The cycle double cover conjecture is easy to state and has been open for fifty years: every bridgeless graph should admit a family of cycles that together cross each edge exactly twice. Szekeres posed it in 1973, Seymour again in 1979, and generations of graph theorists have worn grooves in it since. On Friday OpenAI posted a paper claiming its GPT-5.6 Sol Ultra model has produced a proof.
The reflex question is familiar by now. How much was the model and how much the humans steering it? How many thousands of failed attempts died before this one? Fair questions — and in this one domain, beside the point. Mathematics is the only field where the value of AI output does not hang on the answer to "who really did the work." A proof is valid or it isn't, and its validity owes nothing to its author. The referee process that will now grind over this PDF does not have a field for provenance.
Mathematics has been here before, and it built the machinery on purpose. When Appel and Haken proved the four-color theorem in 1976 by having a computer grind through nearly two thousand configurations, mathematicians objected that a proof no human could read wasn't a proof at all. The theorem outlived the complaint. When Annals referees spent years on Thomas Hales's proof of the Kepler conjecture and emerged saying they were only 99 percent certain, Hales didn't argue — he spent the next decade formalizing the whole thing in the Flyspeck project, until a machine could check every step. The pattern holds each time.
When trust in the author breaks down, mathematics doesn't argue provenance — it escalates verification.
The honest objection is economic, and it's real. Verification is the scarce resource: a proof now costs an API call to generate and a referee's months to check, and "checkable in principle" pays nobody's mortgage. A field flooded with plausible machine-written proofs could drown the small number of people qualified to referee them. Concede all of it. But notice which way it cuts — toward formalization, not despair. The same models that write proofs can be made to write them in Lean, where checking is mechanical and effectively free. A result that arrives with a formal certificate isn't a burden on the referee; it's the end of the refereeing question. Expect proof-assistant certificates to become the submission standard for machine-generated mathematics, and expect the labs to comply, because it's the only receipt that settles the question their own branding raises.
Whether this particular proof survives review is genuinely open, and the failure mode matters as much as the success. If it collapses, it collapses in public, at a specific line, checkable by anyone — exactly the way a human proof would. That symmetry is the milestone. Everywhere else, AI output makes us argue about who did the work. Mathematics gets to ask the only question that was ever worth asking: is it true.