The Jevons Paradox and What It Actually Means for Your Career

Career Landscape Reports: 1 of 5 in a series

In 1865, an English economist named William Stanley Jevons noticed something that didn’t fit the intuition of the time. Steam engines were getting more fuel-efficient, and the assumption was that Britain would start burning less coal. Instead, coal consumption climbed for decades. Cheaper, more efficient engines didn’t reduce demand for coal. They expanded it, by making steam power viable for factories, ships, and industries that couldn’t have justified the expense before.

This is now known as the Jevons Paradox. Economists use the broader term rebound effect for what happens when greater efficiency lowers the effective cost of using something and, as a result, people use more of it. Jevons Paradox describes the strongest version of that effect: demand rises enough that total consumption of the resource increases rather than falls. In other words, making something dramatically more efficient can expand its use by making it economical in situations where it previously was not.

Over the past two years, this 160-year-old idea has become one of the most cited frameworks in the debate over AI and jobs. Apollo Global Management’s chief economist (Torsten Slok) has built what he calls a “Jevons employment effect” thesis around it. Anthropic’s CEO (Dario Amodei) has invoked it while walking back his own more dire predictions about AI-driven unemployment. It shows up constantly in arguments that AI won’t gut the labor market the way many fear.

The idea is genuinely useful. It’s also being used, in a lot of the commentary out there, to tell people something more comforting than the evidence actually supports. If you’re managing a career through this period, it’s worth understanding both what the Jevons Paradox explains and what it doesn’t.

The Pattern Shows Up Clearly in Software

The clearest modern case study is software development, and the data is worth sitting with.

The U.S. Bureau of Labor Statistics tracks two separate occupations that most people lump together. “Computer programmers” translate specifications into code: a narrower, more mechanical task. “Software developers” do something broader: design systems, gather requirements, make architectural calls, coordinate across teams. Programming is part of that job, not the whole of it.

Computer programmer employment has continued a long decline and is now dramatically smaller than it was historically. Software developer employment, over the same period, has barely moved, and the BLS projects it to grow nearly 18% over the next decade, about five times the rate of the average occupation.

So far, AI has not produced the road collapse in software development employment that the collapse in coding costs might have suggested. It collapsed the narrow, codified task and expanded the broader role built around it. That’s Jevons, applied to knowledge work: the cost of producing code dropped, so the volume of software companies decided to attempt went up, and it took more people, doing more ambitious work, to build it.

You can see a version of the same pattern outside software. Spreadsheet automation was supposed to be the end of the accountant in the 1980s. Instead, the number of accountants and auditors in the U.S. grew by roughly 30% between 1990 and 2020, while the profession moved out of manual arithmetic and into analysis, advisory work, and compliance that had always existed but had never been cheap enough to do at scale. In 2016, Geoffrey Hinton told the world to stop training radiologists because deep learning would outperform them within five years. A decade later, radiology residency positions have grown, not shrunk, because AI-assisted screening freed radiologists for more complex diagnostic and interventional work, and total imaging volume rose right along with it.

Even in healthcare documentation, where AI’s ability to draft clinical notes is genuinely strong, the pattern holds in a modified form. One 2026 analysis of 515 medical scribe postings found a drop of 20% in 2025 as ambient AI scribes took over the core documentation task. But roughly a third of the remaining postings now bundle documentation with broader clinical duties, credentialed roles that recruit from a different, more skilled labor pool than the pre-med students who used to fill standalone scribe jobs. The narrow role contracted. The surrounding role expanded and got more demanding.

Where the Comforting Version of the Story Breaks Down

Here’s where a lot of the AI commentary gets ahead of the evidence.

The Jevons Paradox describes what happens in aggregate, over time, once a market fully absorbs a cost shock. It says nothing about who gets caught in between, how long the gap lasts, or whether the person who lost the narrow job is the same person who gets hired into the expanded one.

Look at what actually happened at some of the companies that moved fastest on AI substitution. Klarna announced in 2023 that its AI assistant was doing the work of roughly 700 customer service agents. By 2025, its CEO admitted the company had overcorrected, telling Bloomberg the strategy had produced lower quality that wasn’t sustainable, and Klarna began rehiring humans into a hybrid model. IBM used AI to absorb hundreds of HR roles, then quietly rehired into the judgment-heavy parts of HR that its automation couldn’t handle. Duolingo announced an “AI first” plan to phase out contractors, faced a user backlash within days, and walked the framing back within months. Industry analysts tracking this cycle have started calling it “automated regret”: companies cut faster than the technology could actually support, then buy back the capability at a premium once the gaps show up. Recent reporting citing Hays and Robert Half says about 32% of companies that eliminated positions because of AI in Australia later reinstated them (Courier-Mail, August 27, 2026).

