The debate about AI and science usually gets framed as a question about jobs: will AI replace researchers? I think this framing misses the point. The more interesting question is what happens to the institutions of science—PhD programs, universities, funding agencies, the paper system—when the basic mode of scientific production changes underneath them. This essay traces one causal chain from a simple starting observation to a fairly radical conclusion: research is moving from a system whose basic unit of production is the trained human being toward one whose basic unit is a small number of humans plus AI systems, and when that happens, the entire institutional apparatus we built around scarce human researchers will need to be rethought.
IResearch is, at bottom, a luxury
Start from an uncomfortable observation: research is, in an important sense, a luxury activity.
This is not to say research lacks value. It is to say that most basic research carries no direct pressure to deliver revenue, ship a product, or serve an identified customer. Researchers largely study problems they consider important, not problems that a market has clearly articulated and is willing to pay to have solved right now.
For research to exist at all as a sustained activity, then, one fundamental question must be answered:
Who pays for people to explore, over long horizons, with no guarantee of economic return?
History has offered different answers. Some early researchers were simply rich. Cavendish could devote his life to experiments because he never had to worry about a living. Many mathematicians and natural philosophers depended instead on aristocratic or royal patronage. The forms varied, but the underlying logic was the same: society must first support a class of people—give them money and leisure—before uncertain, long-term, possibly unprofitable inquiry can happen at all.
IIThe modern research system is that patronage, institutionalized
The modern era gradually replaced private and aristocratic patronage with universities, governments, and research foundations. Universities give researchers positions, salaries, laboratories, and students; governments supply grant funding; researchers, in return, teach and train the next generation.
Out of this emerged a stable loop:
Notice the dual role that doctoral education plays in this loop. On one side, PhD students are being trained: they spend years learning how to do research, acquiring knowledge, skills, taste, and professional credentials. On the other side, PhD students are simultaneously the research system's most important labor force. A very large share of the world's papers, experiments, code, data analysis, and theoretical derivation is actually produced by doctoral students and junior researchers.
The whole arrangement rests on a premise so obvious that it is rarely stated out loud:
Research capability lives primarily in people, and producing a person capable of independent research is very hard.
Because "someone who can do research" is a scarce resource, it makes sense to train PhD students, hire researchers, and build large research organizations around them. Scarcity is what justifies the pyramid.
IIIWhat AI actually changes is this premise
The genuinely important thing about AI in science is not "AI can help researchers write papers." The deeper question is this:
If paper writing, literature review, code implementation, data analysis, mathematical derivation, hypothesis generation—and, increasingly, the substance of research itself—can be done by AI, must research productivity remain attached primarily to human beings?
The moment the honest answer becomes "not necessarily," the institutional premise underlying the entire modern research system starts to loosen.
Which means the effect of AI on science is probably not simple substitution of machines for workers. It is something more structural:
Research is shifting from a system whose basic unit of production is the individual human researcher to a system whose basic unit is a small number of humans plus AI systems.
Everything that follows in this essay flows from that shift.
IVFirst consequence: the meaning of the PhD changes
If AI can take over more and more of the work traditionally done by doctoral students, both sides of the advisor–student relationship start asking new questions.
From the advisor's side, training a PhD student is expensive. It takes a student years to come to grips with a field, and the process demands continuous communication, correction, and mentorship—all subject to the frictions of human collaboration: mismatched interests, uneven abilities, differing working styles, and plain communication cost. If a researcher can instead spend unlimited hours in substantive discussion with an AI, and delegate large volumes of concrete research work to it, an uncomfortable question arises naturally:
Why do I need to train many people in order to get research done?
From the student's side, a mirror-image question appears:
If most of the capabilities that the "PhD → research → papers" pathway is supposed to certify can now be supplied by AI at scale, why should I spend five or more years publishing papers to prove I possess them?
None of this means the PhD simply disappears. But the reason for the PhD's existence will need to be redefined. Historically, doctoral education was largely in the business of producing research producers. In the future, its central task may be producing people who can do the things AI systems cannot be trusted to do alone: pose genuinely valuable questions, judge results, orchestrate AI systems, design research pipelines, and take responsibility for scientific claims.
VSecond consequence: the logic of research funding changes
Look at the budget of a typical funded project today and you will find that a large fraction of it is labor. A professor wins a grant and then hires PhD students, postdocs, and research assistants, who carry out most of the actual work.
