Jevons Paradox: Why AI Will Increase Human Jobs

When I was writing this article (May 2026), a sense of apocalypse was hanging over Silicon Valley. Although AI is driving a hot economy and the streets are filled with singing and dancing, many insightful minds believe this is all temporary… because they believe AI will quickly eliminate a massive number of human jobs.

Anthropic CEO Dario Amodei recently sounded like a doomsday prophet. He stated on several occasions [1] that AI will eliminate half of junior white-collar jobs within one to five years, pushing the unemployment rate to 10%–20%; tech, finance, law, consulting, especially entry-level roles, will all be affected. Amodei even appealed like a champion for the people: AI companies and governments must stop sugarcoating the peace!

And you might have some personal experience with this. Over the past two years, several big tech companies in Silicon Valley have laid off workers in the name of AI. Some computer science graduates from prestigious American universities couldn’t find a job. Some believe that junior software engineer positions have completely vanished…

But in this talk, I want to present a bold, counterintuitive claim: AI will not only not reduce jobs, but will greatly increase human employment opportunities.

In fact, early signals have already appeared. Take a look at the chart below, which comes from a report published by Citadel Securities in February 2026 [2] —

The chart compares the trends of software engineers and overall job postings using Indeed job posting data. The most prominent feature is that software engineer job postings already bottomed out and rebounded in May and June 2025.

Of course, the number of new job postings is still lower than the pre-pandemic levels and the tech bubble peak of 2021–2022. However, this is a very good sign. Wasn’t it said that programming is the first job AI would eliminate? Why did software engineer job postings rebound first?

This is likely not an anomaly, but a perfect demonstration of an economic law called the ‘Jevons Paradox’.

What Is the Jevons Paradox? #

To understand the logic, let’s go back to 1865 in the United Kingdom.

The UK was at the peak of the Industrial Revolution, but people at the time also suffered from a kind of doomsday anxiety: coal was the heartbeat of the empire, fueling steam engines, railways, factories, and steamships. What if we ran out of coal? Geniuses of all stripes found ways to improve technology, multiplying the efficiency of steam engines and making them much more fuel-efficient than before. People assumed that total coal consumption would definitely go down.

But right at that moment, a 30-year-old economist named William Stanley Jevons published a book titled The Coal Question [3], saying, ‘You are thinking too simply.’

Jevons’ judgment was: the more fuel-efficient steam engines become, the cheaper steam power becomes → the cheaper steam power becomes, the more industries will adopt it → the more industries adopt it, the higher the total coal consumption in Britain will rise.

This is the Jevons Paradox. And history proved him exactly right.

Moreover, this is a universal law. Let’s look at a similar event that occurred in Xinjiang, China.

As you may know, many parts of Xinjiang are extremely arid. In the 1990s, agricultural ‘drip irrigation’ technology was introduced in the Tianshan region. Watering a mu of land used to require a lot of water, but drip irrigation saved a huge amount. So, should the total water usage have decreased? Not at all.

Farmers saw that since water consumption per mu had decreased, why not plant more? And since drip irrigation increased crop yields, could they plant more profitable cash crops? Since revenues went up, shouldn’t they reclaim more farmland?

The result: since drip irrigation was introduced in the Tianshan region in 1996, water usage rebounded by over 115% in 20 years [4]. In other words, total water consumption is now more than double what it was before.

Isn’t the current AI electricity consumption following the same logic? AI algorithms are becoming more efficient, Nvidia’s chips are getting more power-efficient generation by generation, and single-inference costs are plummeting — yet the total power consumption of data centers is skyrocketing.

These stories should all have happy endings, as humanity is not easily constrained by resources. But the lesson is: efficiency is not the brake; it is the accelerator.

Efficiency Is the Accelerator, Not the Brake #

Once you look at the world through the lens of the Jevons Paradox, you’ll see its shadow everywhere in daily life.

Take writing, for instance. Ten years ago, without AI, I researched on my own, sometimes even reading physical books, drew outlines with a pencil, typed everything by hand, and stared at the screen to edit line by line… If I maintained strict work discipline that day, it would take me about 6 to 7 hours to write an article. Now, with AI, I can use it for research, brainstorm with it, use voice input, and have it proofread. Every step saves time. Can I write an article in 4 hours?

The reality is that the time to produce an article hasn’t shortened at all; in fact, it often takes longer. This is because AI has raised my standards for the article: I explore more ideas, research more literature, incorporate richer case studies, increase the logical density of the text, write longer pieces, become more meticulous about word choices, and occasionally add illustrations and cartoons.

If you have used AI for your own work, you probably feel the same way. Bosses won’t say, ‘Great! Since AI helped you speed up, you can finish early and go home!’ Instead, they will say, ‘Since it only takes ten minutes to draft a version, let’s make three versions first…’ You’ll get bombarded with one requirement after another, one idea after another, until your work time is filled to the brim and beyond.

Isn’t the same true for everything else? You use a cashback app that saves you money on every purchase… but because the purchasing threshold is lowered, you want to buy more and more, and your total spending ends up increasing significantly.

