In short
The Jevons paradox is the finding, first set out by William Stanley Jevons in 1865, that using coal more efficiently led Britain to burn more of it, not less. In AI, each query is becoming far cheaper and more efficient, yet total data-centre electricity use is rising fast. Researchers studying digital technology's climate impact argue that efficiency only delivers real emission cuts when total emissions are capped.
Key points
- The Jevons paradox: making a resource cheaper to use can raise total use of it. Jevons showed this for coal in 1865.
- Rebound can be direct (more of the same task), indirect (savings spent elsewhere) or economy-wide; when it outweighs the saving it is called backfire.
- Google reported a 33-fold fall in the energy of its median Gemini text prompt in a year, while the IEA projects data-centre electricity use will more than double by 2030.
- Freitag, Berners-Lee and colleagues found ICT's climate footprint is underestimated and argue that claims of ICT-enabled savings rest on weak evidence.
- Widdicks and colleagues conclude that efficiencies under emissions constraints are the only way to turn efficiency into true emission reductions.
- In practice: measure totals, set budgets or caps, retire freed capacity and use AI only where it adds value.
AI is getting more efficient very quickly, and its total energy use is growing very quickly. Those two facts are not a contradiction. They are a 160-year-old pattern known as the Jevons paradox. This article explains the paradox, the evidence that it applies to AI, the concerns researchers have raised about treating digital technology as a climate solution, and what actually turns efficiency into lower emissions. For the basics of efficiency and sufficiency, start with efficiency is not sufficiency.
What is the Jevons paradox?
The Jevons paradox is when using a resource more efficiently leads to more of it being used, not less. The economist William Stanley Jevons set it out in The Coal Question in 1865, a book about whether Britain's coal mines would run out.
Jevons disagreed. He wrote that "it is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth." His reasoning was that cheaper steam power made coal worth using for many more jobs, so total demand rose faster than each engine's use fell.
The insight is not that efficiency is bad. It is that efficiency changes the price of doing something, and people, businesses and whole economies respond to prices.
What is the rebound effect?
The rebound effect is the extra use that an efficiency gain triggers, which eats into the saving. The Jevons paradox is its most extreme form. Widdicks and colleagues, in a 2023 paper in the journal Patterns, describe the main types:
| Type | What happens | AI example |
|---|---|---|
| Direct rebound | The same service gets cheaper, so it is used more | Cheaper prompts lead to more prompts, longer answers and more agent calls |
| Indirect rebound | Money or time saved is spent on other things that also use energy (the "income effect" and "time rebound") | Time saved by an AI assistant is spent on more energy-intensive activities |
| Economy-wide rebound | The wider economy adjusts: new products, markets and industries grow on the cheaper input | AI built into search, phones and office software by default |
When the rebound is bigger than the saving, so that total use or emissions rise, it is called backfire. The same authors note that across the global economy, energy efficiencies in most products and services have gone hand in hand with average growth in global carbon emissions of 1.8% a year since 1850. They are careful to say this does not prove efficiency causes higher emissions, only that for it not to, emissions would have had to rise much faster without efficiency.
See the glossary for rebound effect and Jevons paradox.
Is the Jevons paradox happening in AI?
The evidence points that way: energy per query is falling fast while total data-centre electricity use climbs. Two widely cited figures, from different sources, show the pattern:
- Per query. Google reported in 2025 that its median Gemini Apps text prompt uses 0.24 Wh of energy, and that over a 12-month period the energy per median prompt fell 33-fold and its carbon footprint 44-fold.
- In total. The International Energy Agency (IEA) estimates that data centres used around 415 TWh in 2024, about 1.5% of the world's electricity, and that this has grown by around 12% a year since 2017. It projects use will more than double to around 945 TWh by 2030, with AI "the most important driver of this growth".
These figures do not by themselves prove cause and effect, because AI demand also grows for reasons other than price. But they show that big per-query gains have not stopped total use rising. The IEA itself treats the speed of efficiency improvements as one of the main uncertainties in its outlook: across its scenarios, data-centre demand in 2035 ranges from 700 to 1,700 TWh.
AI has features that make rebound especially likely. Cheaper inference lets providers make AI the default in tools people already use, and lets agents make many model calls per task (see agentic AI and token waste). Savings are also often spent on capability, such as bigger models or longer context, rather than banked.
What concerns have Mike Berners-Lee and colleagues raised?
They argue that digital technology's climate impact is underestimated, that its efficiency gains are routinely cancelled by demand, and that claims of climate benefit need far better evidence. Mike Berners-Lee, founder of Small World Consulting, based at Lancaster University, is a carbon-footprinting specialist and co-author of two papers in Patterns on this question.
The footprint is bigger than reported
Freitag, Berners-Lee, Widdicks, Knowles, Blair and Friday reviewed estimates of the climate impact of information and communication technology (ICT) in 2021. They found that peer-reviewed studies put ICT at 1.8–2.8% of global greenhouse gas emissions, but that the share could be 2.1–3.9% once supply chains left out of standard life-cycle studies (the "truncation error") are counted.
Efficiency has not cut the total
The same review found that demand for computation and devices per person had outpaced efficiency improvements, so ICT's energy use and footprint kept growing year on year. All the analysts it reviewed agreed that ICT emissions "will not reduce without major concerted efforts involving broad political and industrial action". The authors gave reasons to expect emissions to rise without intervention, and judged a natural peak from saturated demand unlikely. In fairness, they also report the contrasting view of the researcher Jens Malmodin, given to them in personal communication, that ICT emissions have stabilised for now and could halve by 2030.
"AI for climate" claims lack evidence of net benefit
Industry reports have claimed ICT could save several times its own emissions in other sectors. Freitag and colleagues argue that these estimates give rebound very limited treatment, and that ICT is as likely to enable growth, including in carbon-intensive industries, as to cut emissions. They write that "despite questionable evidence that ICT growth trends will save more carbon emissions than it will introduce", blockchain, the internet of things and AI are seeing increased investment and uptake. ICT only saves emissions elsewhere, they argue, if it replaces carbon-intensive activity rather than being added on top of it.
This does not mean AI cannot help the climate. It means each claim needs to show a net benefit, with the AI's own footprint and rebound counted. See green AI vs AI for green.
Efficiency needs a cap
The 2023 paper by Widdicks and colleagues, drawing on a workshop of 19 experts, concludes that "efficiencies under GHG emission constraints are the only way in which efficiencies for true emission reductions can be realized". The 2021 review argues that without a global carbon constraint, new regulation is needed to keep ICT's footprint in line with the Paris Agreement. Both papers present a cap as good for the sector too: under a constraint, efficiency becomes more valuable, not less.
In his own words
Speaking to government digital leaders in London in February 2026, Berners-Lee summed up the rebound problem as "the more efficient it gets, the more we do", and warned that "a lot of people are throwing a sledgehammer of AI and machine learning at every problem", according to a report by THINK Digital Partners.
What turns efficiency into real savings?
Pair every efficiency gain with a limit on the total, and measure the total. The research above points to five practical levers.
Work under a budget or cap
Set an absolute budget for energy, emissions, compute or tokens for a product, team or project, and keep it as efficiency improves. Efficiency then lets you do more within the budget instead of pushing the total up.
Measure absolute impact, not just per-unit figures
A claim that a model is "ten times more efficient" says nothing about whether total emissions fell. Report total energy and emissions over time alongside energy per task. See how to measure AI energy.
Practise sufficiency
Use AI where it adds real value, and the smallest adequate model when you do. A formula, a search or a database query often does the job with a fraction of the energy. See sufficiency.
Bank the savings
When an efficiency project frees capacity, decide in advance what happens to it. Retiring servers, reserved cloud capacity or licences locks in the saving; leaving them running invites new use to fill them. See how to stop efficiency gains being cancelled out.
Use policy levers that act on totals
Carbon pricing, emissions caps, grid connection limits and reporting of absolute energy use act on total demand in a way that efficiency standards alone cannot. See green AI policy.
What to do next
- Add total AI energy or emissions to your regular reporting, next to per-task figures.
- Set a budget or cap for AI compute or emissions, and keep it when efficiency improves.
- Before each efficiency project, decide whether the freed capacity will be retired or reused.
- Ask whether each new AI feature is needed, and whether it should be on by default.
- Treat any claim that AI will cut emissions elsewhere as a hypothesis to test, with rebound counted.
- Read efficiency is not sufficiency and green AI vs AI for green for the wider picture.
Frequently asked questions
What is the Jevons paradox?
The observation by the economist William Stanley Jevons, in his 1865 book The Coal Question, that more efficient use of coal led to more coal being used, because efficiency made coal-powered work cheaper and so more widespread.
Does more efficient AI reduce emissions?
Per query, yes. In total, not necessarily: cheaper queries encourage far more use. Total data-centre electricity use is still rising quickly even as the energy per prompt falls, so efficiency only cuts emissions overall if total use or emissions are also limited.
What is the rebound effect in AI?
The extra AI use that follows when AI becomes cheaper or easier, which eats into the energy saved. It can be direct (more prompts), indirect (money or time saved is spent on other energy-using things) or economy-wide (new products and industries built on cheap AI).
What is backfire?
Backfire is a rebound effect so large that total consumption or emissions end up higher than before the efficiency gain. The Jevons paradox is an example of backfire.
What do Mike Berners-Lee and his co-authors say about AI and climate?
In peer-reviewed papers they argue that ICT's footprint is underestimated, that efficiency gains tend to be cancelled by growing demand, that evidence for net climate benefits from ICT, including AI, is questionable, and that a cap on emissions is needed for efficiency to deliver real cuts.
How do you stop efficiency gains being cancelled out?
Measure total energy or emissions as well as energy per task, set a budget or cap, decide in advance what happens to freed capacity, and use AI only where it adds enough value.
Sources
- The Coal Question, W. Stanley Jevons, Macmillan (digitised copy, Internet Archive) (1865)
- Systems thinking and efficiency under emissions constraints: Addressing rebound effects in digital innovation and policy, Widdicks, Lucivero, Samuel, Somavilla Croxatto, Smith, Ten Holter, Berners-Lee, Blair, Jirotka, Knowles, Sorrell, Börjesson Rivera, Cook, Coroamă, Foxon, Hardy, Hilty, Hinterholzer and Penzenstadler, Patterns (2023)
- The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations, Freitag, Berners-Lee, Widdicks, Knowles, Blair and Friday, Patterns (2021)
- Energy and AI, International Energy Agency (2025)
- Measuring the environmental impact of delivering AI at Google Scale, Elsworth et al., Google (2025)
- "We're heading for something pretty grim": Mike Berners-Lee warns digital leaders on climate, THINK Digital Partners (2026)
Related
Efficiency is not sufficiency: the rebound effect in AI
Why making AI more efficient does not guarantee it uses less energy overall, what the rebound or Jevons effect is, and what helps keep savings real.
Green AI vs AI for green: the difference
Green AI cuts the footprint of AI itself; AI for green uses AI for climate and nature. How they differ, why they go together and how to judge net benefit.
How to write a green AI policy
A step-by-step guide to writing an organisation's green AI policy, with a template outline covering use, models, measurement, buying, claims and review.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.