On September 11, at 8:01 p.m. Paris time, Le Monde published a declaration signed by 25 Fields Medalists — Terence Tao among them — titled “A Severe Misalignment of AI in Mathematics.” Their complaint is not that the machines can’t do math. It’s that the companies building them have converted the discipline’s most famous open problems into leaderboard entries. Tao’s Caltech slides, Math 2.0, posted this month and sitting on the front page of Hacker News today, are the long version of the same argument: mathematics ran for nearly a century on what he calls proof scarcity, and abundance is about to break the machinery that made proofs mean anything.

The Letter Does Not Ask the Labs to Stop

Worth saying plainly, because the reaction has mostly been guild-protection kayfabe: the signatories concede the tools work. Tao’s own synthesis — an ICM 2026 lecture in July, now an essay — treats this as a second crisis in foundations, not a threat to be litigated away.

The objection is to the sequence. Solve, announce, move on. A result shipped as a benchmark score arrives with no paper, no distillation, no simplification, no doctoral student forced to reconstruct the argument for a seminar and in doing so finding the lemma the announcement skipped. As Le Monde summarized the letter, the mathematical community is a miniature of humanity whose most precious resources are students and ideas, and both take years to grow.

So the signatories did something the standard-issue academic petition never does. They pointed at an alternative: Maththon, November 13–15, a student-run event that aims open-source AI tools at existing results whose proofs are ugly or incomplete. Not famous problems. Unfinished plumbing. That is not Luddism. That is a rival institution being built in public, on a calendar.

Who Mows the Commons

The benchmark race has a cost structure nobody has priced. Compute is paid for. The referee who checks the result is not. The conjectures that never got published because they were going to be somebody’s thesis are not. The four years of a graduate student’s life spent learning to crank prompts instead of proving lemmas are not. Referee labor has been unpaid forever, which is precisely why it is the easiest thing in the building to raid.

There’s a telling artifact from October 7, when a developer-community thread advertised “first look at mathematics manuscripts from an internal frontier model.” The output was described as inert true-or-false statements until humans digest them — the exact point the letter makes. A theorem is not knowledge. Knowledge is a theorem plus a transmission chain: taught, simplified, connected to something else, argued about in a hallway. The labs are optimizing the first half and consuming the second.

A third-year doctoral student in number theory, typing in the group chat that lit up the night the letter dropped, put it more bluntly: “We told ourselves we’d be the ones using these tools. Then a lab cleared a problem in my subfield over a weekend and posted a thread before anyone wrote it up. There’s nothing left for me to work on that won’t be a footnote by spring.”

Why the Free-Market Heckle Gets This Backwards

The predictable right-of-center reflex is that this is buggy-whip nostalgia — incumbents defending a cartel against a cheaper competitor, and haven’t we seen this movie. It’s lazy here, for a reason that should matter to anyone who has spent a decade arguing about externalities.

Normally the incumbent is objecting to a competitor beating him in a market. Here the incumbent is objecting that the competitor is dumping a product into a market with no price mechanism at all — and that the dumping destroys the only thing actually being sold, which is the transmission chain. There is no price signal that will make a lab slow down, because the lab isn’t buying anything from mathematicians. It’s withdrawing from a commons and calling the withdrawal progress. That’s an unpriced externality wearing a benchmark’s clothes, and the correct free-market answer to an externality is not a new agency. It’s a property arrangement: citation norms, embargo conventions, credit rules, and the willingness to enforce them.

Which is the part that ought to make the comfortable reader squirm on both sides. The signatories spent the last two years on labs’ advisory boards and chasing their bounties; several of the loudest names on that letter were, until quite recently, the technology’s most quotable enthusiasts. And the companies, for their part, have hired the field’s best students and now want to be praised for publishing the results of work nobody in the field had time to absorb.

November Is the Only Vote That Counts

The letter has 25 signatures and no enforcement mechanism. Declarations rarely do. But the signatories have done the one thing that converts complaint into leverage: they set a date and named a tool. If Maththon produces proofs a working mathematician would cite, the declaration becomes a founding document — a commons that organized itself before the last of it was scraped. If it produces a weekend of prompt-wrangling and a handful of demos, the letter will be remembered as a guild’s last memo, filed after the machinery had already shipped.

Commons aren’t defended by being valuable. Everything valuable gets taken. They’re defended by being organized, and organizing is dull, slow committee work that produces no benchmarks and wins no grants. That’s why it’s rare. The mathematicians have given themselves roughly a month to find out which kind of community they are.

I’d take the over on the students. Not because they’re idealists — because they’re the ones who’d otherwise be the footnote.

Sources