Twenty-five Fields medalists signed a warning.
Their statement is short and unusually blunt for mathematicians: the goals of the AI companies and the goals of the mathematical community are severely misaligned. They frame it not as a math-specific spat but as part of broader alignment issues hitting other scientific and creative professions. Terence Tao is among the names attached to the effort.
I build bots for a living. My first instinct with any statement like this is to ask what it means for the systems I actually ship. And the answer turned out to be more practical than I expected.
Misalignment here is not the sci-fi kind
When people in my corner of the industry say “misalignment,” they usually mean a model doing something its operators did not intend. That is not what is being described. This is a mismatch of goals between two groups of humans, both of whom are perfectly aligned with their own incentives.
An AI company wants a system that produces correct answers to hard problems, verifiably and at scale. That is a legitimate goal. It is also measurable, which makes it fundable.
The mathematical community, per its own statement of values, cares about nurturing students and ideas. That is the part that does not fit into an eval use. You cannot benchmark a graduate student’s slow development of taste. You cannot score the half-formed idea that goes nowhere for three years and then becomes the key to something else.
Both parties are optimizing honestly. They are just optimizing different objective functions, and one of them is much easier to write down.
The thing that is easier to measure wins by default
Anyone who has built an agent pipeline knows this failure mode intimately. You pick a metric because it is the one you can compute. The system climbs it. Six months later you realize the metric was a proxy, the proxy drifted, and the actual behavior you wanted was never captured anywhere in your code.
That is the shape of the concern the medalists are raising, scaled up to an entire discipline. If the industry’s measurable objective is “solve the problem,” and the field’s unmeasurable objective is “grow the people and ideas that will pose the next problem,” then a sufficiently capable problem-solver does not just help. It quietly reprices the whole training pipeline for humans.
The arXiv paper on mathematics in the age of AI uses a useful phrase for this: impedance mismatches. Two systems that work fine independently, losing energy at the junction. That framing is more honest than “AI good” or “AI bad,” and it is one I keep returning to when I look at my own tooling.
What this actually changes in my builds
I am not going to pretend a statement from Fields medalists rewrites my sprint. But it did sharpen a few habits worth naming.
- Write down the objective you cannot measure. Put it in the README next to the metrics. If your bot is meant to help someone learn, say so explicitly, because nothing in your eval suite will ever say it for you.
- Treat answer-shaped output as a design choice, not a default. A solid math assistant can return a result, a hint, or a question. Most ship only the first because it is the easiest to score.
- Keep the human in the reasoning loop, not just the approval loop. Rubber-stamping a proof is not the same as understanding it. If your interface only offers accept or reject, you have already decided which one the user will pick.
- Log where users disengage. The moment someone stops reading and starts copying is the moment your tool switched from teaching to answering. That transition is measurable, and almost nobody instruments it.
Why bot builders should care more than most
Mathematics is the canary, not the whole story. It has unusually clean verification, a small and vocal community, and a culture that documents its values in public. That combination made this mismatch visible early. Other fields have the same structural problem with worse instrumentation and no signed statement.
Which means the patterns worked out here will propagate. Whatever conventions emerge for AI in mathematics — how credit gets assigned, how students stay in the loop, what a tool owes the person using it — will get copied into code assistants, research agents, and every writing tool built on the same stack. Including mine. Including yours.
The useful move is not to pick a side between the labs and the mathematicians. It is to notice that a specific, well-documented failure mode just got named by people with no commercial reason to name it, and to check whether it is already sitting in your own architecture.
Mine had two instances of it. I found them by looking.
🕒 Published: