Where I stand on AI

A changed landscape

Since the release of GPT-5.6 and Claude Fable 5, the mathematical community has entered a visibly divided period. Each day brings preprints that describe themselves as AI-generated and human-verified, papers that disclose a more limited use of AI, and papers for which the extent of machine involvement is simply unclear. The resulting uncertainty is unhealthy: readers should not have to infer provenance from style.

For me, the central change is not that human mathematicians have become unnecessary. It is that the comparative value of human work is shifting. Formulating one's own questions, examples, concepts, and research programs is becoming more important than being the next person to solve a problem selected by someone else. Human mathematical work is moving toward

This is a statement about where I want to place my effort, not a universal ranking of mathematical worth. Existing conjectures remain important. But when technical execution becomes increasingly shareable with machines, the choice of what to ask—and why it matters—becomes an even more distinctive human contribution.

Three broad attitudes

I currently see three broad attitudes among mathematicians. This is an impression, not a sociological classification, and the boundaries are naturally porous.

Mathematical truth and human authorship

A common objection from the first group is that mathematics is an irreducibly human activity, comparable to art or literature, and that a proof produced without human aesthetic judgment or communal recognition is therefore meaningless. I do not accept this conclusion.

Mathematics certainly has human dimensions: we choose axioms, definitions, questions, standards of explanation, and notions of significance. Yet validity within a fixed formal framework does not depend on the species or identity of the reasoner. A proposed proof is valid or invalid because of its mathematical content, not because its author is human or machine. On my view, whether S⁶ admits a complex structure does not depend on who eventually settles the question. A proof would change what we know, not the mathematical answer itself.

Formal proof assistants make this distinction concrete. Once the intended theorem and its assumptions have been faithfully formalized, a system such as Lean can certify the derivation without a human referee checking every inferential step. This does not eliminate human judgment altogether: someone must still examine whether the formal statement expresses the intended mathematics, whether the assumptions are appropriate, and whether the result is interesting. But it shows that correctness and human authorship are not the same criterion.

Why blanket refusal is the wrong response

I also reject the policy of refusing, as a matter of principle, to read, referee, or help publish any work in which AI has participated. Such a policy does not protect mathematical standards. It replaces evaluation of the mathematics with evaluation of the tool.

When editors, referees, or other gatekeepers impose categorical penalties on disclosed AI use, they create a perverse incentive: authors gain by concealing the very involvement that the community most needs them to describe honestly. The likely result is not less AI-assisted mathematics, but less transparent AI-assisted mathematics, with greater pressure on researchers who disclose their methods. Journals should instead ask about correctness, novelty, significance, provenance, and reproducibility.

The new era has arrived whether or not any individual welcomes it. The task is therefore to build norms that reward candor, preserve rigorous standards, and distinguish mathematical contribution from mere opacity about how a result was obtained.

My disclosure rule

I intend to make the following distinction in my own work.

In joint work, I cannot unilaterally require every coauthor to adopt this exact rule. I will, however, recommend it and argue for disclosure that is specific enough for readers to understand what the AI actually contributed.