AI Proofs and the Race to Skip Understanding
A machine can now produce a proof that no human has time to read. That is not a paradox. It is the exact shape of the problem a group of Fields Medalists put their names to this week, and it is worth sitting with before reaching for either the panic or the applause.
What the letter actually claims
The signatories — among them recipients of mathematics’ most prestigious prize — argue that AI labs are racing each other to crack famous open problems, and that the race itself is the threat. Not the solving. The solving could be a gift. The threat is what gets skipped when speed becomes the only metric that counts. In their words, solutions arrive “in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others.” The letter is careful, and it is not anti-AI. It is anti-collapse — the collapse of the distance between a result and the community that could absorb it.
The gain nobody disputes
Here is the part that gets lost. When an LLM finds a path through a problem that has resisted human attack for decades, something genuinely new enters the world. Not a parlor trick. A structure that may connect to other structures, that may open questions no one thought to ask. The Leiden Declaration, released in June by a working group of mathematicians, takes this seriously rather than defensively: it offers recommendations for mathematicians, institutions, and policymakers precisely because the tools are changing what is possible. Leiden Declaration. The upside is real. The letter’s authors say so. The question is whether the upside survives contact with the incentive to be first.
Attribution is not a technical problem
Tristan Buckmaster, a professor at NYU, accused OpenAI this week of pressuring him not to credit a collaborator who works at Anthropic for solving an important math problem. [1] He also wondered aloud whether the company had used his team’s work with Codex to produce its own proof over a marathon weekend of inference. [2] OpenAI withdrew its sponsorship of a math event at Caltech on Thursday after researchers there criticized the company. [3] None of this is a bug in a citation system. It is the human layer — trust, credit, the willingness to share a half-formed idea — and no benchmark measures it. You cannot patch it. You can only decide, collectively, whether to protect it.
The work around the work
Mathematicians have a phrase for the invisible labor that makes the visible labor possible: the work around the work. The seminar where a graduate student asks the wrong question and someone else realizes it is the right one. The referee report that catches a gap. The slow integration of a new method into the canon so that the next generation inherits it as instinct rather than trivia. The letter names this directly: “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.” A proof that lives only in a model’s weights is not knowledge. It is a rumor with a high confidence score.
Secrecy is the real cost
If a frontier lab sees a useful path to a discovery, it can spend tens of millions of dollars using LLMs to beat the original researchers to the proof. That is the new arithmetic, and it changes behavior before it changes results. Mathematicians are already growing paranoid, wondering whether their own Codex use is being fed back into OpenAI’s new models. The culture of open research depends on a fragile assumption — that sharing your unfinished thinking will not be used against you. Remove that assumption and you do not get faster math. You get quieter math. You get the same proofs, arrived at in isolation, by people who no longer trust each other enough to collaborate.
What the letter is really protecting
The twenty-five signatories are not defending their turf. They are defending the transmission chain — the line of humans who teach, question, and carry ideas forward. Their concern is stated plainly: “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.” That sentence is not about mathematics. It is about every field where a model can now produce the artifact without the understanding. Software engineering knows this. Law is learning it. Medicine is next in line. The letter is a warning shot from the people who saw it first.
Your field is next
The mathematicians are not special. They are early. Their letter is not a plea for exemption. It is a request for a norm: that speed not be allowed to erase the people who make speed meaningful. The voice in the letter is the voice of the discipline itself, asking to remain a discipline — not a dataset.
Sources
1. Tristan Buckmaster, a professor at NYU
2. Anthropic
3. Caltech