Mathematicians' association urges OpenAI boycott after release of AI-written proofs

After OpenAI published hundreds of AI-generated mathematical manuscripts, the Association for Human Mathematics (AHM) has accused the company of violating core norms of scientific research and is urging mathematicians to stop working with it. Fields Medalist Terence Tao, who chairs the group, posted the statement as a guest post on his blog. He also outlined a 'Math 2.0' era on Mastodon, in which solving problems should no longer be the main focus of the discipline. Complexity theorist Scott Aaronson called the situation a 'Mathocalypse' on his blog. The report comes from The Decoder.

The statement escalates a months-long debate. OpenAI had already caused a stir by claiming that an internal model solved more than 100 open math problems in a single month, including an approach to the Navier-Stokes Millennium Problem. After growing criticism, OpenAI set up an advisory group, AGMAI, at the Institute for Advanced Study, though the group explicitly has no say over the pace of the company's internal research. The AHM says OpenAI has already ignored the central premise of AGMAI's advice: that advanced math problems shouldn't be tested on internal models. 'Mathematicians did not ask for this work to be done,' the statement reads, describing the release of more than 700 files at once as 'not a demonstration of scholarship, but a demonstration of power.' Its opening paragraph also mentions the copyright lawsuits OpenAI faces around the world, which The Decoder reads as suggesting, without saying so directly, that the results were possible only because the company trained on mathematicians' work.

The Navier-Stokes episode fed the same dispute. OpenAI recently announced a solution shortly before two mathematicians could present their own AI-assisted one. The two had used ChatGPT, and OpenAI denied suspicions that it used data from those interactions to train its own system and reach a solution faster.

Aaronson contrasts the release with what he calls the 'Anthropic model'. OpenAI drops raw AI proof drafts in one batch, setting off a race to work through them. Anthropic instead partnered with two algorithm researchers whose AI model supplied the key idea for disproving two decades-old conjectures; the researchers were paid and wrote a version of the proof humans could follow. The Decoder notes both approaches have drawbacks: OpenAI's leaves the community doing the unpaid work of making proofs readable, while Anthropic's lets a private company pick which mathematicians act as 'emissaries' for a result.

Tao had earlier joined 24 other Fields Medalists, including Peter Scholze, Maryna Viazovska and Martin Hairer, in a statement warning of a 'severe misalignment' between the AI industry's goals and those of mathematics. Problem-solving, they argue, is only a tool for the real goal of conceptual understanding, and mass-producing solved problems could 'destroy fertile ground instead of breathing life into new ideas.' They also criticized rushed announcements without proper write-ups or citations of prior work, saying this raises 'severe attribution and plagiarism questions.'

On Mastodon, Tao goes further. In traditional mathematics, a proof of a long-standing problem leads to talks, workshops, collaborations and textbook inclusion, drawing in young researchers. Now, he sees problems solved autonomously by AI users with no interest in the field, who can't understand the results well enough to answer questions or give talks, producing far fewer seminars and collaborations. Researchers are holding back promising directions for fear of being scooped. For Tao the damage is irreversible: once a problem is considered solved, it can't be made unsolved, and even knowing a solution exists can 'contaminate' the search for other approaches. Solutions are being 'harvested' at scale, leaving branches of mathematics less fertile. He argues that 'Math 2.0' must stop treating problem-solving as the main measure of progress and value explanation, community-building and new research directions, with matching changes to training, publication and career advancement. This builds on his 2024 vision of 'industrial mathematics', where AI would assist humans much as engines assist chess players. Today, Tao sees the roles reversed: AI solves the deep problems and humans work through the results afterward.

Aaronson describes the personal impact through his wife, Dana Moshkovitz, who has devoted her entire career to the Unique Games Conjecture. The release includes a claimed proof of it. In text messages on the night of the release she wrote, 'It feels like something written by someone who's on psychedelics.' The paper cited numerous works without explaining why they applied, even though earlier results should have ruled that out. 'Basically the paper is so horribly written that it's impossible to read it without AI help,' she said, adding that the proof relied on a bizarre new construction, 'some alien craziness.' Aaronson lists other results in complexity theory, number theory and algorithms, including partial progress on several remaining Clay Millennium Problems; in his view any one of them would have ranked among the year's top results.

Cryptography is notably absent. Citing unnamed sources, Aaronson reports that AI companies are quietly probing weaknesses in cryptographic protocols and primitives. About 8,000 problems were tested overall, with a success rate of roughly five percent and an average of three hours of compute at GPT-Pro level per problem. The model used was likely OpenAI's current internal one, which could ship to paying ChatGPT customers in the coming months.

