The Jevons Paradox for Intelligence
Fears of AI-induced job loss could not be more wrong
The last couple of months has seen an explosion of concern that AI is on the cusp of eliminating a large proportion of white-collar jobs. This is the latest epicycle in a long-run trend whereby new advancements in AI drive huge waves of hype, which then come crashing down as intelligent scrutiny reveals them to be some mix of naïve, impulsive, and astroturfed. The recent exuberance surrounding agentic coding tools such as Claude Code is no different. While these tools represent a genuine increase in capabilities, they are not going to destroy the white-collar labor market any time soon.
David Oks recently published an excellent essay explaining why. Clay Wren published an essay in response to Oks on X, which I would also recommend. While Oks and Wren clash over a wide range of issues, my goal here is to focus on one—the question of a ‘Jevons paradox’ for white-collar work. The crux of the issue is this: will AI decrease or increase the demand for human knowledge work?
The idea behind the Jevons Paradox is that, as the efficiency of a given resource’s use rises, total consumption of that resource rises too, when intuitively we would expect it to fall. The paradox originates with economist William Stanley Jevons’s analysis of English coal consumption in the 1860s. In response to arguments from other economists of his day that, though coal supply was dwindling, its expiration could be staved off by methods for producing more power with less coal, Jevons argued that those methods would actually increase the demand for coal by making coal-powered services much cheaper. That would in turn mean further industrial expansion built on top of those services.
The AI boom has re-inspired conversation about Jevons paradox in part due to questions surrounding rising energy and compute efficiency, and how they could lead to more demand for AI. But the Jevons paradox has also come up in the context of a different resource: human labor, particularly intellectual labor. Here is Oks on the new Jevons paradox for knowledge work:
Simply put … demand for most of the things that humans produce is much more elastic than we recognize today; and as long as humans are complementary to the production process, it won’t be rare for efficiency gains to get swallowed up by demand growth. This is the famous Jevons paradox, the tendency for the more efficient use of a resource to increase total consumption of that resource, rather than decrease it. Energy is the classic Jevons case: we find over and over again that as energy becomes more efficient to produce, people respond not by consuming the same amount of it, but by increasing their consumption—such that overall energy use tends to rise. (Thus the “paradox” part: energy efficiency increases energy consumption!)
And I suspect that the same elasticity of demand that we see with energy applies to a lot of the things that humans create. As a society, we consume all sorts of things—not just energy but also written and audiovisual content, legal services, “business services” writ large—in quantities that would astound people living a few decades ago, to say nothing of a few centuries ago. Demand tends to be much, much more elastic than we think.
I am with Oks here. AI is going to unlock all manner of newly efficient human-machine complementarities, which will increase demand not just for software-engineering labor, but human labor as such. I will elaborate below, but first, consider Wren’s response to Oks, which introduces several complications I want to touch on:
As for the Jevons paradox argument (often cited); that elastic demand will absorb productivity gains. I believe it’s real for some categories of output but cherry-picked as a general principle. Software is Oks’ central example, and it’s well-chosen: software is elastic in demand because it’s a general-purpose tool. But does anyone believe demand for legal document review is infinitely elastic? For tax preparation? For freelance video editors? These are bounded markets where productivity gains translate fairly directly to headcount reductions, and I’m still struggling to understand how we are telling early-wave displaced roles to upskill or find new careers.
Someone commented under Oks’ post another example that I’ll jump on. As global manufacturing shifted toward China and other low-cost production regions, total manufacturing output continued to expand rather than contract, a Jevons-like scale effect where cheaper production increased overall consumption. American manufacturing workers, however, bore concentrated losses. The gains flowed disproportionately to consumers, firms, and capital owners, while many displaced workers (especially in Midwestern industrial regions) faced long-term economic decline that helped fuel a broader political backlash against globalization …
To summarize: Jevon’s paradox in aggregate output is perfectly compatible with catastrophic distributional effects. You can have more total economic activity and still have millions of people whose specific skills and local labor markets are destroyed. The people being displaced right now are not edge cases, they’re illustrators, translators, copywriters, graphic designers, video producers, and 3D artists who were told their skills would always be valuable because they were “creative.” The aggregate framing erases these people, and it will erase more.
