The Jevons Paradox
In 1865, English economist William Stanley Jevons found that steam engines had become far more efficient thanks to technological improvements on the original design. A machine could now produce the same output using much less coal, so the logical expectation was that Britain's total coal consumption would fall. Paradoxically, the opposite happened, consumption soared. The more efficient machines made coal-powered production so much cheaper that vastly more industries started using it, and total consumption ended up higher than before, not lower.
Jevons documented this in The Coal Question, and the finding became known as the Jevons Paradox, the idea that improving the efficiency with which we use a resource can increase, not reduce, total consumption of that resource. The context in which he raised it matters. Britain was in the middle of industrialization, and coal wasn't just another resource, it was the fuel that powered the factories, the mines, the railways, the merchant fleet. The question that was beginning to worry British society was whether a finite resource could keep sustaining a model of progress that depended on it more and more.
Jevons wasn't saying that every efficiency gain automatically increases consumption, he was describing a mechanism that can activate under certain conditions, and one that did activate with British coal. When an efficiency gain generates an additional increase in the use of something, economists call it the rebound effect, and when that rebound outweighs the original savings, that's when the paradox itself appears.
The Examples That Followed
Jevons's idea didn't stay locked inside nineteenth-century coal, it kept reappearing every time a technology became more efficient and, in turn, cheaper. Lighting is a classic case, every technical leap lowered the cost of producing light, and instead of spending on lighting going down, the amount of light we consume as a society multiplied. Something similar happened with transportation, Los Angeles had about ten thousand horses in 1900 and, half a century later, a million cars, because every improvement in engines made moving around cheaper.
When the first ATMs started being installed in the United States, both the public and bank managers themselves assumed human tellers would disappear. Economist James Bessen, of Boston University, looked at the actual data and found something different. The ATM reduced the number of employees needed per branch, but that same savings made opening a new branch so much cheaper that banks opened far more of them, and total teller employment didn't fall, it held steady and even grew. What changed was the type of tasks, from counting cash to handling customer relationships and selling financial products.
Though it predates Jevons by centuries, the case of Gutenberg's printing press shows a similar pattern. Copying books by hand was a trade, that of the scribes, and that specific trade ended up disappearing. What's also true, though, is that total book production and literacy multiplied to a scale that would have been impossible with scribes, and that growth generated entirely new trades, printers, editors, and later the author as a paid profession. The technology substituted a specific trade, but it ended up demanding much more human labor in the broader activity that trade was part of.
Journalist Tim Harford, in his BBC series 50 Things That Made the Modern Economy, revisits research from Planet Money showing that since the arrival of the spreadsheet (VisiCalc, in 1979) there are about 400,000 fewer bookkeeping clerks than in 1980, but about 600,000 more accountants proper. The reading is that the routine part of bookkeeping disappeared, but accounting activity in the broader sense became so much cheaper that it became viable to do far more financial analysis, and that generated demand for a different kind of professional than the one the technology replaced.
From Efficiency to Labor
These cases share a pattern that isn't exactly the Jevons Paradox, but something related worth flagging. Jevons is about the relationship between the cost of a resource and how much of it gets consumed. The question of whether a technology replaces or enhances human labor comes from a different idea, that of labor economics and task-based theory, where the variable at play isn't a physical resource but the demand for labor.
If a technology makes a task cheaper and simply replaces it, we call that replacement, and labor demand for that task falls. But if that cheapening drives the scale of the activity up so much that it ends up requiring more people than before, just for different tasks, more focused on oversight and judgment, we call that complementarity, and this is a Jevons-style rebound effect carried over from resource consumption to labor demand.
The examples show both sides at once, depending on which slice of the activity you look at:
- The ATM: a replacement for counting cash, a complement for banking activity overall.
- The printing press: a pure replacement for the scribe's trade, a complement for the production of written knowledge.
- The spreadsheet: a replacement for routine bookkeeping, a complement for financial analysis in the broader sense, according to Harford's reading.
In no case was the technology purely one thing.
Artificial Intelligence
In late January 2025, DeepSeek released its R1 model, with performance comparable to the most advanced models in the West but at a fraction of the training cost previously assumed necessary. Nvidia lost around 600 billion dollars in market value in a single trading session, because the initial read was that if training powerful models had become that cheap, demand for chips and infrastructure was about to collapse. Satya Nadella, CEO of Microsoft, responded almost immediately, writing that the Jevons Paradox was striking again, and that as artificial intelligence became more efficient and accessible, its use would soar.
What followed proved that read at least partly right, on the investment side. Meta raised its 2025 AI infrastructure spending to a range of 60 to 65 billion dollars just days after the announcement, and combined spending from the major cloud providers kept climbing through the rest of the year. Dario Amodei, co-founder of Anthropic, made a key argument, when you get more capability per dollar invested in training a model, a competitive company doesn't pocket that dollar, it reinvests it to reach a more capable model than the competition's, so the efficiency gain ends up absorbed by larger models rather than lower total spending. On top of that, much of the 2025 debate focused on training costs without looking at inference, and some later analyses suggest DeepSeek's popularity caused server saturation and a constant need for more infrastructure, which points to training savings shifting, at least in part, into higher usage in production.
All of this is about compute and energy consumption, Jevons's original territory. The question of whether AI will act as a replacement or a complement for human labor is the derivative question we saw with tellers, scribes, and accountants, and there the story hasn't been written yet, because it's happening right now. What we do know, from the historical cases, is that both things can coexist within the same technology, and that it's almost never purely one or the other.
What we don't know is which way the balance will tip this time. We don't know whether AI's falling cost will generate a rebound large enough for us to use it dramatically more, though the trend so far points that way, nor whether that greater use will translate into replacing tasks, expanding what people can do, or some mix we can't yet anticipate.