Every major technology that was supposed to eliminate jobs — the steam engine, the spreadsheet, the ATM, the internet — ended up creating more of them. Not because the fears were wrong about what those technologies could do, but because the fears were wrong about what happens next. A 160-year-old economic concept explains why, and it’s the most important framework for understanding where AI and employment are actually headed.
It’s called the Jevons Paradox, and it starts with coal.
The Counterintuitive Logic of Efficiency
In 1865, English economist William Stanley Jevons published The Coal Question, in which he observed something that seemed to contradict basic logic. James Watt’s improvements to the steam engine had made coal dramatically more efficient — less coal required per unit of work. One would expect total coal consumption to fall. Instead, it skyrocketed. Cheaper, more efficient steam power made industrial activity economical at scales that hadn’t been viable before. Demand exploded, and coal consumption went with it.
The principle generalizes: when a technology makes a resource cheaper and more efficient to use, total consumption of that resource tends to rise, not fall — because lower costs unlock uses that were previously unaffordable, creating new demand that more than offsets the efficiency gains.
Now substitute coal for cognition. AI is collapsing the cost of thinking tasks — writing, analyzing, coding, summarizing, planning — toward something approaching zero. In the short run, this unambiguously destroys certain jobs. But Jevons suggests that’s only the first act.
What Happened Every Time Before
The pattern holds across economic history with remarkable consistency. When spreadsheets arrived in the 1980s, they were expected to eliminate accounting jobs. Instead, they made financial analysis cheaper and more accessible, the demand for analysis exploded, and the number of people working in finance grew. When ATMs were introduced, bank teller jobs were supposed to vanish. Instead, ATMs made bank branches cheaper to operate, banks opened more branches, and branch employment rose. When the internet disrupted media, it simultaneously created e-commerce, digital marketing, cloud computing, and an entire startup economy.
A recent Citadel Securities report framing the current moment — which it calls the “2026 Global Intelligence Crisis” — makes the macro case directly: U.S. unemployment sits at 4.28%, software engineering job postings are up 11% year-over-year, and AI capital expenditure is running at roughly $650 billion annually. This is what an investment boom looks like. It is not what a labor market collapse looks like.
The Honest Caveats
The Jevons case for AI is compelling, but several serious economists have pushed back on its easy application — and their objections deserve space.
The paradox is powerful but not automatic. Where AI is a near-complete substitute for a specific human task — not a complement to it — demand for that role can remain permanently depressed regardless of what happens to aggregate employment. Elevator operators didn’t find new elevator-adjacent careers when elevators were automated. The task simply ceased to require a human.
The distributional problem is arguably more important than the aggregate one. Even in optimistic historical scenarios — globalization, for instance — aggregate output expanded while concentrated populations bore permanent losses. The gains from AI efficiency, like those from trade, will likely flow disproportionately to capital owners and highly skilled adaptors. The people currently losing work to AI — illustrators, translators, copywriters, junior coders, entry-level analysts — are not abstract statistics. They are real people whose specific skills are being devalued in real time.
Timing compounds the problem. Job losses from AI tend to be sharp, localized, and immediate. Jevons-style job creation unfolds slowly, unevenly, and often in different sectors and geographies. The gap between those two curves is where actual human suffering lives, regardless of what the long-run equilibrium eventually looks like.
The Question That Actually Matters
A recent Duke University survey of CFOs, conducted in partnership with the Federal Reserve Banks of Atlanta and Richmond, found a striking gap between perceived and actual AI productivity gains — companies report believing AI has made them more productive while the financial results don’t yet bear it out. The job losses being attributed to AI, the study found, are real but not at doomsday scale — and may reflect pandemic-era overhiring corrections as much as genuine automation displacement.
The aggregate story and the individual story are both true simultaneously: more total economic activity, more new roles, and a meaningful cohort of workers whose careers were collateral damage in the transition. History suggests the Jevons outcome is likely. History also suggests that “likely in aggregate over decades” provides cold comfort to the illustrator who lost half their client base last year.