The capability thesis

somethingg builds products whose measure of success is what the person using them can do, rather than how much work has been taken off their hands.

The case rests on three claims. As AI improves, what people contribute to their work becomes more valuable, not less. The way AI is currently built and sold widens the gap between what someone can produce and what they can actually do. And almost nobody is building for the space between the first two.

This is a position rather than a settled result, and it is worth saying so at the top. The evidence below runs in both directions; the strongest field studies point the other way on the first claim, and we have kept them in rather than around. Section 10 lists the four things that would have to be true for any of this to matter, and what would falsify each. If the argument is wrong we would rather find out early, which is most of why it is published.

The gap

There has always been a distance between what a person can produce with tools and what they could do without them. AI has turned that distance into something else. It is now possible to produce work of a standard you have never reached, in a domain you have never studied, repeatedly, with nothing in the process to tell you that this is what is happening.

The gap takes two forms and they are not the same failure.

Capability that never forms. Someone describes an application. It gets built, deployed, and starts taking real users. They have never learned authentication, or data modelling, or what happens to a system under load. The application works, and nothing about the experience distinguishes it from the experience of having known how to build it. Then it leaks user records, or it will not deploy, or a client asks a question they cannot answer.

This is the newer failure and now the more common one. It is also the harder, for a reason worth stating plainly: there is no moment of noticing. Someone who has gone rusty knows what sharp felt like. Someone whose capability never formed has nothing to compare against, so there is no gradual dawning — only the day the work fails in a way they cannot diagnose.

Capability that erodes. Someone three years into a career, competent and well reviewed, whose output has never looked better. They have also not been genuinely stuck on a problem in eighteen months, and have not read a source end to end in longer than that. None of this appears in their appraisal or in their employer’s numbers. They find out when the work changes, or when they interview elsewhere, or when the model is confidently wrong about something they can no longer catch.

The same asymmetry drives both. Not having capability is slow and invisible in its consequences; handing the work over is convenient and immediate. People act far more readily on the second, and so do the companies selling to them.

Three terms, used precisely here. Capability is what a person can do unassisted — not what they know, and not what they can produce with a tool in hand. Productivity is output per unit of input. Output produced with a tool and the capability of the person producing it are two different quantities, and treating them as one is the mistake this whole argument is about. Help means making a person more capable while they do their work, rather than producing their work for them.

The gap is now measurable

Until recently this argument had to be made from laboratory results. It no longer does, because the first population to live entirely inside the gap has produced a public record.

A scan of production applications built with AI coding tools found the share carrying at least one critical vulnerability was 58% (an industry scan of more than 1,400 production applications). Over the same period, tracked CVEs attributable to AI-generated code rose roughly sixfold across three months (Georgia Institute of Technology, 2026). Individual cases make the shape clearer than the aggregates: applications reaching real revenue in weeks with publicly readable databases, private messages exposed through broken access control. In each case the output was real, the revenue was real, and the capability behind it was not there at all.

Two questions you can ask about AI

Almost all commercial AI is built around one question: which parts of a person’s work can the model do instead of them?

That question produces a particular relationship between the tool and the person. The tool becomes more valuable as the person becomes less so. Every improvement to the tool subtracts from the person using it, and taken far enough, the tool holds everything and the person holds nothing.

There is a second question available, and hardly anyone is asking it: which parts of a person’s work can the model make them better at? Under that arrangement the gains accrue to the person and stay with them. Capability, unlike access, does not switch off when a contract ends.

There is an obvious objection and the argument has to survive it. Plenty of tools have removed effort and nobody mourns the loss — mental arithmetic went to the calculator, and nobody wants their compiler back. What separates those cases from this one is the difference between dependence that substitutes for a capability and dependence that compounds one. Arithmetic was an input to accounting rather than the thing accountants were paid for, so automating it removed a step without removing the competence. The trade is different when the step being automated is the thing that keeps the underlying skill alive, or, in the newer case, the thing that would have built it in the first place.

Why now

Four things became true at roughly the same time, and none of them were true when the current generation of AI products was designed.

The models converged. As of March 2026 the leading models from six labs sit within 25 Elo points of one another on the Arena leaderboard, and the US–China frontier gap has fallen to 2.7% from somewhere between 17 and 32 percentage points in 2023 (Stanford HAI, 2026). Once capability differences are in the low single digits, competition moves to cost, reliability and deployment. That diffusion continues until access is close to equal is a forecast rather than a finding, and the argument below depends on it.

Deployment tilted towards handing work over. Directive use — where a task is handed over with minimal back-and-forth — rose from 27% to 39% over eight months to late 2025, and automation briefly overtook augmentation for the first time (Anthropic, 2026). It has since partly reversed, but the level remains well above where it started.

The first measurements of the cost arrived. Nearly every result in the evidence section below was published within the last eighteen months. Two years ago this argument could only have been made from theory.

The labour market moved first where the theory predicts. Using payroll records covering millions of US workers, early-career workers in the most AI-exposed occupations show a relative decline in employment of 13 to 16% while more experienced workers in the same occupations remain stable (Brynjolfsson et al., 2025). The declines concentrate in occupations where AI automates rather than augments — the distinction this document turns on is already visible in payroll data.

The window is about defaults rather than technology. Workflows calcify around whatever was convenient at the start, and the people entering work now are the first cohort whose entire professional formation happens with these tools present. They will not be recovering a way of working they once had.

Why the human contribution becomes worth more

The claim that human contribution grows in value as AI improves sounds like wishful thinking. It follows from how production works when tasks depend on one another.

Kremer’s O-ring model (Kremer, 1993) describes production where every task has to go right for the output to be worth anything, so quality multiplies rather than adds. One badly executed task drags down everything else, however good the rest were. The return to improving any single task is the product of how well everything else is done: improving one task inside a mess is worth little; improving one task in an otherwise excellent process is worth a great deal.

Apply that to AI. As the model improves, the quality of its tasks rises, and the return to human quality rises with it.

Someone shipping an application built by a capable model has a process where nearly every task is executed to a high standard and one is not — the judgement about what the thing needs, which nobody performed.

The critical-vulnerability rate above is not a coincidence or a quality problem with the models. It is the O-ring result observed in production.

One experiment shows the same thing directly. Participants given an AI coding assistant wrote significantly less secure code than those working without one, and were more likely to believe their code was secure (Perry et al., 2023). The participants who produced the fewest vulnerabilities were the ones who trusted the assistant least and interrogated their own prompts most. Those are not two findings. The assistant raised the quality of the code it wrote and lowered the quality of the human check on it, and the second effect determined the output.

The limit of this argument. A task performed perfectly by AI is a task that stops being human. The set of human tasks shrinks even as the value of each one rises, and the O-ring result shows only one of those two movements. Two things limit how far the shrinking goes: much of the judgement at stake turns on private information about a particular firm, client or dataset, which a general model has never seen; and the bottleneck result holds however many tasks are automated. How much resists, and for how long, is genuinely open. It is the risk this company carries, and we would rather state it than argue it away.

Why the gap widens anyway

Capability is built through effortful engagement — retrieving something with difficulty, generating an answer before seeing one, working at the edge of what you can currently do. Forty years of research on desirable difficulties (Bjork, 1994) establishes that conditions which make performance worse in the moment produce better long-term retention and transfer, while fluent conditions produce weaker durable learning. Difficulty is not an obstacle sitting in front of the learning. In large part it is the learning.

A tool that reduces effortful engagement while holding measured output constant causes capability to fall, or never to rise, and nothing in the output shows it. The work gets done, on time and to standard. Current performance is an unreliable measure of learning, and almost all measurement of AI productivity is measurement of assisted output.

Self-assessment is no substitute. Experienced developers working in their own repositories forecast that AI would speed them up by a quarter, believed afterwards that it had, and were in fact slower (METR, 2025).

A second effect matters more inside a workplace. Judging work well requires having produced comparable work yourself. As people stop producing, their ability to evaluate degrades alongside it — and evaluation is the capacity every argument for human oversight depends on. Bainbridge named the general form the irony of automation: the better the automated system, the more critical and the less practised the human operator’s remaining role becomes (Bainbridge, 1983). A workforce that cannot tell correct from plausible is not oversight.

And nobody’s incentives point the other way. Training that makes someone generally better also makes them more valuable to a competitor, so the firm paying for it does not capture the return — which makes a firm’s optimal contribution to general, portable training zero (Becker, 1964). Skills depreciate by default and not slowly: the wage premium to applied STEM skills falls by more than half over the first decade of a career as the content of the work turns over (Deming & Noray, 2020). The concern is no longer a fringe one; beyond a threshold of agent accuracy, an economy can tip into a state where general knowledge erodes even as personalised advice remains excellent (Acemoglu et al., 2026).

What the evidence says

The evidence is mixed, and the mixture is informative — but only if two different claims are kept apart. The first is that human contribution becomes more valuable as AI improves, which predicts a widening gap between more and less skilled users. The second is that the gap between output and capability widens under current tooling. The best-known field studies bear on the first.

On capability itself, most of the strongest results are about formation rather than erosion, which is worth noticing given how the argument is usually stated.

