Keep paying the humans

The debate turns on a single parameter, the elasticity of substitution between human-made and machine-made research. If the two are close substitutes, then once the machine is cheaper and better, spending moves to it and the human share collapses. If they are complements, cheap machine research raises the value of the human input. Research is the second case, for three reasons.

Consider first the distinction between a task and a job. The motion evaluates a bounded artifact, the questions and results a panel can score. Funding buys a bundle: choosing which problem to attack, running and interpreting an experiment, coordinating a group, communicating the result, and standing behind it. Automating one component of that bundle does not dissolve the rest. The assumption that it does is the lump-of-labour fallacy, the premise that there is a fixed quantity of research so that any part the machine takes is subtracted from humans. The historical record runs the other way. For most of the last century the share of national income paid to labour held within a narrow band, roughly 60 to 65 percent, one of Kaldor's stylised facts, which Keynes called one of the best-established regularities in economics, across a century of automation panics. The share has drifted down a few points since the 1980s, which is real and worth watching, but no wave of mechanisation produced the collapse the substitution story predicts. When a task inside a job is automated, demand for the complementary human tasks around it has usually risen.

Second, returns to the time horizon. Here, one piece of direct evidence comes from this community. In RE-Bench, METR compares sixty-one expert ML researchers with frontier agents on seven open-ended research-engineering problems. Given a two-hour budget, the best agent scores about four times the human experts; given thirty-two hours, the humans move ahead and score roughly double the best agent. The agents are fast, and on short, well-scoped problems they win. What they do not yet supply is the human return to long-horizon, open-ended effort, which is the regime research funding exists to support.

Third, the serial structure of the pipeline, which is Amdahl's law in economic form. Chad Jones formalises it as a weak-links, or Baumol, argument: when tasks are complements, total output is governed by the tasks that remain un-automated, and their relative price rises as everything around them gets cheap. Automating an entire category of work, even perfectly, raises aggregate output by a bounded amount, because the un-automated remainder becomes the binding constraint. The remainder is the cluster of research tasks with no verifiable target: setting the agenda, judging whether a result is real, communicating it, taking responsibility for it, and aligning it to human values, the trust and oversight layer of the whole enterprise. Those are the weak links, and their price rises as execution gets cheap.

Underneath all three sits the finding by Bloom, Jones, Van Reenen and Webb that ideas are getting harder to find. Sustaining Moore's law now takes roughly eighteen times the researchers it required in the early 1970s, with research productivity per person falling around seven percent a year. A world in which discovery is getting harder wants more capable searchers at the frontier, human or machine. Substitution requires research to be a fixed pie handed to a cheaper producer. It behaves like an expanding frontier that rewards adding capacity.

The mechanism that ties this together runs opposite to the substitution intuition. Jevons' paradox: when something becomes cheaper to use, total demand for it tends to rise rather than fall. Cheaper research does not shrink the appetite for research; it expands the number of questions worth asking and experiments worth running, each of which still needs a person to choose it, judge it, and stand behind it. Arvind Narayanan's ICML keynote makes this case directly, tracing it through the same automation episodes, ATMs and radiologists, where employment grew rather than collapsed once the task was automated. The human-held share of a project falls while the number of projects rises, and there is no reason to expect the product to shrink toward a skeleton crew.

Grant the machine superhuman output. Someone still has to decide whether to believe a result, sign for it, and answer for it when it is wrong, and that capacity does not stand on its own. The supply of people who can verify research is downstream of the supply of people who do it: a verifier who can catch a subtle error in a superhuman system is someone who has done the underlying work long enough to know what a real result looks like. The oversight layer cannot be a small cadre of watchers detached from practice. It has to be drawn from, and kept inside, a working research pipeline.

Judgment is a practiced skill, and it decays without use. In a 2025 Lancet study, nineteen experienced endoscopists lost detection skill after three months of AI assistance: when the tool was removed, their unaided adenoma detection rate fell from 28.4 percent to 22.4 percent. Three months eroded expert perception in specialists with thousands of procedures behind them. A research workforce that stops running its own experiments for a decade does not only shrink the pool of referees; it degrades the capacity of the ones who remain to referee at all.

