The Jevons Paradox: Why Efficiency Makes Us Use More, Not Less
When resources become more efficient to use, we don't consume less—we find new uses for them and consume more. This counterintuitive pattern, identified by economist William Stanley Jevons in 1865, explains why making things cheaper often increases total demand. It applies strongly to fungible substrates like energy and intelligence, but weakly to specialized goods like hot water. Understanding where Jevons effects are strong or weak is crucial for predicting how AI and electrification will reshape resource consumption.
What Is the Jevons Paradox (and Why It's Not Actually a Paradox)
The Core Claim: Efficiency Increases Consumption
When a resource becomes cheaper or more efficient to use, total consumption of that resource typically increases rather than decreases. This happens because efficiency opens up new economic uses that were previously unaffordable. William Stanley Jevons observed this in 1865 when James Watt's improved steam engine made coal cheaper per unit of work, yet Britain's total coal consumption exploded as new industries (mines, factories, trains) became economically viable.
It's Not Actually a Paradox
A paradox is a logical contradiction (like time travel killing your grandfather). The Jevons effect is simply counterintuitive—efficiency leads to more use because the range of economically viable applications expands. It's a fact about economics, not a logical impossibility.
Historical Context: Jevons and Coal
Economist William Stanley Jevons (1830–1879) published The Coal Question in 1865 when he was 30 years old. Britain dominated industry because of abundant coal. When Watt's steam engine improved efficiency, people expected coal use to drop. Instead, Jevons correctly predicted it would skyrocket because cheaper engines enabled entirely new industries.
When Jevons Effects Saturate: The Hot Water Example
Hot Water: A Case Where Jevons Effect Ran Out
Before home water heaters, hot water was expensive and rare—used only for cooking, occasional bathing, and laundry. Once heaters became cheap and common, demand exploded: daily showers, frequent laundry, dishwashers. But unlike coal or electricity, hot water demand eventually saturated. There are only so many hot-water uses (you don't water lawns with hot water), so consumption plateaued even as efficiency improved.
The Saturation Point: Enough Exists for Some Things
Hot water demonstrates that saturation is possible—there is a ceiling to demand when all economically viable uses have been found. Per capita electricity use in the US was roughly flat from the mid-2000s to around 2020, suggesting a similar saturation point. However, this plateau was partly artificial, caused by efficiency gains and offshoring of energy-intensive manufacturing.
The Fungibility Ladder: When Jevons Applies Hard vs. Soft
Fungibility Determines Jevons Strength
Fungible goods are interchangeable and can be used for many different purposes. The more fungible a resource, the stronger the Jevons effect. Hot water is specialized (limited uses); electricity is highly fungible (lighting, motors, heating, refrigeration, computation). Substrates—energy, information, atoms—are maximally fungible and have no obvious ceiling on demand.
Coal as the Fungible Substrate
Coal is highly fungible—the same amount can be used for many different things (heating, power, transport). This is why Jevons' prediction about coal was so accurate: as engines improved, coal found new applications across industries, and total consumption soared rather than fell.
Electricity: The Most Fungible Energy Form
Electricity is a substrate—it can become motion, heat, light, information, or chemistry. When it was expensive, it had one use: lighting. As it got cheaper, it enabled motors, refrigeration, heating, communications, and computation. Electricity demand is derived from demand for almost everything, making saturation extremely difficult.
The Printing Press Analogy: Predicting New Uses Is Impossible
Gutenberg Could Not Predict Newspapers or Teleprompters
Before the printing press, books were copied by monks and were extremely expensive. When Gutenberg invented the press, people might have guessed demand would double or triple. They could not have imagined newspapers, pamphlets, mass-market paperbacks, or modern digital text (like the teleprompter script Hank is reading). New applications emerge that are unimaginable until the substrate becomes cheap.
Code Is Probably Like That
Code has been expensive because every line required a human programmer. AI is making code dramatically cheaper. We have no idea how code-constrained the world is because code has never been cheap. Just as printed material spawned unforeseen descendants (newspapers, pamphlets, scripts), cheaper code will likely enable uses we cannot currently imagine—not just faster feature shipping, but entirely new categories of software.
Electricity Use in the US: A Temporary Plateau
Per Capita Electricity Flat Since Mid-2000s
From around 2005 to 2020, per capita electricity use in the United States was roughly flat, suggesting we had reached saturation—an 'enough' point. This appeared to contradict Jevons: efficiency was absorbing new demand. However, this plateau was partly artificial.
The Plateau Was Artificial: Offshoring and Efficiency
The flat electricity use was not purely saturation. Part of it came from efficiency improvements (better appliances, LED lighting) that reduced consumption without making electricity cheaper. But a large part came from offshoring energy-intensive manufacturing—demand moved to other countries' grids, not disappearing. The plateau was a statistical artifact.
Electrification Will Push Demand Higher
As cars, water heaters, and furnaces switch from fossil fuels to electricity, electricity demand will increase. Electrification is more efficient but increases total electricity consumption. Additionally, AI compute will add a new, unknown source of electricity demand—turning electricity into things no one previously imagined.
AI and Code: A New Jevons Moment
AI Is Making Code Dramatically Cheaper
Everyone Hank knows who codes is using AI coding tools, writing less code themselves and directing AI to generate code. Anthropic's annualized revenue is over 40 billion dollars (nearly profitable), suggesting massive demand. If AI algorithms continue improving, the cost of code will keep dropping, likely triggering a strong Jevons effect.
Code Is Highly Jevons-Shaped
Code is fungible—it can solve many different problem types. Unlike hot water, there is no obvious ceiling on code demand. If code becomes cheap, we will likely use vastly more of it, not just shipping features faster but creating entirely new categories of software applications (like how Gutenberg could not predict newspapers).
Why Code Is Ideal for AI
Code is a particularly good domain for AI because: (1) there is huge training data, (2) you can test whether code works by running it through a compiler—you get objective feedback. You cannot test whether an essay is good by compiling it or whether a book is good without human readers. Code has verifiable correctness, making it an ideal AI-shaped problem.
Intelligence as a Substrate: The Bigger Picture
Intelligence May Be a Substrate Like Energy
Hank proposes that there are four fundamental categories: atoms, energy, information, and intelligence. Energy manipulates atoms; intelligence manipulates information. Intelligence is fungible—it applies across biology, logistics, design, law, and countless other domains. If true, intelligence would be subject to the same Jevons dynamics as electricity: no obvious saturation point.
Intelligence Has Been Bottlenecked by Brains
We have no idea how intelligence-constrained the world is because intelligence has been limited by human brains and attention for all of Earth's history. If AI systems can manipulate information usefully across many domains, the demand for intelligence could be nearly unbounded—similar to electricity. This is why some AI researchers talk about building data centers in space or Dyson spheres.
The Jevons Question for Intelligence
If AI makes intelligence cheaper, does demand saturate (like hot water) or explode (like electricity)? The answer determines whether we face a bounded or unbounded growth in electricity consumption. Code suggests intelligence is highly fungible, but other domains might be more limited. The constraint is unknown.
What Could Constrain the Jevons Spiral?
Physical Constraints: Electricity Generation and Distribution
We might be capped by how fast we can build power generation and distribution infrastructure. If electricity demand grows faster than we can build capacity, electricity becomes the constraint. This is a physical, hard limit.
Information Constraints: Epistemic Limits
If intelligence becomes cheap, content becomes cheap, but figuring out what is actually true might become expensive. Distinguishing signal from noise in an ocean of cheap information could be a binding constraint. Wet-lab experimentation (e.g., studying cancer in fish) cannot be replaced by pure computation.
Cybersecurity: Attack Surface Explosion
If everything runs on cheap code and cheap intelligence, the attack surface becomes enormous. Defending against attacks might not scale at the same pace as creating new systems. Cybersecurity could be the constraint.
Policy and Law: Regulatory Constraints
We might have all the technology, energy, and intelligence but be legally prohibited from deploying it as fast as technologists want. Public policy and regulation could be the binding constraint.
Social and Distributional Constraints
If abundance is distributed badly enough, people may stop allowing it through democratic or social pressure. Fairness and human dignity could constrain growth. Additionally, human judgment about what to actually do with intelligence might be the limiting factor—AI does not inherently know what humans need.
The Constraint Is Probably Made of Meat
Hank suspects the ultimate constraint will involve human agency and physical bodies. Human judgment about what is worth doing, what is actually good for human flourishing, and how to distribute benefits fairly may be the binding constraint on AI growth.
Why We Cannot Predict the Future
Imagination Is a Function of What Is Cheap
You cannot imagine uses for expensive things. Before printing presses, no one imagined newspapers. Before electricity was cheap, no one imagined refrigeration or computation. Before code was cheap, no one imagined the software applications that now exist. This is why predicting AI's impact is nearly impossible—we cannot imagine what becomes possible when intelligence is cheap.
The Internet Revolution Proved Prediction Is Futile
Over the last 20 years, the communications revolution of the internet has shown that almost everyone's predictions about the future were wrong. Hank argues that too many things interact for meaningful long-term forecasting. Rather than guessing 20 years ahead, it is more useful to understand the present.
Anyone Claiming Certainty Is Selling Something
Anyone who tells you they know what AI will do 20 years from now is probably selling you something. The future is genuinely uncertain. The best we can do is understand present dynamics (like Jevons effects) and recognize constraints as they emerge.
Key Takeaways and Implications
Efficiency Is Almost Never the End of the Story
Making something more efficient does not necessarily reduce its use. If the resource is fungible (can be used for many things), efficiency increases total consumption by enabling new applications. This is true for coal, electricity, code, and likely intelligence.
We Are Living in Interesting Times
We are cursed with living at a transformative moment. The communications revolution (recommendation algorithms) has already reshaped what we see and think. AI is now making code and potentially intelligence cheaper. The Jevons paradox will shape how these changes unfold.
Human Judgment Remains Critical
AI may be good at solving technical problems, but human taste, human understanding of human problems, and human judgment about what is worth doing remain irreplaceable. The constraint on AI growth might ultimately be human wisdom about what to do with it.
Structure the World for Human Flourishing
If any of this is true, society should structure the world so that living humans—not AI or corporations—benefit from efficiency gains and abundance. This is both a moral imperative and pragmatic: human agency and judgment are likely to be the ultimate constraint on growth.
Notable quotes
It is absolutely not a paradox. — Hank Green
When you make use of a resource more efficiently, you often end up using more of it, not less. — Hank Green
Imagination is a function of what is currently cheap. — Hank Green