OpenAI pulled out of sponsoring a mathematics event at CalTech on Thursday, after criticism from researchers at the university. It did so in the same week that Tristan Buckmaster, a professor at NYU, said the company had tried to persuade him not to credit a co-author employed by Anthropic when writing up the solution to an important mathematical problem. Both moves land alongside a new open letter in which mathematicians argue that AI-assisted results are being announced faster than anyone can check them, attribute them or build on them.
One fact governs the rest and is easy to lose in the noise about credit: OpenAI's proof has not yet been verified. The argument over whose name goes on it is running ahead of any determination that it is correct.
The letter's signatories are not opposed to machines solving open problems. They write that the ability of AI models to settle unsolved mathematics could benefit humanity — but only if the mathematical community, and then everyone else, can understand the solutions and pass them on. Their complaint is about the gap between announcement and understanding. A result declared quickly leaves no time to write the proof out properly, to isolate the new methods and ideas inside it, or to cite the prior work it stands on. In their view that practice raises serious questions of authorship and plagiarism, and without mathematicians who develop new ideas and fit them into the shared body of knowledge, AI results have no continuation — the human chain that carries knowledge from one researcher to the next simply breaks.
Buckmaster made a second allegation that is broader than the byline dispute. He asked whether OpenAI had used his group's work with Codex to produce its own breakthrough proof over a single marathon weekend of inference. That question has since spread: other mathematicians have begun to worry that their own Codex sessions could feed into new OpenAI models, and the fear for open research culture has grown alongside it.
The letter spells out the incentive that follows. If a frontier lab spots a promising route to a discovery, it can spend tens of millions of dollars on LLM inference to reach the proof before the people who found the route. The signatories warn that an incentive structure like that pushes researchers toward secrecy.
This is where the dispute stops being about manners. Mathematics has no patents, no embargoes, no registry of priority. Its entire system for assigning credit is social: you say what you are working on, peers check it, the community remembers who got there first. That system was built to withstand slow rivals reading your preprints. It was not built against a counterparty that can compress the distance between a promising direction and a finished proof into a weekend, at a cost only a handful of organisations can pay. The norms the letter defends worked because the gap between having an idea and executing it was large enough to protect the person with the idea. Inference closes that gap, and nothing in the letter can reopen it.
Notably absent from the letter is any account of who would enforce what it asks for. Attribution and time to verify are reasonable requests; they are also requests, addressed to a party with no obligation to grant them. The June Leiden Declaration, prepared by a working group of mathematicians and the document this letter continues, has the same shape — an examination of how LLM-generated proofs change the field, with recommendations for mathematicians, for institutions and for policymakers. Recommendations are the only instrument available, which is itself the finding.
There is also a cost the letter does not name. Its stated enemy is plagiarism, but the practical effect of the incentive it describes is silence: a mathematician now has to weigh whether describing an open problem in public amounts to handing a lab a target list. Secrecy is not a side effect of this dispute. It is the rational response to it, and it will arrive long before any norm does.
The signatories argue that the value of mathematics was never only the proofs or who gets named on them — it is the intellectual structure that trains students, surfaces new questions and places them in a wider context. They also say that anyone indifferent to the closed, high-stakes world of mathematical proof should not treat this as distant, because another professional field is next, and the question underneath it reaches everyone: how to keep hold of the purpose of work while AI changes how the work is done.
They are probably right about the sequence, and that is the uncomfortable part. Mathematics has the strongest verification culture of any discipline and the least ambiguous definition of a correct answer, and it is losing control of its own attribution norms inside a single year. Fields with softer standards of proof have nothing better to defend themselves with.