On October 30, 2026, Caltech is scheduled to begin hosting the Caltech Mathathon — a math hackathon in which participants are encouraged to solve open research problems by prompting large language models. Anthropic and OpenAI committed $2 million in AI credits between them.

On September 10, a coalition of current and former Caltech mathematicians published an open letter asking the organizers to call it off.

"This event is likely to have destructive impacts for the mathematical community."

At publication the letter carried 771 signatories. Signing is restricted: you must either be a member of the Caltech community — undergraduates, graduate students, faculty, researchers, and JPL affiliates — or a member of the mathematical research community at the PhD level or higher. That eligibility rule matters, because it means this isn't an open internet petition. It's a professional field registering an objection about itself.

This is the part most of the coverage got wrong, so it's worth slowing down.

The letter's complaint is about a workflow, not a capability.

"Major AI companies are engaged in an arms race to declare themselves the first to prove prominent open conjectures. New announcements of AI-generated results appear on social media on a weekly basis. These results are often poorly communicated, and have diffuse negative impacts on the careers of human mathematicians who are concurrently proving the same results."

Then the core paragraph:

Strip the heat out of it and this is an externality argument, the same shape as pollution or spam.

Announcing a claimed proof is cheap and produces immediate value for the announcer. Verifying it is slow, difficult, requires scarce expertise, and produces value for everyone except the person doing it. If the two are decoupled — if you can post the claim without paying for the check — you have created a system that generates announcements faster than the field can absorb them.

Peer review is not bureaucracy. It's the mechanism that makes the person making a claim bear part of the cost of substantiating it. Route around it and the cost doesn't disappear. It moves.

There's a passage in the letter that's easy to skim past and shouldn't be, because it's the clearest statement of what the signatories think is at risk:

"Mathematicians do not only prove theorems: they also take measures to ensure that the future of a field remains healthy after the extraction of a result. They do so by sharing their results and explaining their methods to others at talks and conferences. They mentor students and postdocs so that the next generation can advance research further, and take care not to scoop other mathematicians at the last mile."

The word doing the work there is extraction. The letter's charge is that AI companies want the theorem and not the maintenance — "to extract the prestige of newly proven results rather than add value to the community in a way that safeguards its future."

1. The scooping concern is technical, not territorial.

The letter argues human mathematicians working the same problems "may be forced to abruptly abort these research projects, regardless of the additional insights their unique approaches may bring."

This sounds like credit-protection. It isn't only that. In mathematics the method is frequently more valuable than the theorem — a second proof of a known result can introduce machinery that unlocks a dozen other problems. Killing a parallel effort because someone posted a proof first can destroy the more useful of the two outputs.

2. The economics don't generalize.

"There is no reasonable future model of mathematical research in which every research mathematician receives 20000 USD in AI credits to prove a result."

That's a sharp line. If a demonstration only works under resource conditions nobody can reproduce, it demonstrates the sponsor's budget, not a research method.

3. The students are described as the product.

"AI companies will take credit for the effort of talented undergrads; the only relevant question to them is whether Claude or ChatGPT proved the most flashy result."

For a newsletter read by parents and teachers, this is the line to sit with. The participants are undergraduates and PhD students. The letter's claim is that whatever they produce becomes marketing material attributed to a model.

Business Insider reported that OpenAI withdrew its sponsorship on September 11, the day after the letter was published.

Be precise about the rest: as of that reporting, Anthropic had not withdrawn, and neither had a16z or Y Combinator. The tidy version of this story — "OpenAI left, Anthropic stayed" — is too narrow. OpenAI left. Everyone else stayed.

The Mathathon organizers have published a response. And the letter's own comment section carries substantive opposition from inside the field, which deserves airing rather than summarizing away.

The central rebuttal runs roughly like this:

"Slop" is never defined. If it means incorrect, unverified, unreadable, or low-value work, then that standard should apply to human output too. "Mathematical content should not be classified as 'slop' merely because AI contributed to it; likewise, a proof should not automatically be regarded as serious or valuable merely because it was written by a human."

The bar keeps moving. "When AI systems solve IMO problems, we are told that this does not count as mathematical research. When AI begins to make progress on certain Erdős problems, we are told that those problems are not particularly important."

Forty hours can still teach something. A short event doesn't have to produce a publication-ready paper. Participants can learn a conjecture's background, work through the literature, find examples and counterexamples — and models can function as interactive tutors, not just proof generators.

Access is unequally distributed already. Only a small number of mathematicians have sustained access to frontier models. The Mathathon offers that access to PhD students and undergraduates. "Is using one's influence to restrict the choices and development of younger mathematicians really what we should call 'care' for them?"

That last point is the strongest one and got a thoughtful reply in the thread — from a mathematician arguing that caring for young researchers includes giving them a safe environment to fail slowly in, and that an accelerated, high-visibility, sponsor-funded sprint is not that environment. "Burnout is a thing and resilience is much needed for building long-lasting mathematics."

Both things can be true. Access is real. So is the pressure that comes attached to it.

Two claims are tangled together in this story and they have different strengths.

The capability claim — that AI-generated mathematics is low quality — is contested, and the rebuttal above lands some real hits on it. It will also age quickly in one direction or another.

The process claim — that announcing results without paying for verification transfers cost onto people who didn't consent to it — doesn't depend on capability at all. If the models get dramatically better, the claim gets stronger, not weaker: more claimed results, arriving faster, still needing human checking, still unbudgeted.

That's a governance problem with a dull fix. Don't announce a claimed result until someone qualified and compensated has verified it. Credit and pay the verifiers. Publish what failed verification, not only what passed.

None of that requires anyone to decide whether AI "really" does mathematics.

This is an unusually good classroom example, because the disagreement is honest on both sides and the subject matter is not politically charged.

The question to put to a student isn't "is AI good at math?" It's: who checks the answer, and what does checking cost?

That question transfers to essentially every AI claim a young person will encounter. A result that nobody has verified isn't wrong — it's unverified, which is a different and much more common condition. Learning to hold that distinction is more durable than any opinion about a particular model.

And there's a second lesson sitting right in the comment thread: the strongest objection to the letter came from someone the letter was ostensibly written to protect. Reading the dissent is part of reading the story.

Disclosure: Anthropic, the company that makes the AI model used to help assemble this newsletter, is one of the two AI companies named in the letter's opening paragraph, is named again through "Claude" later in it, and had not withdrawn its sponsorship as of the reporting cited above. It seemed better to say so than to leave it out.

Source: "Open Letter about the Mathathon," Proofs and Prompts, a coalition of current and former Caltech mathematicians: https://proofsandprompts.com/2026/09/10/open-letter-about-the-mathathon/

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter