Unfortunately, You Need to Know What the Jevons Paradox is

Source: Unfortunately, You Need to Know What the Jevons Paradox is, Hank Green, 33:22, uploaded 2026-07-06, playlist index 16.

Hank Green opens by objecting to the name. The Jevons paradox is useful, he says, although it is not a paradox in the strict sense. It describes a counterintuitive economic pattern: when a resource becomes cheaper to use, people often find more uses for it and total consumption rises. The pattern matters because people imagine with the materials they can afford. Cheapness changes the range of things that seem possible.

Coal and the first efficiency problem

The claim comes from William Stanley Jevons’s 1865 book The Coal Question. Britain was the dominant industrial power of its time, and coal supplied much of that power. Jevons worried that the country would exhaust its coal. Other people saw efficiency as the answer. James Watt’s improved steam engine could perform the same work with less coal, so coal consumption should fall.

The improved engine changed the economics of the resource. Deeper mines became workable. Smaller factories made sense. Railways became possible at a scale that earlier engines could not support. Industries appeared whose demand for coal had been hidden by the high cost of steam power. Each engine used less coal per unit of work, while Britain burned much more coal overall.

Green calls this the Jevons paradox and then removes some of the drama from it. A time-travel contradiction can qualify as a paradox because its terms cannot coexist coherently. Jevons describes an economic effect that can look surprising until the new uses are included. The effect varies in strength. Sometimes efficiency leaves consumption close to its former level. Sometimes new demand overwhelms the savings.

Hot water and the point of saturation

The household water heater shows why the effect has limits. Before hot water became a normal feature of homes, it cost fuel and time. People reserved it for tasks that needed heat, such as cooking, occasional bathing, and some laundry. A home water heater lowered the cost, and households developed new habits around the supply. Daily showers became ordinary. Laundry happened more often. The dishwasher became a practical household appliance once hot water could be produced and stored cheaply.

The demand eventually reached a ceiling. Tanks stopped getting larger because hot water has a specialised set of uses. A cheaper supply could support more showers and more washing, although a household still had little reason to heat its lawn or fill every room with steam. The Jevons effect ran until people had found most of the things for which hot water was uniquely useful. The resource then approached saturation.

That distinction gives Green a way to ask how strongly the effect will apply in other cases. A resource with a short list of uses can reach enough. A resource that can become many different things keeps creating new demand.

Code after the printing press

Code has been expensive for most of its history because a trained person had to write each line. The world therefore contains an unknown amount of code-shaped value that never existed because the labour cost exceeded the expected benefit. Lowering the cost of code may reveal that hidden demand.

Green compares this with the printing press. A person living before Gutenberg might predict that cheaper books would lead people to read twice as many books, or perhaps three times as many. That person could not foresee newspapers, pamphlets, mass-market paperbacks, or a teleprompter script. Those forms descend from the book while serving uses that the expensive manuscript economy kept out of view. Green adds that Martin Luther would also have been difficult for Gutenberg to predict.

Code has a similar breadth. The same resource can produce a web application, a data tool, an automated process, or a small device. Green uses fungible for things that can serve many interchangeable purposes. A dollar can replace another dollar. Coal can produce heat through different kinds of work. Code has more possible functions than hot water, which makes its demand harder to bound.

The current fall in the cost of code comes from AI coding tools. Green says that many people he knows who write software now direct these tools and type less code themselves. The tools remain expensive to run, which limits present demand. Algorithms can become more efficient, though, and cheaper code could make software engineering a Jevons-shaped resource. The demand would include faster delivery of familiar features and uses that have yet to be imagined.

Green keeps this claim local to software engineering. Code is unusually suited to machine assistance because large amounts of training data exist and a computer can test whether a program performs the requested task. A compiler can provide a rough answer to a question that an essay cannot answer mechanically. A system can check whether code runs; it cannot establish with equal ease whether a piece of writing helps people understand something. Other industries may therefore follow different paths.

Electricity and the temporary idea of enough

Electricity began as an expensive way to provide light. Cities promoted electrical infrastructure as a cleaner and safer replacement for gas lamps, which produced soot and could cause deadly accidents. Once electricity became cheaper, it acquired other uses. Motors, refrigeration, heating, communications, and computation all became possible through the same basic supply. Electricity can become motion, heat, light, information, or chemical work. It is a substrate for much of the modern economy.

Green says that per-person electricity use in the United States stayed roughly flat from the mid-2000s for about two decades. Efficiency gains absorbed some new demand, and appliance standards made ordinary devices use less electricity. The plateau also reflected a geographic shift: energy-intensive manufacturing moved abroad, so part of the demand appeared on other countries’ grids. The apparent ceiling therefore described a particular accounting boundary rather than a settled human need.

Electrifying cars, water heaters, and furnaces will raise electricity demand even when the new systems use energy more efficiently and produce less carbon. AI compute adds a second source of demand. It turns electricity into a kind of work that people had not previously asked electrical infrastructure to perform at this scale.