But if AI becomes a primary engine of research production, funders will ask a very practical question:
Why should I keep paying for dozens of salaried researchers, when I could fund a small number of exceptional people and equip them with AI, compute, data, and automated experimental infrastructure to do the same work—or more?
The structure of research funding then shifts. The old equation was, roughly:
The emerging equation looks more like:
Governments may well remain the dominant funders of science, but the way money is organized changes, and with it the shape of research organizations. The large professor-and-students pyramid that defines today's labs could flatten into something quite different: a small number of humans who pose questions, set direction, and marshal resources, with the bulk of concrete research work executed by AI.
VIThird consequence: a genuine "paper market" may emerge
On top of this, a more radical inference becomes plausible.
If AI drives the marginal cost of producing papers and research results down far enough, research may develop a real market mechanism—call it a paper market.
Under the current system, participating in research effectively requires joining the academy: enter a university, become a researcher, join a lab, invest years in training. In the future, another path may open.
Imagine someone with money who is not a professional researcher. They care about a particular problem. They want a conjecture settled, a theory built, perhaps a result associated with their own name. They could simply fund a researcher—or an AI research system—to attack that problem directly.
Concretely: one party poses a mathematical problem; another supplies funding; a researcher designs the attack and judges the route; AI carries out the bulk of the search, proof, and experimentation; and the final result is credited to the participants under some new convention for allocating contribution.
What this describes is a market that has never clearly existed before: a market in which problems, capital, research skill, AI compute, and reputation are traded and matched against one another.
VIIThis forces a sharper answer to "why does research exist?"
Here we reach the deepest layer of the argument.
Today's research system, for all its flaws, grants researchers a remarkable privilege: they may study problems they themselves judge important, without directly answering the questions who needs this research? and who is willing to pay for it? Universities, governments, and foundations answer those questions on the researcher's behalf.
But if the cost of research production falls, and more and more actors can fund research directly, science may differentiate into two much more explicit forms.
One is research as market activity: someone has a concrete problem and is willing to pay to have it solved.
The other is research as explicit altruism, or public-goods production: a government, a foundation, or an individual judges that a question is worth exploring for society or for human knowledge even without direct economic return, and deliberately shoulders the cost.
In other words, the science of the future may be forced to answer, far more explicitly than today:
Who wants this research? Why is it worth doing? Who is willing to bear its cost?
VIIIWhat we are really discussing is not "AI replacing researchers"
The core claim, then, is this:
What AI brings is not straightforward job displacement, but a change in the relations of production of science.
The modern research system of the past two or three centuries was built on one basic fact: research capability is scarce, and it lives primarily in people who have undergone years of training. On that fact, society constructed the doctorate, the university laboratory, the grant system, the paper system, and the academic career. If AI makes research capability something that can be supplied at scale and at low cost, the fact itself changes—and the question stops being "can AI write a paper?" and becomes:
When doing research no longer requires large numbers of long-trained humans, how much of the institutional apparatus we built around humans is still necessary?
Why does the PhD exist? Why should universities organize research in their current form? Why is funding allocated the way it is? Why are papers authored the way they are? Who owns a research result? And how should credit be divided among the person who funds a problem, the person who poses it, the person who runs the AI, the person who verifies the output, and whatever genuine intellectual contribution lies in between?
The argument, as one causal chain
- Research inherently needs to be supported.
- Historically: wealthy and aristocratic patronage; later, universities and government funding.
- The modern system organizes production through professors, PhD students, and papers.
- The system presumes research capability comes from scarce human researchers.
- AI begins to decouple research capability from individual humans and scale it.
- PhD students become less necessary as the primary research labor force.
- The way advisors organize research changes.
- Funding shifts from "supporting people" to "a few people + AI + compute."
- Production costs fall; private capital can participate in research directly.
- A paper market — a research market — may form.
- Research enters a new paradigm: more market-driven, more explicitly public-goods-oriented, open to many more participants.
None of this will happen overnight, and much of it may happen unevenly across fields. But the direction of pressure seems clear. The institutions of science were designed for a world in which the scarce resource was the trained human mind. We are entering a world in which that is no longer quite true—and institutions, sooner or later, follow their premises.
How this was writtenThis essay was written by the author in collaboration with AI. The argument and the views in it are the author's own; AI assisted with drafting, editing, and preparing the English and Chinese versions.