Confirmed in a primary source 11 Sept 2026 Research

Twenty-five Fields Medal winners signed a single page saying the AI companies racing to solve famous problems are damaging mathematics.

On Friday a declaration titled A Severe Misalignment of AI in Mathematics appeared at mathandai.org carrying 25 winners of the Fields Medal, mathematics' most prestigious prize, spanning awards from 1978 to 2026. Their argument is that solving stated problems is only a proxy for what the field is for, which is understanding, and that announcing machine-generated solutions without writeups, attribution or peer review breaks the chain by which methods pass from one mathematician to the next.

The caveat is that a declaration compels nobody: it asks for a discussion, sets no rule, names no company and carries no sanction.

Until now

Pushback on AI mathematics came from individual mathematicians one blog post at a time, and from institutions with a stake in the event being sponsored.

Now

Put more weight on the discipline these companies most want validation from organising against the way they seek it, and on peer review becoming the contested ground; put less weight on a proof announced by a lab being a result the field has accepted.

The declaration went up on Friday at mathandai.org, and Terence Tao posted the same text on his own blog. Twenty-five Fields Medal winners signed, from awards given in 1978 through to 2026. The text does not claim AI cannot do mathematics. It says the thing AI companies measure, solving stated problems, is "only a tool and proxy for achieving the primary goal of conceptual understanding and insight," that solutions arriving without documentation or attribution raise "plagiarism questions," and that on this path "the human transmission chain between mathematicians would be lost." It asks for a discussion among mathematicians, companies and society. It names no company and sets no rule, and anyone with an academic email can add their name.

The week behind it explains the timing. OpenAI published a proof of one of the seven Millennium Prize problems on September 8 and credited a model it has not released. On September 10, 771 mathematicians asked Caltech to suspend an event funded with 2 million dollars of Anthropic and OpenAI credits, and OpenAI pulled its sponsorship within hours. The same day, a second mathematician in a week asked OpenAI to show it had not trained on his private conversations with ChatGPT. The people who would have to check any of these results are the people who signed this page. Put more weight on the discipline these companies most want validation from organising against the way they seek it, and less on a proof announced by a lab being a result the field has accepted.

How it got here

Sep 8
OpenAI publishes a proof of the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize problems, and credits an unreleased model. Carded in Edition 016.
Sep 10
An open letter signed by 771 mathematicians asks Caltech to suspend its October 30 Mathathon, funded with 2 million dollars of Anthropic and OpenAI credits. OpenAI withdraws its sponsorship the same day.
Sep 10
Andreas Thom becomes the second mathematician in a week to ask OpenAI to show it did not train on his private ChatGPT conversations.
Sep 11
Twenty-five Fields Medal winners publish A Severe Misalignment of AI in Mathematics at mathandai.org, open to further endorsement by ORCID or academic email.
Who has to react
OpenAI and Anthropic, both of which have published machine-generated proofs in the past two weeks and need referees drawn from this group; the journals and arXiv moderators who decide what a submitted machine proof must contain; Caltech, whose Mathathon is still scheduled for October 30 with one sponsor gone; the Clay Mathematics Institute, which owns the Millennium Prize rules and has said nothing about what an AI-assisted claim has to satisfy; hiring and tenure committees, which count publications; and every doctoral student whose thesis problem a model might finish first.

The evidence

ObservedA declaration titled A Severe Misalignment of AI in Mathematics was published at mathandai.org on Friday September 11, 2026, and Terence Tao posted the same text on his own blog the same day. We read both.

The whole observation, as recorded

It lists 25 Fields Medal recipients as endorsers, with awards ranging from 1978 to 2026; named signatories include Artur Avila, Manjul Bhargava, Caucher Birkar, Pierre Deligne, Simon Donaldson, June Huh, James Maynard, Peter Scholze and Tao. The medal is given to at most four mathematicians under 40 every four years, so 25 signatures is a large share of the living holders. The central line is that "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight." The text objects that AI solutions arrive without documentation or attribution, which it says raises "plagiarism questions"; that the rush to announce prevents "writeup, the isolation of new methods and ideas"; and that on the current path "the human transmission chain between mathematicians would be lost." It calls for urgent discussion among mathematicians, technology companies and society, and collects further endorsements through ORCID authentication or academic email confirmation, which means non-medallists can sign. It names no company, states no rule and proposes no sanction. The dated week behind it: OpenAI published a proof of the Navier-Stokes existence and smoothness problem on September 8, credited to a model it has not released, carded in Edition 016. On September 10 an open letter signed by 771 current and former Caltech mathematicians and PhD-level researchers asked Caltech to suspend its October 30 Mathathon, funded with 2 million dollars of AI credits from Anthropic and OpenAI, and OpenAI withdrew its sponsorship the same day, recorded in Edition 018's Undercurrents. Also on September 10, Andreas Thom became the second mathematician in a week to ask OpenAI to show that it had not trained on his private ChatGPT conversations. No AI company has responded to the declaration as of this morning.

Established

A declaration titled A Severe Misalignment of AI in Mathematics was published at mathandai.org on September 11, 2026, listing 25 Fields Medal recipients as endorsers.
Its stated objection is to problem solving being used as the measure, and to announcements made without writeups, attribution or peer review.
Terence Tao published the same text on his own blog the same day.
The declaration names no company, states no rule and asks for a discussion, with further endorsements collected by ORCID or academic email.

Not established

That any announced AI proof is wrong. The declaration does not contest a single result.
That anything changes. No journal, funder, prize body or university has adopted it.
That the field is united. Twenty-five is a large number of medallists and a small number of mathematicians.
That AI companies trained on private work. That is a separate allegation by two mathematicians and it is unresolved.

Why it weighs this much

Signal strength

CostLow

CommitmentHigh

ReversibilityHard

PersistenceYear

A signature costs nothing to produce, and we set cost low rather than pretend otherwise. What it does cost is optional: these are people with nothing to gain from a public fight with well-funded companies that fund mathematics departments, and several of them are cited by those companies as validation. Commitment is high because the names are attached in public with no hedge and no anonymous option. Reversibility is hard for the same reason; a signature can be withdrawn but not unseen. Persistence is a year rather than structural because a declaration with no rule behind it decays unless a journal, a funder or a prize body turns it into a requirement, and none has.

What to watch

strengthens A mathematics journal, arXiv, the Clay Mathematics Institute or a national funding body publishes a rule or submission requirement covering AI-generated or AI-assisted proofs, by June 30, 2027.
weakens No journal, funder, prize body or university department adopts any AI-proof requirement and total endorsements of the declaration stay below 500, by March 31, 2027.

The words

Fields Medal: Mathematics' most prestigious prize, given to at most four mathematicians under 40 every four years. There is no annual award, which is why 25 holders is a lot of them.
Misalignment: A gap between what a system is optimised to do and what the people using it actually want. Here the mathematicians apply the word to the companies, not to the models.
Peer review: Other specialists checking a result before it counts as established. It is unpaid work, and it is the step the declaration says is being skipped.
Benchmark: A fixed set of test problems used to score a model. Solving them is what AI labs publish; the declaration's point is that the score is not the goal of the field.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.