The mathematical community has mounted its sharpest collective response yet to AI companies' involvement in research. Within two days in mid-September, two separate letters challenged the premise of a student-organised AI maths competition at Caltech. OpenAI responded by pulling its $1m in funding, though Anthropic maintained its equal commitment. The dispute centres not on whether AI models work, but on how companies use their results.

The first letter and 'slop mathematics'

On 10 September, mathematicians from Caltech and beyond published an open letter bearing 771 signatures, restricted to PhD-level researchers and above. The signatories called for the event's suspension, introducing the term "slop mathematics" to describe work produced rapidly and left for verification by others.

The letter's core argument went beyond terminology. It contended that research mathematicians shoulder the burden of verifying, disseminating and correcting AI-generated results, labour that "goes uncompensated, uncredited, and unacknowledged". The signatories characterised this arrangement as research misconduct. They highlighted the economic implausibility of a sustainable model in which every research mathematician received $20,000 in AI credits to validate findings.

OpenAI exits, Anthropic remains

On 10 September, OpenAI withdrew its sponsorship. Dan Roberts, a research lead at the company, posted on X that OpenAI recognises the disruptive pace of AI progress in mathematics and seeks dialogue with the mathematical community on integrating the technology. The company reclaimed its $1m in credits.

Anthropic took the opposite path, maintaining its $1m pledge without comment. In the preceding fortnight, Anthropic had pursued a different strategy, working to build relationships with mathematicians. The company used Claude to formalise Fermat's Last Theorem in Lean within 11 days, subsequently expanding free access and research credits for formalisation projects.

The Fields Medallists weigh in

On 11 September, 25 Fields Medallists released a declaration titled "A Severe Misalignment of AI in Mathematics". The group included Terence Tao, Peter Scholze, Maryna Viazovska, Cédric Villani, Pierre Deligne, Manjul Bhargava and Martin Hairer, with medals spanning from 1978 to 2026.

Their statement pursued a narrower but more fundamental critique than the Caltech letter. Problem-solving represents merely a tool and proxy, they argued, while the genuine objective remains conceptual understanding and insight. They cautioned that accelerating production of statements—whether true or false—risked destroying "fertile ground" for mathematical development. Notably, the declaration avoided naming the Mathathon specifically and made no call for cancellation, framing itself instead as a statement about benchmarks.

The competition itself

The Mathathon was announced on 4 September and will run at Caltech from 30 October to 1 November. One hundred teams will each receive approximately $20,000 in credits and 40 hours to tackle an open problem. Work deemed significant by judges advances to a six-month second phase requiring an arXiv preprint, video presentation and auto-formalised solution. An educational track also exists for course materials, explanatory content and streamlined proofs of known theorems.

By 10 September, organisers had received over one thousand applications spanning undergraduates through faculty. Beyond AI credits, the competition lists backing from DARPA expMath and Cognition.

Student organisers respond and adapt

The five named organisers are students. On the same day as the letter's publication, they released a response acknowledging that several concerns held validity and that their initial framing had overlooked them.

They implemented revisions. The arXiv requirement was added following the letter's request, and the educational track expanded. They also removed a website line questioning the role of mathematicians when AI could solve conjectures faster—a removal the letter itself documented in a footnote.

Cancellation was not on the table. All five organisers have signed the Leiden Declaration, the June statement in which mathematicians demanded that AI companies cease using their work without permission. Caiman Moreno-Earle, a 19-year-old Caltech sophomore who initiated the event, told Business Insider he approached Anthropic requesting $1m, anticipating an offer of $50,000. He explained that younger mathematicians cannot afford to wait for rules to crystallise: "We don't have the luxury of waiting around to find out." The team is now pursuing alternative credit sources to replace OpenAI's funding.

What triggered the crisis

The immediate catalyst came the week prior. On 8 September, OpenAI announced that a collection of its agents had solved the Navier-Stokes problem, publishing the proof the following day while declining to claim the $1m Millennium Prize.

Tristan Buckmaster of NYU, pursuing the same problem alongside Levent Alpöge of Anthropic, disclosed that OpenAI had proposed authorship terms that would have excluded Alpöge. Sébastien Bubeck of OpenAI disputed any attempt to remove Alpöge from credit. Buckmaster subsequently told the ABC that the competitive rush to publish has become futile and that the situation has fundamentally changed.

The distinction both letters defend matters: AI mathematics results most frequently cited tend to be counterexamples rather than proofs, a pattern documented in August reporting.

Unresolved questions

  • Whether the Mathathon secures replacement sponsors and under what conditions
  • Whether OpenAI's withdrawal signals a broader shift in its mathematical research strategy or applies only to this event
  • Whether the Clay Mathematics Institute, which continues to list Navier-Stokes as unsolved, will recognise the proof

Source: The Next Web