What is Jevons paradox?

Jevons paradox describes a situation in which using a resource more efficiently leads to greater total consumption of that resource. Each task requires less, but the efficiency improvement encourages so much additional activity that overall use rises. [1,2]

The idea comes from William Stanley Jevons’s 1865 book The Coal Question. He argued that more efficient steam engines made coal-powered production cheaper and opened up new uses. Industry expanded, increasing demand for coal rather than simply performing the same work with less fuel. [1]

The mechanism is not just that existing users consume more. Lower costs can bring in new users and make previously uneconomic activities worthwhile. Efficiency changes what people choose to do, not only the resources needed to do it. [1,2]

Consider a simplified AI example. Suppose an improvement halves the electricity needed for a comparable task. If the resulting lower cost causes usage to triple, total electricity consumption becomes 0.5 × 3 = 1.5 times its original level: a 50% increase despite using half as much electricity per task.

This is stronger than an ordinary rebound effect, where additional use takes back some of the expected savings. If electricity per task halves but usage increases by only 50%, total consumption falls by 25%. There is rebound, but not Jevons paradox. When additional use more than cancels the savings, economists also call this backfire. It is a possible outcome, not an inevitable law. [2]

For AI, cheaper inference—the use of a trained model—could encourage more requests, more experimentation and more automated work. The International Energy Agency’s 2026 analysis describes rapid efficiency improvements alongside growing adoption and more energy-intensive uses, including reasoning and agents. A cheaper individual model call does not necessarily mean a cheaper or less energy-intensive service overall. [3]

However, rising consumption alone does not prove Jevons paradox. The increase must be attributable to the efficiency improvement, compared with what would otherwise have happened. Growth driven by unrelated demand is different. Spending, electricity and emissions are also different quantities: a lower price is not itself evidence of lower energy use. [2,3]

In health care, a related workload analogy is worth considering. An AI tool might make each patient message quicker to prepare, yet encourage a service to send many more messages. Reviewing drafts and handling replies could then consume the time saved. This is an illustration, not evidence that AI necessarily increases clinical workload.

For leaders, the question is therefore not only “How much does each task save?” but “How will the amount and type of work change?” More activity may be valuable when it meets unmet need. The aim is to measure total resource use alongside better care—not assume that efficiency automatically produces either savings or waste.

Sources: [1] William Stanley Jevons — The Coal Question, Chapter VII: Of the Economy of Fuel (1865); [2] Gillingham, Rapson and Wagner — The Rebound Effect and Energy Efficiency Policy (2016); [3] International Energy Agency — Key Questions on Energy and AI (2026).

To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.

AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy