The Permanent Underclass

The comforting story about technology and work goes like this: machines take the old jobs, the displaced move into new ones, and after a painful transition everyone ends up richer. It has been true often enough to feel like a law. But it is not a law. It is an outcome that depended on specific conditions — that the new jobs arrived in time, that ordinary people could learn them, that the value of human effort had somewhere to go. The honest question about AI is not whether those conditions held in the past. It is whether they will hold this time. If the value of some kinds of labor falls and does not come back, the difference between this transition and the others will not be a footnote. It will be the whole story.

This essay is not a prediction that it will go badly. It is an argument that we should be able to look directly at the version where it does, without flinching, because flinching is how societies sleepwalk into shocks they could have softened. Optimism that cannot survive examining its own failure case is not optimism. It is a refusal to look.

What "permanent" would actually mean

The frightening word is not "underclass." It is "permanent." Unemployment that comes and goes is a hardship; economies have machinery for it. What would be new is labor whose market value falls and stays down — not because the people are idle by choice, but because the thing they can do can be done better, faster, and for almost nothing by a machine that keeps improving. Previous automation replaced specific tasks and left whole categories of human judgment standing. The unsettling property of general-purpose AI is that it reaches into the categories we assumed were safe: writing, analysis, coding, design, the cognitive middle that a generation was told to retreat into when the factories closed.

If that happens broadly, the usual escape hatch — retrain and move up the ladder — narrows, because the ladder is being automated from several rungs at once. You cannot reliably tell people to flee into "more cognitive" work when the cognitive work is precisely what is being absorbed. The advice that protected the last displaced generation may not be available to the next one, and pretending otherwise is its own kind of cruelty.

It is worth being precise: this is a conditional. It depends on how far capability goes, how fast, and how the gains are distributed. But the conditional is plausible enough that a serious society should have a plan for it, the same way it plans for other low-probability, high-consequence events it hopes never to need.

The shock does not vanish — it relocates

Here is the part that gets skipped. When labor loses value, the loss does not evaporate. It has to land somewhere. In the simple economic story it lands as cheaper goods and higher overall wealth, and everyone benefits. But "everyone benefits on average" is compatible with large numbers of specific people being worse off, and people do not live in the average. They live in their own situation, and they vote, organize, and act from there.

So the shock relocates into politics. A large group of people who can no longer trade their effort for a decent life is not an economic statistic. It is a political force, and an unstable one. History is fairly clear that societies with a permanent, visible gap between those who own the productive machinery and those locked out of it do not stay calm. They get populism, scapegoating, and a hunger for someone to blame. The danger of a permanent underclass is not only that it is unjust to the people in it. It is that it is corrosive to everyone, including the winners, because it dissolves the shared sense that the system is worth defending.

This connects to a question raised by what Ilya saw: if the capability is real and arriving, the interesting problems stop being purely technical and become questions of who benefits and who decides. A technology powerful enough to reshape the labor market is powerful enough to reshape the political settlement that sits on top of it. Pretending the second is someone else's department is how the first one blindsides you.

Why the optimistic answers may not arrive on time

There are real, honest reasons to think it could go well. New industries we cannot yet imagine. Human work that becomes more valuable precisely because it is paired with capable machines — the centaur pattern, where the human supplies judgment, accountability, and taste while the machine supplies leverage. Whole categories of human attention — care, presence, craft, the things people want from a person — that no amount of capability makes obsolete. These are not fantasies. They are plausible, and some version of them will probably be part of the answer.

The problem is timing. Transitions are measured against human lifetimes and political patience, not against the long run in which everything works out. The long run is cold comfort to a forty-five-year-old whose skill stopped paying in the decade they needed it most. New industries may arrive, but maybe not where the displaced live, not at the wage they had, not before the anger does. The gap between "it will be fine eventually" and "it is fine for me now" is exactly the gap where political shocks are born. Betting that the good outcome will arrive fast enough, on its own, for the specific people hit hardest, is a bet — and dressing a bet as a certainty is how you end up unprepared.

Honest responses do not require pessimism

To take this seriously is not to be a doomer. It is to do what any competent system does with a known failure mode: build for it before it triggers. That starts with refusing the two easy exits. One is denial — insisting the old pattern will repeat because it always has, which mistakes a historical contingency for a guarantee. The other is fatalism — deciding the outcome is fixed and nothing can be done, which is just denial wearing a darker coat. Both let you avoid the actual work.

The real responses are unglamorous and political, because the shock is political. They include thinking now, before the pressure peaks, about how the gains from automation are shared — whether through how ownership is structured, how the surplus is taxed and redistributed, or how access to the productive machinery itself is widened rather than concentrated. They include treating the dignity question as separate from the income question, because a society can keep people fed and still leave them without a place to stand, and people deprived of a place to stand do not stay quiet. And they include builders making choices about leverage — whether the tools they ship concentrate power in a few hands or distribute it, whether they replace people wholesale or augment them, whether the default is the centaur or the empty chair.

None of this is solved in an essay. The point is narrower and, I think, defensible. The optimistic case for AI and labor might be right, but it is not guaranteed, and the failure case is severe enough that we should hold it in view rather than assume it away. A society that looks honestly at the version where some labor loses its value for good is far more likely to keep that version from arriving — and far better prepared if it does — than one that keeps repeating that everything has always worked out before. It usually has. "Usually" is not a plan.