Day 958 — Why The Jevons Paradox Won't Save You
Or: why the most popular reassurance about AI and jobs is a half-truth wearing economist's clothing.
Every time I post something about AI displacing knowledge workers, a particular reply shows up in the comments. Sometimes it’s polite, sometimes it’s smug, but it always arrives:
“You’re forgetting Jevons paradox. When something gets cheaper, we use more of it. Demand for cognitive work will explode. Everyone will be fine.”
I want to take this seriously because the people making this argument aren’t stupid, and the paradox is real. But “Jevons will save us” has become a kind of intellectual comfort blanket, pulled out whenever the unemployment data gets uncomfortable, then put away before anyone asks how exactly it’s supposed to work. So let’s pull it out properly and look at it.
The case, stated fairly
William Stanley Jevons noticed in 1865, in The Coal Question, that as steam engines got more efficient, Britain burned more coal, not less. Cheaper coal-per-unit-of-work meant coal was suddenly economic for applications that had been marginal before. Total consumption exploded.
The pattern repeats. Cheaper bandwidth didn’t kill the internet; it gave us Netflix, TikTok, and video calls with grandma. Cheaper computation didn’t end programming jobs; it gave us a software industry the size of a small continent. Cheaper air travel didn’t empty the airports; it filled them.
The optimistic application to AI: as AI makes cognitive work radically cheaper per unit, latent demand expands until total employment in cognitive fields holds steady or grows. Legal services that previously cost $400/hour become affordable to small businesses, individuals, the chronically underserved. Medical second opinions, tutoring, accounting, design — all the things people wanted but couldn’t afford — suddenly become accessible. The pie grows faster than the per-slice productivity gain, and human workers ride the expansion.
There’s even a beloved case study: ATMs. The conventional wisdom in 1990 was that ATMs would destroy bank teller jobs. Instead, teller employment grew for two decades after ATM rollout, as James Bessen documented in his work on automation and labour. Why? Because ATMs made branches cheap to run, so banks opened more branches, and tellers shifted from cash-handling to relationship work the machines couldn’t do.
It’s a good story. It’s even partially true. But it leans on three assumptions that are doing more work than the storytellers admit.
Assumption 1: Demand is highly elastic
Jevons only kicks in if there’s enormous untapped demand waiting to be unlocked by lower prices. Sometimes there is. The world genuinely wants more legal advice, more therapy, more medical diagnostics, more tutoring, more software systems — vast unmet need at current prices, gated almost entirely by cost.
But sometimes there isn’t. Does the world want 50x more middle-management memos? 100x more quarterly earnings analyses? 200x more enterprise sales decks? Demand for many specific knowledge-work outputs is bounded by the real-economy activity they serve. A company with $200M in revenue doesn’t need ten times more financial reports just because reports got cheaper. It needs the reports that match its actual operations. The supply curve shifts; the demand curve mostly doesn’t.
Cheap coal found new uses because coal is an input to making physical things, and we always want more physical things. Cheap memos don’t find new uses because memos aren’t an input to anything most people care about. This distinction matters, and Jevons-optimists routinely paper over it.
Assumption 2: Expanded demand requires humans
This is the assumption that gets smuggled in without examination, and it’s the one I want to nail to the wall.
Jevons paradox is about demand for the task. It says nothing about demand for humans doing the task. Demand for coal exploded in the 1860s. Demand for human coal-shovelers… did not, particularly. The work got mechanised faster than demand grew.
The ATM story works because ATMs couldn’t do relationship banking, mortgage advice, or trust services. They were a partial substitute; they handled the routine, freeing humans for the complex. If a future ATM could do everything a teller does, including building rapport with the elderly widow opening her first investment account, the analogy collapses. Demand for “banking interactions” might still explode; demand for human bankers wouldn’t.
So when someone says “AI will expand demand for legal services,” the honest follow-up is: expand demand for human lawyers, or for legal services delivered by AI? These are very different questions, and the answer determines whether Jevons is your friend or your replacement. Coal demand growth was great for coal companies. It was less great for the men with shovels.
