I am not a mathematician. I work in the philosophy of technology and in law, and I have written with Susana Checa Prieto an article on the accountability of artificial intelligence that Law, Innovation and Technology has accepted for 2027. I am writing in response to the question Terence Tao put to his community at the ICM: what mathematics is for. His essay ends with a recommendation from the Leiden Declaration: to cooperate with researchers in other fields who face similar challenges. I write from one of them. I adopt his Working Hypothesis for this conditional analysis — that AI tools will reasonably soon perform a reasonable fraction of research-level work, at reasonable cost and with reasonable supervision — and I add no claim of my own about what the machine can do.
The lecture’s diagnosis is right, and I want to state it in its own terms. A result passes through five stages: generation, verification, exposition, publication and canonicalization. Only the first was ever an explicit goal of the community; everything that follows generation and verification, Tao calls digestion. For centuries, solving a problem was so costly that whoever solved it also, as a by-product, understood it and could explain it: solving stood in for everything else. Goodhart’s law now applies, and a measure turned into a target ceases to be a reliable measure. Hence the rule the lecture proposes: a result should not be published unless its authors can convincingly show that they are able to give a clear, correct and properly attributed expert-level talk on it; a proof that no human can properly explain is incomplete, even if it has been formally verified.
There is something the lecture, and the Declaration it takes as its starting point, leave undeveloped. Both give reasons for responsibility and authorship to remain human: credit, the collective human labor behind every result, the understanding the community must preserve. What they do not say is what it takes to answer for a result: what the person who presents it as their own commits to. Automating the operations of a practice — generating, verifying, even explaining — does not answer that question; it leaves it open. Stated as a preference, the requirement can be revised; stated as a condition, it obliges one to say what it consists of. What I propose is one way of saying it: to carry over to mathematics the conditions of accountability we develop in our article, and to specify what the person who presents a result commits to, even when its formal correctness is already established.
Our article, developed in the context of legal accountability, holds that to answer for something three conditions must be met, each externally assessable, and all three together: having been able to refrain from acting, giving reasons for the particular case, and having something to lose. What this adds to understanding, attribution and expert review is not one more requirement; it is the question of who meets them and what they commit to in meeting them. The lecture supplies a case: the Erdős problems database contains AI-generated proofs, many of them probably correct, submitted by people who declare that they are not qualified to vouch for them. The case shows that a result with a name attached does not by itself establish competent backing, and Tao’s rule already catches that shortfall. My question begins where the rule is satisfied: what does the person who does understand the result, and decides to present it to others, commit to? Formal verification does not settle that. Of the three conditions, the one that matters here is the third, and it has to be stated precisely. Suffering a loss does not, by itself, amount to undertaking a commitment. Such a commitment requires someone who could have refrained, can give reasons for acting, and has something of their own at stake. That is why recording sanctions or withdrawing resources is not enough to establish that there is a subject to whom those consequences can be imputed. Under the criterion developed in our article, current systems do not undertake such commitments in their own right. I do not claim that no system could ever satisfy the conditions. What I examine here is what they allow us to require of those who present results obtained with these tools.
Mathematics is the hardest case for this thesis, and therefore the best. Tao himself notes that once a proof is formally verified, its correctness no longer rests on its author’s reputation or diligence. And still he requires that the author be able to explain it. So what the rule leaves tied to the author after verification is not correctness. It is explaining the result and defending it before others. It will be said that this is competence and nothing more: that whoever claims credit must understand what they present. True; that is what the rule establishes. What I add is the question of what that person commits to in presenting it, and to whom they are answerable. And the question has consequences for tools and procedures: they must allow the person presenting a result to understand it, explain it and expose it to criticism. That understanding may be reached with the system’s help, by independent reconstruction, or by working with others. My preference is that the system itself should make it possible, helping to explain the proof the way a mathematician explains a result to referees; but what can be required is that someone be able to stand behind the result, and the commitments of presenting it, on reasoned grounds — by whatever route. A verified proof and a file with a provenance chain guarantee different things — a correct derivation within a formal system; an origin and a sequence of transformations — and that is no small thing. But they leave the same thing open: who is answerable for the result, and to whom.
What about the stages where there is no single author? Digestion is, as the lecture says, slow, human and external: the community does it. Nobody authors the canon, and nobody needs to. What we call the canon is a chain of acts attributable to identifiable people, each with its own object and scope: the referee answers for the report, the editor for the decision, whoever teaches a result for how it is taught, whoever builds on it for what is built. No one is responsible for everything. For each act, it should be possible to identify who is answerable, to whom, and for what. The useful question is not whether the community answers, but who answers for what, to whom, and to what extent. The danger of proof abundance is not that the canon loses its author, because it never had one. It is that this chain of attributable acts gets replaced by unattributed accumulation. The line, then, does not run between proofs made with AI and proofs made without it: it runs between results someone answers for and results no one answers for. And to answer is not to be punished. It means there is someone who can acknowledge an error and take responsibility for addressing it. Seen this way, the expert-talk rule is exactly where it should be: it is the attributable act by which a result enters digestion.
One question remains that I cannot answer from outside, and that I would like to put to Terence Tao and to his community: when generating and verifying a proof can be automated, who is answerable for it to the community at each step of digestion, and what do they commit to in doing so — what must they be able to explain, correct and lose?
I ask for nothing else. The article is open access (https://ssrn.com/abstract=6264938), and corrections from inside the field are welcome.
References. S. B. Checa Prieto and J. Fernández Tamames, “Why AI Cannot Sign: Deliberative Suspension, Reflective Judgment, and the Verification Tax in Generative Legal Technology”, Law, Innovation and Technology 19(2), 2027, forthcoming (accepted 15 September 2026). T. Tao, “Mathematics in the age of AI”, arXiv:2608.16753 (2026), an essay based on the ICM 2026 public lecture. The Leiden Declaration on Artificial Intelligence and Mathematics (June 2026).
Disclosure. The first draft and translation of this note were prepared with AI assistance and revised by me; the argument is from the article cited.
ORCID. 0009-0007-8851-9833. Archived version. https://doi.org/10.5281/zenodo.24144542
Received 25 October 2026.

Add to the discussion