The Jevons Paradox

What exactly is being claimed?

A great deal of confusion here comes from four different claims travelling under one name.

1. Rebound exists. Efficiency improvements produce less saving than the engineering calculation suggests. Essentially undisputed.
2. Rebound is often large. Enough to matter for policy design. Widely accepted, with the size contested.
3. Backfire occurs. Some efficiency improvements increase total consumption. Accepted for specific historical cases; contested as a general pattern.
4. The Khazzoom–Brookes postulate. Economy-wide, in the long run, energy efficiency improvements systematically increase energy consumption above what it would otherwise have been. This is the strong claim, and it is the genuinely disputed one.

Someone who says "the Jevons paradox is real" and someone who says "the Jevons paradox is overstated" are frequently asserting (1) and denying (4), and therefore agreeing.

The Khazzoom–Brookes camp

Daniel Khazzoom argued in 1980, in the Energy Journal, that US appliance efficiency standards would deliver much less than promised because the standards lowered the cost of the service. Leonard Brookes made a broader argument through the 1970s and 1980s: energy efficiency is a form of productivity growth, and productivity growth expands output, and expanded output uses energy. Harry Saunders gave the position its name and its formal shape in 1992, showing that in a standard neoclassical growth model with plausible parameters, energy efficiency improvements raise long-run energy consumption.

The core of the argument is that treating efficiency as a way to cut energy use confuses a supply improvement with a demand restraint. Making energy services cheaper is economically identical to discovering more energy. Nobody expects discovering an oil field to reduce oil consumption.

The measured-rebound camp

Sorrell, Gillingham, Rapson, Wagner and most empirical energy economists accept the mechanism entirely and dispute its magnitude. Their objections to the strong version:

• Energy is a small share of costs. For most sectors of a modern rich economy, energy is a few percent of input costs. A large proportional fall in that few percent produces a small fall in total costs and therefore a small output expansion. The Jevons mechanism was powerful in a nineteenth-century economy where coal was a dominant input; it is weaker in an economy where it is not.
• Saturation. Rich-country demand for many energy services is close to satisfied. Elasticities below 1 are the norm, and the table above shows that below 1 there is no backfire.
• The general-equilibrium estimates are model artefacts. Economy-wide rebound figures above 100% tend to come from models whose substitution elasticities were assumed rather than estimated, and the results move a great deal with those assumptions.

Gillingham and colleagues make a further point that is often lost: even a rebound of 60% leaves 40% of the savings intact, and the rebound itself consists of people consuming more of something they value. Rebound is partly a benefit being counted as a failure. An efficiency measure that lets poor households heat their homes properly has not failed because the energy saving was smaller than projected.

Where the honest disagreement is sharpest

The unresolved question is about long-run, economy-wide, general-purpose technologies. Direct rebound on a single appliance is settled and modest. Whether a century of pervasive efficiency improvement across an entire economy raises or lowers total energy use is not settled, because the counterfactual is unobservable and the models that construct it disagree.

Two facts are worth holding together. Global energy consumption has risen more or less continuously for two centuries alongside continuous efficiency improvement — which is suggestive but proves nothing, since population and output rose too. And energy intensity per unit of GDP has fallen steadily in most rich countries since the 1970s, with total consumption roughly flat or falling in several of them — which is also suggestive and also proves nothing, since some of that is offshored manufacturing.

Serious people read this evidence differently. The entry's status is debated for that reason.

1865. William Stanley Jevons, then thirty, published The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. The efficiency argument is chapter VII, "Of the Economy of Fuel". The book made his reputation; John Stuart Mill and William Gladstone both cited it, and it contributed to a Royal Commission on coal supplies and to Gladstone's argument for retiring the national debt while Britain could still afford to.

Jevons's forecast was wrong in its particulars. He predicted British coal production would become uneconomic within a century as seams deepened. British coal output peaked in 1913 and declined for reasons he did not anticipate — oil, gas, imports, and the collapse of the industries that consumed it. The mechanism he described has aged far better than the prediction he used it for.