Or consider applications like WeChat, Lark, and AI meeting summaries. The cost of communication between people has dropped dramatically. So, should our meeting times decrease? Not at all. With groups created instantly, constant synchronization, and alignment meetings, the threshold for starting a meeting is close to zero, leading to far more meetings than before.

The insight here is: when you increase efficiency and reduce the per-unit resource consumption, you simultaneously lower the barrier to action. Once the barrier drops, previously suppressed or even unimagined demands will unleash like a flood.

How Automation Creates More Jobs #

The Jevons Paradox originally stated that higher efficiency consumes more resources. But if you replace ’efficiency’ with ‘automation’ and ‘resources’ with ‘human labor,’ the paradox becomes: automation increases human work.

History has validated this law time and again —

When textile machinery first appeared in the 19th century, the so-called ‘Luddites’ feared that weavers would lose their jobs and went around smashing machines. Little did they expect that the plummeting price of fabric would result in ordinary people going from owning only one set of clothes to owning dozens, and the number of weavers actually increased dramatically.

In the 1970s, Automated Teller Machines (ATMs) became widespread, and everyone predicted that bank tellers would disappear. In reality, while the number of tellers required per branch did decrease, the reduced operating costs prompted banks to aggressively open new branches on every street corner. As a result, the total number of tellers actually rose. More importantly, tellers were no longer just ‘money-counting machines’; they were freed up to handle more complex and higher-value tasks, such as account openings and financial advising.

Fast forward to a few years ago when computer vision first became popular, and scholars asserted that radiologists would soon be out of a job. The subsequent reality was that because AI slashed the cost of reading scans, hospitals began ordering massive amounts of preventive MRIs and CT scans, leading to a surge in scanning volume. AI helped doctors screen out 90% of normal scans, while the remaining 10% of complex cases and the final sign-off responsibility still required humans. Consequently, radiologists worldwide are not unemployed but in severe shortage [5].

Similarly, spreadsheets did not eliminate accountants; instead, they expanded businesses like financial analysis, budget management, and business modeling. Search engines did not eliminate researchers; they birthed new roles such as SEO, content operations, data analysis, and digital marketing… and the list goes on.

A recent example comes from China. There is an art outsourcing platform called ‘Mihuashi.’ Its operating model is that clients post orders, such as ‘I want an avatar’ or ‘I want a certain anime character,’ and artists on the platform bid for the jobs. Now that anyone can draw with AI, does this business still exist?

Quite the contrary. In October 2022, due to a famous leak of an AI drawing model, near-free AI illustration capabilities suddenly became accessible to the general public. This caused the average price of a single image on Mihuashi to drop by 64% — yet the number of orders skyrocketed by 121%, resulting in a 56% increase in total revenue [6]. Existing creators were not squeezed out; they retained most of the market share, while the growth primarily came from low-end personal orders that were previously deemed ’not worth doing.’

AI Eliminates Tasks, Not Jobs #

This law is not just about ’lower price → more orders → larger overall market.’ It also has another key characteristic: ’task transformation.’

A job is not a single action; a job is a bundle of tasks. An accountant is not a ’number inputter,’ a doctor is not a ’lab report reader,’ a lawyer is not a ‘statute searcher,’ and a programmer is not a ‘code typer’… These jobs all contain judgment, aesthetics, responsibility, and other tasks that cannot be replaced by AI, and they can also encompass various new tasks generated because of AI.

AI eliminates tasks, not jobs. Labor economics has long had a ’task-based model’ [7] which states that while automation indeed has a ‘displacement effect’ — moving certain tasks from humans to machines — new technologies also create new tasks, allowing labor to re-enter the production process. This is called the ‘reinstatement effect.’

In 2025, the World Economic Forum released a jobs report [8] predicting that by 2030, global macro trends are expected to create 170 million new jobs and replace 92 million old jobs, resulting in a net increase of 78 million jobs. The US Bureau of Labor Statistics also explicitly noted [9] that AI may lower the cost of software products, thereby increasing the demand for software development, AI business solutions, and AI system maintenance. Consequently, they project a 17.9% employment growth for US software developers from 2023 to 2033.

Of course, we still have to wait and see. But looking at it from this moment, AI is indeed changing the world and may bring about a singularity — yet economic theory and historical experience do not support the doomsday hypothesis of ‘AI causing massive unemployment.’

What New Jobs Will the AI Era Create? #

If we are to align with the historical tide of the Jevons Paradox, we must not stand on the side of old tasks, but on the side of new demands. Let’s boldly imagine what new jobs AI will create.

Historical experience can still guide us. Simply put, when new technology lowers the barrier, the tasks that were previously impossible, unaffordable, or not worth doing become new business opportunities.

