In 1865, a young English economist named William Stanley Jevons noticed something awkward about steam engines. The more efficient they became, the more coal Britain burned. Efficiency, he argued, would not conserve the resource. It would industrialise the appetite for it.
A hundred and sixty years later, we are having the same conversation about cognition, with roughly the same level of composure, which is to say not much. Goldman Sachs estimates 300 million full-time jobs are exposed to generative AI. The World Economic Forum, in its January 2025 report, calmly predicts 92 million roles displaced and 170 million created by 2030. Dario Amodei has reportedly warned that half of entry-level white-collar work could vanish within five years (I haven't verified the exact figure independently, but the remark circulated widely). Everyone has a chart. Nobody has an answer.
The honest position is a dialectical one, so let us do this properly.
This time, the machine eats the escape route
Every previous automation wave came with career advice. Mechanisation hit the farm, so you moved to the factory. Robotics hit the factory, so you moved to the office. The office, specifically the cognitive bit of it, was the escape route. Thirty years of education policy rested on one premise: thinking was the thing machines would leave us.
That premise is now being stress-tested at roughly $0.07 per million tokens. The Stanford AI Index tracked the cost of querying a GPT-3.5-class model falling about 280-fold between late 2022 and late 2024. Klarna announced in early 2024 that a single assistant was handling the workload of 700 customer service agents. IBM paused hiring for thousands of back-office roles it expected AI to absorb. These are not projections. They are procurement decisions.
I'll confess something from my consulting years. I have written more automation business cases than I care to remember, and the "FTE savings" line was always a polite fiction. It meant: same people, more throughput, everyone keeps their badge. For the first time, the line is technically honest. When a fine-tuned model drafts the deliverable, the deliverable genuinely needs fewer drafters. The fear is not Luddite. It is arithmetic.
Efficiency has never once shrunk appetite
And yet. Jevons has been right for 160 years, in coal, in electricity, in compute itself. GPU performance per watt has improved by orders of magnitude over a decade, and the result was not conservation. Nvidia's data centre revenue went from about $15 billion to over $100 billion in roughly two fiscal years. Nobody used less compute. We used the efficiency to justify using absurdly more.
Labour markets have their own version. James Bessen's classic example: ATMs spread across America from the 1970s onward, and the number of bank tellers roughly doubled over the following four decades, because cheaper branches meant more branches, and tellers shifted from counting cash to selling products. Spreadsheets did the same to accounting. VisiCalc arrived in 1979, bookkeeping clerks declined, and the ranks of accountants and financial analysts swelled, because once analysis got cheap, everyone wanted more of it.
I run a small local rig at home: Ollama, a few Qwen and DeepSeek models, the usual tinkerer's setup. Inference there is effectively free. So has my token consumption gone down? It has gone up by orders of magnitude. I have not automated myself out of a weekend. I have found fifteen new things to throw tokens at, most of them pointless, two of them genuinely useful. Multiply that by every knowledge worker on earth and the antithesis becomes visible: cheap cognition does not reduce the demand for cognition. It explodes it, and the work migrates to specifying, verifying, integrating, and taking responsibility for what the machine produces.
The resource being economised is you
Both positions survive, because they operate at different units of analysis. Resolve the dialectic and it looks like this: the optimists are right about volume, the pessimists are right about price.
Total consumption of cognition will explode. Jevons virtually guarantees it. And the unit price of routine cognition will collapse, because that is what efficiency means. Both things. Employment may hold up remarkably well in aggregate while wages compress in the exposed middle and job content mutates beyond recognition. The paradox saves the economy. It does not save you personally. It operates on aggregates with the indifference of a tide table.
The sharpest edge is distributional, and it sits at the bottom of the career ladder. I saw a preview on a large industrial programme in France years ago. We automated the tedious data-quality checking in the engineering document flow. Nobody was fired; the senior engineers were delighted. But two junior posts quietly never opened, because the juniors' traditional work, the checking, the reconciling, the learning-by-tedious-repetition, was gone. Five years later the programme wondered aloud where its mid-level people were. We had sawn off the bottom rungs and were surprised the ladder no longer reached the ground.
That, I suspect, is the real synthesis. AI does not create mass unemployment so much as it creates a repricing event and an apprenticeship crisis, arriving together, each wearing the other's coat.
What to watch while staying calm
Three indicators matter more than the think-pieces. First, entry-level hiring in exposed occupations: if the bottom rungs keep disappearing, wage data will lag the damage by years. Second, token consumption per knowledge worker: if it keeps doubling, Jevons is winning and new tasks are being born somewhere. Third, whether agentic tooling produces genuinely new job categories or merely cheaper versions of old ones. History says new tasks appear because humans hold a comparative advantage somewhere. The open question of the decade is where that somewhere is now.
Britain never lost its appetite for coal services. It lost it only when it found something better than coal. Whether something proves better than human cognition at the things humans still get paid for is, mercifully, an empirical question. We will find out. Panicking early will not improve the answer, and panic, it should be noted, is also a form of labour. It too will be automated eventually.