Suppose engineers invent a computer chip that performs the same calculation using half as much electricity. The obvious conclusion is that electricity consumption should fall. Now suppose that cheaper computation makes it economical to put those chips into millions of additional devices, run larger models, process more video and automate tasks that were previously too expensive. Each calculation uses less energy, yet the world performs so many more calculations that total electricity demand rises.

This is the unsettling logic associated with the Jevons paradox: improvements in efficiency can sometimes increase total resource consumption rather than reduce it.

The idea is more than 160 years old. It began not with computers but with coal, steam engines and Victorian Britain's fear that the fuel powering the Industrial Revolution might eventually run out. Today the same argument appears in debates about cars, air conditioning, data centers and artificial intelligence.

But there is an important distinction that popular explanations often lose. Efficiency does not automatically cause higher total consumption. The paradox is the extreme case of a broader phenomenon economists call the rebound effect. Whether consumption actually rises depends on what people and businesses do with the savings.

Jevons noticed something strange about coal

In 1865, British economist William Stanley Jevons published The Coal Question, an investigation into Britain's dependence on coal and the consequences of resource depletion. The country had spent decades improving the steam engine. James Watt's designs and later innovations could extract much more useful work from a given quantity of fuel than earlier engines.

It seemed reasonable to think that better engines would conserve coal.

Jevons argued that the opposite tendency could emerge. Greater efficiency made steam power cheaper and commercially attractive for more activities. Industry expanded its use of engines, production increased and coal consumption could grow rather than shrink.

Historical scholarship on Jevons and The Coal Question identifies this argument as an early foundation for what is now called the rebound effect or Jevons paradox. Jevons' larger concern was national: Britain could not assume that engineering efficiency alone would solve the problem of finite coal reserves. citeturn0search1

The insight was counterintuitive because it separated two measurements that are easy to confuse: resource use per unit of service and total resource use.

A factory can burn less coal per product while producing so many more products that its total coal consumption increases.

The hidden mechanism is often price

Efficiency effectively makes a service cheaper. If a car travels twice as far on a liter of fuel, the fuel cost of driving one kilometer falls. If a server can answer twice as many requests with the same electricity, the energy cost per request falls. If an AI model becomes dramatically cheaper to run, companies can afford to invoke it in many more places.

Lower effective prices can stimulate demand.

Economists generally distinguish several layers of rebound. A direct rebound occurs when people consume more of the service that became efficient. A household buys a more efficient heating system, for example, then chooses a warmer indoor temperature because heating is cheaper.

An indirect rebound occurs when the savings are spent elsewhere. Money saved on electricity might pay for an airline ticket, a new appliance or another energy-consuming activity.

At the broadest level, efficiency can alter prices, production, investment and economic growth throughout an economy. These economy-wide effects are the hardest to measure and the most controversial.

If an efficiency improvement would theoretically reduce resource consumption by 100 units but behavioral and economic responses add 30 units back, the rebound is 30 percent. Consumption still falls by 70 units.

Only when the rebound exceeds 100 percent does total resource consumption rise above where it started. That stronger outcome is often called backfire and is the version most closely associated with the Jevons paradox.

Efficiency does not always backfire

This qualification matters because the Jevons paradox is sometimes presented online as a universal law: make anything more efficient and society will inevitably consume more of it.

The evidence does not justify that claim.

A major review by energy researcher Steve Sorrell concluded that economy-wide rebound effects may be larger than conventional estimates assume, but that empirical evidence for full Jevons-style backfire is far from conclusive. Measuring the effect is difficult because researchers must determine what consumption would have been without the efficiency improvement while separating efficiency from income growth, changing prices, new technologies and other forces. citeturn0search2turn0search4

The size of rebound also depends on demand. If making a service cheaper barely changes how much people want, efficiency can produce substantial net savings. If demand responds dramatically to lower costs, rebound can be much larger.

Consider refrigerators. Once a household has enough refrigerated space, making refrigeration cheaper does not necessarily persuade it to install twenty refrigerators. Efficiency improvements can therefore reduce electricity consumption significantly.

Computation is different. Humanity appears to have an enormous appetite for additional computing when its price falls.

Computing may be a near-perfect rebound machine

The history of computers is one of extraordinary improvements in efficiency. Modern chips perform calculations at energy and monetary costs that would have seemed impossible to early computer engineers.

Yet cheaper computation did not lead civilization to say, “Excellent, we can now perform the same number of calculations using less electricity.” It led us to invent entirely new categories of computational demand.

Personal computers spread into homes. Then came laptops, smartphones, cloud computing, streaming video, digital maps, recommendation systems, cryptocurrency, connected devices and machine learning. Tasks that were once too expensive became routine; tasks that were unimaginable became products.

This is the Jevons mechanism in an especially intuitive form. Reducing the cost of one unit of computation expands the number of economically worthwhile uses for computation.

A modern smartphone uses vastly less energy per calculation than a room-sized early computer. But billions of people now carry computers continuously, communicate with distant data centers and generate enormous quantities of data.

The relevant question is therefore not whether computers became more efficient. They unquestionably did. It is whether efficiency gains outran the explosion in demand.

Artificial intelligence has revived the 1865 argument

AI has made Jevons fashionable again because model efficiency is improving at the same time that demand for AI computation is expanding rapidly.

Researchers writing in Joule have explicitly warned that efficiency improvements in AI need not translate directly into lower total energy consumption. They point to the possibility that more efficient models make AI accessible to more users and devices, increasing aggregate demand for computation. citeturn0search8

A 2025 paper by Alexandra Sasha Luccioni, Emma Strubell and Kate Crawford similarly argued that environmental analysis of AI must account for rebound effects rather than focusing only on improvements in energy or water use per operation. Their point is not that every efficiency gain necessarily causes backfire, but that cheaper AI can stimulate wider deployment and therefore offset some of the savings. citeturn0academia24

This is particularly relevant to inference — the repeated use of a trained model. Suppose an engineering breakthrough makes each AI query ten times cheaper. A company may keep its existing workload and save resources. But it may instead add AI to customer support, search, coding, advertising, document processing, personal assistants and millions of automated workflows.

If usage increases twentyfold after the cost per query falls tenfold, total computation rises.

That is textbook rebound logic.

But rising AI demand is not automatically proof of Jevons

There is another subtlety. If AI electricity consumption rises while chips simultaneously become more efficient, that correlation alone does not prove the Jevons paradox.

Demand might have exploded even without the efficiency improvement because AI itself is a rapidly expanding new technology. Researchers must establish that efficiency lowered costs and that this lower cost induced enough additional consumption to offset the original savings.

This distinction is important because almost any growing technology can appear Jevons-like if efficiency and demand happen to increase at the same time.

Recent scholarship on AI therefore treats rebound as one mechanism among several. Efficiency lowers resource use per unit, while falling costs, new applications, economic growth and greater scale can push total demand upward. Which force wins is an empirical question, not something the word “Jevons” can answer by itself. citeturn0search6

The paradox does not mean efficiency is useless

Another common mistake is to conclude that if rebound exists, improving efficiency is pointless.

That does not follow.

Even with a 30, 50 or 80 percent rebound, an efficiency improvement can still produce substantial resource savings relative to what consumption would otherwise have been. Efficiency can lower costs, increase productivity and allow society to obtain more useful services from the same resources.

The policy lesson is narrower: efficiency alone does not guarantee that total resource consumption will decline.

If the objective is specifically to reduce absolute energy use, emissions or another environmental burden, policymakers may need measures that address the total quantity as well as efficiency per unit. Those measures can include carbon pricing, clean electricity, emissions standards, resource constraints or other mechanisms depending on the sector.

Efficiency answers the question, “How much resource do we need for each unit of service?” Sustainability often requires a second question: “How many units of the service will we consume?”

A steam-engine problem in a digital world

The durability of Jevons' insight is striking. He was writing about coal at a time when electricity grids, automobiles, airplanes, computers and data centers did not exist. Yet the economic mechanism he noticed can appear whenever technological progress lowers the effective cost of using a resource.

Today a data center can process more information per kilowatt-hour than its predecessors. A processor can perform more operations per joule. An AI model can be compressed or optimized so that each response requires less computation.

Those are real gains. But efficiency changes behavior because it changes what is affordable.

The cheaper computation becomes, the more places society discovers to use computation. The cheaper transportation becomes, the farther people and goods can travel. The cheaper illumination becomes, the more extensively we can light homes, streets, buildings and screens.

Jevons' paradox is therefore not really a paradox of engineering. Engineers can make machines genuinely more efficient. The surprise comes afterward, when human demand responds to what engineering has made possible.

Sometimes we use the efficiency gain to save the resource. Sometimes we use it to do more. And sometimes we do so much more that total consumption rises.

That was the uncomfortable possibility Jevons identified in the age of steam. In the age of AI, it has become uncomfortable all over again.