The field is divided. Commenters on Tao's post include supporters of the boycott, who propose that researchers stop using OpenAI products and call for better protection of preprint servers like arXiv against mass collection of training data. Others call the AHM's position unrealistic, arguing OpenAI won't stop working on mathematics and that public results are better than private ones. One commenter asks whether a problem counts as solved when neither the authors nor anyone else fully understands the proof. For Navier-Stokes, it remains unclear whether the result deepens anyone's understanding of fluid dynamics. AGMAI takes a more diplomatic line than the AHM, calling the release a first step with mathematical understanding of the results only now beginning, but warns that the future of math research can't consist of working through AI-lab results. It writes that mathematicians must be able to formulate their own questions, develop their own approaches and explore directions not selected as showcases of an AI system's capabilities.

Key facts

The Association for Human Mathematics, chaired by Fields Medalist Terence Tao, accuses OpenAI of violating core research norms and urges mathematicians to stop working with the company.
The statement describes the release of more than 700 files at once as 'not a demonstration of scholarship, but a demonstration of power.'
The release includes a claimed proof of the Unique Games Conjecture, which Dana Moshkovitz, who has devoted her career to it, found nearly unreadable without AI help.
Tao calls the damage irreversible: once a problem is considered solved it can't be unsolved, and solutions are being 'harvested' at scale.
Commenters on Tao's blog are split between backing the boycott and calling it unrealistic; OpenAI's own advisory group AGMAI takes a milder line.

Why it matters

This is a public clash between a mathematicians' body and a leading AI lab over what counts as a solved problem. The AHM's complaint is not only about quality but about process: a batch of more than 700 files lands at once, and the community has to work out what is in them. Tao's argument is that problem-solving was never the discipline's real goal; conceptual understanding was. If AI produces proofs faster than people can understand them, he argues, the incentives of training, publication and careers need to change. It also revives a 2024 picture, Tao's 'industrial mathematics', in which humans set the pace and AI assisted. Tao now sees those roles reversed.

Who it affects

Research mathematicians, first of all. Those working on long-standing open problems, like Dana Moshkovitz on the Unique Games Conjecture, face claimed proofs of their life's work that they struggle to read. Young researchers who would traditionally be drawn in by talks, workshops and textbooks are affected too, as are people who now hold back promising directions for fear of being scooped. OpenAI is directly in the line of fire, as are the AGMAI advisory group at the Institute for Advanced Study and, as a point of comparison, Anthropic. Maintainers of preprint servers like arXiv come up in the comments, where some call for better protection against mass collection of training data.

How to use it

Nothing here is a product to adopt. For readers in the field, the practical content is the proposals on the table. The AHM urges mathematicians to stop working with OpenAI, and some commenters go further and suggest stopping use of its products. Tao's 'Math 2.0' agenda asks the field to reward explanation, community-building and new research directions, with matching changes to training, publication and career advancement. Aaronson sketches an alternative he calls the 'Anthropic model': pay the researchers involved and have them write a version of the proof humans can follow. Per The Decoder, the model behind the release could reach paying ChatGPT customers in the coming months.

How solid is it

The AHM statement and its quotes come via Tao's blog and The Decoder's account of it, and the Mastodon points are Tao's own. The Decoder's reading that the statement hints at training on mathematicians' work is the outlet's interpretation, not something the statement says directly. The figures on about 8,000 problems tested, a success rate of roughly five percent and three hours of GPT-Pro-level compute per problem come from unnamed sources relayed by Aaronson. The source gives no confirmation that any of the 700+ proofs is correct or has been verified. The count is also loose: the lead says 'hundreds' of manuscripts, the AHM quote says more than 700 files, and the two are not reconciled.

Risks and caveats

The dispute is unresolved and the field is split. Critics of the boycott argue OpenAI won't stop working on mathematics and that public results are better than private ones. The source does not say how many people have joined the boycott, nor what it involves beyond 'stop working with the company', and OpenAI's response to the AHM statement is not reported. Tao's concern is that damage cannot be undone: once a problem is considered solved, that status can't be reversed, and knowing a solution exists can 'contaminate' the search for other approaches. Both release models have drawbacks: OpenAI's leaves the community doing unpaid work to make proofs readable, while Anthropic's lets a private company choose which mathematicians serve as 'emissaries'. For Navier-Stokes, it is unclear whether the result deepens anyone's understanding of fluid dynamics.

“Mathematicians did not ask for this work to be done”

— Association for Human Mathematics statement