What Becomes Valuable When Intelligence Is Cheap?

For most of industrial history, the expensive input in knowledge work was the thinking. Drafting a contract, diagnosing an illness, debugging a program, or building a financial model required a scarce person who had spent years acquiring the skill. Cheap intelligence changes which input is scarce. The interesting question is not whether machines can think, but what a market and a society do when a formerly expensive cognitive step becomes a commodity — and which human capacities grow more valuable precisely because the machine does not supply them.

The honest answer is narrower than either the optimists or the pessimists usually offer. When analysis becomes abundant, the scarce skills are not mysterious. They are the ones that decide what analysis is for: choosing a problem worth solving, setting a standard of quality, verifying a result against reality, accepting responsibility for a decision, and sustaining the relationships through which any of it becomes legitimate. These are not mystical endowments. They are teachable, and the evidence about how they behave under abundant AI is already available — in randomized studies of writers, consultants, support agents, programmers, and patent lawyers.

This article sets out what the measured science actually shows about where AI helps, where it quietly subtracts, and why the qualities that decide the outcome are usually the ones a model cannot verify for you.

The economics of a collapsed input price

A useful starting point is comparative advantage, but stated carefully. If one input — call it reasoning, synthesis, or analysis — becomes very cheap, it does not stop being used. It stops being the thing that distinguishes good work from bad. The margin of competition moves to whatever remains scarce. That is a general prediction, and the question is empirical: which complements actually stay scarce?

Three candidates stand out, and each has been measured in real workplaces.

The first is judgment about which task is worth doing. The second is the ability to evaluate whether an impressive-looking output is actually right. The third is the willingness to own consequences — to sign, ship, prescribe, publish, or refuse. A model can generate all three convincingly in prose. None of the three is settled by the model’s own confidence.

What the productivity experiments actually measured

The cleanest early evidence came from Noy and Zhang, who ran a randomized experiment with 453 college-educated professionals on realistic writing tasks (Noy & Zhang, 2023). Access to a generative model cut time on task by about 40 percent and raised rated quality by roughly 18 percent. Two details matter more than the headline. The gains were largest for the participants who started with lower ability, narrowing the spread between strong and weak writers. And the time savings came less from drafting good text from scratch than from removing the friction of getting started and roughing out structure — the model substituted for effort, not for the writer’s decision about what the document needed to accomplish.

Customer support produced a similar story with a sharper texture. Brynjolfsson, Li, and Raymond randomized AI assistance across 5,172 agents at a single firm and found a roughly 15 percent increase in issues resolved per hour (Brynjolfsson, Li & Raymond, 2025). The distribution is the interesting part: novice and lower-skilled agents gained the most — on the order of 30 percent — while the most experienced agents saw little improvement and, in some measures, a slight decline in quality. The authors also found evidence of durable learning, because resolution rates stayed elevated in the days after a system outage that had forced agents back onto the tool. Assistance appeared to accelerate the accumulation of tacit skill rather than merely paper over it.

Consulting showed that abundance has a boundary. Dell’Acqua and colleagues ran a study with 758 BCG consultants on a set of tasks carefully chosen to fall either inside or outside the frontier of what the model could reliably do (Dell’Acqua et al., 2026). Inside the frontier, consultants using AI completed about 12 percent more tasks, roughly 25 percent faster, at more than 40 percent higher quality. Outside it — on tasks the researchers had deliberately designed to look similar but contain a trap — AI-assisted consultants were 19 percentage points less likely to reach the correct answer than those working unaided. The authors coined the “jagged frontier” to describe a capability surface that is not smooth: adjacent tasks can sit on opposite sides of it, and a user cannot tell which side they are on simply by how fluent the output feels.

That last finding is the pivot of the whole subject. Abundance of capability is real, but it is not uniform, and the cost of mistaking the boundary for the interior is high.

Where verification and taste become the scarce inputs

If the frontier is jagged and self-report cannot locate it, then the scarce skill is diagnosis: knowing, before you trust an answer, how you would know if it were wrong. That is the skill the consulting study rewarded and the one the customer-support study partially taught.

Anthropic’s usage data gives a complementary view from the outside. Mapping millions of real conversations onto the United States Department of Labor’s O*NET task taxonomy, the research found that AI use is concentrated in software and writing, that roughly a third of occupations show use on at least a quarter of their tasks, and that augmentation (learning, iterating, validating) is about as common as straightforward automation (Handa et al., 2025). The tasks where AI presence is thin are revealing: negotiation, physical manipulation, and managing people. Those are exactly the tasks whose value does not reduce to producing an answer.

The exposure estimates point the same way but invite a common misreading. Eloundou and colleagues estimated that around 80 percent of workers have at least 10 percent of their tasks affected by large language models, and about 19 percent have at least half affected, with exposure rising at higher wages (Eloundou et al., 2024). Exposure is not displacement. A task can be exposed to AI and become more valuable, less valuable, or unchanged, depending on whether the technology substitutes for it or complements the person doing it. Reading exposure as a job-loss forecast is the most frequent error in popular summaries, and the paper’s authors are explicit that it is an error.

The case that expertise is the thing that keeps paying

The most careful argument about what stays scarce comes from David Autor, who frames the issue as the future of expertise. Expertise commands a market premium when it is both necessary and scarce. He contrasts an air traffic controller, whose median pay was about $132,000, with a crossing guard, whose median pay was about $33,000, even though both make rapid decisions to prevent collisions (Autor, 2024). The difference is not the importance of the decision but the scarcity of the skill required to make it safely.

Autor’s proposal is deliberate about its own status: it is an argument about what is possible, not a forecast. AI, used as a decision-support tool, could extend the reach of expertise — allowing a nurse practitioner, a paralegal, or a junior engineer to handle higher-stakes work that was previously reserved for a scarce elite, restoring a middle stratum of well-paid cognitive jobs. But the same tool, used to substitute rather than extend, could deepen the gap between those who direct the machine and those who merely feed it.

That conditional is testable, and recent work tests it directly. In a pre-registered three-month randomized trial, Autor and colleagues gave 133 practicing patent lawyers at eleven firms access to a custom AI drafting assistant (Autor et al., 2026). With the tool, lawyers produced better drafts — about a third of a standard deviation higher on blinded scoring, with the largest gains among juniors. Then came the crucial second measurement, designed to separate performance from learning. After three months, every lawyer was asked to mark up a flawed patent application without AI — a task that foregrounds expert judgment. The treated group outperformed controls overall, but the entire advantage sat with the senior lawyers; juniors showed no average gain, and their scores fanned out, with more very good and more very poor results. The authors describe AI as a springboard for some junior lawyers and a hammock for others: a tool that makes adequate work easy can remove the pressure that used to build judgment.

This is the most important result for the question of scarcity, and it cuts against the simplest version of both optimism and pessimism. Abundance of drafting did not make expertise irrelevant. It made the acquisition of expertise more fragile for the people who need it most, while rewarding the people who already had it. Foundational expertise appears to be the precondition for extracting durable skill from AI-assisted practice.

Where the machine’s presence can quietly subtract

The most counterintuitive evidence comes from experienced software developers. METR ran a randomized trial in which 16 veteran open-source maintainers completed 246 real tasks in their own mature repositories, with each task randomly assigned to allow or forbid AI tools (Becker et al., 2025). The developers expected AI to make them about 24 percent faster before they began, and still believed, afterward, that it had sped them up by about 20 percent. The measured result was the opposite: tasks took 19 percent longer when AI was allowed. Time saved on active coding and searching was more than consumed by prompting, reviewing, waiting, and correcting output that was “often directionally correct, but not exactly what’s needed.”

This is a single study in a demanding setting, and its authors are careful to say it does not establish that AI fails developers generally; later follow-up work by the same group found evidence of a small positive effect and flagged selection problems that made the newer measurement unreliable. But the study is valuable precisely because it isolates a phenomenon the abundance narrative hides: a capability can feel helpful while lowering measured productivity, and the perception gap can be large. The developers were not lazy or gullible. They were working in a regime with many implicit requirements and high quality standards, where the marginal value of a fluent but slightly wrong suggestion is negative.

The same pattern appears in the creative domain. Doshi and Hauser randomized AI idea generation across 293 writers, with 600 evaluators scoring the results (Doshi & Hauser, 2024). Access to AI ideas made individual stories more creative, better written, and more enjoyable — particularly for writers who began with less creativity. But the AI-assisted stories resembled each other more than the unaided stories did. Individually, each writer was better off; collectively, the pool of new work grew less diverse. That is a scarcity story with a twist: the machine adds value at the individual level and subtracts it at the population level by concentrating output around a common center. When everyone’s analysis is drawn from the same cheap source, distinctive judgment and taste become the thing that stands out, and also the thing the tool most tempts you to abandon.

The three-way distinction that keeps this honest

It is worth stating plainly which claims in this article are which kind.

Demonstrated, in controlled settings: AI assistance raises measured output on many bounded professional tasks; the gains skew toward less-experienced workers; the frontier is jagged, and unassisted users can be more accurate outside it; sustained assistance can help or harm the acquisition of unaided skill depending on how it is used.

Plausible but not yet established by measurement: that the patterns generalize to most occupations; that the value of expertise rises in aggregate; that the desirability of the underlying work improves rather than merely the throughput.

Speculative: that any of this constitutes a permanent theory of human value. Nobody has run the experiment on a mature, AI-abundant economy, because no such economy has existed long enough to study.

Keeping these categories distinct matters because the temptations run in both directions. Advocates inflate a bounded productivity result into a claim that expertise is obsolete. Critics inflate a slowdown in one demanding setting into a claim that the technology is hollow. Neither reading survives contact with the actual studies.

What the scarcity argument implies for institutions

If the scarce inputs are problem selection, verification, ownership, and taste, then the institutions that produce them become the leverage points.

Education is the first. The evidence from patent drafting suggests that making the producing part of a task easy can starve the skills that come from doing it badly and getting corrected. A curriculum that lets students delegate the reasoning will produce people fluent at prompting and unable to judge. A curriculum that uses AI as a tutor while requiring unaided performance on the parts that build judgment is a different bet. The distinction is not about whether to allow the tool; it is about which capability the tool is allowed to stand in for.

Hiring and promotion are the second. If entry-level work is where judgment used to be built, then automating entry-level work changes the talent pipeline even when it improves output. Firms that measure only throughput will find that they have optimized away the bench. Measuring whether a junior employee becomes more capable over time, not just faster today, is the corrective.

The third is verification as a professional norm. The jagged-frontier result implies that the safe way to use an abundant analyst is to keep an independent test of correctness — a source, a measurement, a second method, a person with standing to disagree. Where verification is cheap and expected, abundance is a gain. Where it is skipped because the output looks confident, abundance is a liability that accumulates silently.

None of this requires believing that AI cannot be trusted. It requires recognizing that trustworthiness is a property of the system around the model — the checks, the incentives, the accountability — and that those are human-built and human-owned.

The qualities that do not get cheaper

The central question was what becomes scarce. The measured answer is narrower and less romantic than the usual list, and more useful for being narrow.

Attention that notices when a fluent output is wrong. Taste that can tell a good problem from a merely tractable one. Judgment that knows which errors are survivable and which are not. The willingness to put your name on a decision the model helped produce. And the relationships through which expertise is transmitted from people who have it to people who need it — the part of learning that a springboard can accelerate and a hammock can replace.

These are all capacities that AI can support. Each of the productive studies shows a version of assistance that extends reach without displacing ownership. The danger is not that the machine takes the scarce skills away; it is that the abundance of everything adjacent to those skills makes them easy to neglect. Cheap analysis does not make judgment obsolete. It makes the absence of judgment much harder to notice until it is expensive.

The reason to insist on the three-way distinction — demonstrated, plausible, speculative — is that the same evidence is used to sell two opposite futures. One says expertise is finished. The other says the technology is hollow. The studies support neither. They support a narrower and more durable claim: when the thinking gets cheap, the value moves to whoever decides what the thinking is for, and to whoever can tell whether it was right.