There is a lot to say here. While Wren is right to push Oks on the distributional effects of a prospective AI shock, I think his response misses the mark.
Go west, desk jockey
The first thing to say is that Wren is dead-wrong that there are no major demand elasticities in legal document review, tax preparation, or video editing. With respect to law, there are already numerous examples of law firms integrating AI into their internal practice, as well as expanding into new AI-powered business lines already seeing widespread uptake—a strong signal that there is demand for new AI-enabled services in the legal industry.
For example, the law firm A&O Shearman launched an AI-powered contract drafting and negotiation tool called ContractMatrix in December 2023 that is already used by hundreds of firms. Now they are on the verge of launching a new SaaS product leveraging AI agents for more complex legal tasks in areas like antitrust and cybersecurity. More recently, the VC-backed Justpoint, a consumer research company that “uses proprietary AI to analyze data, medical studies, safety reports, and patient records” just incorporated its own law firm—Justpoint Law—that it plans to use to represent clients in mass tort litigation based on its research findings. This is an early example of AI unlocking opportunities for horizontal expansion that would have previously been infeasible.
Tax services and video production are undergoing similar shifts. Far from a narrow focus on slimming and efficiency, tax services providers plan to use AI to “enable more granular and accurate financial forecasting, offer more expansive scenario-modeling capabilities, and help firms provide more insightful tax strategies and decision-making support based on real-time analysis of much larger sets of data” according to research by the Thomson Reuters Institute.
In video production, Wren points to video-generation models like Veo 3.1 and the newly released Seedance 2.0 to argue that AI advancement has cut out the “intermediary layer” of “producers, directors, editors, camera operators, sound technicians, [and] VFX artists” once responsible for transforming ideas into watchable content. That consumers can cheaply generate whatever video they want, he says, means destroying thousands of jobs in the entertainment industry.
In my view, this could not be further from the truth. Yes, the average person can now generate gimmicky A.I. slop with ease. That might have posed an enormous threat to the creative economy, not to mention the creative spirit, were humanity’s desire for art and storytelling satiable by 10-second videos of SpongeBob evading arrest, or George dropkicking Jerry on Seinfeld. But obviously it is not. There have so far been no wholly AI-generated entertainment products with comparable cultural or economic impact to Hollywood movies or T.V. shows. I doubt there ever will be absent substantial human intermediation. Moreover, what interesting AI video-art so far exists clearly requires substantial human ingenuity to produce, and is also far from being integrated into the corporate entertainment ecosystem. For that to happen will require large new investments in cultivating and operationalizing human talent. It will also require the deep pool of tacit knowledge had by more experienced practitioners in fields like traditional editing, sound design, VFX, and the like.
And this is just for existing product-types; as with law and tax services, AI will also create new whole new entertainment verticals. For example, we may soon see the rise of small consumer-facing production companies that use AI to create bespoke, high-quality video content, whether on commission or for traditional film and television markets. Maybe you want to show a Lord of the Rings-style fantasy film starring your son at his graduation party. Maybe you want to watch a three-season anime entirely about caterpillars. You may soon be able to hire a relatively cheap production team to create these for you, one whose output will be much higher-quality than whatever you could make on your own.
Wren’s, and many others’ main mistake is to assume that firms will only use AI to more efficiently produce the exact same products. But that is not what they are going to do, nor has it ever been firms’ main reaction to the rise of new general purpose technologies (or short of that, large efficiency gains in production). Rather, they have and will continue to do what they used to do faster and better, while also and much more importantly expanding into new products that once would have been inconceivable.
This is very difficult to do while also cutting headcount en masse. What little evidence we have so far agrees. A January 2026 survey by EY-Parthenon found that 69% of CEOs believed investments in AI would lead to maintaining or increasing headcount. According to KPMG, as of December 2025, 92% of CEOs were planning to increase headcount, even as 69% were allocating up to a fifth of their budget on AI. And a new February 2026 survey of 12,000 European firms by the CEPR found that, while firms who adopted AI saw productivity rise by 4%, they did so without reducing heacount.
In customer service—a sector often treated as ground zero for AI-driven disruption—a survey by Gartner found that only 20% of leaders have so far reduced staffing due to AI, while 55% report both stable staffing levels and higher customer volumes. 42% are also “hiring specialized roles—including AI strategists, conversational AI designers, and automation analysts—to support AI deployment and management.”
Other fields that were allegedly going to be devastated by AI, such as software engineering, translation and radiology, have all seen employment increase since the arrival of GPT-3. As the Georgetown Center for Security and Emerging Technology’s Jack Karsten explains, “(AI) is not only not replacing [radiologists], but it’s actually increasing the amount of work they can do and increasing demand for their services.” This is just as the Jevons paradox would predict.
More granularly, IBM has announced that they are tripling entry-level hiring, not in spite, but because of AI. IBM HR chief Nickle LaMoreaux anticipates that AI is going to create “totally different jobs,” with entry-level coders moving up the ladder of abstraction to focus more on marketing, client relations, and building totally new products.
While many, Wren included, have pointed to rising graduate unemployment and sluggish hiring overall as evidence of AI-driven disruption, the prevailing evidence does not support that these trends are related to AI. They are much more likely the result of monetary tightening and broader labor supply shifts, such as a rising number of graduates generally, and a tighter labor market for non-graduates.
As of now, what little evidence there is on AI’s emerging labor market effects points toward rising, not falling white-collar employment. While we can only speculate about mechanisms at this point, this may be the Jevons paradox in action: knowledge work has plausibly become much more efficient, increasing the demand for intellectual labor.
Future shock
What about Wren’s ‘China shock’ analogy? Whatever their effects on output or employment, trade or technology shocks can still have “losers” who suffer for being displaced, whether by cheaper workers or machines. Autor, Dorn, and Hanson (2021) found that, from 2000 to 2019, between ~6% and ~33% of Americans lived in areas that saw personal income per capita decline on net due to the employment effects of exposure to cheap imports from China (p. 45-46; to be clear, that doesn’t mean these areas got poorer overall, just that the effect of trade with China was net negative). Similar distributional effects could result from AI adoption, whether because whole jobs are automated away, or because they become cheaper to perform, perhaps even by remote workers stationed abroad.
I am not worried about such a scenario, first because if the China shock is any guide, an analogous AI shock would still be beneficial on net, and second because the China shock is not any guide—there are several reasons that it is a poor analogy for AI.
As the figures from Autor, Dorn, and Hanson imply, the vast majority of the country benefitted from the China shock, at least in terms of personal income, because Chinese imports lowered prices. Those benefits also tended to favor the poor, who spend more in more traded sectors. The China shock was also a net job creator; while it led to job losses in manufacturing, it led to job growth in services, oftentimes within the same firms.
That doesn’t mean the shock was harmless—many smaller, manufacturing-dependent communities did not fare well, and the shock probably contributed to political polarization—but it is hard to argue it was bad for the American economy, all things considered. That the China shock has loomed so large in our politics has more to do with the psychological advantage had by small, but concentrated harms relative to large, but diffuse benefits in human apperception than it does with the severity of the harms themselves. If the ‘AI shock’ is similar to the China shock, we have (relatively) little to worry about.
