Mathematicians are not worried that AI will replace them. That’s the part the coverage keeps getting wrong. When 25 Fields medalists put their names to a statement in 2026 saying the goals of AI companies and the goals of the mathematical community are severely misaligned, they weren’t filing a job-loss complaint. They were pointing at something considerably more awkward for the tools I spend my days testing: the products are optimizing for an outcome the field doesn’t actually value most.
I review AI toolkits. My job is to figure out what works, what doesn’t, and where the marketing outruns the software. This story lands squarely in that territory, because “misalignment” here isn’t a safety abstraction. It’s a product critique from the most credentialed user group any AI lab could hope for.
What the medalists actually said
The statement is short on hedging. The goals of the AI companies and the goals of the mathematical community are severely misaligned, and the signatories framed this as part of broader alignment issues affecting other scientific and creative professions. Not a math problem. A pattern.
The counterweight they offered is the interesting bit. The mathematical community’s stated core values are nurturing students and nurturing ideas. Read that next to how math AI gets sold and the mismatch is immediate. Nobody ships a model with a slide about nurturing graduate students. You ship benchmark scores. You ship problems solved, competition results, theorems closed.
Why this is a tooling problem, not a philosophy problem
Every AI product encodes a theory of what its user wants. Most math-focused systems encode this one: a mathematician is a device that converts open problems into closed problems, and faster conversion is strictly better.
That theory is testable, and the people best positioned to test it just returned a poor review. If your value system centers on developing people and developing ideas, then a tool that hands you finished answers is solving for the wrong variable. The struggle isn’t friction to be engineered away. It’s where the training happens.
I see the same failure mode in developer tooling constantly, just with lower stakes. An assistant that writes the whole function is genuinely faster and genuinely produces a junior engineer who never learned to debug. The difference is that software teams argue about this in Slack. In mathematics, the argument now has 25 Fields medals behind it and a public statement.
The evaluation gap
Here’s what makes this hard to fix with a patch. The things the mathematical community says it cares about are the things nobody has figured out how to score.
- Did a student’s understanding deepen? No benchmark.
- Did an idea get more generative, opening paths rather than closing one? No benchmark.
- Did the work leave the field healthier for the next generation? Definitely no benchmark.
- Did the model produce a correct proof of a stated theorem? Easy benchmark, and so that’s what gets optimized.
This is Goodhart’s law wearing a lab coat. The measurable proxy becomes the target, the target becomes the roadmap, and the roadmap quietly redefines what the tool is for. The medalists appear to be objecting to that redefinition before it hardens.
What I’d want to see from the labs
Skepticism aside, this is actionable. If a math AI product wants to align with its most serious users, the design brief writes itself from the statement’s own values.
Build tools that explain rather than answer. Build tools that show the space of approaches instead of the single shortest path. Build tools a supervisor can put in front of a first-year graduate student without worrying that the student stops thinking. That’s a harder product than an answer machine, and it demos worse, which is precisely why it doesn’t exist yet.
The alternative is a generation of systems that are excellent at the version of mathematics that fits in an eval use and indifferent to the rest of it.
My honest read
I’m not calling this a crisis and I’m not calling the tools useless. I’ve used enough of them to know the capability is real and improving. But capability and fit are separate axes, and the reviewers who matter most just flagged a fit problem in public.
The medalists were explicit that this extends past mathematics into other scientific and creative professions. If you build or buy AI tools for any field with an apprenticeship structure, that’s your warning label. Ask what your tooling optimizes for. Then ask whether anyone bothered to check that against what your field says it values. In mathematics, someone finally did, and the answer was no.
🕒 Published: