Hatim RihThe Jevons paradox has a second half
In 1865 Jevons argued that efficient engines would make Britain burn more coal, and that cheaper products would mean more work. He was right for about fifty years. What coal, light, farms, bank tellers and switchboards say about software jobs now that AI makes code cheap.
On 31 October I started a calculator for the taxes and fees you pay when you buy a home abroad. It covers Portugal, four regions of Spain and England, and every line of the bill names the law it comes from. It went live on 2 November as Buying Costs. In the same three days, a child support calculator for Texas, New York and Virginia went from a prototype to a live site, Coparent Math, with every step of each state’s formula written out next to its section of the statute.
Meanwhile, the agent that builds a project on my machine every night kept going. In its first three weeks, from 14 October to 3 November, it built 22 programs. Some are toys I love: a lambda calculus evaluator that draws every term as a diagram, and a QR code encoder that lets you scratch the code and watch the error correction heal it. Some are tiny tools for one problem: a bill splitter for housemates who moved in on different days, a tracker for subscriptions that renew in different months, a seating planner for weddings.
None of that would have been worth building at 2020 prices. A tax calculator for three countries was a team’s quarter, and a wedding seating planner was a spreadsheet and an argument. The price of software fell, and I didn’t save the difference. I spent it on a lot more software.
Economists have a name for that: the Jevons paradox. I mentioned it in passing in the compiler essay, as the reason I don’t expect AI to leave fewer software jobs. That was one paragraph doing a lot of work. So I went back to the book where it comes from, and it turns out the usual quote leaves out the part about jobs, and history adds a part Jevons didn’t see.
The coal question
In 1865 William Stanley Jevons published The Coal Question. Britain ran on coal, and Jevons wanted to know how long it could keep doing that. A common hope at the time was that more efficient engines would make the coal last longer. Jevons thought that hope had it backwards:
It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.
He even used the word himself: “Nor is it difficult to see how this paradox arises.” (If you look the passage up, most citations say Chapter VII. That’s the 1866 second edition. In the first edition it’s Chapter VI, which is the kind of thing I now know and can’t unknow.)
His mechanism is worth reading in full, because it’s the whole idea in one sentence. If a blast furnace burns less coal for each ton of iron, “the profits of the trade will increase, new capital will be attracted, the price of pig-iron will fall, but the demand for it increase; and, eventually, the greater number of furnaces will more than make up for the diminished consumption of each.”
He had the numbers to back it. In Scotland, the hot blast cut the coal used per ton of iron to less than a third, and total coal used in Scottish iron went up tenfold. Steam engines told the same story. Jevons printed a table of their “duty”, the water an engine could lift with a bushel of coal: 5.6 million pounds raised one foot for the old atmospheric engines of 1769, and 54 million for the average Cornish engine of 1859. Engines got about ten times better, and Britain mined more coal every year. Output was about 93 million tons in 1864. It peaked at 292 million tonnes in 1913.
The part nobody quotes
Here’s what surprised me. Two sentences after the famous line, Jevons makes the same argument about people:
The economy of labour effected by the introduction of new machinery, for the moment, throws labourers out of employment. But such is the increased demand for the cheapened products, that eventually the sphere of employment is greatly widened.
His example is the sewing machine. Seamstresses, he wrote, “have perhaps in no case been injured, but have often gained wages before unthought of”.
So the optimistic case for software jobs is 161 years old, and it was made by the man the paradox is named after. Make something cheaper, people want much more of it, and the people who make it end up busier. That’s exactly the claim I made about software, and he made it with sewing machines.
The optimists have light on their side, too. Roger Fouquet and Peter Pearson worked out the price and use of light in the United Kingdom over seven centuries. Between 1800 and 2000, light got more than 3,000 times cheaper. Use per person went up 6,500 times, and total use 25,000 times. We didn’t pocket the savings on candles. We lit everything.
When it doesn’t happen
Then there’s food.
In 1900, 41% of the American workforce worked in agriculture. In 2000 it was 1.9%. Farming got enormously more productive (an acre of wheat went from more than two weeks of work to about two hours), and food got cheaper. But nobody eats five dinners. A 2006 article from the Bureau of Labor Statistics put it in one line: “Sharply rising farm productivity, together with limited appetites for farm products, caused the decline.”
Economists measure this with the price elasticity of demand: the percentage by which purchases of something rise when its price falls by 1%. Above 1, a cheaper product means more total spending on it, and more work making it. Below 1, the savings leave the industry. A review of 160 studies put food between 0.27 and 0.81. Light, on Fouquet and Pearson’s numbers, behaved as if it sat above 1 for two centuries (incomes rose fifteenfold over the same years, which helped).
That’s the whole question for software jobs, and you can play with it:
A two-number model, not a measurement. If labour is the whole cost of software, a developer who works g times faster makes a feature cost 1/g as much. People then want g^e times as much software, where e is the elasticity, and the hours of work come to g^e / g. Nobody has measured e for software.
The research on energy is also more careful than the slogan. Steve Sorrell’s 2007 review for the UK Energy Research Centre found that for household heating, cooling and driving, the “rebound” (the share of the saving eaten by extra use) is likely under 30%. Backfire, where total use goes up, isn’t routine. It shows up for “pervasive general-purpose technologies”, adopted by businesses, early in their spread. Sorrell’s examples are steam engines and electric motors.
AI for writing software fits all three: general-purpose, used by companies, and early. If you were designing a case for the Jevons paradox, you’d draw something like it.
The second half
Now go back to the coal, because this is the part Jevons couldn’t see from 1865.
