The Jevons Paradox in The Age of AI

The faster AI makes you, the more work finds you.

There are a lot of new subscribers here so I want to say Hello and Welcome. I’m working on finishing my manuscript to send to my editor on Tuesday. I can’t wait to focus more time here. Until then, here’s something from my upcoming book. Leave a comment below and let me know what you think. And Thank you for being here. Share this with someone on your network.

Before AI, I had two unfinished projects. Now I have 57.

An economist named William Stanley Jevons noticed something paradoxical about coal-powered steam engines. As engineers made engines more efficient and able to produce the same power with less coal, Jevons expected coal consumption to decrease. The basic logic was that if you need less coal per unit of output, you’ll burn less coal overall.

The opposite happened.

Coal consumption exploded. Efficiency made coal cheaper, which made it economically viable in applications where it had previously been too expensive. Steam power expanded into new industries, new uses, new markets. Jevons wrote about this in 1865 in a book called The Coal Question, and the principle became known as the Jevons Paradox.¹

We are living inside it right now.

The promise of AI is simple and seductive. AI will handle the drudgery, freeing you up for the work that matters. The creative work. The strategic work. The human work. Leaders believed it. Vendors sold it. People went for it—from fourteen-year-old teenagers starting app-based businesses to people in corporate using it to make them more efficient. And to be fair, the efficiency gains are real. Measurably, replicably, impressively real. Projects that used to take me months now take me minutes.

When researchers gave 453 college-educated professionals writing tasks and randomly assigned half of them access to AI, the AI group finished 40 percent faster and produced work rated 18 percent higher quality.² When a Fortune 500 company rolled out an AI assistant to over 5,000 customer-support agents, they resolved 15 percent more issues per hour.³ New hires using AI reached the performance level of eight-month veterans in just two months.⁴ In a field experiment with 758 knowledge workers, those using AI completed 12 percent more tasks, 25 percent faster, and delivered substantially higher quality. (However, for more complex managerial tasks subjects using AI were 19% less likely to produce correct solutions compared with those without AI.)⁵

None of this is disputed. The efficiency is not a myth.

And that is exactly the problem.

Because Jevons didn’t say efficient engines don’t work. He said they work so well that they create more demand, not less. The better the engine, the more coal you burn.

The faster AI makes you, the more work finds you.

Paradox #1: More Productive, More Busy

There’s a well-known experiment—sometimes called the bottomless bowl study—in which diners ate from soup bowls that secretly refilled from the bottom. They consumed 73 percent more soup without realizing it. They didn’t feel fuller. They just kept eating because the bowl never emptied. Digital work environments carry the same logic. There is no natural stopping point. No signal that says enough. The AI keeps generating. The inbox keeps filling. The task list keeps expanding. And the worker keeps consuming their own time without ever feeling full.

Despite leaders believing AI would reduce workload, the actual pattern shows something darker. Instead of creating space for creativity or reduced hours, the efficiency gains are creating pressure for more output, especially for productive people. The time we “save” often gets filled with more.

More requests. More iterations. More channels. More meetings. More reading. More tweaking. More checking. More double-checking. More work, fewer people.

When technology saves time, people rarely get that time back.

In February 2026, UC Berkeley Haas published an eight-month ethnographic study of a 200-person technology company after AI tools were rolled out across the organization. The researchers expected to find reduced workloads. Instead, they found the opposite. Employees worked faster, took on a broader scope, and extended work into hours that used to be free — often without being asked.⁶ The scope of what counted as “my job” expanded. Work seeped into lunch breaks and evenings. Multiple AI processes ran simultaneously in the background while people were in meetings. What started as excitement became something harder to sustain.

The reason is straightforward: without deliberate structural choices to translate AI capability into reduced demands, the default path is that work expands to fill the capacity made available to perform it—a principle known to organizational researchers as Parkinson’s Law. That’s the Jevons Paradox in action: efficiency gains get reinvested in more work, not more life.

Leaders are the only ones who can interrupt this pattern.

History repeats itself when it comes to technology. The Financial Times‘ Tim Harford traced this pattern through every supposedly liberating technology. Email was faster than a letter—but it spawned, as he put it, “a profusion of low-quality, low-value messages bleeding into the evenings and weekends.” PowerPoint meant that “highly paid and skilled professionals started wasting time making their own slides badly.”⁷

Sociologist Judy Wajcman makes this point powerfully in her book Pressed for Time: The Acceleration of Life in Digital Capitalism.⁸ She studied how supposedly time-saving technologies actually reshape our work and found that they rarely save time in practice.

Despite decades of automation, knowledge workers now average 47 hours per week—up from 43 in the 1990s.⁹

Technology doesn’t give us time back. It raises the bar for how much we’re expected to do with it. Unless we make a conscious, deliberate choice to protect the time AI gives us back, and our brains, that’s exactly what will happen. Every tool that promised to save time created new categories of work that didn’t exist before. AI is doing the same thing, only faster and at a greater scale.

The CTO of Dun & Bradstreet told Fortune what this looks like in practice: “I got the eight hours to two hours, but now I can get 20 hours of work.” His teams aren’t going home early. They’re shipping product development cycles that once took two to three years in six months. The saved time frees up time for more projects. The CEO of KPMG didn’t even pretend the savings would go to workers. “That means I can put more volume through my business,” he said.¹⁰

Exactly.

Paradox #2: More Time, Less Leisure

Using nearly two decades of time-use data from the American Time Use Survey, researchers found that higher AI exposure is associated with longer work hours and reduced leisure time. Moving from the 25th to 75th percentile in AI exposure corresponds to an additional 2.2 hours of work per week. After large language models went mainstream, workers in the most AI-exposed occupations—computer systems analysts, credit counselors, logisticians—were working roughly 3.15 hours more per week than those in low-exposure jobs. Leisure time declined, particularly for non-screen activities like socializing and entertainment.¹¹ Read that again. When the hours got cut, we didn’t give up scrolling. We gave up each other. TV watching held steady. Going out with friends didn’t. The screen kept us. The people lost.

Two forces drive this.

The first is economic. When AI makes each hour of your work more productive, that hour becomes more valuable—to your employer and to you. The rational move, in cold economic terms, is to work more hours, not fewer. Workers in AI-exposed occupations did see wage increases. But they traded those gains for longer hours, not more freedom.

The second force is surveillance. AI doesn’t just do your work. It watches you do your work. Digital monitoring tools expanded rapidly during the shift to remote work, and they didn’t retreat when people returned to the office. Occupations with high AI-driven surveillance saw longer work hours regardless of where the work happened. Among the self-employed—people who set their own expectations—this effect disappeared entirely. The technology isn’t the constraint. The power dynamic is.¹²

The Adecco Group published a 2025 survey with 37,500 workers across 31 countries. They found workers believe they save two hours per day on average (up from one hour in 2024), but added the caveat that “evidence suggests that this is not reflected in productivity gains recorded by organizations.” ¹³ But where does that time go? Only about a quarter of workers used it for better work-life balance. The rest poured it back into output—more creative work, more strategic thinking, more of the same workload compressed into tighter deadlines. Twenty-three percent tackled the same workload. Twenty-one percent spent the time on personal activities. The efficiency gains were reabsorbed into the system like water into sand.¹⁴

Paradox #3: The Illusion of Productivity

A survey of 2,500 workers and executives found that 96 percent of C-suite leaders expected AI to boost productivity. Meanwhile, 77 percent of employees said AI had actually increased their workload. Not decreased it. Increased it. Thirty-nine percent spent more time reviewing and moderating what AI produced. Twenty-three percent were investing hours into learning the tools themselves. Twenty-one percent were simply told to do more because the tools made it possible. Seventy-one percent reported burnout.¹⁵

At the macroeconomic level, the productivity miracle is invisible. A survey of over 5,000 executives across four countries found that more than 80 percent of firms reported no measurable productivity gains or employment from AI over the prior three years.¹⁶

The engines are running. The coal is burning. And the mines are somehow not producing more heat. Or maybe they are—but it’s the workers who are absorbing it.

Paradox #4: Better Tools, More Stress

A meta-analysis from researchers from Auburn University, Old Dominion University, and the University of Illinois Urbana-Champaign, based on 515 studies spanning six decades, analyzing data from almost 800,000 workers, found that the real antidote to stress at work is clearer role definitions and responsibilities.¹⁷ But in the AI age where things change on a minute-by-minute basis, it’s getting more and more difficult to have any clarity as to what your job really is.

The leaders saw a faster engine. The workers felt the extra coal.

Paradox #5: AI Brain Fry — Too Many Tools, Too Little Thought

Researchers found that productivity improves when people use a few AI tools. But once you cross the threshold of four or more tools, something breaks. Workers in one study reported brain fog, slower decision-making, more small mistakes, and intent to quit—a phenomenon the researchers called “AI brain fry.”¹⁸

Researchers at Carnegie Mellon, MIT, Oxford, and UCLA gave two groups a set of problems: fraction math, SAT reading comprehension. One group had AI assistance. One didn’t. Then the researchers pulled the AI mid-test without warning. The assisted group’s solve rate dropped 20% below people who’d worked alone the whole time. Ten minutes was enough to erode their own ability to think.¹⁹

Half of workers are now using AI, with 62% of Gen Z saying they are using it as a crutch that could cost them their entire careers - because of everything they are losing while depending on AI—cognitive ability, critical thinking, decision-making, human skills.²⁰

A separate Microsoft study found the same pattern: the more they trusted AI to do the task, the less critical thinking they applied, and the harder those skills became to access when needed.²¹

We need to learn better human-AI collaboration skills without losing the human skills that give us an advantage. The question isn’t whether AI is powerful. It is. The question is what it’s doing to the one thing machines still can’t replicate: you.

It’s the cognitive equivalent of Jevons’ coal mines running at full capacity. The engines are efficient. The system is overheating. The miners are burning out.

The Other Side

Let me be honest about the upside: it’s real, and ignoring it would be dishonest. AI closes the gap between novice and expert faster than any tool in history. It gives a new hire the output quality of a veteran. It gives a struggling writer a clean first draft. It gives a small business the analytical capability that used to require a consultant. The lower you start, the more it lifts you. That’s not nothing. That’s meaningful. For individuals, in specific moments, on specific tasks, AI is genuinely transformative.

But the Jevons Paradox was never about whether the engine worked. It was about what happens to the system when the engine works too well. Individual gains get absorbed by institutional appetites. And the humans inside the system work harder, longer, and more anxiously than they did before the miracle arrived.

Jevons watched engineers build better steam engines and expected the coal mines to empty. Instead, they dug deeper. We built tools that could do our work in half the time, and then we found twice the work to do. So here’s the question the Jevons Paradox actually asks — the one Jevons himself never quite answered:

If efficiency always creates more demand, is that bad?

It depends entirely on what the demand is for.

When coal got more efficient, yes, consumption exploded. But that explosion powered the Industrial Revolution. It heated homes that had been cold. It moved goods that had been stuck. It built cities, railroads, economies. The problem was never that we used more coal. The problem was that we used more coal without thinking about what we were building with it. We got the factories but poisoned the rivers. We got the railroads but broke the workers. The efficiency wasn’t the villain. The lack of intentionality was.

AI is coal. The question is whether we’re building factories or cathedrals.

Right now, most organizations are building factories. They’re taking the time AI saves and feeding it back into the machine — more output, more volume, more throughput, same vision. That’s Jevons at his most predictable. A CTO compresses eight hours into two and then assigns twenty hours of work. A CEO sees efficiency and reaches for volume. The coal burns hotter and the mine gets deeper and nobody asks what are we digging toward?

That’s not a technology problem. That’s a leadership problem.

My friend Christopher Lochhead put it bluntly in a social media post: only the tech industry could take the single greatest innovation in human history and market it so badly that people want to protest it in the streets.²² He’s right. When executives brag about headcount reduction instead of human possibility, when the dominant narrative is AI will take your job instead of AI will expand what your job can become, the backlash is earned. The Luddites didn’t smash machines because they hated progress but because the people who owned them offered nothing in return.

And the data supports the backlash—to a point. Pew Research found that as of late 2025, half of Americans were more concerned than excited about AI in daily life, up from 38% just two years earlier.²³ When commencement speakers get booed for mentioning AI and communities revolt against data centers, that’s not ignorance. That’s people correctly sensing that the efficiency gains are flowing in one direction.

But here’s the other side — the side that doesn’t get page-one headlines.

Books and music have exponentially increased since the release of ChatGPT, together with scientific papers and self-filed lawsuits. Maybe some of them are AI slop. Maybe some aren’t. ²⁴

DeepMind’s AlphaFold predicted the structure of virtually every known protein — over 200 million — a problem that had consumed biology for fifty years. ²⁵Google DeepMind’s AI-driven weather models now outperform traditional numerical weather prediction at a fraction of the cost.²⁶ AI-powered diagnostics are detecting diabetic retinopathy, tuberculosis, and certain cancers with accuracy that matches or exceeds specialists—often in settings where specialists don’t exist.²⁷ Drug discovery timelines that once stretched a decade are compressing to months. Insilico Medicine used AI to identify a novel drug target and generate a candidate molecule for pulmonary fibrosis in a fraction of the traditional timeline — and moved it to Phase II clinical trials. No AI-discovered drug has yet received full FDA approval, but over 100 AI-generated candidates are now in clinical trials.²⁸

Global GDP will be 14% higher in 2030 as a result of AI - the equivalent of $15.7 trillion, more than the current output of China and India combined according to PwC research.²⁹

The Jevons Paradox is real. Efficiency will always create more demand. The coal will always burn hotter. But Jevons missed something. Coal was just energy. It didn’t care what you built with it.

AI is different. AI is a thinking tool. It doesn’t just make you faster — it makes you capable of things you couldn’t do before. But the caveat is that you still have to do the thinking.

The question isn’t whether we’ll use more of it. Of course we will. The question is whether we’ll use it to do more of the same — more emails, more reports, more meetings, more checking, more double-checking — or whether we’ll use it to do something we haven’t imagined yet and create different jobs, categories, and innovations that don’t even exist yet.

Every transformative technology in history faced this fork. Electricity could have been used solely to make candle factories run longer shifts. Instead, someone invented the refrigerator, the radio, the assembly line, the hospital ventilator. The Internet could have been used solely to send faster memos. Instead, someone invented e-commerce, streaming, search engines, the mobile app economy. The technology didn’t decide. People did. Leaders did. Meaning makers did. Creative thinkers did.

The historical pattern is consistent: general-purpose technologies take 20 to 30 years to produce their full economic impact. Electricity was commercialized in the 1880s but didn’t transform manufacturing productivity until the 1920s, when factory layouts were redesigned around electric motors instead of steam shafts. The modern computer was introduced in the 1970s and 1980s, but economy-wide productivity didn’t surge until the late 1990s, after businesses reorganized workflows around networked computing. We are almost certainly in that same lag period with AI — the technology works, but the institutions haven’t restructured yet and we really don’t know how to use it well. But since everything is going faster these days, we probably have less than twenty years to figure this out.

That adaptation is a choice. And right now, most organizations are choosing the lazy version: same work, faster. That’s Jevons winning. That’s coal burning hotter in the same furnace.

But some are getting this right.

In a study tracking AI adoption across firms from 2010 to 2023, MIT Sloan researchers found that companies using AI extensively grew faster — with roughly 6% higher employment growth and 9.5% more sales growth over five years compared to firms with less AI use. Losses in highly exposed roles were largely offset by gains in other jobs and by overall hiring growth at firms that became more productive through AI by giving people back time back for “critical thinking or coming up with new ideas.”³⁰ That requires more unhustling, not less.

The desire path is harder and less obvious. It means redesigning workflows instead of accelerating them. It means redesigning work completely instead of just doing more of the same. It means asking what should we stop doing? instead of how can we do more? It means — and this is the part that requires actual leadership — giving some of the efficiency back. Not all of it. But some.

Enough to let people think. Enough to let people rest. Enough to let people do the creative, strategic, deeply human work that AI can’t do and that never gets done when every freed hour gets immediately reabsorbed into the system.

Jevons was right that efficiency increases demand. But demand for what is a human decision. The coal is burning. It’s going to keep burning. The only question is whether we build another factory or something the world hasn’t seen yet.

Your assignment today: Watch the documentary “The AI Doc: Or How I Became An Apocaloptimist,” where Sam Altman, the CEO of OpenAI, himself says: “I can’t promise you that this is going to go well.”

I’m not a pessimist, and I see a lot of great things coming from AI - in education, healthcare, and the workplace if we put the right guardrails in place and slow down enough to proceed safely.

My hope is that we do the right thing in time.

To learn more about what I do, invite me to speak, apply for one of my programs, or get 1:on:1 support, visit unhustle.com.

