Can AI governance models from mathematics work across scientific fields?

Nature's editorial treats the new Leiden Declaration on Artificial Intelligence and Mathematics as a sequel to the 2014 Leiden Manifesto, the ten-principle guide to the "responsible use of metrics in research" that, together with DORA, "have been adopted around the world." The editorial reports that last September's Leiden meeting reprised that role, this time for AI in mathematics, and that the resulting declaration "has been gaining endorsements from researchers across the discipline, which includes those who are deeply sceptical of AI and those who are much more optimistic." Nature states it "wholeheartedly endorses both the declaration process and its conclusions," and closes by extending the claim past mathematics: "Now it's time for the discussion to become wider and stretch to other fields."

The editorial's reasoning rests on a concrete test case. An 80-year-old geometry problem, the unit-distance problem first proposed by Paul Erdős, "was solved by mathematicians at the US technology firm OpenAI using only a single prompt to a chatbot." OpenAI posted the proof publicly and "the findings have been verified by a group of mathematicians who are independent of the firm" — so correctness was established. But "OpenAI has so far not disclosed the name or details of the software used to solve the conjecture," and, "despite several requests," has "not fully disclosed which data sets its models are trained on." For Nature, that gap between verified output and undisclosed method is exactly what the declaration's disclosure principle, and its line that "no proprietary knowledge or equipment should be required to understand" results, are meant to close. The editorial also ties the stakes to an autonomy argument: AI integration "will change the kinds of problems that are pursued and the forms of proof that are valued," and it cites emerging cross-science evidence that "the use of AI correlates with a narrower breadth of research topics," which it says disadvantages researchers without access to proprietary tools or who decline to use them.

This sits alongside Can AI-generated proofs ever replace human mathematical understanding?, which covers what the declaration itself asks of mathematicians (disclosure sections, human responsibility for correctness). This note covers Nature's framing of the declaration as a reusable governance template, backed by the 2014 precedent, plus the editorial's own evidentiary case — OpenAI's partial disclosure — for why the template is needed now. The topic-narrowing claim Nature cites in passing is the same finding developed at length in Does AI help individual scientists while narrowing scientific focus?; here it functions as a stake in the governance argument rather than as the subject itself. The editorial's description of AI as "rapidly changing mathematicians' job descriptions" echoes Tao's own account in Can opaque machine learning models help prove new mathematics?, where usefulness depends on independent validation of output — the same validation the unit-distance proof received, even as its method stayed opaque.

As an editorial, the piece argues rather than measures: it does not establish that mathematics' disclosure norms would transfer to other fields' different risk profiles, nor does it quantify how much narrower AI-augmented research topics have become — it cites that figure secondhand rather than reporting its own analysis. It also does not establish why OpenAI withheld the software and training-data details; commercial confidentiality, competitive pressure, and simple non-response to the request are all consistent with the silence it reports. What the case does establish, at the strength the evidence allows, is that verified correctness and methodological transparency can come apart in practice — the specific gap the editorial wants other fields to close before their own version of the unit-distance case arrives.