What if the alignment problem we should be worried about has nothing to do with a machine wanting the wrong thing, and everything to do with a company wanting the wrong thing?
In 2026, a group of mathematicians, including 25 Fields medalists, put their names to a statement that deserves more attention from the agent-architecture community than it has received. Their claim is blunt: the goals of the AI companies and the goals of the mathematical community are severely misaligned. Not subtly out of step. Severely misaligned. And they framed it explicitly as one instance of a broader alignment issue affecting other scientific and creative professions.
I want to take that word seriously, because in my field it has a technical meaning, and the mathematicians are using it correctly.
Misalignment without malice
Most public discussion of AI alignment imagines a system that acquires an objective its designers did not intend. That framing lets the designers off the hook. The mathematicians’ complaint describes something different and, I would argue, more tractable to reason about: an optimization target that was chosen deliberately, specified precisely, and happens to be the wrong one for the domain it is being applied to.
Consider what an AI lab can actually measure in mathematics. Benchmark performance. Problems solved. Proofs verified. Competition scores. These are all legible, gradeable, and improvable. They make excellent reward signals. An agent architecture built around them will get very good at producing correct answers to well-posed questions.
Now consider what the mathematical community says it values. The statement points at nurturing students and nurturing ideas. Neither is a scoreable output. A graduate student who spends two years on a wrong approach and emerges with taste and judgment has produced no benchmark-eligible artifact. An idea that reframes a problem so it becomes obvious in hindsight does not register as a solved problem at all, because it dissolves the problem instead of answering it.
This is a specification failure, not a capability failure. The two objectives are not opposed in principle. They come apart because one is measurable and the other is not, and gradient descent goes where the gradient is.
What agents optimize for is what the field becomes
The architectural point that interests me most: the reward function does not stay inside the model. It propagates outward into the practice.
Once a field has a system that reliably closes problems, the incentive structure around that field reorganizes to feed it. Problems that are formalizable get attention because they are the ones the system can attack. Problems that require inventing a language before you can state them get less. Grant committees, hiring committees, and journals adjust to what is now cheap to produce. Nobody decides this. It emerges from a change in relative cost.
The same dynamic threatens the human pipeline. If the fastest path to a solved problem routes around the slow apprenticeship where mathematical judgment gets built, the apprenticeship becomes economically indefensible long before anyone concludes it was unnecessary. You do not need to believe AI will replace mathematicians to worry about this. You only need to believe that the training of mathematicians is expensive, slow, and produces value on a timescale that quarterly planning cannot see.
Why this generalizes
The mathematicians were right to say this extends past their field. Mathematics is simply the cleanest test case, because correctness is checkable. That checkability makes it the ideal proving ground for agent systems, and it also makes it the place where the gap between the measurable objective and the actual objective is most visible.
Everywhere else, the same structure holds with murkier feedback. Software engineering has tickets closed. Science has papers published. Design has assets shipped. In each case there is a legible proxy, an illegible thing the proxy stands in for, and now a system capable of saturating the proxy.
Agent architects should read this as a design constraint rather than a political complaint. If your system’s objective is defined by what the domain can currently score, you are not building a tool for the domain. You are building a tool that will reshape the domain into something scorable. Those are different products with different consequences, and only one of them was on the roadmap.
The uncomfortable version
I do not think the labs are acting in bad faith. I think they are doing exactly what optimization pressure rewards, which is precisely the concern. When 25 Fields medalists describe a misalignment as severe, they are not saying the technology does not work. They are saying it works, at something other than what they were trying to do.
That distinction is the whole of alignment research compressed into one sentence. The mathematicians have handed the field a live case study, with named stakeholders and stated values, in a domain where the objective mismatch can be examined rather than speculated about. Treating it as a labor dispute rather than a technical warning would waste it.
🕒 Published: