OpenAI Posted 722 AI-Written Math Papers With the Proofs Attached. The Mathematicians Answered With a Boycott.

In August we wrote that generating the answer got cheap and testing it did not, and that when one side of a loop gets cheap the profit moves to whoever owns the other side, the test. Mathematics is the one field where that rule should fail. A proof written in Lean is checked by a program, in seconds, for nothing. This week OpenAI tested the exception, and the result says something uncomfortable about what the argument over AI in science was actually about.

## What was published

On October 6 OpenAI published 722 mathematics manuscripts, grouped into 372 families of results, produced by an internal frontier model. They went into a public GitHub repository with Lean formalizations and methodological notes, so mathematicians can audit, extend or refute them. By the figures OpenAI disclosed, the model was posed roughly 4,000 problems, and The Rundown reports the catalogue includes a claimed proof of a "quasi Riemann hypothesis". Chinese coverage picked it up within a day, and the story is now a trend in our records.

Do the division. Three hundred and seventy-two families from roughly 4,000 problems is fewer than one in ten. The model is not solving mathematics; by its maker's own count it produces something publishable on under a tenth of what it is asked. That is still a great deal of mathematics. It is also a number a human department can compare itself against, which is exactly the comparison the release invites.

The model itself was not released. That is the odd part of the week: a capability disclosure with every result checkable, about a model nobody outside OpenAI can use.

## What the answer was

Two days later the Association for Human Mathematics published a statement condemning the release and urging mathematicians not to work with OpenAI. Its objections, as reported by the Indian Express: OpenAI used unreleased internal frontier models on advanced problems despite an advisory panel's warnings, and the release model undermines norms of verification, credit and the public interest in mathematics.

Read the list again. None of the three says the proofs are wrong. The company shipped the checkable artefact, and the objection moved to who decides which problems are worth attacking, who gets the credit, and whether a panel that was consulted was heeded.

This is the second round. On September 12 the Guardian reported that OpenAI claimed an internal model had solved the Navier-Stokes Millennium Prize problem, using some 10,000 agents at an estimated cost of around $15 million. Researchers worried about being scooped, about credit, and about becoming de facto proof checkers for opaque machine-generated arguments. A week later El País and Cinco Días were debating the claim alongside a 10 percent extinction estimate, seven outlets in our records. "Opaque" was the September complaint. October's release answers it with Lean files, and the objection survived. When the verification complaint is answered and the objection stays, the objection was about the profession, not the proof.

## Why this is a technology story, not an etiquette story

In September we wrote about Anthropic disclosing that Claude leads 26 percent of its own research, and said the operational question was who can stop a thread the model started. Here the thread was roughly 4,000 problems posed by a lab to its own model, and the brake on offer was an advisory panel whose warnings, by the mathematicians' account, did not hold. A month before the release, OpenAI's own chief scientist wrote that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed. The 722 papers are what maximum speed looks like in the one domain where the output can be checked.

The same week, Biohub, the Department of Energy, the NIH, Google DeepMind, Isomorphic Labs and Meta put $1.8 billion into open biological datasets for models that predict how cells respond to interventions. That is the August rule working as written: in biology, checking a prediction costs a bench, as we argued when Anthropic built a wet lab, so the money goes to the test loop. In mathematics the test loop is free, and the money, the model and the fight all went somewhere else.

## Hedges

Every figure is OpenAI's, and "roughly 4,000" is approximate; a family can hold several manuscripts, and a problem can yield nothing publishable for reasons other than failure. We have not audited the repository and do not know what share of the 722 is fully formalised in Lean rather than accompanied by it. The "quasi Riemann hypothesis" claim reaches us through one newsletter. The Association for Human Mathematics statement reaches us through one outlet, and we cannot see how many working mathematicians it speaks for.

## What to do with this

When you next read an AI-for-science claim, ask the question we set in August: where is the checkable artefact. A Lean file, a reproducible notebook, a bench result. If it is there, the remaining argument is about credit and norms, which is a negotiation with a field rather than a claim against a vendor, and a better model will not settle it.

Then watch the repository, not the statement. OpenAI invited mathematicians to extend or refute the 372 families. The first human-authored refutation or extension that lands there, and how OpenAI credits it, will tell you more about how this settles than any boycott. OpenAI is on our tracker, and the trend page will collect the records as they arrive.