Responsible Sharing of AI Math Tools

Key Takeaways:

OpenAI announced an independent Advisory Group on Mathematics and AI on September 21, 2026, with nine members hosted at the Institute for Advanced Study, after its model reportedly resolved more than 100 open problems.

The group advises on release but has no power to slow or redirect research, a limitation the Institute itself confirmed in writing.

On September 29, 2026, the group published general guidelines for responsible release, informed by more than 600 survey responses, covering attribution, provenance, formalization, and funding for human understanding.

Twenty-five Fields Medalists signed a September 11, 2026 declaration calling the goals of AI companies and the mathematical community “severely misaligned.”

Only one of the nine advisory members, Camillo De Lellis, had signed that protest letter, which critics cite as evidence the group is not as independent as its name suggests.

The Break That Forced the Question

On September 8, 2026, OpenAI announced that an internal model had produced a proof resolving the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, using roughly 10,000 concurrent AI agents, according to VentureBeat’s coverage of the announcement. We covered the mathematical substance of that claim and the priority dispute around it in our analysis of the Navier-Stokes result.

The claim was poorly received by parts of the mathematical community. On September 11, 2026, 25 winners of the Fields Medal released a declaration titled “A Severe Misalignment of AI in Mathematics,” arguing that the push by AI companies to solve famous problems as a benchmark harms the science of mathematics and the mathematical community. The declaration, published at mathandai.org with a Zenodo DOI, states that the goals of AI companies and the goals of the mathematical community “are severely misaligned.” Signatories include Terence Tao, Wendelin Werner, and Martin Hairer.

The declaration’s main argument concerns process rather than capability. It notes that research mathematics advances through talks, discussions, and simplification, a long and difficult human chain that turns a result into something a graduate student can learn. Solutions announced in haste, it says, leave no time for proper writeup, isolation of new methods, or citation of prior work. Scientific American reported that the declaration closed what it called perhaps the most consequential week in the history of mathematics.

The Advisory Group: Independence Without Decision Power

The Advisory Group on Mathematics and Artificial Intelligence, described on its own site at agmai.org, states its purpose clearly: to advise AI companies on their interactions with mathematical research, including responsible presentation and release of mathematical results. Its nine named members include Timothy Gowers, Martin Hairer, Ravi Vakil, Edward Witten, and Melanie Matchett Wood.

What the September 29 Guidelines Actually Require

On September 29, 2026, the group published its first substantive output: a document titled “Responsible Release of AI-Generated Mathematics,” informed by over 600 replies to a community survey. The guidelines divide release into two paths depending on whether a human understands the result.

For papers a mathematician fully understands, the recommendation is to follow traditional norms: post a preprint, submit for peer review, and give talks. For results not yet understood by anyone, the document lays out a stricter sequence. It asks labs to scan the literature for related ideas and cite papers where those ideas first appeared, even if the model rediscovered them independently. It asks that proofs be rewritten in conventional paper style, with precise theorem statements rather than wordy reasoning and non-standard terminology. It recommends depositing results in repositories not controlled by any AI lab that guarantee persistent citable identifiers.

The provenance requirements are the most operationally specific. For each result, the lab should make public the model name, the prompts used, a summarized chain of thought, the time taken, and the estimated cost of computation. Where possible, proofs should be formalized, with artifacts meeting community standards including copyright headers, a comparator challenge file, and a formalization.yaml file. The document also asks labs to document how many comparable problems the model tried and failed to solve, and how problems were chosen, to avoid the selection effect of announcing only successes.

One recommendation stands out. The group states that it does not endorse testing advanced mathematical problems on proprietary models inaccessible to the broader community, and asks labs to stop. It also recommends that labs fund the human work of understanding their output, through conferences, summer schools, postdocs, or expository books, while insisting that the development of human understanding must remain organic and community led.

The Credit and Data Dispute Behind the Guidelines

The guidelines respond directly to the Navier-Stokes controversy. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been pursuing the same smooth forcing route to the problem, and Buckmaster released a public statement questioning whether OpenAI’s model had benefited from their unpublished work. The two had been putting drafts into private Codex sessions, and Buckmaster wrote: “I do not know whether our data was used. I am not accusing anyone of anything.”

OpenAI’s response included a qualification that matters beyond this dispute. In a statement posted to X, the company said researchers and agents did not see any of their work through any means until they released it publicly, then added that while unlikely, they cannot rule out that de-identified data derived from their usage of OpenAI’s products helped improve the models, as reported by VentureBeat. The distinction between retrieving a user’s private work and learning from de-identified data derived from product usage is exactly the gap the group’s provenance recommendations address.

For enterprises, the practical question is narrower: what protections actually prevent proprietary research fed into an AI system from becoming part of future model improvement? The OpenAI statement does not answer this, and the group’s guidelines address release norms rather than the training-data pipeline.

A Release Workflow You Can Implement

The guidelines translate into a checklist any lab or research team can adopt when releasing AI-assisted results. The diagram below maps the two paths.

The provenance block is the part teams often skip. Here is a minimal manifest that captures the fields the guidelines ask for, written as plain structured data rather than a specific library:

The `problems_attempted` field is the one most releases omit. Publishing only the 100 successes, without the number of failures and how the problems were selected, produces a biased picture of capability. The guidelines ask for that denominator explicitly.

Two Release Paths Compared

The guidelines define two courses of action. They differ in who verifies the result and what the lab must supply.

Requirement	Path A: Human understands	Path B: Not yet understood	Source
Primary verification	Journal peer review	Formalization plus community review	AGMAI guidelines
Literature attribution	Standard citation norms	Lab must scan and cite prior work, even for independent rediscovery	AGMAI guidelines
Provenance disclosure	Not required beyond normal authorship	Model name, prompts, chain of thought, time, cost	AGMAI guidelines
Repository control	Any journal or preprint server	Neutral repository not controlled by the lab, with persistent ID	AGMAI guidelines
Formalization	Optional	Strongly recommended, with comparator challenge file	AGMAI guidelines
Funding for human understanding	Not required	Lab should fund conferences, postdocs, or expository writing	AGMAI guidelines

The choice involves speed versus comprehension. Path B releases results sooner but shifts the cost of understanding onto the community, which is why the guidelines pair it with a funding obligation.

What to Watch Next

The advisory group’s influence is reputational, not structural. It can publish recommendations and go public with dissent, but it cannot stop a release. Whether that is enough depends on whether labs find it costly to ignore their own advisors. The September 29 document already tests this: it asks labs to stop testing advanced problems on proprietary models, a request OpenAI has not publicly agreed to follow.

The community is creating parallel organizations. The Association for Human Mathematics, launched the same week as the declaration, is one channel for mathematicians who want a voice independent of any company’s advisory board. The two efforts overlap in personnel and purpose, which raises the question of whether an advisory group tied to a lab and an association independent of all labs can coexist without one lending legitimacy to the other.

For teams applying AI to research outside mathematics, the release checklist itself is relevant. The provenance fields, the selection-effect denominator, and the neutral-repository requirement apply to any domain where a model produces results faster than humans can verify them. Generation and verification should be separate steps, and the verifier should be independent of the generator.

I expect the group to publish at least one additional formal recommendation document beyond its September 29 general guidelines by April 30, 2027, given the number of open questions its first document left unresolved.

More in-depth coverage from this blog on closely related topics:

Sources and References

Sources cited while researching and writing this article: