Association for Human Mathematics tops 1,000 members pledging not to work for AI firms

What happened

More than 1,000 research mathematicians have signed up to a group that asks them not to work for AI companies and not to publish mathematics written by AI models. The Association for Human Mathematics (AHM) lists 1,008 members on its public membership page, read by AI Lookout on 21 October 2026. Of those, 654 have also joined its AI-free caucus, whose members say they do not use AI models in their research at all.

The milestone was reported by New Scientist on 20 October, which described the two tiers with a food analogy: "vegetarians" who refuse to work with AI companies or publish AI-written texts, and "vegans" who also keep AI out of their research. New Scientist calls the group the "Association of Human Mathematicians"; its own website uses "Association for Human Mathematics".

Membership is free and public, and the AHM says it updates its list about once a week, so the exact count will move. It is open to people in the mathematical research community at graduate level and above. Undergraduates and people without a research role are directed elsewhere, because the group describes itself as a professional organisation.

The three commitments

Every member signs two commitments. The first is not to work with commercial AI companies by providing "technical labor, mathematical knowledge, consultation, or publicity." The second is not to publish mathematical texts generated by AI models, a category the AHM spells out as papers, reviews, referee reports, lecture notes, teaching materials and books. "Members only publish their own writing," the page says.

The third commitment is optional and defines the AI-free caucus. Its members "do not use AI models in a research capacity," which the AHM describes as finding proofs, exploring examples and posing research questions without prompting AI tools.

The group's FAQ makes the lines less absolute than the headline numbers suggest. It says the commitments are "expressions of principle rather than precise legal agreements" and that members should follow their conscience. Working for an AI company covers taking a job there, helping its staff advertise or improve products, endorsing proprietary models and knowingly giving mathematical expertise or data for training. Talking to AI industry staff about the effects on mathematics, or debating them in public, is explicitly allowed.

On writing, the FAQ says the ban targets "prompt-based writing, where the human element consists mainly of feeding instructions to a machine." It names one example of the kind of text the second commitment rules out: the Navier-Stokes document produced by OpenAI. Disclosing AI use in a paper is not treated as promoting a company, and journal editors who sign up can still decide for themselves whether to publish AI-generated submissions. People who have worked with AI companies in the past can join.

Why the group formed now

The only statement on the AHM's website is a response to OpenAI. On 6 October OpenAI released 722 AI-produced mathematical manuscripts with Lean proof files in a single repository. The next day the AHM's communications working group posted a statement saying that "mathematicians did not ask for this work to be done" and that "releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power." It urged mathematicians to stop working with OpenAI.

The statement also accuses OpenAI of ignoring the Advisory Group on Mathematics and Artificial Intelligence, an independent panel set up to advise AI companies. According to the AHM, the panel's first statement said frontier AI companies should not test advanced mathematical problems on internal models. New Scientist reports that OpenAI had said it would take the group's advice on board, but that its latest release appears to go against many of its suggestions.

New Scientist links the AHM's growth to that release and to criticism from Terence Tao of the University of California, Los Angeles, who said last month that the way AI results are being "dumped" on mathematicians could harm the field. Even when such results are formally verified, the magazine notes, they are often written in a hasty, impenetrable way that leaves mathematicians struggling to understand what was actually learned.

What members say

Sam Raskin of Yale University told New Scientist he joined but did not sign the third pledge, because he was unsure whether his occasional use of AI for simple tasks would count. He called the pledges "pretty mild" and said members join for different reasons, from environmental worries to job security. He also raised a disclosure concern: "I keep hearing about people who are in positions of leadership in the mathematical community ... and are quietly being paid by either Anthropic or OpenAI and are not disclosing this information necessarily. And I think that's toxic."

Steven Sivek of Imperial College London, who is in the AI-free caucus, pushed back on the idea that the group is simply anti-AI. "We're not denialists," he said. "It is happening, it's real. We need to figure out how to adapt the profession to this new emerging reality." He added that he is not comfortable "putting out things that I don't understand, that I can't vouch for." Hervé Le Dret of Sorbonne University cited what he called AI's "scandalous waste of resources and environmental impact."

OpenAI and Anthropic did not respond to New Scientist's request for comment. AI Lookout did not contact the AHM for this article; quotes come from New Scientist and the group's own pages.

Working groups and a first recommendation

The AHM's activities page says it is run collectively, with an organising committee of volunteers and working groups on teaching, journals, preprints, careers, coalitions with other organisations, and communications. Most meet weekly. The preprints group is discussing how to protect preprints from AI training and how to label them. The careers group is looking at hiring and promotion from the point of view of applicants, reviewers and departments. The teaching group is working on recommendations for undergraduate and graduate education.

Its first formal output, posted on 10 October by the journals working group, is a preliminary recommendation on author choice and consent. It says journals, conferences and preprint servers should tell authors before submission whether and how large language models will be used in refereeing, typesetting or any other stage, including by editors, referees and outside copy-editing services. Authors should be able to refuse "without being prejudiced in any way in the reviewing process."

The recommendation anticipates the obvious objection that journals are flooded with submissions and need LLMs to cope. Its answer is that the pressure does not remove the duty to be transparent. It allows AI tools for narrow jobs such as plagiarism checks, as long as authors are told, while warning that AI-writing detectors have many false positives. It also suggests journals could offer a "menu" of options instead of a yes-or-no choice, and that accidental exposure to AI, for example through search summaries, should be ignored under a good-faith standard.

How it fits the wider fight over AI maths

The pledge lands in a month when AI labs have been pushing hard into mathematics. Besides OpenAI's manuscript dump, Google DeepMind published AlphaProof Nexus results on open Erdős problems in Science, and OpenAI's earlier Navier-Stokes blow-up proof is still under review, with researchers finding that its Lean code does not match the paper.

The labs are also offering money. On 15 October the Simons Foundation said it, OpenAI and Anthropic would commit $15 million to help mathematicians evaluate AI-assisted results, with model access and compute for supported researchers. Under the AHM's first commitment, members would presumably need to think hard before taking part in anything the labs run directly, though the FAQ does not address foundation-managed programmes.

Other mathematicians have taken a middle path. In an earlier New Scientist piece, Alessandro Della Corte of the University of Camerino argued that it would be healthy for a small minority of mathematicians, perhaps 5 to 10 per cent, to stay AI-free while most engage with the tools critically. Inside the AHM the AI-free tier is no small minority: 654 of its 1,008 listed members, nearly two thirds, say they keep AI out of their research entirely.

What is still unclear

The membership list is self-reported and the commitments are not enforced; the AHM says so itself. It is not clear how many members are senior researchers, how many are graduate students, or how the figure compares with the size of the research community. The group has not published a full policy programme beyond the journals recommendation, and its working groups are still at the discussion stage.

It is also unclear how the pledges will work in practice. Many mathematicians already use AI-powered search, autocomplete in editors or formal proof assistants with AI features, and the FAQ leaves the boundaries to individual judgement. Universities, funders and journals have not said whether they will respond to the AHM's demands, and AI companies have not commented publicly.

Why it matters

For AI companies, mathematics has been a showcase: results that can be checked by machine make capability claims easier to believe. The AHM is an organised sign that part of the community those results are meant to impress does not want to supply expertise, consulting or publicity for that showcase, and wants a say in how journals and preprint servers use AI.

Its first practical demand, that authors be told about and able to refuse LLM use in peer review, is narrow and concrete, and could spread beyond mathematics if other fields copy it. Whether the group changes how labs release mathematical results will depend on whether journals, societies and funders take up its recommendations, and whether the labs start treating mathematicians' consent as part of how they publish.