More than a thousand mathematicians have joined the Association for Human Mathematics, pledging to avoid work for AI companies and publication of AI-written texts, with 654 also pledging to forgo AI in their research

More than a thousand mathematicians have joined the Association for Human Mathematics, pledging to avoid work for AI companies and publication of AI-written texts, with 654 also pledging to forgo AI in their research

New Scientist reported on the association on October 20. Its members include Peter Scholze, a recipient of the Fields Medal, the leading award in mathematics. The association announced its launch in early September. On October 7, the day after OpenAI released 722 mathematics papers written by an internal model, it called on colleagues to stop working with the company.

The association identifies several problems: the AI industry has chosen mathematics to promote its products, journals face a surge in papers written with AI assistance, and students are losing the opportunity to develop a deep understanding of the subject. Members make two commitments: to withhold their labour, expertise, advice and promotional support from commercial AI companies; and to refrain from publishing papers, reviews, teaching materials and books produced by AI models. A third commitment is optional: those who make it, including Scholze, form an AI-free faction and refrain from using models to search for proofs, examine examples or choose problems.

In its answers to frequently asked questions, the association explains that prohibited consulting includes supervising a project in which a company tests its model's capabilities. Discussions with industry about AI's impact and public debates are encouraged. The second commitment concerns texts for which the human mainly supplies prompts and the model does the writing. The commitments express principles, and members follow them according to their own conscience.

According to archived copies of the website, the membership list contained 138 names on September 10, 703 on September 27 and 1008 on October 19: more than seven hundred people had joined nine days before OpenAI's release.

In its October 7 statement, the association writes that OpenAI cites an independent advisory group on mathematics and AI as a source of legitimacy while ignoring its central recommendation: companies should refrain from testing internal models on difficult problems. The group of nine mathematicians formed after OpenAI approached several of them about establishing an external advisory body. According to participants in an August meeting with mathematicians, OpenAI promised to avoid releasing the results all at once; the company says it is unaware of such a promise. The association wrote:

“Releasing more than 700 files at once is a display of power rather than a demonstration of scholarship.”

In its September 29 recommendations, the group asks companies to disclose the model, prompts and computing costs when releasing results. In a collection of members' comments, the association writes that the group sought to reduce the harm caused by bulk releases, but that its recommendations risked giving companies a “road map” for further mathematical extraction, now with the “fig leaf” of community-approved practices.

The association has established working groups to agree on what mathematics should look like following the arrival of AI, Steven Sivek of Imperial College London told New Scientist: “We are not denialists… This is happening, it is real.” On October 10, the journals working group issued preliminary recommendations that journals and other publishing platforms tell authors in advance how they use AI and allow authors to opt out of AI use in peer review without affecting the assessment of their work.

Alessandro Della Corte, a member of the AI-free faction, considers it less than ideal for all mathematicians to use a single technology as an unavoidable intermediary between themselves and mathematics. He told New Scientist on October 1:

“I think it would be better if a small minority, say 5–10 percent, remained AI-free, while most of the community worked with these tools, hopefully critically and within academia.”