AI Did the Maths Homework. Now OpenAI and Anthropic Are Paying Humans to Mark It

Every teacher knows the feeling. A student hands in a stack of homework so thick it could stop a door, and somebody now has to check all of it. Mathematicians have been living that feeling for the past fortnight, and the AI companies have finally offered to help pay for the red pens.

On October 15 the Simons Foundation announced a partnership with Anthropic and OpenAI worth $15m in combined funding and resources. The money will support workshops, conferences and programmes that help researchers evaluate recent AI-assisted results, learn new tools, share views and shape shared standards for how AI is used in mathematics. The foundation will run an open call for proposals and decide who gets funded. OpenAI said it is contributing $5m, and both companies will give Simons-backed researchers access to their models and computing power.

How we got here

Earlier this month OpenAI published a broad set of new mathematical results produced by one of its internal frontier models, releasing them in a GitHub repository rather than through the usual journals. According to The Independent, the release of hundreds of experimental AI-generated proofs at once drew a sharp backlash from mathematicians, who argued it swamped traditional peer review and leaned on open research without enough collaboration with the community.

OpenAI’s own update says it wants to help mathematicians “examine, understand, and build on results produced by AI,” and that it will keep improving citations, exposition and presentation in future releases. In plain terms: the papers will get easier to read, and humans will get money to read them.

Why this matters beyond maths

Mathematics is an unusually good test bed for AI because a proof is either right or wrong. Tools like the Lean proof assistant, which the Simons Foundation helped fund, let computers check each logical step. That makes maths one of the first fields where AI output can be verified at scale, and also one of the first where it can arrive faster than people can absorb it.

The same problem is coming for other fields. If AI can produce research faster than experts can review it, the bottleneck in science moves from discovery to checking. Somebody has to decide what is true, what is new and what is merely long. Journals, universities and funders have barely started asking who pays for that work. Funding workshops and shared standards is a sensible, if modest, first step. Fifteen million dollars is small change next to what AI labs spend on chips, but it is real money for a field that runs on grants and goodwill.

The Snark take

There is something very modern about a company producing a mountain of work, being told it is too much, and responding by funding a committee to look at it. Still, credit where due: paying the people who mark your homework beats insisting that the homework marks itself.

This is commentary. Facts are as reported by the sources linked below.