Jevons Paradox in Intelligence

The classical paradox, applied

The classical Jevons Paradox, from 1865, holds that when a resource becomes dramatically more efficient, total consumption increases rather than decreases, because the resource becomes useful enough to apply to vastly more problems. William Jevons observed that more efficient steam engines made coal so useful that coal consumption exploded rather than contracted. Garry Tan applies this directly to the current AI moment: making intelligence cheap will not reduce demand for intelligence-intensive work, it will create demand for more of it than currently exists.¹

Today's anxiety about AI assumes substitution: AI replaces intelligence work, therefore fewer intelligence-intensive jobs. The Jevons framing says that assumption is wrong. Intelligence becoming cheap is not the end of intelligence-intensive work, it is the opening of a new demand frontier: instead of one team analyzing one market, a thousand; instead of one product, a hundred; instead of one customer interaction, a million.

The implication for ambition

If the paradox applies to intelligence, fear of AI becomes a function of ambition size. Small ambitions make AI terrifying, since a machine that does your current job faster and cheaper reads as an existential threat. Large ambitions make AI the best possible news, since the same machine gives you headroom to pursue something you could not previously attempt. The behavioral implication is that the right response to AI is not defensive optimization, doing the same thing more cheaply, but offensive expansion, attempting a bigger thing that has newly become possible.¹ Tan adds that the paradox does not activate automatically; it requires capital and management to actually raise their ambitions, since the demand frontier only expands if someone decides to pursue it.

Ryan Petersen is cited within this argument as evidence from trade and logistics: "the human desire for more things is absolutely limitless." His broader frame on global trade expansion, that lowering the cost of moving goods multiplies rather than saturates demand for moving them, is the same logic applied to a physical industry rather than a digital one.

Operator evidence

Brendan Foody of Mercor supplies a concrete labor-side data point rather than a theoretical one: coding tools make his company hire more engineers, not fewer. "Every engineer on our team is loving using the coding tools, and it's making us hire even more engineers because the productivity of each person is going up." Cheaper code, in his account, produces more code worth writing, which produces more engineers.²

Foody extends the argument with a historical analogy to Keynes, who a century ago predicted productivity gains would deliver a roughly fifteen-hour work week. Instead, rising productivity across the economy let people do far more rather than far less. Near-term displacement in specific roles is real, in his view, but the demand of unsolved problems, plus the long tail of things humans can do that models cannot, keeps labor demand high for decades.²

Contrast with a leisure-liberation view

Brad Jacobs' utopian AI scenario offers a popular counter-narrative to this argument: that humanoid robots will replace nearly all jobs, "including mine," and that this is acceptable because it frees up time for people to enjoy and be happy. That is a leisure-liberation read, where automation removes work and returns time. The Jevons view instead holds that cheap intelligence expands the frontier of work worth doing, so demand for intelligence-intensive output grows rather than collapsing into free time. The two views are partially reconcilable, since robots could absorb today's tasks while humans migrate to newly tractable higher-order work, but they differ sharply in emphasis: Jacobs foregrounds time freed, Jevons foregrounds demand created.

Open question

The Jevons Paradox is an empirical claim about demand elasticity, not a law, and some technologies have produced genuine demand saturation rather than explosion. What the falsification condition looks like, the point at which making intelligence cheaper stops exploding demand, remains unspecified. It is also worth noting that the strongest evidence for the expansion thesis tends to come from people with a direct commercial stake in it being true, an AI labor marketplace, a venture fund, an AI-native operator, which does not make the evidence wrong but does mean it is not disinterested.