1. Jevons, W. S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. Macmillan.

2. https://www.science.org/doi/10.1126/science.adh2586

3. https://www.gsb.stanford.edu/faculty-research/publications/generative-ai-work

4. https://mitsloan.mit.edu/ideas-made-to-matter/workers-less-experience-gain-most-generative-ai

5. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838

6. Ye, X.M. & Ranganathan, A. (2026). AI Doesn't Reduce Work—It Intensifies It. Harvard Business Review. Published February 2026. newsroom.haas.berkeley.edu/ai-promised-to-free-up-workers-time-uc-berkeley-haas-researchers-found-the-opposite

7. https://timharford.com/2026/04/what-if-ai-just-makes-us-work-harder/

8. Wajcman, J. (2015). Pressed for Time: The Acceleration of Life in Digital Capitalism. University of Chicago Press.

9. https://news.gallup.com/poll/175286/hour-workweek-actually-longer-seven-hours.aspx

10. “AI just gave you six extra hours. Your boss already took them.” Fortune, March 10, 2026. Includes interviews with Mike Manos (CTO, Dun & Bradstreet) and Tim Walsh (Chair and CEO, KPMG U.S.). The Walsh quote also appears in the companion piece: “CEOs are using one number in the AI age to decide how many workers they need,” Fortune, March 10, 2026.

11. Jiang, W., Park, J., Xiao, R. & Zhang, S. "AI and the Extended Workday: Productivity, Contracting Efficiency, and Distribution of Rents." NBER Working Paper No. 33536 (February 2025). Also available on SSRN (abstract ID: 5119118). Summary column: "As AI's power grows, so does our workday," CEPR VoxEU, March 28, 2025.

12. https://www.nber.org/papers/w33536

13. https://www.unleash.ai/future-of-work/news/one-in-three-workers-future-ready-as-comfort-with-ai-based-work-grows-adecco-group

14. https://www.prnewswire.com/news-releases/ai-saves-workers-an-average-of-one-hour-each-day-302278019.html

15. https://investors.upwork.com/news-releases/news-release-details/upwork-study-finds-employee-workloads-rising-despite-increased-c

16. https://cep.lse.ac.uk/_NEW/PUBLICATIONS/abstract.asp?index=12096

17. https://www.sciencedirect.com/science/article/abs/pii/S000187912600028X

18. https://www.bcg.com/news/5march2026-when-using-ai-leads-brain-fry

19. https://gizmodo.com/spending-just-10-minutes-with-ai-can-fry-your-brain-researchers-find-2000755701

20. https://fortune.com/2026/05/19/gen-z-workers-over-reliance-ai-early-career-jobs/

21. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

22. https://www.technologyreview.com/2026/03/02/1133814/i-checked-out-londons-biggest-ever-anti-ai-protest/

23. https://www.pewresearch.org/short-reads/2026/03/12/key-findings-about-how-americans-view-artificial-intelligence/

24. https://www.washingtonpost.com/technology/2026/05/20/data-shows-that-ai-slop-is-taking-over-books-lawsuits-music-science/

25. https://www.engineering.org.cn/engi/EN/10.1016/j.eng.2024.12.003

26. https://www.science.org/doi/10.1126/science.adi2336

27. For diabetic retinopathy: Gulshan, V. et al. “Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy.” JAMA, Vol. 316, No. 22 (2016), pp. 2402–2410. For tuberculosis: Qin, Z.Z. et al. “Using artificial intelligence to read chest radiographs for tuberculosis detection.” Scientific Reports, Vol. 9, Article 15000 (2019); see also World Health Organization, Operational Handbook on Tuberculosis: Module 2 – Screening (2024 update), which formally recommends AI-assisted TB screening in low-resource settings. For cancers: Liu, X. et al. “A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging.” The Lancet Digital Health, Vol. 1, No. 6 (2019), pp. e271–e297. For a general overview: Topol, E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again (Basic Books, 2019)

28. https://pubs.rsc.org/en/content/articlelanding/2023/sc/d2sc05709c

29. http://preview.thenewsmarket.com/Previews/PWC/DocumentAssets/476830.pdf

30. https://mitsloan.mit.edu/ideas-made-to-matter/how-artificial-intelligence-impacts-us-labor-market