The Permanent Underclass

The core of it

The argument rests on four mechanisms.

1. AI replaces thinking work, and there's no next rung to step onto. The Industrial Revolution moved farmhands into factories. Mechanization moved factory workers into services. Automation moved assembly lines into knowledge work. Every wave had a destination. When AI automates thinking work itself, what is the next sector? There may not be one. The ladder runs out of rungs.
2. The capability gap compounds. People who use AI get far more productive — building systems, automating their own workflows, adding to the advantage daily. People who don't fall behind faster and faster. This is not a straight-line gap; it's exponential divergence (the distance doubles rather than adds). My own cos- system is the proof of concept: every day spent building with AI makes the builder more capable of building with AI. Someone starting from zero next year is further behind than someone starting from zero today.
3. Access barriers reinforce each other. Cost ($20/month for frontier AI), education (knowing what to ask), digital literacy (finding your way around the interface) and the judgment to evaluate what the AI hands back all gate access. Each one props up the others: you can't build AI literacy without access, and you can't justify the cost without the literacy. The loop locks people out.
4. AI improves faster than people can retrain. Earlier waves gave humans decades. Agriculture to industry took a century. Manufacturing to services took a generation. AI capability doubles on a scale of months. By the time a displaced worker finishes a retraining program, the skills they learned may already be automated. The target moves faster than the archer can aim.

Why "permanent"

That word is the load-bearing claim. Temporary displacement is normal. Permanent displacement is a civilizational problem. The case for permanence:

• The skills gap reproduces itself across generations. Parents who can't use AI raise children in AI-illiterate households. Schools in low-income areas adopt AI tools last. The thinking environment itself becomes stratified. This is "it takes money to make money," applied to intelligence — see intelligence inequality.
• There is no precedent for automating the destination sector. Every previous wave automated the origin sector — farming, manufacturing — and workers moved to the destination sector: manufacturing, then services, then knowledge work. AI is automating the current destination without a clear next one. "New jobs will emerge" is an assertion, not an argument. It may be true. It may not.
• Retraining half-lives are shortening. If a skill stays useful for 5 years today and 2 years tomorrow, continuous retraining becomes a full-time job. You can't retrain for a living while also earning a living. The math breaks.

Who argues it

• Yuval Noah Harari — the "useless class," from 21 Lessons for the 21st Century and Homo Deus. His argument is that AI and biotech will produce people who aren't just unemployed but unemployable — economically irrelevant. It's the most provocative framing and the most contested. The word "useless" does a lot of work, and critics call it dehumanizing. The economic logic is still worth taking seriously even if the language isn't.
• Kai-Fu Lee — AI Superpowers, with displacement predictions built around the US-China rivalry. His number: AI will displace 40% of jobs within 15 years. His proposed fix is an economy organized around human empathy and care — the work AI does badly. Worth testing: is that a real alternative sector, or is it the "learn to code" of the 2030s?
• Andrew Yang — universal basic income (a regular cash payment to everyone), argued from automation. His contribution is turning displacement into a policy conversation rather than an economic theory. His framing: the truck drivers are coming and we have no plan. This thesis extends it — even if UBI delivers income, it doesn't deliver purpose, and a purposeless class is a politically unstable one.
• Erik Brynjolfsson and Andrew McAfee — The Second Machine Age and Race Against the Machine. Careful empirical work on how technology produces abundance and inequality together. The "bounty" (more stuff) and the "spread" (more inequality) are both real. This thesis says the spread wins if the bounty never reaches the bottom.
• Daron Acemoglu — the skeptic's skeptic, critical of AI productivity claims and focused on what actually happens in labor markets. His research suggests much AI adoption is "so-so automation": it displaces workers without producing a matching gain in productivity. If he's right, the permanent underclass isn't a side effect of progress. It's the main event.

The debate

Two camps, and the truth may not be in the middle.