The first category consists of jobs directly related to AI —

• For example, AI Workflow Architect: Instead of just writing prompts, this role redesigns a company’s sales, customer service, data, legal, and financial processes into systems that are executable, auditable, and accountable by AI.
• For example, Agent Supervisor: Managing a group of AI agents, letting them divide work, collaborate, upgrade, and review, much like managing human interns in the past.
• For example, Model Evaluator & Red Teamer: Specializing in finding hallucinations, biases, privilege escalations, vulnerabilities, and dangerous behaviors in AI. The more models there are in the future, the more valuable the people who accept and audit models will become.
• For example, Knowledgebase Gardener: Maintaining internal corporate data, permissions, semantic structures, versions, and sources. AI feeds on data, so data needs chefs as well as food safety inspectors.
• For example, Robot Fleet Manager: As cleaning, delivery, inspection, and care robots enter cities, someone must be responsible for scheduling, maintenance, anomaly handling, and human-robot conflict resolution.

The second category is life-enhancement jobs —

• For example, Personal Learning Director: Not a traditional tutor, but someone who uses AI to customize long-term learning paths, daily feedback, error tracking, and project challenges for a student.
• For example, Eldercare Life Enhancer: Using AI, sensors, and robots to help the elderly manage medication, exercise, socialization, family contacts, and emergency response. This does not replace family affection but reduces the chaos surrounding it.
• For example, Micro-Experience Planner: Generating scripts, music, visuals, routes, and interactive games for a family gathering, a birthday party, a road trip, or a community festival. In the past, only large events were worth planning; in the future, small-scale life moments will also deserve customization.
• For example, Personal Digital Archivist: Helping individuals organize photos, videos, chat histories, articles, and family stories into searchable, inheritable, and presentable digital life archives.
• For example, Solo Film Producer: Ordinary people will not only be able to make films, but can also be the subjects of films. A single person, working with AI, manages storyboards, characters, voiceovers, editing, and special effects to produce a movie directly for a small group, a family, or an educational scenario.

The third category consists of trust & responsibility jobs. These tasks have always been necessary but are now becoming so critical that they deserve to be upgraded into dedicated positions —

• For example, AI Output Auditor: Specifically auditing legal documents, medical advice, financial reports, and scientific abstracts written by AI to find errors, identify discrepancies, and allocate risk.
• For example, Human Responsibility Signer: In high-risk scenarios such as healthcare, finance, law, and education, AI can provide recommendations, but a knowledgeable human must ultimately sign off, explain, and take responsibility.
• For example, Chief Aesthetics Officer: AI can generate ten thousand images, but selecting the one that has the most soul and touches the human heart can only be done by you.

When capabilities become cheap, demands become complex. When tasks are automated, responsibilities become personalized.

The Only Limit Is Our Imagination #

Both the Jevons Paradox and Baumol’s cost disease tell us that modernization is good news. Improving efficiency is good news; others improving efficiency when you haven’t is also good news for you. People always have concerns, but even the things you worry about will ultimately turn out to be good news.

The deeper insight of the Jevons Paradox is that human society is not a resource-saving machine — human society is more like a desire engine. When technology lowers a barrier, humanity doesn’t say ’that’s enough’; humanity says ’then we also want this, that, and that other thing.’

During the 2026 GTC conference, Nvidia CEO Jensen Huang said in an interview [10]: ‘Companies that lay off workers for AI are out of imagination. Truly imaginative companies should use AI to expand, not contract (do more with more).’

Jevons would fully agree with him. AI is a liberation of humanity, not a limitation. Only our imagination is our limit.

Notes #

[1] VandeHei, Jim, and Mike Allen. “AI Jobs Danger: Sleepwalking into a White-Collar Bloodbath.” Axios, May 28, 2025.

[2] Frank Flight, “The 2026 Global Intelligence Crisis,” Citadel Securities, February 24, 2026. https://www.citadelsecurities.com/news-and-insights/2026-global-intelligence-crisis/

[3] Jevons, William Stanley. The Coal Question. London: Macmillan, 1865.

[4] Wang, Yanyun, et al. “The Verification of Jevons’ Paradox of Agricultural Water Conservation in Tianshan District of China Based on Water Footprint.” Agricultural Water Management 239 (2020).

[5] UDS Health. “AI in Radiology: Why Demand for Humans is Growing 9%.” UDS Health Blog, February 15, 2026. https://udshealth.com/blog/ai-radiology-demand-for-humans-growing/.

[6] Zhang, Kaichen, Zixuan Yuan, and Hui Xiong. “The Impact of Generative Artificial Intelligence on Market Equilibrium: Evidence from a Natural Experiment.” arXiv, 2023. https://arxiv.org/abs/2311.07071

[7] Acemoglu, Daron, and Pascual Restrepo. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33, no. 2 (2019): 3–30.

[8] World Economic Forum. The Future of Jobs Report 2025. Geneva: World Economic Forum, 2025.

[9] Machovec, Christine, Michael J. Rieley, and Emily Rolen. “Incorporating AI Impacts in BLS Employment Projections.” Monthly Labor Review. U.S. Bureau of Labor Statistics, February 2025.

[10] Huang, Jensen. Interview by Jim Cramer. “CNBC Exclusive: Transcript: Nvidia Founder & CEO Jensen Huang Speaks with CNBC’s Jim Cramer on ‘Mad Money’ Today.” CNBC, March 17, 2026.