI can do math. It is about what happens to the field when AI companies optimise for speed of solution as a marketing metric.

Tao writes that "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight." The declaration argues that AI-generated solutions arrive "in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work." The human process of turning a solved problem into shared mathematical knowledge (peer review, writeup, teaching, integration into the canon) is the part that actually advances the field. Skip it, and you have answers nobody understands.

The declaration's sharpest line: "Without the willing mathematicians who must take care of their development and integration into the mathematical canon, AI-conceived ideas would never become fully alive."

The argument, at its strongest on both sides

The signatories are not Luddites. Tao actively uses AI tools in his own research, and several signatories have published work leveraging AI for proof verification. Their objection is institutional, not technological. When companies announce that their model "solved" an open problem and use it to justify a funding round or a capability claim, the incentive is speed of announcement, not depth of understanding. The mathematical community then inherits an unverifiable result with no writeup, no pedagogical value, and no credit to the prior work that made the solution possible.

The strongest counter-argument is pragmatic. If AI models can solve problems that matter, the solutions have value whether or not the mathematical community processes them on its preferred timeline. Open-source models are closing capability gaps rapidly, and the mathematics that enables applications (cryptography, optimisation, machine learning itself) does not require human comprehension of every proof step. A correct answer is a correct answer.

What Hacker News argued about

The thread generated six distinct subthreads worth reading.

"Even incomprehensible AI proofs would generate productive academic discourse, conferences, and student problems."
tmhn2 on Hacker News, drawing an analogy to how the mathematical community processed Mochizuki's ABC conjecture proof, which was also initially incomprehensible to most mathematicians.

"It doesn't oppose AI use but rather advocates for alignment with mathematical community values."
GPerson on Hacker News, pushing back on headlines that framed the declaration as anti-AI.

"Terence Tao himself actively uses AI for mathematics. Dismissing him as a gatekeeping Luddite misreads the whole thing."
qlte on Hacker News

The deepest technical debate centred on whether the word "misalignment" even applies. User zozbot234 argued it misuses the term, since current AI struggles reflect capability gaps, not deliberate sabotage. User omnicognate countered that misalignment properly describes the divergence between corporate incentives and the mathematical community's actual needs. Both are right about different things: the machines are not misaligned, but the institutions using them are.

"Solving math problems is not the same as increasing understanding of math concepts."
thayne on Hacker News

"Applications don't require human comprehension of proofs. Historical technological disruption ultimately benefited populations."
eru on Hacker News, representing the acceleration camp, joined by Marha01 and others.

How the platforms diverged

This is where the story gets interesting. Hacker News split roughly 60/40 between measured sympathy for the signatories and pragmatic dismissal of their concerns. The top-rated comments generally defended Tao while acknowledging the tension. Comments dismissing all concerns or celebrating AI replacement got heavy engagement but contentious replies.

On X, the trending topic ran hotter. The framing there was binary: Fields Medalists versus AI labs. Much of the X discussion collapsed the declaration's nuance into "mathematicians afraid of being replaced," which is specifically not what the text says. The coverage from outlets like CryptoBriefing and OfficeChai leaned into the conflict framing, because conflict framing gets clicks.

The gap between the platforms is the story. Hacker News, for all its problems, produced a thread where people actually quoted the declaration before arguing about it. X and the media coverage mostly argued about what they assumed it said.

The comment worth reading twice

User robotpepi pushed back on the chess analogy ("this is like when engines beat grandmasters, and chess survived") with a point that deserves more attention: mathematics and chess differ fundamentally in economic structure and practical application. Chess survived engines because nobody builds bridges based on chess games. But mathematical understanding underpins engineering, cryptography, and the AI models themselves. If the pipeline that produces mathematicians breaks because students cannot learn from AI-generated proofs they cannot read, the downstream effects reach far beyond academia.

"Tech companies ignore externalities on research culture. Understanding, not mere solutions, constitutes mathematical progress. Institutional incentive misalignment threatens the field's foundation."
robotpepi on Hacker News

Our read

The declaration is right about the incentive problem and wrong about nothing, because it carefully avoids claiming anything falsifiable. It does not say AI cannot do math. It does not say AI should not do math. It says the speed at which companies announce AI-solved problems, without writeups, without credit, without integration into the teaching pipeline, is bad for the field. That is a narrow, defensible claim.

The harder question is whether it matters. AI companies will not slow down because 25 mathematicians asked them to. The models will get better at proofs. The announcements will keep coming. The real test is whether the mathematical community can build institutions fast enough to process AI-generated results on their own terms, rather than on the timeline of a product launch cycle.

The declaration is not asking companies to stop building smarter models. It is asking them to stop treating unsolved mathematical problems as marketing material. That distinction matters, and most of the coverage has already lost it.

Tao's credibility makes this harder to dismiss than a typical open letter. He uses the tools. He publishes with them. He is not arguing from ignorance or fear. He is arguing from a position where he has seen both sides and concluded that the speed of announcement is outpacing the speed of understanding. Whether the AI community listens is a separate question. Whether it should is not.