greenai.guide

Home › Articles › Efficiency is not sufficiency: the rebound effect in AI

Article

Efficiency is not sufficiency: the rebound effect in AI

Updated · 5 min read

In short

Making each AI task cheaper does not mean AI as a whole uses less energy, because cheaper tasks tend to lead to more of them. This is the rebound or Jevons effect. Green AI needs efficiency (doing each task with less) and sufficiency (doing only the tasks worth doing), with the savings measured at the level of the whole system.

Key points

  • Efficiency lowers the energy per task; total energy depends on how many tasks are run.
  • When efficiency makes something cheaper, demand often grows, sometimes enough to cancel the saving. This is the rebound or Jevons effect.
  • Computing has become vastly more efficient for decades, yet its total electricity use has kept growing.
  • Sufficiency asks whether a task needs AI, or a model this large, at all.
  • Efficiency gains count as savings only when total use is capped, measured or priced.

Green AI is often presented as an efficiency problem: make models smaller, chips better and data centres cleaner, and the footprint falls. Efficiency matters, but on its own it does not guarantee that AI uses less energy. This article explains why, and what has to go alongside it.

What is the difference between efficiency and sufficiency?

Efficiency is doing a task with less; sufficiency is doing only the tasks worth doing. Both reduce environmental impact, but they answer different questions.

Efficiency Sufficiency
Question How do we do this with less energy? Do we need to do this, at this scale?
Examples Pruning, quantisation, better chips, cooling Not using AI where a search or a formula will do; using a small model instead of a frontier one
Measured as Energy per task Total energy, and value per unit of energy
Who acts Engineers, hardware makers, data-centre operators Users, product owners, buyers, policy makers

The two are not rivals. A 2026 UK green AI sandpit of mathematicians, computer scientists and industry ran a poll on whether green AI should focus on efficiency or on only using AI where needed; the most popular answer was "both equally". See what is green AI? for the wider definitions.

What is the rebound effect?

The rebound effect is the extra use that follows when something becomes cheaper or easier. If the extra use is large enough to outweigh the saving, total consumption rises. That extreme case is called the Jevons paradox (also known as the rebound effect in its milder forms).

The economist William Stanley Jevons described it in The Coal Question in 1865. He argued that more efficient steam engines would not reduce Britain's coal use, because cheaper steam power would be put to many more uses. He was right: coal consumption kept rising.

Rebound comes in degrees:

  • Direct rebound: a cheaper task is done more often. A faster, cheaper chatbot gets asked more questions.
  • Indirect rebound: money or time saved is spent on something else that also uses energy.
  • Economy-wide rebound: cheaper capability enables whole new products and industries.

Has this happened in computing?

Yes. Koomey and colleagues found in 2011 that the number of computations per unit of energy had doubled about every 1.57 years since the first electronic computers of the 1940s. Over the same decades, computing's total electricity use grew, because the cheaper computation was put to far more uses.

Freitag and colleagues' 2021 review of ICT's climate impact put the sector at 1.8–2.8% of global greenhouse gas emissions, or 2.1–3.9% once supply chains are fully counted. They argued that ICT's footprint has been growing in spite of large efficiency gains, and that rebound effects need to be accounted for before efficiency is counted as a climate saving.

Is AI following the same pattern?

The early signs suggest it is. The energy per AI task is falling fast, while total demand is rising.

  • Google reported in 2025 that the energy of its median Gemini text prompt fell 33-fold over twelve months, through better models, software and hardware.
  • Over a similar period, the International Energy Agency projected that data-centre electricity use would more than double by 2030, from around 415 TWh in 2024 to around 945 TWh, with AI the most important driver of the growth.

Both can be true at once. Cheaper prompts make it economical to put AI into search results, documents, phones and software tools by default, and to run agents that make many calls per task (see agentic AI and token waste).

Luccioni, Strubell and Crawford examined this in a 2025 paper on AI's polarised environmental debate. They argued that making AI more efficient will not by itself guarantee lower environmental harm, because efficiency gains can raise total use, and that the debate neglects these indirect effects. They call for life-cycle assessment combined with analysis of business incentives, markets, governance and social norms.

Why doesn't efficiency automatically become a saving?

Because the person who gains from efficiency usually spends the gain. Three mechanisms are common in AI:

  • Price. Providers pass on lower costs, so usage grows.
  • Capability. A saving at one size is spent on a bigger model or a longer context, because quality sells.
  • Defaults. When AI is cheap enough to switch on for everyone, it is switched on for everyone, including people who did not ask for it.

None of this is a reason to stop improving efficiency. Without it, total demand would be higher still. It is a reason not to count efficiency gains as climate savings until total use has been measured.

What does sufficiency look like in practice?

It means matching the tool to the task, and sometimes choosing no AI at all. Practical forms include:

  • Use a deterministic method when one exists. A calculator, a database query or a classical numerical method is usually cheaper and more reliable than a language model.
  • Right-size the model. Route simple requests to small models. See choosing the right model.
  • Make heavy AI opt-in, not default. Features that generate content on every page view multiply energy use.
  • Iterate less. Repeatedly regenerating an image or video until it looks right multiplies the cost of each attempt. Better prompting and editing tools reduce the number of tries.
  • Reuse instead of regenerating. See orchestration over generation.

At the sandpit, one theme was that the system, not each user, should make these choices where it can, for example by routing to the smallest adequate model automatically, because asking every user to judge the energy cost of every request does not scale.

How can rebound be kept in check?

By measuring, capping or pricing total use, not just use per task. Options include:

  • Budgets. Set an energy, token or carbon budget for a project or product, so efficiency gains free up headroom rather than automatically expanding use.
  • Whole-system measurement. Report total energy over time, alongside energy per task. See measurement is the bottleneck.
  • Honest claims. A claim that a product is "X times more efficient" says nothing about total impact. See making honest green claims.
  • Policy. Carbon pricing and grid constraints act on total demand in a way that efficiency standards alone do not. See green AI policy.

What to do next

  • Track total AI energy or spend over time, not only per-request figures.
  • Before adding an AI feature, ask whether a non-AI method would do, and whether it should be on by default.
  • Route requests to the smallest model that does the job: see use AI efficiently.
  • When you make an efficiency gain, decide in advance whether it will reduce total use or fund new use.
  • Read the business guide for how to build these choices into procurement and product design.

Frequently asked questions

What is the Jevons paradox?

The observation, first made by William Stanley Jevons in 1865 about coal, that using a resource more efficiently can increase total consumption of it, because efficiency makes it cheaper and so more widely used.

Does the rebound effect mean efficiency is pointless?

No. Efficiency is still needed, and without it total use would be higher. The point is that efficiency alone does not guarantee lower total use, so it has to be paired with sufficiency and measurement.

What is sufficiency in green AI?

Using AI only where it adds enough value, and using the smallest model and the least computation that does the job well. It asks 'should we do this?' as well as 'how do we do this more cheaply?'

Sources

  1. The Coal Question, W. Stanley Jevons, Macmillan (1865)
  2. From efficiency gains to rebound effects: the problem of Jevons' paradox in AI's polarized environmental debate, Luccioni, Strubell and Crawford, ACM FAccT (2025)
  3. 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)
  4. Implications of historical trends in the electrical efficiency of computing, Koomey, Berard, Sanchez and Wong, IEEE Annals of the History of Computing (2011)
  5. Energy and AI, International Energy Agency (2025)
  6. Measuring the environmental impact of delivering AI at Google Scale, Elsworth et al., Google (2025)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy