A coalition of leading mathematicians, including Fields medallists and Terence Tao, published a statement on 11 September arguing that the goals of AI companies and those of the mathematical community are gravely misaligned. The text dominated technical discussions this weekend. Its argument is not that the models are weak: it explicitly concedes the opposite.
Four days ago we covered the announcement of a solution to the Navier-Stokes problem and the dispute over its authorship. Here is the collective response of the discipline concerned, and it is more interesting than one might have expected.
What the statement does not say
Let us begin by setting aside the easy reading. This text is not a corporatist defence by mathematicians who fear being replaced.
It explicitly acknowledges that the mathematical capabilities of the models have improved dramatically, and that they now solve open problems in several fields. That is a considerable concession from this community, and it makes the criticism much harder to dismiss.
The objection lies elsewhere, and it is methodological.
The argument: what a race for score does to a discipline
The central complaint is that treating mathematics as a performance contest damages mathematics. Three mechanisms are identified, and each deserves to be understood.
Hasty announcements. A result published in haste has not benefited from the writing time that allows it to be presented properly. Yet in mathematics, writing is not formatting: it is the moment when one understands what one has done.
Methods not isolated. This is the most important point and the least obvious. The value of a proof lies not only in the theorem obtained, but in the new ideas it introduces. A well-written paper separates those ideas, names them, shows how to reuse them elsewhere. A proof delivered raw gives the result and loses the method.
Prior work not cited. A proof always builds on decades of work. Failing to cite it does not merely deprive its authors of credit: it renders the result unintelligible, because one can no longer see where it comes from.
One could read the question of attribution as a matter of ego. That would be to miss the point. As we wrote in our article on the authorship of a discovery: attribution designates a person responsible, indicates whom to talk to, and directs resources. It fulfils an epistemic function before being a reward. A discipline in which one no longer knows where results come from can no longer verify or extend them.
The point where the criticism is strongest
There is an asymmetry that this statement lays bare, and it is structural.
An AI company has an interest in announcing quickly: the announcement has value in itself, independently of what it produces for the discipline. We noted this in relation to extraordinary claims: whoever asserts captures attention, whoever says they are waiting loses it.
A scientific community has an interest in publishing slowly: slowness is what makes verification, clarification and reuse possible. We detailed this in our article on the acceptance of proofs — twenty-six years, a single millennium problem validated.
These two temporalities are not reconcilable. The word "misaligned" is therefore technically accurate, and that is what makes the statement difficult to refute.
What can be argued against it
Two serious objections, for the sake of balance.
Slowness is not a virtue in itself. A discipline can also be slow out of conservatism, out of protection of vested positions, or because its journals are saturated. Defending the long timescale presupposes showing what it produces, not merely demanding it.
Companies are publishing more and more formally. We welcomed a formally verified proof as the right configuration, and the Navier-Stokes announcement also came with a formalisation. That is exactly what the community could have asked for, and it was done.
What to take away
This statement matters because it shifts the debate from capability to method, which almost nobody does.
The question is no longer "can these systems do mathematics" — this text concedes that they can, increasingly so. It is: can a discipline remain a discipline if the pace of production exceeds its capacity to understand, verify and transmit what is produced?
It is a question other fields will ask themselves, and probably sooner than they think. Mathematics encounters it first because it has a mechanical criterion of validity. For all the disciplines that do not, it will arise with fewer means of answering it.