Technology, Automation, AI, and Innovation

Here is the single most important fact in all of long-run economics. For almost all of human history, the typical person lived on roughly the same income as their great-great-grandparents. A peasant in 1700 was not much richer than a peasant in 700. Then, starting around 1750 with the Industrial Revolution, living standards exploded. Plotted on a graph over thousands of years, the line is flat, flat, flat — then suddenly shoots straight up. Economists call this the "hockey stick."

What caused the blade of the hockey stick? Not more workers, and not just more machines. The answer is technology — new ways of producing more from the same effort. This chapter explains how innovation drives growth, why progress is so disruptive, why new technologies pay off slowly, what automation really does to jobs, what AI might do, and — the deepest question — who actually keeps the gains.

27.1 Innovation: the real engine of growth

In the 1950s the economist Robert Solow (Nobel Prize, 1987) tried to measure what makes economies grow. He added up the growth from more capital (machines, factories, tools) and more labor (workers and hours). Then he compared that to actual growth in output per person. A huge chunk was left over — unexplained by capital or labor at all.

Total factor productivity (TFP)

The "leftover" growth that isn't explained by simply having more machines or more workers. It measures how cleverly we combine inputs — better technology, better know-how, better organization. Often called the "Solow residual."

Productivity

Output per unit of input — for example, goods produced per hour of work. Rising productivity is the only sustainable way to raise wages and living standards.

The cause→effect chain is short and powerful: innovation → more output per worker → higher real wages → higher living standards. A country gets rich not mainly by working harder or piling up more machines, but by inventing better ways to do things. That is why ideas, not just stuff, sit at the center of growth.

27.2 Schumpeter and "creative destruction"

So how does technology actually enter the economy? Through entrepreneurs who break the old way of doing things. The Austrian-American economist Joseph Schumpeter named this in 1942: creative destruction — "the process of industrial mutation that incessantly revolutionizes the economic structure from within, incessantly destroying the old one, incessantly creating a new one."

His key insight: the most powerful competition isn't two firms shaving a few cents off the price of the same product. It's a new product, process, or business model that makes the old one obsolete entirely. The entrepreneur earns big profits as a temporary reward for innovating — and then competitors copy it and the profit gets competed away, pushing the next innovator to leap again.

The uncomfortable implication: bankruptcies and layoffs are not a flaw in the system — they are the mechanism of progress. Old industries must die for resources (workers, capital, attention) to flow to better ones. Economists Aghion and Howitt turned this into formal "Schumpeterian growth theory" in 1992.

27.3 General-purpose technologies (GPTs)

Some technologies are bigger than others. A general-purpose technology (GPT) (Bresnahan and Trajtenberg, 1995) has three traits: it is pervasive (spreads into almost every industry), it keeps improving over time, and it spawns complementary innovations (it triggers a cascade of other inventions built on top of it).

The classic trio is the steam engine (1700s–1800s), electricity (from the 1880s), and the computer / information technology (later 1900s). Many economists argue AI is the next one. The crucial point: a GPT's value comes less from the core invention itself than from the flood of follow-on reinventions it makes possible. Electricity mattered not because of the dynamo, but because of refrigerators, assembly lines, elevators, radio, and a redesigned world.

27.4 The productivity paradox: why payoff comes slowly

In 1987 Solow joked: "You can see the computer age everywhere but in the productivity statistics." Businesses were pouring money into computers, yet US productivity growth had actually slowed after 1973. Computers everywhere, growth nowhere. This is the productivity paradox.

The resolution came from historian-economist Paul David in 1990, in a famous essay called "The Dynamo and the Computer." He looked back at electricity. Factories had run on one giant central steam engine, with belts and pulleys carrying power to every machine. When electric motors arrived, factory owners simply bolted a single big electric motor where the steam engine used to be — and saw almost no gain. The big payoff came only in the 1920s, after the "unit drive" redesign: one small motor per machine. That freed factories from the central shaft, allowing single-story layouts, flexible arrangement, and bright lighting. It took roughly 40 years from invention to full payoff.

The lesson: a GPT needs costly complementary investments — new organization, new processes, retrained workers — before it pays off. Technology alone does nothing; you have to rebuild around it.

THE GPT DIFFUSION CURVE (productivity over time)

 productivity
   ^
   |                                   ____ explosion phase
   |                                 /      (firms redesign
   |        ___________            /         around the tech)
   |       /  adjustment \________/
   |      / (flat: tinkering, learning)
   |  ___/ investment phase
   | /   (costs up, productivity DIPS)
   +-------------------------------------------> time
     new GPT      ~years of slog        big payoff

This pattern repeated. As IT finally diffused, US productivity growth roughly doubled from about 1995 to 2004 — exactly the GPT explosion phase. (It then slowed again after about 2005, and economists still argue about why — Robert Gordon says the easy gains are gone; Erik Brynjolfsson says we're just early in the next wave.)

27.5 Automation and jobs: the great fear

Every wave of automation triggers the same fear: the machines will take all the jobs. The classic version is the lump-of-labor fallacy — the mistaken belief that there is a fixed amount of work in the world, so if a machine does a task, that work is gone forever.

Why is it a fallacy? Because automation cuts costs, and that sets off ripples: lower costs → lower prices → people buy more and have more money left over → they spend it elsewhere → demand and jobs grow in other sectors. And machines often complement workers, making the human part of the job more valuable.

The pattern is old. In England from 1811–1816, the Luddites smashed textile machines to save their jobs — and the long boom in textile employment that followed gave us the term "Luddite fallacy." More recently, in 2013 Frey and Osborne estimated that 47% of US jobs were "at risk" of computerization within two decades. It became one of the most-cited numbers ever — and one of the most criticized. They scored whole occupations as automatable, when really only certain tasks within a job are. A task-level redo by the OECD found only about 9–14% at high risk. The predicted mass unemployment did not arrive; even insurance underwriters, scored as highest-risk, saw employment rise about 16% from 2013–2021.

A stunning fact from Autor (2018): roughly 60% of the jobs people did in 2018 did not exist in 1940. Automation reinvents work more than it erases it.

So is the fear baseless? No — and here is the honest part. The real damage is not mass unemployment but transition pain. When automation hits, the people displaced are often mid-skill, mid-career, and concentrated in particular towns. They face long spells without work, "wage scarring" (lower pay for years even after re-employment), and communities that hollow out. Acemoglu and Restrepo's 2020 studies of industrial robots found robots did cut jobs and wages in the specific US local labor markets exposed to them. The economy-wide jobs hold up; the distribution of harm is brutally uneven.

27.6 Why the middle gets hollowed out

To see who gets hurt, we need the task-based model (Autor, Levy, Murnane, 2003). Older theory — skill-biased technological change — simply said technology helps educated workers and raises the "college wage premium." The task model is sharper. Computers are great at routine tasks (rule-based, repeatable: bookkeeping, filing, repetitive assembly) and so they substitute for the middle-skill workers who did them. But computers complement non-routine tasks at both ends: high-skill analytic and managerial work (a spreadsheet makes an analyst more powerful) and low-skill manual work that's hard to codify (cleaning, caregiving, food service).

The result is job polarization: growth at the top, growth at the bottom, and a hollowed-out middle. Middle-class clerical and factory jobs shrink; their wages stagnate. This is a leading explanation for rising US wage inequality since about 1980 (though others argue weakened unions and a falling minimum wage matter just as much).

27.7 AI's potential impact — a genuine dispute

Will AI be the next great GPT? Experts sharply disagree, and you should treat this as honestly unsettled.

The optimists: Goldman Sachs (2023) estimated generative AI could raise global GDP by about 7% (roughly $7 trillion) over a decade and lift US productivity by around 1.5% per year. McKinsey reached similar figures.

The skeptic: Daron Acemoglu, in "The Simple Macroeconomics of AI" (2024), used the task framework and got a tiny number — total productivity gains under 0.66% over ten years (about 0.05% a year). His logic: only a modest share of tasks are cheaply automatable by current AI, and AI may not create many valuable new tasks. About 40% of labor income is "exposed" to AI — but, crucially, exposure is not the same as replacement.

The 2024–2025 reality looks like another productivity paradox: hundreds of billions invested, little economy-wide productivity lift yet — exactly what Paul David's diffusion-lag story predicts. Yet controlled studies of individual workplaces do show gains: customer-support agents using AI were about 14% more productive (Brynjolfsson, Li, Raymond, 2023), with the biggest help going to the least experienced workers — a possible leveling effect, the opposite of the old skill-bias story.

27.8 Network effects and winner-take-all

Digital technology has one more twist: network effects. A product gets more valuable as more people use it. One telephone is useless; a million connected telephones are priceless. The same holds for social networks, payment systems, and operating systems.

Metcalfe's Law says a network's value grows roughly with the square of its users (n²) — a rough heuristic, not a literal law. The consequence is winner-take-all markets that "tip" toward one dominant platform: Windows/Office, Google search, Facebook, Visa. Once a network is big, newcomers struggle because the value is the existing users.

27.9 Who captures the gains?

Now the deepest question. Technology almost certainly grows the total pie. But who gets the new slices? The gains split three ways: consumers (lower prices), workers (wages — but very unevenly across skill levels), and capital owners (firm profits).

Recent decades tilted toward capital and the top. The labor share of income — the fraction of national income going to wages rather than profits — fell in the US from roughly 64% to 58% between the 1980s and 2010s. Karabarbounis and Neiman (2014) linked part of this to cheaper machines replacing workers. Gains also concentrated in a few "superstar firms" — the most productive companies grabbing an outsized share (Autor et al., 2020).

The core lesson, argued forcefully by Acemoglu and Johnson in Power and Progress (2023): in history, broad-based prosperity from new technology arrived only when institutions — unions, laws, public policy, countervailing power — forced the gains to be shared. Technology grows the pie; politics and institutions decide the slices. That distribution is a choice, not an automatic outcome.

Key Takeaways

• Long-run growth in living standards comes mostly from technological progress (TFP), not just more machines and workers — that's the "hockey stick."
• Schumpeter's creative destruction means progress works through the death of old industries — disruption is the mechanism, not a malfunction.
• General-purpose technologies pay off slowly: their gains arrive only after costly redesign of organizations and skills (the productivity paradox; electricity took ~40 years).
• Automation reallocates and reinvents work far more than it permanently destroys it (the lump-of-labor and Luddite fallacies) — but transition pain for displaced workers is real and concentrated.
• Computers substitute for routine middle-skill tasks and complement the top and bottom, producing job polarization and rising inequality.
• AI's macro impact is genuinely contested (Acemoglu's ~under-1% vs. Goldman's ~7% GDP); early data fits the slow-diffusion pattern, with firm-level gains already visible.
• Network effects create winner-take-all platforms — but who keeps the gains from any technology is decided by institutions and policy, not by the technology itself.