In the effort to mitigate climate change, technological efficiency is often framed as a primary solution. The logic is intuitive: if we make a process—such as training a neural network or delivering a package—more efficient, we should use fewer resources and emit less carbon. However, a recurring pattern across diverse research fields suggests this intuition is frequently wrong. This phenomenon, known as the rebound effect or Jevons Paradox, occurs when improvements in the efficiency of a resource lead to a reduction in its effective cost, which in turn stimulates a surge in demand that offsets the initial savings.
The Mechanism of Efficiency-Induced Demand
The core of the rebound effect is a systemic feedback loop. As Rolnick et al. explain, technological optimizations that reduce the resource intensity of a system also reduce the cost of using that system. When the cost of a service drops, users tend to consume more of it.
This effect manifests in two primary ways:
- Partial Rebound: Efficiency gains reduce consumption per unit, but increased total usage negates some of the savings.
- Jevons Paradox: The increase in demand is so significant that total resource consumption actually rises, despite the efficiency improvements.
For example, Przybyła et al. observe this in aviation: while aircraft fuel efficiency has improved, the carbon footprint of academic travel has not decreased because researchers now travel significantly longer distances. The efficiency gain was "consumed" by expanded mobility.
The Efficiency Paradox in Artificial Intelligence
The field of AI is currently experiencing a "Red AI" arms race, where state-of-the-art results are achieved through massive increases in compute. While many researchers focus on "Green AI" to make models more efficient, Luccioni argues that these technical optimizations often backfire.
When algorithmic speedups or hardware improvements lower the cost of computation, the savings are rarely "saved." Instead, they are typically reinvested into scaling. Schilter et al. note that hardware efficiency gains often lead to the development of exponentially larger models rather than a reduction in total energy consumption. Similarly, Selvan et al. point out that making model training more efficient may encourage researchers to run a higher volume of experiments, potentially increasing the total carbon footprint of the research process.
This paradox extends to model inference. Sel et al. demonstrate that increasing the token efficiency and speed of Large Language Models (LLMs) can trigger a rebound effect where total energy consumption rises because the lower per-task cost makes the models viable for a much wider range of frequent, high-volume applications.
Digitalization as a Double-Edged Sword
Digitalization is often touted for its potential to "dematerialize" the economy, but Creutzig et al. argue it is a double-edged sword. By reducing the friction and cost of services, digital tools can drive radical efficiency while simultaneously fueling rebound effects that increase total consumption.
A clear example of this tension is found in the transportation and logistics sector. Sun et al. describe an "informatization paradox" where digital infrastructure intended to optimize logistics actually inhibits carbon productivity. The efficiency gains in delivery systems accelerate e-commerce demand faster than technology can mitigate the resulting energy consumption. As Guo et al. suggest, understanding the magnitude of this rebound is critical for assessing whether digitalization truly supports long-term carbon neutrality.
Beyond Technical Optimization
If efficiency alone cannot guarantee sustainability, how should researchers and policymakers respond? The consensus across the literature is that technical optimizations must be integrated into a broader systemic framework.
- Systems Thinking: Easterbrook argues that traditional computational thinking, which focuses on optimizing individual components, must be supplemented with systems thinking to account for feedback loops and emergent behaviors like the rebound effect.
- Policy and Pricing: The authors of Tackling Climate Change with Machine Learning emphasize that efficiency improvements must be paired with regulatory frameworks, such as carbon pricing, to ensure that cost reductions do not simply drive a surge in demand.
- Holistic Metrics: Relying on narrow metrics can be misleading. Cao et al. demonstrate that Power Usage Effectiveness (PUE) in data centers can actually improve (numerically increase) even as total energy use decreases, suggesting that such metrics should not be the sole indicator of environmental performance.
Ultimately, as Wu suggests, scaling AI and other digital technologies sustainably requires moving "beyond efficiency" to address the aggregate consumption and deployment patterns that determine the net environmental impact.