The consensus has a name. AI is electricity. A general-purpose technology, slow to diffuse, then it lifts everyone: the steam engine, the dynamo, the internet, and now the model. It is a comforting story. It is also wrong.
AI is not a second electricity. It is a capital good that substitutes for judgment, and its gains accrue to whoever owns the model. The rising tide is a category error.
AI is not a second electricity. It is a capital good that substitutes for judgment, and its gains accrue to whoever owns the model.
A single tool release repriced roughly $285 billion of software and legal firms in the days after: Thomson Reuters down 16 percent, RELX down 14, Wolters Kluwer down 13, LegalZoom down nearly 20.
Workers aged 22 to 25 in the most AI-exposed occupations trail their less-exposed peers by 19 percent, a gap that has widened steadily since it was first measured and reflects reduced hiring, not layoffs.
A Ph.D. in industrial sociology is offered $37 an hour to teach a machine to teach, against a minimum wage of $2 and youth unemployment of 47.4 percent.
A county manager emails thousands of public employees asking them to conserve power, because 37 data centers with 17 more planned have raised government and school rates by 25 percent, an estimated $5 million next fiscal year.
Five hyperscalers own 71 percent of global AI compute, and training a frontier model costs 3.5 times as much each year.
The mechanism is not mysterious. A Yale economist's model splits human work into two classes. Bottleneck tasks are essential to growth: energy, infrastructure, science, security. Supplementary tasks are ones the economy can expand without: arts, support, hospitality, design, research, economics. AI automates the bottlenecks with compute, because that is where the return on compute is highest. The supplementary work is left alone, not because it is sacred, but because it is not important enough to bother replacing. Once the bottlenecks run on compute, expansion depends on adding compute, not labor. Human work is then priced at the opportunity cost of the compute required to replicate it. Wages decouple from GDP. Labor's share of GDP converges to zero, in that model and under its assumptions: a result, not a measurement. Most income accrues to the owners of computing resources. The deepest uncertainty is whether taste and trust are durable invariants or only the hardest things to automate for now. I cannot yet tell which.
The thesis is not that AI makes everyone poorer. Total output can rise. Average wages can be higher. The claim is narrower and harder: the gains concentrate in whoever owns the compute, even while the total grows, and the distribution is a choice, not a technological necessity.
The opposing idea has a name. It is Jevons Paradox, and its cousin Amdahl's Law. The argument runs like this. If you automate 90 percent of a job, everyone ends up doing the 10 percent that remains, and that 10 percent "kind of expands to be 100% of what people do and kind of 10xs their productivity." Markets are good at recreating jobs. Agriculture, electricity, the internet: each displaced some work and created other work, though not on the same terms or timelines. The chief executive of a major bank, speaking on May 5, 2026 alongside the chief executive of an AI company, put it plainly: the capitalist society is very good at recreating jobs and recreating things. The AI company's chief executive, who had warned months earlier that AI could eliminate half of entry-level white-collar knowledge work within years, added the caveat himself: AI is moving faster than all these previous technologies.
The caveat is the whole problem. Jevons needs time. Electricity coincided with high productivity from 1900 to 1940 and still needed the Rural Electrification Act of 1936 to reach the places the market had skipped; its benefits took decades to arrive. The internet created new categories of work and also concentrated ownership in platform landlords. Agriculture displaced farm labor and created industrial and service jobs, on a different timeline again. Labor's share of income held roughly constant for much of the twentieth century, then fell in most developed countries after the 1980s. The rising tide was real, for a specific mid-century window. It was never a law of technology.
There is a second, quieter mechanism that does the same work. Jobs are bundles, not task lists. A radiologist reads scans, interprets edge cases, consults clinicians, and signs off. In weak bundles, such as support tickets and predictable code, AI splits off the automatable part and leaves the human with the leftovers. Output per worker rises, prices fall, and headcount follows. In strong bundles, around judgment, context, and responsibility, the worker gets faster and better paid. The danger is that strong bundles do not stay strong. The leftover work is the part that is hardest to price and easiest to crowd. Faster and better paid, for now, is not the same as safe.
The costs do not land on the owners. They land on the people in the scenes. A Ph.D. offered $37 an hour to teach a machine to teach, interviewed for 45 minutes by an AI, given feedback by email and a suggested retake, still looking for a classroom. A county manager who reportedly asked thousands of public employees to close the blinds and turn off their computers because 37 data centers raised the power bill. Twelve utilization-review nurses laid off in July at Montefiore in the Bronx and replaced with software. At Dignity Health in 2025, a supervisor told a nurse to follow an AI protocol against a dialysis patient's needs. At an unnamed hospital, a sepsis tool adopted without consulting nurses alerted too often on patients who were not septic.
The savings accrue to the system. The alert burden lands on the bedside, along with a twenty-minute wait for pain medicine and a nurse on the stairs, hand to mouth, frowning.
The honest response to concentration is not to fight the technology. It is to choose what cannot be subsumed and back it. Taste, but only if it means judgment under uncertainty with consequences, not aesthetic preference. Trust, but only in a person who is accountable, not in a feeling. Physical contact, in healthcare and the trades, for now. Accountability, the real kind, not a human front rubber-stamping machine output. And distribution, which is not a skill at all but the question of who owns the compute: universal income, or compute treated as a public resource akin to land or natural capital.
Bet on what cannot be subsumed. And keep the question asked by a man still looking for a classroom, after teaching a machine to teach: "What does it mean for us to hand over the judgment that defines our character and our professions to a machine?"