OpenAI releases 722 machine-generated proofs, and nearly 900 mathematicians sign a letter of protest

OpenAI published a batch of 722 mathematical proofs generated by its models. The reaction from the field was immediate: according to the reporting, close to 900 mathematicians signed a letter of protest, expressing fear for the future of their profession. That is the full set of confirmed facts, and it is worth stating plainly before anyone extrapolates from it.

Context matters here, because mathematics is not an ordinary benchmark. It is one of the very few domains where a machine's output can be checked to the last step — a proof is either valid or it is not, and formal verification systems can settle the question without a human vote. Compare that to the benchmarks we usually argue about: coding leaderboards drift, chat arenas can be optimized for voters, and composite indices flatter whichever version was submitted. A proof cannot be optimized for applause. So the headline figure of 722 is not the result; it is the input to a verification process that has barely begun.

That is also why the protest deserves to be read carefully rather than dismissed or amplified. A signed letter from hundreds of working mathematicians is a strong professional signal, and the fear it voices — that the craft of producing proofs is being automated — is the same anxiety we have tracked sector by sector, from legal associates to bank back offices. But it is a signal about perceived threat, not a verdict on the proofs themselves. Both things can be true at once: the mathematicians may be right about the direction of travel, and the 722 may still contain errors, restatements of known results, or work whose originality is contested. Nobody should assert either outcome today.

The deeper pattern is one we keep seeing: the bottleneck is moving from production to verification. Generating candidate work is now close to free; establishing that the work is correct and genuinely new is where the scarce human effort goes. In mathematics that shift is unusually visible because the discipline already has the machinery — referees, formal proof assistants, a culture of adversarial checking. Most professions facing the same wave have nothing comparable.

Our read: this is a transition story with a hard short term and a genuinely good long term. If the proofs hold up, the immediate pain is real — fewer routine positions, a narrowing base of the professional pyramid, and a legitimate grievance about who captures the value of work trained on a community's published output. We do not sugarcoat that. But mathematics is the engine room of everything we care about on the optimistic side of the ledger: protein folding, drug discovery, materials, climate models, the formal methods that make critical software safe. A field that can generate and verify proofs faster is a field that unblocks other fields. The mathematician whose value was transcribing known techniques is exposed; the one who chooses which problems are worth attacking, and who can tell a correct proof from a plausible-looking one, becomes more valuable, not less. That judgment role is exactly what does not automate.

What to watch next, and what would change our view: independent verification rates on the 722, whether any of them constitute results the community considers new, and whether OpenAI publishes enough detail for outsiders to audit the claim rather than take it on trust. Until those numbers exist, 722 is a press release figure. Treat it as a hypothesis under test, not a milestone already passed.