On Thursday, 25 Fields Medallists published a declaration. Not 25 mathematicians — 25 people who hold the Fields Medal, among them Artur Avila, Pierre Deligne, Alessio Figalli, Maxim Kontsevich, Peter Scholze, Terence Tao, Maryna Viazovska and Cédric Villani. Tao writes on his blog that the text came out of discussions over the past week. There was no time for a proper consultative process like the Leiden declaration had. They judged the situation urgent enough to publish anyway.
The title is “A Severe Misalignment of AI in Mathematics”. The word misalignment is not an accident. It is the term the AI labs use about their own models, and the mathematicians are handing it back.
The charge
The declaration does not dispute that LLMs can now solve major open problems. It says so up front. The objection comes after that. Famous problems have always worked as landmarks, a way to measure improved understanding of the mathematical landscape. Solving one almost always meant new ideas and methods, and those then got worked through for years in talks, discussions and simplifications, until eventually there was a write-up a graduate student could read.
On that view a proof is not the result. It is a by-product. The result is the understanding.
When AI labs run problem-solving as a benchmark, that relationship flips. The declaration puts it bluntly: mass-producing true/false statements at an ever faster pace could destroy fertile ground instead of breathing life into new ideas.
Then there is how the work gets announced. Solutions are rushed out with no time for a proper write-up, no isolation of the new methods, no citation of relevant prior work by others. The declaration names the consequence directly: this raises severe attribution and plagiarism questions, as it would in any creative profession.
Why this goes beyond mathematics
The most interesting paragraph is the second to last, where the signatories step outside their own field. In many professions, years of training were never only about producing a final answer. They built understanding and the ability to ask new questions. AI systems, standing on generations of human work, now deliver the end product directly — and the two goals come apart.
The question that follows is aimed at everyone: how do we make sure that, as AI changes the way work gets done, we don’t lose sight of what that work was meant to achieve?
What set this off
The timing is not random. A week ago OpenAI announced a proof of the Navier-Stokes problem, and the argument about whether it holds up — and whose prior work it leaned on — has been running ever since. The declaration names neither OpenAI nor any specific case. It doesn’t need to.
Further signatures are explicitly invited, via ORCID, straight from the page.
My take
What makes this declaration worth reading is what it doesn’t argue. No jobs, no fear. It says a tool can be turned against its own purpose if you mistake the intermediate result for the goal. That argument still stands in two years when the models are twice as good.
And it transfers. Anyone who has had an agent write code knows the version of it: the thing works, you understood none of it, and at the next problem you’re back at zero. That’s usually the moment I start actually reading the diff. 25 Fields Medallists are saying the same thing about their field right now, with rather more weight behind it.
Sources: Declaration – Math and AI, Terence Tao: A Severe Misalignment of AI in Mathematics, Techmeme roundup on the declaration