Two AI tutors were given to around a thousand school mathematics students. One behaved like a standard chat interface; the other was prompted to give teacher-designed hints instead of answers. Students with the standard interface did better during practice and then did worse than the control group once access was removed. The safeguarded version largely eliminated the harm (Bastani et al., 2025). Same model, opposite effects, and the difference lay entirely in how the interaction was designed. This is the most important result here, because it moves the question from what AI will do to people towards what we are choosing to build.

In randomised trials where an AI assistant was withdrawn without warning, participants who had used it solved significantly less and gave up significantly more often than the control group, with effects appearing after roughly ten minutes (Liu et al., 2026). People who learned a topic through model synthesis developed measurably shallower knowledge than those who learned the same facts through ordinary web search, and adding live source links did not fix it — only about a quarter clicked even one (Melumad & Yun, 2025).

The clearest case of true erosion is also the most confounded: detection rates in unassisted colonoscopy fell within months of routine AI assistance arriving (Budzyń et al., 2025).

The compression results measure what happens when a tool is introduced. The capability results measure what happens to the person. These are different quantities over different horizons rather than contradictory findings. Our bet is that the second compounds and the first does not, and the evidence for it is not yet conclusive.

This has happened before, and it went both ways

The factory system shows how it goes wrong: craft knowledge held by individual artisans was broken into divided tasks demanding less of each worker. Capability was not lost to the technology. It was designed out, because designing it out was cheaper.

Skill-biased technical change shows the other outcome. For most of the twentieth century, new technology raised the return to education rather than lowering it — a race between education and technology in which inequality widened when technology moved faster than educational attainment, and narrowed when it did not (Goldin & Katz, 2008). The technology did not decide the outcome. The relative speed of human capital formation did. And the residual is large: within narrowly defined industries, the plant at the 90th percentile of productivity produces roughly twice the output of the plant at the 10th from the same measured inputs (Syverson, 2011).

The strongest optimistic case is that AI could extend the reach and value of human expertise, letting more people do work currently reserved for elite experts (Autor, 2024). We do not dispute that it is possible. Which of the two happens is a design question, and it is currently being answered at enormous scale by companies competing for revenue — building products that visibly increase output, because increased output is what buyers can see and pay for — and by firms maximising short-run profit against accounts in which human capital does not appear at all. Neither party is behaving irrationally. Both are optimising against a measure that cannot see the cost.

Why this cannot be added to an existing product

If the difference between a tool that builds capability and one that does not comes down to how the interaction is designed, the reasonable question is why every existing AI company does not simply make that change.

The counterexample is already shipping and makes the point better than the argument does. Over 2025 every major lab built something close to the safeguarded tutor — study modes, guided learning, learning modes — each designed to give hints rather than answers. The insight was never scarce, and several of these organisations published it themselves.

What shipped was a toggle. Opt-in, off by default, one click from a product whose ordinary behaviour is to produce the work, and available precisely to the user who already wanted to be taught, which is not the population whose capability is at stake. Nothing else moved: not the default, not the pricing, not the metrics the product is judged by. A safeguard bolted onto a system optimised for output is subordinate to that system, and the numbers decide which one wins every time they disagree.

Building for capability and building for output conflict at almost every decision. The version that makes someone better is usually the version that takes longer, feels worse in the moment, and shows worse on every metric the industry runs on. It is also not a feature: a constraint has to be adopted at the start, because taking one on later means subtracting from the thing you already sell.

Four load-bearing beliefs, and what would falsify each

Everything above rests on four things being true. Each could turn out false, and it is better to find that out in year one than in year five.

1. The gap between output and capability is real and costly, rather than capability being reallocated to higher-order work. False if people offload the routine and reinvest the freed capacity in judgement. If that is what happens, the problem solves itself and this company is not needed.

2. The multiplier runs in the direction predicted — a widening gap between more and less skilled users as models improve. False if the gap keeps narrowing once everyone has had the tools for years. The current evidence is consistent with the opposite, which makes this the most uncomfortable item on the list.

3. Design decides the outcome. False if the gap widens regardless of how the interaction is built: if safeguards only delay the effect, or people route around them to the unrestricted tool in the next tab. This is the belief everything else rests on, and it currently rests on a small number of studies. Everything else here could be correct and there would still be nothing to build.

4. People will act on it. The cost is slow and invisible; the convenience is immediate. False if nobody chooses the harder path when the easier one is one click away. The thesis could be right and the company still fail.

The measure of success that follows is a change in what a person can do unassisted. Not output, not time saved, not tasks removed. The first product is knowhow, which is the narrowest available test of the third belief: the same models everyone else has, the opposite design choice, and a measurable answer to whether the people using it end up more capable.