He forecast 466.9 million tons of coal for 1911. The real peak was 292 million tonnes in 1913, when 3,024 British mines employed 1,107,000 people. Then output fell for most of a century. In 2025, Britain’s deep-mined output rounded to zero.
The paradox held for coal for about fifty years after Jevons wrote it down, and then it stopped. The miners found out first.
Bank tellers are the version I keep coming back to, because it’s usually told as a happy story. ATMs arrived in the 1970s and spread fast from the mid-1990s, and teller jobs didn’t disappear. James Bessen worked out why: the number of tellers an average urban branch needed fell from 20 to 13 between 1988 and 2004, so a branch got cheaper to run, and banks opened 43% more branches in urban areas. The teller’s job changed from counting cash to talking to people. Jevons would have nodded.
In 2010 the Bureau of Labor Statistics projected that tellers would grow from 600,500 jobs in 2008 to 638,000 in 2018. By 2012 there were 545,300. In 2025 there were 339,200, and the bureau now projects a further 13% fall by 2035, “as more people use online banking tools, such as mobile check deposits”. The ATM made tellers cheaper. The phone in your pocket made the visit unnecessary.
Telephone operators had it worse, and faster. In the early 1900s, operating a switchboard was one of the most common jobs for American women, and the number working in the telephone industry peaked in 1930 at 180,000. Then AT&T replaced operators with mechanical switching in more than half of its network. When a city cut over, the number of young women working as operators fell by 50 to 80%, immediately and for good. People kept making phone calls. The dial just didn’t need an operator to connect them.
(One good note from that research: automation didn’t reduce how many of the next generation of young women had jobs, because clerical and service jobs grew to take them. The operators who were already in the job were the ones who lost.)
So here’s how I read the history now. The Jevons paradox is about the thing: coal burned, light used, calls made, software written. Whether it also holds for the people who make the thing depends on what the machine does to their job. A sewing machine made a seamstress faster. An ATM made a teller cheaper per customer. A dial telephone did the operator’s whole job. The first two kept people at work for decades. The third didn’t.
Where software is now
On the output side, software already looks like coal in 1865.
On the evening of Sunday 26 January 2025, Seattle time, after DeepSeek released a model that was cheap to run, Satya Nadella posted: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” The next day Nvidia lost about 17% of its value, close to $600 billion, as investors bet the other way: that cheaper models would need fewer chips.
So far Nadella looks right. In 2024, Andreessen Horowitz worked out that a model of a given ability was getting about ten times cheaper to run every year. Use grew faster still: Google went from processing 9.7 trillion tokens a month in May 2024 to more than 3.2 quadrillion in May 2026, about 330 times as much in two years. GitHub counted 121 million new repositories in a single year. Appfigures counted 24% more new apps submitted to Apple’s App Store in 2025 than in 2024, and in the first quarter of 2026 new app releases across both stores were up 60%.
(Google’s tokens include AI features it switched on for people who didn’t ask for them, so treat that 330 as an upper bound on demand. It’s still a big number.)
On the jobs side, the picture is cloudier. The Bureau of Labor Statistics projects software developer jobs ten years ahead, and each edition I could find has been lower than the last: 25% growth for 2021 to 2031, then 17%, then 15%, and now 10% for 2025 to 2035. For “computer programmers”, the narrower job of turning a design into code, it now projects a 7% decline, partly because companies will use AI “to automate repetitive programming tasks”. The young are already feeling it. Erik Brynjolfsson and his colleagues at Stanford find that employment of 22-to-25-year-olds in jobs exposed to AI is 19% below where it would be if it had kept pace, mostly because companies are hiring fewer of them.
Radiology is the hopeful case. In 2016 Geoffrey Hinton said that “people should stop training radiologists now”. Ten years later, American diagnostic radiology programs offered 1,083 advanced residency places in this year’s Match, up from 997 in 2022. Works in Progress, which looked into why, mostly found that AI hasn’t taken over the work yet, rather than proof that cheap reading created a flood of new scans. So it’s good news, with an asterisk.
Put together, it reads like this: the amount of software is going up much faster than the number of people paid to make it.
Three guesses
Here’s what I’d bet, and what would prove me wrong.
There will be far more software, of a kind nobody used to make. Software for one problem and one person, like the tax calculator and the seating planner. Most of it will be small, a lot of it will be thrown away, and the total will grow faster than anything we’ve measured. I’m wrong if app releases and new repositories flatten out while the tools keep getting cheaper.
The developer jobs that grow are the sewing-machine part. Deciding what should exist, writing the spec and the tests, checking what came back: that’s where Jevons can keep people busy, because every cheaper feature creates more of those decisions. The switchboard part is the narrow ticket with a clear answer, and that’s the part an agent can take whole. That’s why I expect the programmer line to keep falling while the developer line keeps rising, slowly. I’m wrong if the developer projection drops below zero in the next edition or two.
Software gets a second half too. I can’t tell you the year. Coal had about fifty good years after Jevons, and tellers had about thirty after the ATM. The demand for software can be bottomless and the work can still move somewhere else, the way phone calls kept growing after the operators left. The question to keep asking is whether the agents are making developers faster, or doing what developers do.
Back to October
My October says demand for software is elastic, at least mine. I wanted twenty-two programs nobody asked for and two calculators that read the law, and I got them because they suddenly cost evenings instead of quarters.
Jevons would have liked that. He’d also have pointed out, correctly, that it doesn’t tell you how many people get paid to build them in 2040. The ones still doing this work will be the ones the machine made faster, which is the seamstress’s side of the story.
Is your work getting more like the sewing machine or the switchboard? Tell me in the comments. I’d like to know which side of the dial you’re on.
A thought to add?
A good question, another perspective, or something this made you think about. There’s room for it here.
Loading the conversation…