On 11 September 2026, Terence Tao published a declaration titled "A Severe Misalignment of AI in Mathematics", signed by 25 initial signatories, all Fields Medallists — among them Tao, Peter Scholze and June Huh.

The sentence at its centre: "The goals of the AI companies and the goals of the mathematical community are severely misaligned."

Tao says he facilitated the statement after discussion among the signatories, and that the urgency called for a faster process than mathematicians would normally use.

The objection is about the proxy

It would be easy to read this as senior figures in a field objecting to being automated. That is not what the text says.

The declaration's argument is that "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight."

That is a precise claim, and it is one this site keeps running into in other forms. A discipline picks a measurable thing that correlates with the thing it actually wants. Open problems solved is a good proxy for mathematical understanding, in the same way that patch counts are a proxy for security or benchmark scores are a proxy for capability. Optimise the proxy hard enough and it separates from what it was standing in for.

An AI system that announces a solution to a famous open problem scores maximally on the proxy. Whether the field understands anything new afterwards depends on things the proxy does not measure.

What actually gets skipped

The declaration is specific about what is lost when results arrive at speed. Solutions announced hastily skip "proper writeup, the isolation of new methods and ideas, and citing relevant previous work."

Each of those is the part that makes a result useful to anyone other than its author:

  • The write-up is how a proof becomes checkable by people who were not there.
  • Isolating the method is how a one-off answer becomes a technique other mathematicians can apply to different problems.
  • Citation is how a result is placed in a line of work, which is both a credit mechanism and a map for whoever comes next.

Strip those away and you have an answer without a discipline around it. The declaration calls for urgent action from the mathematical community, from AI companies, and from society, to keep sight of the original purpose as the tools change the practice.

The week it landed in

The declaration did not appear in a vacuum. The same week, OpenAI withdrew its sponsorship of a Caltech mathematics competition after criticism from mathematicians about the quality of machine-generated proofs.

It also lands in a month when AI researchers themselves have been arguing publicly about pace — OpenAI went as far as asking Congress whether an industry-wide slowdown would be legal. The mathematicians are making a narrower point than the safety debate, and a more concrete one: not that the technology is dangerous, but that the scoreboard being used to demonstrate progress measures the wrong thing.

The fair counter-argument

Worth stating, because the declaration does not settle it.

Machine assistance has already produced real mathematics, formal verification systems have caught errors humans missed, and a solved problem is not worthless merely because a model solved it. Several signatories have themselves written about using these tools productively. The complaint is about what gets rewarded and announced, not about whether the tools work.

The open question is whether AI companies have any incentive to optimise for the slower thing. Conceptual insight does not produce a launch post.

What is not established

  • Whether the 25 signatories represent a broader consensus. The declaration describes them as initial signatories.
  • What, concretely, the declaration asks companies to do. It calls for action without specifying mechanisms.
  • Whether any AI company responds.
  • How many recent machine-assisted results would fail the write-up, method-isolation and citation tests the declaration sets out.