Second, even putting the benefits of the China shock aside, it is too disanalogous with a prospective shock from AI to be informative. The China shock was mainly a story of cheap labor competition; China’s large supply of cheap, low-skill labor is what allowed its manufacturers to handily outcompete ours on price. But the work exposed to AI is high-skill, not low, and while a small share of high-skill work is moving abroad, particularly to South Asia, it is not clear why AI in particular would hasten this trend. Plus, if AI’s economic future depends on leveraging workers’ tacit knowledge, it will be difficult to substitute foreign remote workers for Americans better-embedded in the local contexts productive of that knowledge.
Another key disanalogy is that the parts of the country hit hardest by the China shock were those most specialized in manufacturing, and with the fewest college-educated workers (Autor, Dorn, and Hanson 2021, p. 18). But since a hypothetical AI shock would hit white-collar workers the hardest, it would affect areas with more economic diversity and more college-educated workers, precisely those best prepared to weather economic disruption.
All this in mind, the U.S.’s experience with the China shock does not suggest we should be worried about an analogous AI shock. The China shock is largely a story about competition from cheap labor, not labor-saving technology, and even if it was, AI is not just labor-saving; it also makes tons of new labor possible.
AI as a resource blessing
This is the fundamental lesson of the new Jevons paradox for knowledge work. Generative AI represents an utterly vast, new, and widely accessible vein of intelligence. To imagine that we will only use it to do what we already do faster represents a catastrophic failure of imagination.
The same goes for thinking it will be possible to mine that vein without relying on codified and tacit knowledge had only by human beings, the large majority of which is not recorded on the open internet and therefore absent from LLMs’ training data. This is not to speak of our other comparative advantages like physical mobility, greater neurological complexity, continual learning, superior generalization, seamless multimodality, more plentiful ‘training data’ and so on.
Far from the recent focus on labor supply, I suspect AI’s economic impacts are best modeled as a positive resource supply shock. And if artificial intelligence is a resource, it follows that whoever is responsible for extracting it will benefit greatly from its discovery. If you can prompt a large language model, then that is you, and while that extraction has never been cheaper, Jevons tells us it will be demanded like never before.
Although, if we want to be true to Jevons’s original argument, to say we have ‘discovered’ this new intelligence is not quite right. Before AI, we were already plumbing the digital archive stored in LLMs’ trillions of weights. But for the most part it took us much more time and effort to do so. With AI, we can now extract the diffuse intelligence of that archive with ease, in purified, fungible form. We can ask that archive what it knows, and greedily, we are told.
This is not to deny that AI poses serious risks as well as opportunities. But to fear for our economic prospects is to radically misunderstand both. You just became rich. You can celebrate it, you can fear it, but you must recognize it to be so. You are now an intelligence baron. Your country is an intelligence empire. It is no time to fear for your desk job.
If anything, fear the profusion of new jobs that very soon you are going to have.
I suggest that Jevon's Paradox applies to the product of AI. So what is that product? I maintain that it is words: not meaningful or useful content, just verbiage. Indeed, it's arguable that the failure to realise the productivity gains expected over the past couple of decades is due to the increasing demand for verbiage: that is, a significant proportion of the white collar class is employed in producing and consuming verbiage with little to no effect whatsoever on the production of economic goods. This relates to David Graber's Bullshit Jobs hypothesis. I suggest that verbiage, not knowledge, is what many "knowledge workers" produce.
As an example: the Lower Thames Crossing has required 2,000 documents, 360,000 pages, about 100 million words: costing £267 million to produce before permission could be obtained to start work. What will happen when the cost of producing and consuming such documents falls by a factor of a million? Will we see the cost of producing the planning documents for the next such project fall to £267? Or will the total amount of verbiage produced rise to 100 trillion words? My guess is that the latter will be closer to the truth.
Thank you for this positive vision!
For Jevon’s paradox to apply to the use of intelligence of humans for humans, there must be a limit to how much AI intelligence can replace human intelligence applied for human goals. Otherwise, even if there are multiplying uses for intelligence as increasing amounts become available, those uses may be for humans or for AI, and either way, rely on increasing use of AI intelligence, not human intelligence.