Demand for oversight rises from both ends at once. Automated researchers generate far more experiments, code, and decisions per unit time than any fixed set of humans can review, which raises the volume to be checked, while cutting the workforce shrinks the pool qualified to check it. The review layer of our own field already shows the strain. A 2024 study by Liang and colleagues estimates that between 6.5 and 17 percent of the text in peer reviews at recent AI conferences was substantially produced by language models. That is a demand curve for practiced human referees.

Alignment sits inside this layer, and it is the part I find hardest to imagine automating. Deciding whether to trust a high-stakes claim that has no clean ground truth is a judgment about values and consequences, and it cannot be delegated to the system under evaluation without circularity, because the optimiser cannot certify itself. My own research is on privacy and contextual integrity, where the ground truth is a contested, outcome-based judgment about which norms apply and whose interests count, and where no benchmark settles the question. Those judgments are made well only when the people affected are represented among the people making them. Aligning a system to human values is not a step you run once and store; it is a standing negotiation that requires humans to do the negotiating, and it does not have a fixed answer to hand to a model.

The pipeline is itself the oversight infrastructure, and we already treat training this way in other domains. In 1914 a Sperry autopilot flew hands-off, and by 1947 a US Air Force aircraft crossed the Atlantic, takeoff to landing, entirely under automatic control; large aircraft still carry two trained pilots today, because regulation and hard experience treat the human as the accountable agent when the automation reaches its edge (history of flight automation). We still teach children arithmetic in a world full of calculators for the same reason: the training exists to build the judgment that lets a person catch a wrong answer, not to compete with the tool on speed. Teaching research, and funding the people who do it, is how a field keeps producing verifiers who can still recognise when the machine is wrong; remove the doers and within a generation there is no one left qualified to referee. The deskilling result above is the near-term edge of exactly this problem.

Aggregating many judgments improves on a single one only when the errors are independent, a condition that goes back to Condorcet's jury theorem. Copies of one model do not meet it. Across more than 350 models, Kim and colleagues find that when two of them answer a question wrong, they choose the same wrong answer about 60 percent of the time, and the correlation is highest among the largest and most accurate systems, even across different providers and architectures. Adding a tenth model to a panel of nine does not add a tenth independent judgment; it restates a shared prior. Ensembling buys little when the members converge underneath.

Human researchers fail in less correlated ways, because their formation is not shared: different advisors, languages, subfields, and failures produce different blind spots. There is evidence that this heterogeneity is not just present but productive. Analysing 1.2 million US doctoral theses, Hofstra and colleagues find that researchers from underrepresented backgrounds introduce novel conceptual combinations at higher rates; across nine million papers, AlShebli and colleagues find that ethnically diverse teams produce higher-impact work. The diversity of a large research community works as error-correction: it covers regions of the hypothesis space that a correlated system leaves dark. The matching result on the machine side is that an algorithmic monoculture can lower the quality of collective decisions even when the shared model is individually more accurate (Kleinberg and Raghavan, 2021). Concentrating the world's research judgment into a few correlated systems is that monoculture at the largest possible scale.

There is a further concern about concentration, and it is about power rather than accounting. If research is fully automated, the capacity to set the agenda, to decide which questions get asked and which results are trusted, moves from a distributed public system of universities and agencies to the small number of firms that own the frontier models and the compute they run on. That capacity is currently spread across many institutions and countries, which is part of why science can correct itself and serve interests beyond any single owner. A publicly funded research base keeps the questions plural and the answers accountable to more than whoever controls the datacenter. This is the sense in which the motion is a normative claim and not only an empirical one, and it is why the concentration point is central rather than decorative.

Grant that execution is solved. Two things stay scarce on the supply side.

The first is taste: choosing what is worth working on before the field agrees that it was. This resists the current training paradigm, because the reward signals available, citations, acceptance, downstream use, are all lagging measures of consensus. Optimising against them produces work aimed at what the field already values, which is the opposite of the contrarian bet that defines an important direction. The recent reasoning gains are consistent with this reading: reinforcement learning on verifiable rewards mainly sharpens a model's ability to reach answers its base model could already sometimes reach, and whether it expands the frontier of new directions is contested. For now, the taste for the unproven direction is supplied from outside the reward loop.

The second is exploration. A large, uncoordinated population of researchers with idiosyncratic paths is a wide portfolio of low-probability bets, most of which return nothing and a few of which are penicillin. Optimisation prunes the low-probability branches, because in expectation they look like waste. Some of that population works on questions with no near-term payoff, in astrophysics or pure mathematics, because they want to understand something, and that motive keeps the search broad in a way a returns-maximising allocation would not. The point of funding people to follow curiosity is that you cannot know in advance which unpromising line becomes the important one.

There is also a demand-side scarcity, and it is measurable. In controlled experiments (Horton and colleagues, 2023), people assign a lower value, including a lower willingness to pay, to an identical artifact once it is labelled machine-made, even when they cannot tell the difference. Around research this is beginning to appear as a premium on human-authored work. It is evidence that human-made research is a distinct good with its own demand, which is the substitutability claim from the first argument arriving from the consumer side. The uncertainty-of-outcome hypothesis from sports economics, that people pay to watch a contest precisely because a human might fail, has survived seven decades of testing. A helicopter reached the summit of Everest two decades ago, and the number of people paying to climb it, and the price of the permit, have only risen since. The result was never the whole of the value.

The final argument holds even if the first four fail. Part of the case for a large human research base is insurance, and insurance is priced by two quantities: how correlated the primary system's failures are, and how long the backup takes to rebuild once it has been cut. Both point the same way, and both are worth spelling out.

Start with correlation, because the AI research stack is correlated at every layer. The models are trained on substantially the same corpus, so a systematic gap or bias in that corpus is a systematic gap everywhere at once, with no independently-trained population to catch it. Training models recursively on model-generated output degrades them across generations, a failure mode documented as model collapse, and it is arrested only by a continuing supply of fresh human data. The physical layer is a single point of failure of a different kind. The large majority of leading-edge chips are fabricated on one island, and much of the world's oil still moves through one strait, so a disruption in the Taiwan Strait or the Strait of Hormuz would not be a local shock to be routed around; it would propagate through a network that everyone now depends on simultaneously. The pattern is local robustness with global fragility, the same structure that makes a genetic monoculture efficient right up until one pathogen finds the shared weakness.

Now reversibility, where the asymmetry bites. A human research base cannot be switched off and on. The path from a first-year PhD student to an independent investigator runs the better part of a decade, and the tacit knowledge that makes a working scientist, the parts learned by apprenticeship and never written down, disperses far faster than it can be regrown. A version of this is happening now: when a large share of a national research budget is frozen or cut in a single year, thousands of grants lapse and a substantial fraction of the affected scientists begin planning to leave, and capacity that took a generation to build is damaged in a single budget cycle. You cannot repurchase it at short notice at any price, which is exactly why an uncorrelated, slow-to-rebuild backup has to be held before it is needed rather than reconstituted afterward.

The premium is unusually cheap, and it returns a profit even if the insured event never occurs, because public research has a large measured social return on ordinary grounds. Fieldhouse and Mertens estimate the social return to non-defense public R&D at roughly 140 to 210 percent, and attribute about a fifth of postwar US business-sector productivity growth to it, concluding that it is substantially underfunded. The amounts are also small relative to what they are weighed against. The entire annual NSF budget is on the order of nine billion dollars, which pales beside the roughly half a trillion dollars of AI capital expenditure now spent in a single year. The choice is not human salaries against compute; compute already outspends the whole public-science enterprise by an order of magnitude. Maintaining the human base is, in budget terms, a small hedge with a positive expected return.