At the same time, a separate and less reversible dynamic is playing out. In 2026, Meta cut roughly 8,000 jobs and cancelled thousands of open positions. That’s not AI performing the work of laid-off employees. It’s AI capital expenditure crowding out headcount, a different mechanism with the same visible result. That suggests a different labor-market mechanism: AI doesn’t have to perform the work of a dismissed employee to affect employment. Capital committed to AI infrastructure can also compete with labor for corporate investment.

And underneath both of these stories sits a distribution problem that aggregate statistics tend to hide. Burning Glass Institute data shows that between 2018 and 2024, the share of jobs requiring three years of experience or less fell sharply across AI-exposed fields: from 43% to 28% in software development, from 35% to 22% in data analysis, from 41% to 26% in consulting. The traditional apprenticeship ramp, junior people doing routine work while they build judgment, is narrowing precisely because that routine work is what AI now does. Aggregate labor demand in a field can grow at the same time entry into that field gets harder. A profession with more partners and fewer first-year associates is not obviously good news if you’re the one trying to become a lawyer.

What This Means for How You Should Actually Think About Your Career

The honest version of the Jevons Paradox is not “AI will create more jobs than it destroys, so don’t worry.” It’s something narrower and more useful: efficiency gains expand the total volume of viable work in a domain, but they do not guarantee that any particular person, role, or career stage rides that expansion. Whether you benefit depends on where your work currently sits.

A few things are worth watching in your own situation, regardless of your field.

Notice which part of your job is the narrow, codified task and which part is the broader judgment around it. The programmer/developer split, the scribe/medical assistant split, the associate/partner split: in every version of this pattern, there’s a routine core that gets absorbed and a surrounding layer of design, coordination, oversight, and relationship work that expands. Your job almost certainly contains both. The question worth asking regularly is which one is growing as a share of what you actually do day to day, and which one is shrinking.

Don’t treat “the aggregate numbers look fine” as a guarantee about your specific position. BLS projections for software developers, accountants, or radiologists as categories tell you about the field, not about you. The people getting hurt in the transition are often the people at the entry-level end of the pipeline, or the people whose entire job was the narrow task rather than the surrounding one. If your role is closer to the narrow end, that’s information worth acting on before the market forces the issue.

Expect volatility, including reversals, rather than a smooth trend line. The Klarna and IBM stories are a useful reminder that companies routinely overestimate what AI can do, cut too fast, and correct course. That volatility can work in your favor if you stay in the field and adaptable through the correction, but it also means the ground under any given role can shift more than once, in either direction, over a fairly short period.

Treat this as an ongoing thing to monitor, not a one-time judgment you make about your job and then forget. Today’s expanded, judgment-heavy role is not guaranteed to stay that way. The physical, regulatory, or relational aspects of work that currently protect a role from full automation are themselves subject to further technological change. What looks like a durable position in 2026 is worth re-checking again in a year, not because you did anything wrong, but because the boundary between “task AI does” and “task a person does” keeps moving.

None of this means panic, and it doesn’t mean chasing every new AI tool to stay ahead of a wave that may not be coming for you specifically. It means paying honest, ongoing attention to where the actual line is falling in your own field, and building the parts of your capability that sit on the expanding side of it rather than assuming the aggregate story protects you personally.

I use AI tools to help research sources and draft these posts. I review, edit, and take responsibility for everything published here, including any errors.

---

Sources

This piece drew on the following articles, in addition to background research on the original 1865 Jevons Paradox and BLS occupational data:

• Myoung Cha, “AI Jevons Paradox: Why AI May Create More Work, Not Less,” Be Curious, Not Judgmental, February 2026. Original analysis of BLS programmer/developer data and a 515-posting study of medical scribe job listings.
• George Hoyem, “AI, Jobs, and the Jevons Paradox: Why the Void Rarely Stays Empty,” George Hoyem (Substack), February 2026. A general framing of the paradox applied to historical technology shocks and current AI adoption.
• Eduard Lupacescu, “Jevons Paradox AI: Why Companies Are Hiring More Developers, Not Fewer,” BinarCode, March 2026. Industry-side data on developer hiring and productivity tooling.
• OneResume Team, “AI Is Not Replacing Your Job: How Jevons Paradox Proves Technology Creates More Work, Not Less,” OneResume.ai, April 2026. Historical case studies on spreadsheet automation and accounting employment, and on the 2016 Hinton radiology prediction.
• James Robbins, “The Efficiency Trap: AI, the Jevons Paradox, and the Future of the Human Workforce,” Employer Branding News, May 2026. Detailed reporting on the Klarna, IBM, Duolingo, and Meta layoff/rehire cycles, and Burning Glass Institute data on the narrowing entry-level hiring ramp.
• “A 160-year-old paradox explains why AI will create more lawyers and accountants,” Fortune, April 2026, on Apollo economist Torsten Slok’s “Jevons employment effect” thesis.