The most important question AI raises about work is not "will there be jobs." It is "who captures the value." When an agent performs a task a person used to perform, the economic value of that task does not disappear — it moves. It moves from the worker who used to be paid to do it toward whoever owns the agent, the compute it runs on, and the data it was trained on. The anxiety everyone feels about AI and employment is real, but it is aimed at the wrong target. The jobs may well survive. The question is who gets the surplus once the work is done by something you can own rather than someone you have to pay.
Hold two things apart, because almost every public argument collapses them. One is the level of employment: how many people have work. The other is the distribution of income: who receives the value that work produces. These are different variables, they move independently, and the honest historical record says something reassuring about the first and something unsettling about the second.
The wage channel, and why it is narrowing
For roughly two centuries, the way most people captured a share of economic output was by selling labor. You owned little productive capital — no factory, no fleet, no seat on an exchange — but you owned your time and skill, and firms needed them badly enough to pay a wage that let you claim a slice of what the economy produced. The wage was the channel. It was how someone with no capital nevertheless participated in the growth of a capital-intensive economy.
That channel works because human labor is a genuine input firms cannot get any other way. Remove the "cannot get any other way" and the channel narrows. This is the mechanism, stated plainly, and it is worth walking one transaction at a time.
Take a firm that runs a task — drafting a first-pass contract, reconciling invoices, writing the boilerplate half of a marketing brief, triaging support tickets. Before automation, the firm splits the value that task creates with the worker who does it. The worker's wage is their share of the surplus; the firm keeps the rest as margin. The split isn't charity — it reflects that the firm needs the worker and the worker has an outside option. Now the firm replaces the task with an agent whose all-in cost is a fraction of the wage. The task still gets done. The value it produces is unchanged, maybe higher. But the portion that used to flow to the worker as wages now stays with the firm as margin. Nothing was destroyed. The surplus simply stopped being split and started being kept — by the owner of the agent.
Multiply that across enough tasks and you get the aggregate pattern economists have tracked for decades: labor's share of national income drifting downward across many advanced economies since around 1980, even as output per worker rose. You do not need to invoke AI to see the trend; it predates it. AI is best understood as a powerful accelerant of a capital-biased shift already underway, not a break in kind.
Why "learn to prompt" is necessary but not sufficient
The standard advice — become the person who directs the agents, learn to prompt, move up the stack — is correct as far as it goes, and I don't want to wave it away. It is the modern version of the centaur insight from chess. After Garry Kasparov lost to Deep Blue in 1997, he championed "advanced chess," where a human paired with an engine could, for a while, beat either alone; in the freestyle tournaments that followed, the human-plus-machine "centaur" was the strongest configuration on the board. The lesson generalizes: complementarity is real, and the worker who wields the tool can be worth more, not less, than before. David Autor's work on task-based automation and labor-market polarization makes the same point with more rigor — automation hits some tasks and not others, and workers whose tasks are complemented rather than substituted can see their wages rise. That is a serious counterargument to any flat "labor loses" story, and it deserves to be taken at full strength.
Here is why it is insufficient. "Learn to prompt" keeps you employable. It keeps you inside the wage channel, and possibly at a higher wage. What it does not do is change which channel the compounding returns flow through. If the returns to owning the agents are pulling away from the returns to supplying labor — even skilled, agent-directing labor — then staying maximally employable is optimizing the variable that is losing relative ground. You can be the best centaur on the team and still watch the value of the game accrue to whoever owns the engines, the tournament, and the audience. And the centaur edge did not hold: as engines improved, the human's marginal contribution to the pairing shrank toward zero, and strong engines running alone came to dominate freestyle play. Complementarity is not a fixed property. It is a race between how fast the tool improves and how fast your distinctive contribution to the pairing erodes.
There is a real learning-science reason to keep struggling with hard problems yourself rather than offloading everything to the agent. The "desirable difficulties" Robert Bjork identified — the productive struggle and retrieval practice that build durable skill — are exactly what you lose if the agent does all the reps. And Bloom's "2-sigma" finding, that one-on-one tutoring could lift the average student roughly two standard deviations over conventional instruction, hints at how much cheap, personalized instruction from an agent might raise human capability. But raising your human capital is, again, an investment in the labor channel. It makes you a better seller of labor in a market where the price of labor is under structural pressure. Worth doing. Not the whole answer.
The reframe: ownership as the channel that compounds
If the wage channel is narrowing, the question becomes: what other channel lets an individual participate in the upside? The answer, offered as analysis of where the returns are moving rather than as financial prescription, is ownership — of assets that compound.
The distinction that matters is between value that stops when you stop and value that keeps accruing after you stop. A wage stops when you stop. Equity does not. An audience you control does not. A proprietary dataset that accumulates as a byproduct of work you're already doing does not — I've argued separately that your byproduct is often more defensible than your product, precisely because it compounds silently at near-zero marginal cost while a competitor would have to live through your entire operating history to reproduce it. A tool or a productized service that runs without your hourly presence does not stop when you stop. These are the ownership analogues of the wage — ways to hold a claim on output that does not require you to keep selling your time.
This is not a get-rich scheme, and I won't sell it as one. Most individual assets don't pay off; equity is risky; audiences are fickle. The claim is narrower and structural: in a world where labor's share is under pressure and capital's share is rising, the rational response for an individual is to shift some weight from accumulating employability to accumulating ownership, because that is the channel the returns are migrating toward. The same logic operates at the frontier of the technology itself. When AI starts improving AI, whoever owns the self-improvement loop captures the compounding, and the advantage concentrates regardless of whether it ever reaches anything like superintelligence. Ownership of the compounding asset is the whole game at every scale, from a solo operator's audience to a lab's compute cluster.
Running an AI product myself — Velya — makes the asymmetry concrete in a way a spreadsheet doesn't. The value the product creates does not accrue to the hours anyone spends operating it. It accrues to the thing that persists: the model relationships, the accumulated user data, the equity. That is the difference between selling labor and owning an asset, felt from the inside.
The counterarguments, at full strength
Three serious objections cut against the pessimistic reading, and intellectual honesty requires stating them well.
First, new work historically appears. Every wave of automation for two centuries destroyed tasks and created others no one had imagined, and total employment rose. Betting against that track record is betting against the single most robust regularity in the economics of technology. The strongest version of this objection is not "trust me, jobs come back" — it is that we are systematically bad at imagining tasks that don't exist yet, so the absence of visible replacement jobs is weak evidence they won't materialize.
Second, commoditization passes value to consumers. If agents get cheap and competition compresses margins, the surplus doesn't stay with owners — it flows to everyone as lower prices and vastly expanded capability. A person with a phone gets, for free, what used to cost a salary. This is the genuine optimistic mechanism, and it is not fantasy.
Third, policy can redistribute. If the market allocates the surplus to capital, the political system can reallocate it — through taxation, through public ownership stakes, through direct transfers. The distribution markets produce is not the distribution a society must accept.
I take all three seriously. My response to the first two is the same: both operate on consumption, not on income and ownership. New jobs and cheaper goods raise how much you can consume; neither changes whether you hold a compounding claim on the economy's output. A world of full employment at flat wages, with abundant cheap goods and steadily rising capital income concentrated among asset owners, is fully consistent with every optimistic fact — and it is still a world where the distribution question is the live one. The third objection is the real one, and it points exactly where the debate should go.
What to actually do, and what to actually debate
For an individual, the move is to build ownership, not only employability — to route a deliberate slice of effort toward assets that keep throwing off value after the effort stops, while continuing to sell labor at the best price you can get. Employability keeps you in the game. Ownership is how you get a share of the pot.
For a society, the reframe is the whole point. Arguing about whether AI will "destroy jobs" is arguing about a variable history suggests will mostly take care of itself, and it lets the decisive question slip by unexamined. That question is distributional: as the wage channel narrows and returns concentrate in the owners of agents, compute, and data, how does a society keep most of its members holding a real claim on the output — through broadened ownership, through the tax base, through public stakes in the compounding assets, through something not yet invented? That is the debate worth having, and "will there be jobs" is the decoy that keeps us from having it.
Watch the labor share, not the unemployment rate. The unemployment rate can look fine while the answer to "who captures the value" quietly changes underneath it.