Camp 1: displacement is temporary

New jobs will appear, as they always have. AI creates more value than it destroys. Education and retraining close the gap. The loom killed weaving jobs and made textile factory jobs. The car killed horse jobs and made auto industry jobs. The computer killed clerical jobs and made tech jobs. Every time we've predicted technological unemployment, we've been wrong.

Strongest version: AI makes intelligence cheap, which makes services cheap, which raises living standards for everyone. The jobs of the future are the ones nobody can picture yet — "social media manager" was unimaginable in 1990. Retraining works if you fund it properly.

Camp 2: this time is different

Automating thinking is a different kind of thing from automating muscle. The pace of change outruns human adaptation. The gap hardens within a generation. Every earlier wave left humans something machines couldn't do — physical dexterity, judgment, creative thinking, emotional intelligence. AI is coming for all of those at once. There is no safe harbor.

Strongest version: the "new jobs will emerge" argument only counts the survivors — we remember the transitions that worked out. We don't count the civilizations that collapsed during technological transitions, the generations that suffered through them, or the populations that never recovered. And this transition is moving too fast to leave room for a generation to adapt.

Rich's take

Both, so far. The democratization is real — that's the Keys argument — but it may not reach everyone, which is this one. The tension between universal access and structural exclusion is the actual research question. Not "will AI create an underclass?" but "under what conditions does displacement become permanent, and what intervention prevents it?"

Where it connects

• The Man Handed Over the Keys — the counter-thesis. Keys says AI was accidentally handed to everyone. This one asks whether the handover actually reaches everyone. What if $20/month is still too much, or the thinking skills needed to use AI are their own barrier? Keys describes the supply side. This describes who gets locked out on the demand side. Both can be true at once, and the tension between them is the core question.
• Intelligence inequality — related but distinct. That one describes the mechanism: it takes intelligence to use intelligence, and AI-augmented people compound their advantage. This one describes the outcome: that mechanism, unchecked, produces people who don't just fall behind for a while but get locked out structurally. Intelligence inequality is the how. Permanent underclass is the so-what.
• Livelihood sovereignty — that framework assumes you can build income systems with AI. What about people who can't? Sovereignty is the aspiration; the permanent underclass is the population for whom it's structurally out of reach. If sovereignty requires agency, and agency requires the intelligence to use intelligence, then sovereignty itself becomes an elite good.
• Token capital — Satya Nadella's moat argument. If tokens (the AI processing you can afford to spend) become the scarce resource and the unit of production, then people without token access are the new poor. Concentration at the infrastructure layer could coexist with genuine accessibility at the application layer — but only if the application layer actually reaches everyone. If it doesn't, token capital becomes the new capital-K Capital and the underclass is defined by token poverty.
• Automation displacement — this thesis is the logical endpoint of that one, if the displacement turns out to be structural rather than cyclical.
• Lift the bottom — the direct counter-argument. If AI raises baseline capability for everyone through better healthcare, cheaper education and automated government services, the underclass may live better even without AI fluency. The question: does a rising floor matter if the ceiling is rising faster?

The frame problem

There's an assumption buried in this whole thesis: that the capitalist-tech world is the only world. It isn't.

Intelligence — specifically the need for AI-boosted intelligence — is relative to the context a person lives in. In the fast-moving tech economy, cognitive capability is the price of admission. You must use AI to stay relevant. The race is real, the stakes are real, the displacement is real. Step outside that frame — a rural community, an artisan trade, a subsistence economy, a contemplative life — and the hierarchy dissolves. Intelligence is still valuable to anyone who wants it. It just isn't mandatory.

• The underclass may only be permanent inside the capitalist-tech frame. A farmer in a traditional agricultural community isn't in an underclass — they're in a different class system entirely. A craftsman building furniture by hand isn't falling behind AI-augmented manufacturing unless he's trying to compete in that market. The underclass isn't defined by lacking intelligence. It's defined by being in a context where AI-boosted intelligence is the survival requirement, and not having it.
• The pressure gradient matters more than the absolute gap. In the tech world the pressure to augment is enormous — every day you don't use AI you fall further behind competitors who do. In traditional worlds that pressure barely exists. The same person could be "underclass" in San Francisco and perfectly functional in a small town in Portugal. The gap is real; how much it hurts depends on whether you're somewhere that punishes it.
• There is more than one kind of intelligence. The capitalist-tech world privileges the analytical kind and treats the rest as nice-to-haves: bodily-kinesthetic, naturalistic, emotional, intrapersonal, wisdom. In other contexts those are primary. The fisherman's feel for the natural world, the midwife's emotional intelligence, the elder's wisdom — none of these are boosted by AI, and none of them need to be. The thesis only holds if you define intelligence as "the kind AI can augment or replace."
• Some of life resists optimization. The contemplative, the embodied, the relational. I chose the high-pressure cognitive world and am increasingly drawn to simpler modes — trail running, smaller circles, presence over productivity. That isn't regression. It's recognizing that the augmentation race is one game among many, and not everyone has to play it.
• The uncomfortable question. Is the permanent underclass a problem to solve, or a frame that only exists because we assume everyone wants to be, or should be, inside the capitalist-tech world? If someone opts out and lives well somewhere that doesn't require it, are they underclass, or just elsewhere?
• The real danger. Not that an underclass forms, but that the capitalist-tech frame becomes inescapable — that markets, policy and infrastructure design remove the option of standing outside it. The farmer who doesn't need AI today needs it tomorrow, when AI-optimized supply chains set the price of grain below what traditional farming can produce. The artisan who doesn't need AI today needs it tomorrow, when AI-generated designs flood the market. The underclass becomes permanent not because people can't learn AI, but because the world reorganizes until every context demands it.

Open questions

• What share of the workforce is structurally unable to use AI? Is it 10%? 30%? 50%? How would you even measure that?
• Does UBI solve the money problem and create a purpose problem? Is a well-funded but purposeless life actually better than what we have?
• Is there a cognitive floor below which AI tools simply don't help? If so, where is it, and how many people are under it?
• How does this interact with inequality that already exists — race, class, geography, age? Does AI amplify the existing sorting or create new lines?
• What does "permanent" actually mean — one generation? Two? Forever? Is there a tipping point after which the gap reproduces itself?
• Does the deflationary spiral partly solve this by making a dignified life cheaper?
• What's the difference between "underclass" and "differently classed"? If AI covers everyone's basic needs, is the inequality that's left tolerable?
• What role does dignity play? Harari's "useless class" is an economic category. But people need to feel needed. Economic security without social purpose may be its own crisis.

Evidence still to gather

• Labor market displacement data from early AI adoption, 2023–2026: which jobs are actually being eliminated versus assisted?
• Education and retraining success rates: what share of displaced workers actually transition, and at what age does retraining stop working?
• Historical parallels: Industrial Revolution displacement timelines, the mechanization of agriculture, the deindustrialization of the Rust Belt. Which populations recovered, which didn't, and how long did it take?
• Current AI tool adoption by income, education, age, geography and race. Where are the adoption cliffs?
• Acemoglu's data on "so-so automation" — displacement without a matching productivity gain.
• UBI pilot programs: what do they show about purpose, dignity and social stability once income is separated from work?
• Leapfrog patterns in the developing world: did mobile phones leave a permanent underclass of non-adopters, or did adoption eventually reach everyone?

The uncomfortable version

The most uncomfortable version of this thesis is that the permanent underclass isn't a policy failure. It's a market outcome. Markets optimize for efficiency, and an AI-augmented worker is more efficient than one without. No amount of goodwill changes that math. The question isn't whether the underclass forms. It's whether society decides to intervene before it hardens, and whether "intervene" means retraining, redistribution, or redefining what it means to contribute.

All theories · clickbridge.com · Rich Price · last reviewed September 2026 · People linked here have their own pages at people.clickbridge.com