Jevons is better remembered as one of the three independent discoverers of marginal utility, with Carl Menger and Léon Walras, in the early 1870s. He drowned while swimming near Hastings in 1882, aged 46.

1970s–1980s. The oil shocks made energy conservation a central policy question and revived the argument. Leonard Brookes, an economist at the UK Atomic Energy Authority, argued from the mid-1970s that efficiency policy would not deliver the projected reductions.

1980. Daniel Khazzoom published "Economic Implications of Mandated Efficiency Standards for Household Appliances" in the Energy Journal, formalising the direct rebound argument for appliance standards.

1992. Harry Saunders published "The Khazzoom–Brookes Postulate and Neoclassical Growth" in the Energy Journal, naming the position and deriving it within a growth model.

1990s–2000s. The debate became empirical. Amory Lovins and the efficiency advocacy community argued the effect was small and being used as an excuse for inaction; Brookes and Saunders argued back in the same journals. The exchange in Energy Policy through the 1990s is unusually direct.

2005–2007. Blake Alcott's survey of the history in Ecological Economics, and Sorrell's UKERC review, established the current framing: rebound unambiguously real, backfire situational, magnitude the open question.

It changes what an efficiency policy is for. If the goal is reducing total energy use or emissions, efficiency alone is an unreliable instrument, because the savings it produces are partly spent. Policies that work on quantity or price directly — carbon pricing, emissions caps, fuel taxes — do not have this problem, because they raise the cost of the service rather than lowering it. The standard policy conclusion in the literature is not "abandon efficiency" but "pair efficiency with a price or a cap", so that the freed-up demand meets a constraint.

It explains a persistent forecasting failure. Engineering estimates of savings from efficiency programmes have a long record of overshooting. Utility programme evaluations routinely find realised savings well below projections. Rebound is not the only reason, but it is a systematic one, and the gap is now large enough that regulators build discount factors into their accounting.

It reframes the relationship between technology and resource use. The intuitive model is that technology is how a fixed appetite is satisfied with less. Jevons's model is that technology is how the appetite grows. Which model is right determines whether you expect innovation alone to solve a resource problem. Anyone arguing that a technology will reduce consumption of something owes an account of why the demand for the cheapened service will not expand.

It is an argument that has been repeatedly rediscovered. Induced demand in transport planning, the observation that faster computers get filled with heavier software, and the finding that agricultural intensification does not reliably spare land for nature — sometimes called the Borlaug hypothesis and its counter-evidence — are all the same structure arrived at independently. That recurrence is itself evidence the mechanism is general.

Data centres and AI. Every generation of hardware and every algorithmic improvement has cut the energy needed per unit of computation by a large factor, and total data-centre electricity consumption has risen. Cheaper inference means more inference: uses that were not worth the compute become worth it. The projected electricity demand from AI training and serving is the current highest-profile case of the argument, and the elasticity of demand for computation appears to be well above 1.

LED lighting. Roughly an order of magnitude more efficient than incandescent. Global installed lighting has expanded correspondingly, satellite imagery shows outdoor lit area and radiance increasing year on year, and the light-pollution literature treats this as a settled example of rebound in action.

Electric vehicles. Cheaper per mile to run than petrol cars, which by the mechanism means more miles driven. The direct rebound here is measurable and modest, and the emissions arithmetic still favours electrification substantially — but the miles-driven increase is real, and it is why some analysts argue road pricing has to accompany the transition.

Cryptocurrency mining. More efficient mining hardware does not reduce the energy used to mine, because the difficulty adjusts and the equilibrium is set by the price of the coin against the cost of power. This is a limiting case where rebound is essentially exactly 100% by construction — an unusually clean natural experiment in the mechanism.

Water and land. Efficient irrigation was expected to leave more water in rivers. In several basins it has not, because farmers responded by irrigating more land or switching to thirstier crops. The same argument runs in agriculture over whether higher yields spare land or, by making farming more profitable, expand the frontier.

Defining rebound precisely

Let $S$ be the quantity of an energy service consumed, $E$ the energy input, and $\epsilon = S/E$ the efficiency. Then $E = S/\epsilon$, and in logarithmic changes

$$\hat{E} = \hat{S} - \hat{\epsilon}$$

The price of a unit of service is $p_S = p_E/\epsilon$, so an efficiency improvement is a proportional fall in the service price: $\hat{p}_S = \hat{p}_E - \hat{\epsilon}$. Holding the energy price fixed, $\hat{p}_S = -\hat{\epsilon}$.

Let $\eta = -,\partial \ln S / \partial \ln p_S \ge 0$ be the own-price elasticity of demand for the service. Then $\hat{S} = \eta,\hat{\epsilon}$, and

$$\boxed{;\hat{E} = (\eta - 1),\hat{\epsilon};}$$

The rebound is $R = \eta$, usually quoted as a percentage. The regimes fall out immediately:

• $\eta = 0$: $\hat E = -\hat\epsilon$. Energy falls one-for-one with efficiency.
• $0 < \eta < 1$: energy falls, but by less than the efficiency gain.
• $\eta = 1$: $\hat E = 0$. Exactly cancelled.
• $\eta > 1$: $\hat E > 0$. Backfire.

So the entire question of whether Jevons's strong claim holds for a given service reduces to whether demand for that service is elastic. Stated this way it is neither mysterious nor especially paradoxical — which is why the entry's nature is a counterintuitive result rather than a contradiction.

Why the measured elasticity is not the whole story

The formula above holds the energy price fixed. It does not. If efficiency improvements reduce aggregate demand for energy, the energy price $p_E$ falls, which reduces $p_S$ further and induces more service demand. Writing $\sigma$ for the elasticity of energy supply, the general-equilibrium rebound exceeds the partial-equilibrium figure whenever supply is not perfectly elastic. This is one of several reasons economy-wide estimates run above direct ones.

Three further terms are omitted by the simple decomposition:

• Substitution in production. Cheaper energy services shift factor proportions toward energy. The magnitude depends on the elasticity of substitution between energy and capital, a parameter estimated across a very wide range in the literature and the single largest driver of disagreement between models.
• Output effects. Lower costs raise output, and higher output uses energy even at unchanged intensity.
• Induced innovation. Cheap energy services change what is worth inventing. This is the largest and least tractable channel; it is where the railway comes from in Jevons's account, and no model handles it convincingly.

Saunders's growth-model result

Saunders (1992) embedded energy-augmenting technical change in a neoclassical growth model with a constant-elasticity-of-substitution production function

$$Y = \left[\alpha (A_K K)^{\rho} + \beta (A_L L)^{\rho} + \gamma (A_E E)^{\rho}\right]^{1/\rho}$$

with elasticity of substitution $\sigma = 1/(1-\rho)$, and showed that an increase in energy efficiency $A_E$ raises equilibrium energy use $E$ for a range of plausible $\sigma$ and factor shares. With Cobb–Douglas production ($\sigma = 1$) the result is backfire in the long run under standard assumptions about capital accumulation.

The result is real but conditional, and its critics have concentrated on exactly that conditionality: the conclusion is driven by $\sigma$ and by the treatment of efficiency gains as costless, and neither is established empirically. This is the technical core of the disagreement, and it is a disagreement about parameter values in a model rather than about whether the mechanism exists.

Empirical magnitudes worth carrying

Channel	Typical estimate	Confidence
Direct, household space heating	10–30%	reasonably firm
Direct, personal transport	10–30%	reasonably firm
Direct, lighting (developed)	5–20%	moderate
Direct, lighting (developing / historical)	can exceed 100%	case-specific
Indirect (re-spending)	5–15%	weak
Economy-wide	20% to >100%	model-dependent

These are drawn from Sorrell's review and the subsequent literature. The pattern is consistent: the better identified the estimate, the smaller and the more partial it is; the broader the question, the larger the estimate and the weaker the identification. That relationship is itself the most important thing to know about this literature.