25 Fields Medalists sign an AI attribution declaration

On September 11, Terence Tao published a declaration signed by 25 Fields Medalists arguing that AI labs racing to solve famous math problems are damaging the field they are benchmarking against. The word they chose is "misalignment," and the core complaint is not that the machines are wrong. It is that nobody can tell who did what. AI attribution is now a live dispute at the top of mathematics, and the same failure mode is sitting in your pull request history.

What actually happened

The statement, A Severe Misalignment of AI in Mathematics, argues that "solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight," and that mass production of true/false statements "could destroy fertile ground instead of breathing life into new ideas."

The operational complaint is sharper. Solutions are "announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work" — which the signatories say raises "severe attribution and plagiarism questions." The initial 25 span Fields classes from 1978 to 2026, including Tao, Peter Scholze, Maryna Viazovska, Maxim Kontsevich and this year's medalist Yu Deng. The declaration is open for further signatures at mathandai.org.

The timing is not subtle. TechCrunch reports that NYU's Tristan Buckmaster accused OpenAI of pressuring him not to credit a collaborator employed by Anthropic, and that OpenAI withdrew sponsorship of a CalTech math event after researcher criticism. TechCrunch also notes mathematicians questioning whether their own Codex sessions fed the models that then raced them. We have no confirmation of that last one, and neither does anyone else — which is precisely the complaint.

Why AI attribution matters for your business

Strip the Fields Medals out and this is a procurement question you already have. Two parts.

First, credit inside your team. When an agent writes 300 lines and an engineer reviews and ships them, your commit log says one human did it. That is fine until a client audit, an acquisition diligence, or a dispute over who owns a contribution. Decide now what a coding agent's involvement looks like in your history — a trailer, a PR label, something greppable — and apply it consistently. Retrofitting provenance across 18 months of commits is not a task anyone completes.

Second, credit leaving your building. Read the retention and training terms for every AI tool your team runs, not the marketing page. You want three answers in writing: how long prompts and outputs are retained, whether they can be used to train or improve models, and whether a business tier changes either. Enterprise plans frequently default to no-training; free and personal tiers frequently do not, and your engineers signed up for the free tier in March. That gap is where your unreleased work walks out.

The mathematicians have a public reputation system to defend theirs. You have a contract. Go read it.

Sources: Terence Tao: A Severe Misalignment of AI in Mathematics, TechCrunch.