If AI is a complement to human cognitive labour, making us 10x more productive but still needed, Jevons probably helps. If it’s a substitute — capable of doing the whole job — the expanded demand flows to more AI, not more humans, and Jevons becomes the polite name for your unemployment.
The current evidence is mixed and shifting fast. In some domains AI looks complementary (radiology, where it improves diagnostic accuracy alongside humans). In others it looks increasingly substitutive (routine document drafting, customer service tier 1, code generation for well-specified tasks). The mix will determine which sectors get the friendly Jevons and which get the British coal shovelers’ Jevons.
Assumption 3: The transition is smooth
Even granting that Jevons works in the long run, the transition can still be catastrophic. Falk and Tsoukalas’ AI Layoff Trap paper (which I covered recently) makes this point sharply: firms can rationally fire workers ahead of AI capability, betting on tools that don’t quite exist yet, and the displaced workers don’t smoothly retrain into the new expanded-demand roles. A ten-year transition with 15% knowledge-worker unemployment is a generational disaster, even if year twenty looks fine.
We can already see this mechanism warming up. Challenger, Gray & Christmas recorded 1.2 million U.S. job cuts in 2025 — the highest annual total since the pandemic — with hiring plans at their lowest since 2010. AI was directly cited in roughly 55,000 of those cuts, less than 5% of the total. You can read that two ways: AI is barely a factor (the AI-washing critics’ read), or AI is the named tip of a much larger iceberg of “restructuring” and “market conditions” cuts that are functionally the same bet. Either way, the door to new hires is closing faster than the door to existing ones, exactly as Falk and Tsoukalas predicted.
And the speed matters. Previous Jevons-style adjustments unfolded over decades across narrow sectors: coal across the 19th century, ATMs across two decades, spreadsheets across one. Labour markets had time to reallocate. AI hits most cognitive work simultaneously over a few years. Even if every affected sector reaches a new equilibrium with healthy human employment, the simultaneous shock has no historical precedent to reassure us.
There’s also the distributional question. Cheap legal services helping small businesses are great, but they don’t help the paralegal who lost her job last March. Even when Jevons creates new work, it goes to different people in different cities with different skills. The aggregate statistics can look fine while individual lives are wrecked.
Where Jevons does help
I want to be fair to the argument, because it isn’t wrong, just oversold. There are sectors where Jevons will probably soften the blow significantly:
Healthcare, where latent demand is enormous and regulatory & trust requirements keep humans firmly in the loop.
Childcare and early childhood education are similar to eldercare. Physical presence is the product; latent demand is enormous (cost is the binding constraint for most families), and no one is sending their toddler to be cared for by robots running on GPT-6.
Skilled trades that touch AI tooling, where physical presence is irreducible, and AI is a force multiplier.
Advisory work with high accountability stakes — the kind where someone needs to sign their name, take the call at 11 pm, and be sued if they’re wrong.
These are the sectors I’d retrain into if I were 25.
Where it doesn’t
And where I wouldn’t:
Routine cognitive work with bounded real-economy demand (most middle-office corporate roles).
Work where AI is a near-complete substitute (a lot of document production).
Anything where the value being delivered is information transformation rather than judgement, accountability, or relationship.
The honest frame
Jevons paradox is a useful concept. It is not a panacea, a guarantee, or a serious answer to “what happens to displaced workers.” It’s a partial mechanism that works in some sectors and not others, on timescales that may or may not match the urgency of the transition.
The framing I keep coming back to in this newsletter is: who captures the surplus? When AI makes a task cheaper, the savings go somewhere. Sometimes, to consumers (cheaper services), sometimes to firms (higher margins), sometimes to workers (higher wages for the humans still needed), and sometimes to AI providers (concentrated rent). The Jevons story implicitly assumes the surplus flows in a way that creates more human work. But “implicitly assumes” is doing a lot of lifting, and history shows the surplus can go anywhere.
If you find yourself reaching for Jevons as reassurance, ask the harder question instead: in this specific sector, with this specific AI capability, who captures the surplus, and does that flow create or destroy human work?
That’s a question with answers. They’re just not always the comforting ones.
— Olaf