Jevons' Paradox, applied to AI

theory · clickbridge · Rich Price

The claim: when technology makes a resource cheaper to use, we don't use less of it. We use more. Lower cost pulls in so much new demand that it swamps the savings. If that holds for AI, cheap cognitive work means far more cognitive work gets done, not the same amount with fewer people.

Where it comes from

William Stanley Jevons spotted it in The Coal Question (1865). James Watt's steam engine got far more work out of every ton of coal. The obvious prediction was that Britain would burn less coal. It burned much more. Cheaper coal power made factories, railways and steamships worth building, and those new uses dwarfed the old demand.

The logic runs in three steps:

  1. Efficiency lowers the effective price. More output per ton means each unit of work costs less.
  2. A lower price pulls in demand two ways. Existing users use more (direct rebound). New users and new uses become affordable (indirect rebound).
  3. If demand grows faster than the price falls, total use goes up. Economists call this "backfire": the efficiency gain increases consumption instead of cutting it.

Jevons wrote about coal. Later economists generalized it. Khazzoom (1980) and Brookes (1990) argued that economy-wide energy efficiency raises total energy use, an idea now called the Khazzoom-Brookes postulate. Saunders (1992) showed the conditions under which rebound exceeds 100%.

Applied to AI and knowledge work

If AI makes analysis, writing, code and design cheaper and faster, Jevons predicts organizations won't just do the same work with fewer people. They'll do dramatically more of it. The same three rebound channels show up:

  1. Direct rebound. Work already being done gets done more. A team that shipped 10 campaigns a quarter ships 50, because each one costs a fraction of what it did.
  2. Indirect rebound. Work that was too expensive becomes viable. A small business with no legal team runs contracts through AI. People who never paid for design make visual assets daily.
  3. Economy-wide rebound. Cheaper cognitive work lowers the cost of nearly everything in a services economy. That raises real incomes, which pulls in still more demand.

It has happened before

Where the bottleneck moves

That makes Jevons the strongest counter to the automation displacement thesis. It also backs Erik Brynjolfsson's augmentation argument: if AI-assisted workers get far more productive, demand for what they make can grow faster than efficiency displaces them.

Who argues this, knowingly or not

Does it hold for labor? Four objections

The paradox is well established for energy. Applying it to jobs is genuinely contested.

1. Labor isn't coal. Coal is interchangeable and endlessly divisible. People have fixed hours, training costs, locations and their own plans. A laid-off accountant doesn't automatically become a curation professional. Carl Benedikt Frey's history of the Industrial Revolution is the warning: decades of flat wages separated the new machines from broad gains.

2. Speed. Even if new demand eventually outruns displacement, the gap in between can do real damage. Earlier waves took decades. If AI moves faster than institutions adapt, the transition means sustained unemployment and falling wages. Jevons says nothing about timing. Nikhyl Singhal sees it at the role level: senior product managers get more leverage, junior ones get squeezed out, and nothing automatically turns the second group into the first.

3. Institutions decide. Daron Acemoglu argues the rebound depends on how technology is deployed: labor markets, schools, safety nets, company incentives. Policy can amplify it or suppress it. It's a tendency, not a law of nature.

4. Nowhere left for the bottleneck to go. David Shapiro points out that past cycles worked because humans kept something machines couldn't do. Cognition survived the mechanization of strength. Creativity survived the computerization of routine thinking. Current AI competes on strength, dexterity, cognition and empathy at once. If there's nowhere for the bottleneck to shift, demand can grow output without growing employment.

The honest read

Jevons is probably right about direction: cheaper cognitive work will raise total demand for cognitive output. The size of the rebound, the timeline and who gets the gains are all open. It's a strong argument against "AI ends knowledge work." It doesn't settle the jobs question on its own, because it says nothing about speed, distribution, or who captures the value.

Where the value moves

One tension to hold onto: Jevons says total output rises. It doesn't say wages for that output rise. You can get a demand explosion and falling pay at the same time if machines produce the extra output instead of newly hired people.

Sources

All theories · clickbridge.com · Rich Price · last reviewed September 2026 · People linked here have their own pages at people.clickbridge.com