Will AI destroy the economy?

ELI5/TLDR

Gary takes the standard economist line — “AI raises productivity, so AI raises wages” — and walks it through actual industrial history. The Industrial Revolution, the cleanest natural experiment we have, produced 150 years of hellish poverty before living standards moved. Productivity grew enormously the whole time. The gains went to factory owners until labor movements and post-WWII redistribution forced a different split. His punchline: whether AI raises or crushes wages is not a technology question. It’s a distribution question, and the answer depends on who owns the model weights and who has bargaining power.

The Full Story

The two horses in the race

The orthodox case for AI being good for wages comes from one of the first ideas you learn in an economics degree — marginal productivity. If a worker can now produce 100 pounds of biscuits a day, somebody will eventually bid them up to roughly 100 biscuits’ worth of wages, because if the current employer doesn’t, a competitor will poach them. AI raises productivity, therefore AI raises wages. The Economist, Fortune (citing the IMF), and Fox Business have all been running variations of this line — including the stronger claim that AI will compress wage inequality by helping low-paid workers more.

The intuitive case runs the other direction. Companies need fewer juniors. Hiring slows. Especially at the bottom of the ladder, where AI is “good at doing things badly but very cheaply.” More unemployed workers, more slack in the labor market, wages drift down.

“On the one hand, we don’t need to hire so many workers… On the other hand, this technology has increased worker productivity. So I should hire them and it will make me profitable. So these are the two horses in the race.”

Gary’s project for the rest of the video is to show that the orthodox case rests on a misreading of the only big historical episode that’s anything like what’s happening now.

The Luddite story, retold

The standard reading of the Industrial Revolution is the one most economists internalize as a kind of folk parable. New technology came in, some workers — Luddites — broke the machines because their wages fell, but in the long run living standards rose enormously, so they were on the wrong side of history. Quod erat demonstrandum: don’t fight the machine, accept some short-term friction, the long run takes care of itself.

Gary’s argument is that the long run was extremely long. The first stage of the Industrial Revolution, beginning in late-1700s Britain, ran on the back of the Enclosure Acts pushing farmers off common land into a desperate urban labor pool. That pool made the factory model viable — wages were low because there were too many people and not enough land. The factories themselves were then genuinely productive marvels. He cites the Bryant & May matchstick factory in Bow, East London, producing 300 million matches a day from a single building. A worker’s jaw would gradually disintegrate from white phosphorus exposure if they stayed long enough.

Living standards for ordinary workers did not meaningfully rise until late in the 1800s, and the real lift only came after WWII — close to two centuries after the revolution began. He uses his grandmother as the anchor point: born in the late 1920s in London, several siblings dead of tuberculosis. That’s the very tail end of the “things worked out” curve.

“It took 200 years, you know, do you want, you know, 200 years of hell, you know, is it worth it?”

Where the productivity actually went

The mechanism Gary wants to highlight is straightforward and gets glossed over in the folk version. Productivity rose enormously. Wages didn’t. The arithmetic identity says the surplus has to land somewhere. It landed on a small British industrialist class.

Two factors made that distribution possible and stable for so long. First, British factories didn’t just outcompete British artisans — they outcompeted the entire world, propped up by the Royal Navy and forced trade. Indian textile workers were displaced at scale; he ties the resulting unemployment to the recurring famines. Second, the labor pool inside Britain was unorganized and replaceable until it wasn’t.

The thing that eventually shifted distribution wasn’t technology getting better. It was workers getting concentrated in cities and factories, realizing they could shut production down, and using that leverage. The labor movement is the missing variable in the orthodox story. Post-war redistribution — the welfare state, progressive taxation, mass housing — was the political consolidation of those gains.

“Were it not for these worker movements, I think it’s very, very possible that living standards would never have increased.”

Why marginal productivity theory needs a customer

This is the most useful piece of the video. Gary’s challenge to the orthodox model isn’t that it’s wrong on its own terms. It’s that it implicitly assumes a missing actor — a customer big enough to absorb the new output.

The marginal-productivity story says: a competitor will rush in, build a new factory, hire your workers at higher wages, and undercut you. That only works if the new factory has someone to sell its extra biscuits to. The two big productivity surges of modern history both had giant external demand sinks. The British factories had a colonial empire to absorb output. The Chinese export boom of the last 30 years had the Western middle class on the other side of the trade.

Where is the equivalent customer for AI-driven productivity gains? The Western working class is squeezed. The middle class can’t afford housing. Governments are cash-strapped. The only group with serious excess spending capacity is the very wealthy, and they don’t scale consumption with income.

“Good luck getting these machines to buy your cars.”

That’s the (probably apocryphal) Henry Ford / union rep exchange he uses to land the point. If the productivity gain has nowhere to go on the demand side, you don’t get a hiring boom, wage competition, and rising real wages. You get the gain captured as profit by whoever owns the technology.

The distribution mechanism — why “just work less” doesn’t happen

Keynes predicted in the 1930s that his grandchildren would work 15-hour weeks. Productivity per hour has risen massively since. Hours worked have not fallen — and once you account for women entering the paid workforce while childcare and housework remained, total hours per household have probably gone up.

Gary’s explanation for why Keynes was wrong sits inside a distinction most mainstream models skip over. Standard economic models implicitly assume a perfectly equal society where workers own a meaningful share of the productive capital. In that world, a productivity boost is a genuine choice — work the same and earn more, or work less and earn the same.

In an unequal society where capital, land, housing, and debt claims are concentrated at the top, work plays a different role. It’s the only way most people pay their ongoing rent on the assets the wealthy own — mortgage interest, rent, food, energy, transport, the interest your government pays on its debt. You can’t choose to work fewer hours, because the bills are denominated in whatever the asset owners decide to charge. Productivity gains owned by someone else don’t translate into your leisure time. They translate into their margin.

“Work is how you balance your basically, like from birth, debt to the rich.”

The actual question to ask about AI

By the end Gary collapses the whole thing into a single variable: who owns the technology. If model weights, compute, and the application layer are concentrated, AI is a wage suppressor and an inequality amplifier — the productivity gains pool at the top and the workers below get the displacement without the surplus. If ownership is broader, or if labor coordinates well enough to claim a share through bargaining or tax policy, AI can deliver Keynes’s leisure dividend.

The closing pivot is rhetorical — Gary moves from analysis to organising-pep-talk. The Industrial Revolution generation built the post-war social democratic settlement from a starting point worse than ours, and they had no precedent to point to. We do. The point lands cleanly enough that you can take or leave the call to action without hurting the argument.

Key Takeaways

The economist case for AI raising wages is the marginal-productivity argument: more output per worker forces competitors to bid wages up. The intuitive case for AI lowering wages is displacement: fewer hires, especially juniors, more slack, lower wages.
The Industrial Revolution is the precedent the orthodox view leans on, but the timeline is brutal — roughly 150-200 years between the technology arriving and ordinary living standards rising. Real gains came after WWII.
One factory in East London (Bryant & May) made 300 million matches a day. Workers’ jaws disintegrated from white phosphorus. The productivity miracle and the worker misery coexisted for generations.
During the British industrial boom, the surplus didn’t go to wages — it built an industrialist elite. The Indian artisan class was wiped out by British factory output backed by colonial trade enforcement; famines followed.
Living standards rose because of organized labor and post-war redistribution, not because productivity eventually “trickled down.” Take out the labor movement and the trickle never arrives.
The marginal-productivity argument silently requires a large solvent customer base to absorb extra output. Britain had colonies. China had the Western middle class. AI has — who exactly? Working class is squeezed, middle class is squeezed, governments are cash-strapped. The only excess spending sits with the very wealthy, who don’t scale consumption with wealth.
Keynes predicted 15-hour work weeks by now. Hours worked per household have risen, not fallen, despite massive productivity gains. The reason is distribution: in an unequal society work isn’t optional — it’s how you service ongoing claims (rent, mortgage, energy, food) on assets owned by someone else.
The framing question for AI isn’t “will it raise productivity?” (yes) but “who owns it and who has bargaining power over the surplus?” That’s a political question, not a technological one.

Claude’s Take

The analytical core here is genuinely strong and worth the time. The point that the marginal-productivity argument requires a demand sink — and that the two historical cases that actually delivered rising wages (post-Industrial-Revolution Britain after labor organized, the Chinese export boom) had unusually large external customer bases — is the kind of mechanism most “AI will raise wages” pieces gloss over. The reframing of work as the servicing mechanism for asset claims in an unequal society is also a clean way to explain why Keynes’s leisure dividend didn’t show up.

Where Gary plays loose: he flattens “the orthodox economist view” into a single naive position, when there’s actually a sizable literature (Acemoglu, Restrepo, Autor) that makes essentially his argument — that the distribution of gains from automation depends on institutional and political conditions, and that historical episodes where labor benefited were specifically the ones with strong worker bargaining power. He’s reinventing skill-biased technical change critiques without flagging that he’s doing so. That’s fine for a YouTube format but worth noting.

The Industrial Revolution comparison also smooths over real disanalogies. AI doesn’t yet have a single unambiguous productivity signal in aggregate data — measured productivity growth in 2024-2025 has been notable but not industrial-revolution-scale. The displacement is also concentrated differently (cognitive juniors, not handloom weavers), and reskilling pathways look different from “move to the factory town.” None of this kills the argument; it just means the analogy is doing more work than Gary admits.

The closing third drifts into political exhortation, which is on-brand for the channel and not really analysis. The 7/10 score reflects: strong on the demand-side critique of marginal productivity theory, strong on distribution-as-the-real-variable, weaker on engaging with the actual modern literature that says similar things, and noticeably better than the median macro-on-AI take precisely because he insists on the historical mechanism rather than waving at it.

For a finance audience the most portable insight is the demand-sink point. If you’re modeling AI as a productivity shock, ask the same question Gary does: who’s the customer for the extra output? In a world where the wage share is already low and household balance sheets are stretched, productivity gains landing as profit rather than wages is the base case, not a tail risk.