Green cites Anthropic’s annualised revenue as more than $40 billion and says that the figure does not establish profitability. He uses the spending as evidence that people already find enough value in AI coding tools to pay for them. The cost may fall as the systems and their algorithms improve. His uncertainty concerns the demand that would follow. The internet’s last twenty years made confident forecasts look poor in retrospect, and he prefers to reason from the present conditions rather than pretend to know the eventual shape of the communications and AI revolutions.

The fungibility ladder

Green places resources on a ladder according to how many different things they can become.

• Specialised goods such as hot water, plastic, and a particular fertiliser have a limited set of uses. Their demand can saturate once people discover those uses.
• Broad outputs such as code, transport, printed text, and manufacturing serve many kinds of problems while retaining some practical limits.
• Substrates such as energy, information, and atoms feed almost everything else. Saturating them would require people to run out of things to do.

The Jevons effect grows stronger lower down the ladder. Hot water can reach enough because its useful applications are bounded. Electricity has no comparable ceiling because the economy can turn energy into almost any kind of work. Green leaves open whether human wants can ever saturate. He also leaves open the possibility that population, wealth, and consumption could keep moving the ceiling.

Intelligence as a substrate

Green’s provisional model treats intelligence as another substrate. He proposes four broad categories: atoms and energy, information and intelligence. Energy manipulates atoms. Intelligence manipulates information in ways that produce something useful in the world.

He presents this as a thinking tool rather than a complete theory. The definition of intelligence is deliberately broad and separates it from consciousness, qualia, and sensation. Those may belong to further categories. The point is to ask whether intelligence has the same cross-domain reach as electricity. Biology, logistics, design, and law share little subject matter, yet intelligence can work on each of them.

If AI systems can manipulate information usefully across many domains, intelligence will behave like a substrate and the Jevons effect will become intense. The world has been constrained by the number of brains and the amount of human attention available to it. Cheap machine intelligence would expose demand that current scarcity keeps invisible. This is the logic behind forecasts about space-based data centres and Dyson swarms. The forecast may be absurd, or it may describe the direction of a real resource problem. There is no obvious enough for intelligence in the way that a household can reach enough hot water.

The present evidence stays narrower. Green sees code as a case where AI has clear utility and substantial demand. He does not know whether the same systems can make general problem-solving cheap. The technology may keep growing rapidly, or it may reach the flat part of an S-curve earlier than its advocates expect. Anyone who claims certainty about that future is also selling a view of it, and Green includes himself in that warning.

The constraints that may appear later

Jevons described coal as a constraint, although the next constraint often comes from somewhere else. Green recalls England’s need for bones as a source of phosphorus for agriculture. The limiting material can remain outside the question people are asking until the system reaches it.

Electricity may constrain AI through the speed at which generation and distribution can be built. The supply of sunlight could become the limit in a far-future energy system. Physical research creates a different boundary. Science acquires new information by doing things in the world. A scientist still needs wet fingers in a laboratory to learn how much cancer fish develop. Thought can organise information, while experiments create observations that thought could not supply alone.

Cybersecurity could constrain a world filled with cheap code. If every process runs on inexpensive software, the attack surface grows, and defence may fail to scale with creation. Public policy and law can limit deployment even when technology and energy exist. Social distribution can become a constraint when people refuse an abundance that treats them without respect. Cheap content can also make truth harder to find, shifting the scarce resource from information to the work of checking what is true.

Green’s own guess is that the constraint will involve human bodies. AI may solve many technical problems while remaining poor at identifying the lives people want to lead. Human agency and judgement decide which goals deserve resources. He gives priority to people who can feel and have sensations, and he keeps a stated preference for humans rather than letting abstract philosophical thought experiments decide public policy. The preference is part of his moral position, not a result proved by the Jevons effect.

Human taste and knowledge of human problems also remain part of the present picture. Applications that direct AI towards useful tasks may arrive sooner with the technology than they would have otherwise. Green sees no reason to expect the systems to invent those applications faster than people can understand their own needs. He also points out that recommendation algorithms have already changed the communication environment by deciding much of what people see. That is a related argument for another video, while the present one stays with the way cheaper code and intelligence may create further demand.

A useful warning about efficiency

Green ends with the modest claim that gives the video its reach. The Jevons effect is a description of a recurring economic pattern, and its strength depends on the range of uses available to the resource. Efficiency often moves demand rather than removing it. When a resource can serve many purposes, the new purposes remain hidden until the resource becomes cheap enough to support them.

The video’s description adds a limit to the AI comparison. Green says that AI has so far shown its clearest utility in organising existing information, including through the statistical distillation held in model weights. A large part of science consists of acquiring new information, and AI can make that work more efficient without performing the observations itself. That distinction sits beside the coding example. Software can be tested inside a computer; many facts about the world require bodies, instruments, places, and time.

The source’s reported figures about electricity use and Anthropic’s revenue remain claims made in the video. The argument also depends on forecasts about AI capability, code demand, future energy systems, and the point at which human wants saturate. Green names those limits openly. The durable lesson is therefore a question about constraints rather than a forecast: when efficiency makes a broad resource cheaper, which new uses become possible, and what material or human limit appears after them?

Further reading / references

• William Stanley Jevons, The Coal Question (1865).