A joint statement from 25 Fields Medalists on AI’s impact on mathematics — published September 11 under the title “A Severe Misalignment of AI in Mathematics” — has taken many people by surprise.
If AI can solve problems faster, even push the research frontier forward, isn’t that progress for mathematics? Why would the people standing at the very top of the field show such strong alarm?
The Worry Isn’t Jobs. It’s the Soil.
But the mathematicians’ collective concern touches the same problem I wrote about on March 27, 2026, in a piece on how AI and related technologies are hollowing out the skills pyramid layer by layer, using the transformation of hotel concierge work as the case. Frankly, the mathematicians’ worry is not unfounded.
First, we need to draw a precise line between the professional interests of mathematicians and the development of mathematics as a discipline. One person’s competitive edge eroding does not mean the discipline is going backward. But if what’s being affected is the discipline’s capacity to produce its next generation of researchers, that is a different matter.
In the concierge piece, I argued that a senior concierge can use AI to become more efficient, because he already has the judgment, the network, and the accumulated trust. But the everyday tasks a newcomer once used to build those very abilities are being taken over by technology. The senior expert in front of you performs better than ever, while the newcomer’s path to becoming a future expert narrows.
Mathematics may be facing a similar situation. Young researchers solve problem after problem, attempt proofs, find mistakes. These processes produce results, but they also cultivate research intuition. The ability to ask good questions is often formed precisely in this concrete work.
From an organization design perspective, a large share of the cost of developing people has historically been hidden inside the process of getting work done. AI finishing the work quickly does not mean the development of talent was finished along with it.
Mathematicians Are Human Too
Second, there is a more down-to-earth problem: mathematicians are human, and they care about incentives too.
Curiosity and a sense of mission are no substitute for research opportunities, funding, and recognition of individual contribution. For a person to commit years to research, they need to believe the commitment is worth it, and they need the conditions to keep committing. We cannot assume that because mathematics is a noble pursuit, the people who pursue it need no return.
If AI gets large volumes of research tasks done quickly, while evaluation and resource allocation continue to reward mainly final results, then a researcher’s intermediate contributions may find it harder and harder to buy continued support for their work. Young people’s room to survive gets squeezed. Even those who still love mathematics may choose to leave, or never enter the field at all.
The talent pyramid therefore faces two pressures at once: the tasks needed for growth are shrinking, and the incentives that sustain long-term commitment are weakening.
This is a textbook case of draining the pond to catch the fish. Existing knowledge and mature researchers have their productivity amplified by AI, producing a leap in output. At the same time, the growth path and the living space of the next generation of researchers are contracting. Results keep piling up in the present, while the talent base that would support future results may be quietly coming loose.
What Does This Statement Actually Prove?
This concern carries an implicit judgment about the limits of AI’s capability: continuing to pose important questions and open new research directions still requires human leadership. As long as the human role remains indispensable, talent development and incentive mechanisms bear on the future of the discipline.
That said, a statement like this can produce another effect as it spreads: the top mathematicians’ vigilance gets read by the public as a certification of near-limitless AI capability.
In my view, the statement does constitute an endorsement of capability. But there is still a great distance between “AI is enough to disrupt the ecosystem of mathematical research” and “AI can already lead the mathematical frontier on its own.” The leap resembles the one in the “one-person company” narrative that swept China a few months ago: AI lets one person do more work, and that gets extrapolated into one person sustaining every function of an entire company.
The AI narrative today slides too easily toward the poles: talk about capability, and it sounds as if humans are about to exit entirely; talk about limits, and it sounds as if the professional value of people will be untouched.
My own judgment is closer to a middle state that may last a long time: AI can substantially raise efficiency, improve the quality of deliverables, and replace many kinds of work, but continuing to lead in new directions still requires people. And it is precisely in this middle state that the talent problem deserves the most vigilance: AI is already enough to squeeze people’s living space, yet it cannot on its own carry the development process those people sustain.
Everyone Is Being Rational. The System Pays.
For companies and HR, once AI is embraced, what needs redesigning is not just the workflow but the growth path and the reward mechanism. Newcomers must still have the chance to form capability, and long-term commitment must still be worth sustaining.
But this accountability cannot be left to companies and individuals alone. An employer adopting AI to cut cost is a rational response to the competition in front of it; a researcher who finds their investment harder and harder to recoup, and who invests less or exits, is equally understandable. No one can be asked to absorb today’s losses alone for the sake of what society will need ten years from now.
From an organization design perspective, the problem is this: every participant makes the choice that is reasonable for themselves, and the system as a whole may gradually lose the ability to develop the next generation. Short-term incentives do not stop working just because long-term goals matter more.
This is also why government’s coordinating role in the direction and pace of AI development is hard to replace. The choices of companies and individuals need to be reconciled on a longer time scale and across a wider social scope. One key accountability government must take on is to turn society’s long-term needs into choices that companies and individuals have reason to make today.
We cannot narrow the road to expertise, weaken the reward for walking it, and still expect experts to keep appearing.