By asahifile 
What this is about
This review covers the Economist’s leader “EurekAI!” (October 17th 2026, p.10) together with the Science & technology special report “AI scientists” (pp.69–74). It uses the moment when AI suddenly solved a batch of mathematical problems to ask, head-on, what it means to hand over science itself to machines. The author is unnamed.
Key points
The event began on October 6th 2026, when OpenAI released more than 700 academic papers at once, claiming progress or solutions on 372 problems that had baffled scholars for decades. The same model reportedly solved the Navier–Stokes existence and smoothness problem—a Millennium Prize problem concerning the motion of fluids—in 88 hours. It was only the second of the seven prizes to fall, and the first to a machine.
The reaction in mathematics has been complicated. The Association for Human Mathematics, a group of more than 900 scholars, warns that “mathematicians did not ask for this work to be done,” arguing that the value of mathematics lies not in the proof but in the toil. Yet the article also finds “an element of special pleading” in this backlash: no profession has a right to monopolise the production of knowledge, and a guild of scholars is no exception.
The article also notes AI’s achievements in the empirical sciences—better weather forecasts, millions of proposed new materials, help controlling nuclear-fusion experiments, and predicting the shape of nearly every known protein, work that won a Nobel prize. On this basis it voices a wariness of a future in which machines become incomprehensibly intelligent and humans are left merely writing the cheques.
What bothered me
I broadly agree with the leader’s argument that science must always keep a human in the loop, but I think the real crisis sits one step away from where the article places it. The true danger is not that AI takes scientists’ jobs. It is a structure in which humans accept results without understanding—a science that gives up on verification and comprehension and merely consumes answers.
To me, the backlash from mathematics mixes a legitimate point with a guild’s vested interest. The legitimate part is the claim that proofs without understanding do not advance human understanding: a student handed the answers learns nothing, and comprehension is born precisely of effort. But the phrase “we did not ask for this work to be done” carries the implication that only an existing community may produce knowledge, and that undercuts the public character of science. I accept the first point and reject the second outright.
What the article should have pressed harder is the question of who guarantees the verifiability of results. Mathematics, it says, can be verified mechanically because it is derived from axioms. But in empirical fields such as biology and materials science, the cost of confirming the reproducibility and validity of AI’s output is far higher. The moment unverifiable results begin circulating as “papers,” the foundation of trust in science erodes. My position is plain: the condition for accepting AI’s output is not whether a machine solved it, but whether humans can verify and understand it.
There is a deeper point that the leader gestures at but does not finish. The article recalls the Royal Society’s motto, Nullius in verba—”take nobody’s word for it”—and says the same spirit must now apply to machine-made science. I would go further. Scepticism is not a luxury that slows science down; it is the mechanism by which science stays science rather than becoming a black box we merely trust. If the price of speed is that no one can check the answer, then the speed is an illusion, because a result no one can verify is not knowledge. It is, at best, a claim we have chosen to believe. That is a fine basis for faith, but a poor one for science.
What it means for the world
This is not a Japanese problem alone; it is a global one. Wherever research evaluation is weighted toward publication and citation counts, AI-generated results that cannot be verified will find a way in as “numbers.” The more quantity is rewarded, the more the labour-intensive work of verification is neglected. That is a structural gap, not an efficiency gain.
The remedy is the same across borders: reward verifiability rather than volume. Put resources into peer review and reproducibility, not just novelty. The point is not to stop machines from solving problems, but to give the slow, unglamorous work of actually understanding what was solved its proper status and career reward. Science will remain trustworthy only as long as some human, somewhere, actually understands it.
Who should read this
Researchers, and especially anyone involved in designing how research is evaluated. It is also worth a pause for the general reader who has begun to accept AI’s output as automatically correct.
Quotations
“Mathematicians did not ask for this work to be done.”
“Science will always need a human in the loop.”
Source: The Economist, October 17th 2026, p.10
Source: The Economist, October 17th 2026 (Leader, p.10 / Special report pp.69–74)