Twenty-five Fields medalists signed a statement saying the goals of AI companies and the goals of the mathematical community are “severely misaligned.” Twenty-five. For context, the Fields Medal is awarded to at most four people every four years. That signature list is a meaningful slice of every living winner, and they all agreed to put their names on the word “severely.”
I review AI tools for a living. I have watched a lot of people complain about a lot of models. This is different. This is not a group of practitioners griping that the output quality is inconsistent or that the pricing page is confusing. This is the most decorated group in a field telling the industry that the industry has the wrong objective function.
What the complaint actually is
The statement, from mathandai.org, frames the problem as a mismatch of goals and values, not a mismatch of capability. That distinction matters and most coverage will flatten it. The mathematicians are not saying the models are bad at math. They are saying that even if the models get very good at math, the thing being optimized is not the thing the field cares about.
The medalists named their core values plainly: nurturing students and nurturing ideas. Those are process values. They describe how a field reproduces itself and how understanding gets built and transmitted. Neither one shows up on a leaderboard. You cannot benchmark “this graduate student developed taste” or “this line of inquiry turned out to matter fifteen years later.”
The signatories also positioned this as part of a broader alignment problem hitting other scientific and creative professions. That framing is the smart move. Mathematics is a useful test case precisely because it has clean verification. Answers are checkable. If the incentive mismatch shows up here, in the field where “did the model get it right” is least ambiguous, it is going to show up everywhere with worse instrumentation.
Why the tool reviewer in me cares
Here is the pattern I keep seeing in AI tooling across every category. A vendor picks a metric that is easy to measure and easy to demo. Then the product gets optimized against that metric. Then the metric becomes the definition of the work. Code assistants optimized for lines accepted. Writing tools optimized for output volume. Support agents optimized for tickets closed.
The math case is the same shape with much higher stakes, because the downstream product is not a feature. It is the next generation of researchers. A student who learns to prompt their way past a hard problem does not develop the ability to sit with a hard problem. That capacity is the entire pipeline. You cannot buy it back later with more compute.
What strikes me about the phrasing “severely misaligned” is who chose it. Alignment is the AI industry’s own vocabulary. The medalists picked up the term and pointed it back at the companies. That is not accidental. It is a group of people who understand precision using the industry’s terminology to say: your alignment problem is not hypothetical and not in the future, it is operating right now, and we are the ones experiencing it.
The uncomfortable middle
I want to be fair to the other side, because the discourse around this is going to get stupid fast. Working mathematicians have written about AI’s usefulness in their practice, and the arXiv paper on mathematics in the age of AI treats these as “impedance mismatches” — friction points to be diagnosed and responded to, not walls. That is the more useful framing than either “AI will finish mathematics” or “AI will destroy it.”
The people signing this statement are not technophobes. They are specialists in the most abstract reasoning humans do, and they are describing a specific structural problem: the entities building these systems answer to product timelines and capability demos, while the field they are building into answers to a multi-decade apprenticeship model. Those two clocks do not sync. Nobody has to be acting in bad faith for the outcome to be bad.
What I’d actually watch for
Watch whether any lab responds by changing an evaluation, not by publishing a blog post. A statement of shared values costs nothing. Changing what you measure costs something, because it means admitting your old numbers described the wrong thing.
And watch whether the mathematical community can articulate what alignment would look like concretely. “Nurture students and ideas” is a real value but not yet a specification. Somebody has to translate it into things an evaluation suite can see, or the industry will keep optimizing what it can see and calling that progress.
Twenty-five people with the highest credential in their field just told the AI industry it is solving the wrong problem. Whether anyone in a position to change an objective function was listening is the actual open question here.
🕒 Published: