In the quest for sustainable technology, the prevailing mantra is "efficiency." If we can make a Large Language Model (LLM) use 50% less energy per inference, or train a Vision Transformer in half the time, we assume the environmental footprint of the industry will shrink accordingly.
However, a growing body of research suggests this intuition is often wrong. Across fields ranging from edge computing to urban logistics, researchers are documenting a phenomenon known as the rebound effect or Jevons Paradox: technical improvements that increase the efficiency of a resource often lead to a net increase in its total consumption.
The Mechanism of Induced Demand
The core of the rebound effect is economic. As Noman Bashir and Prashant Shenoy argue in Sustainable Computing - Without the Hot Air, computing resources are highly price-elastic. When algorithmic optimizations lower the marginal cost of a single computation, the "saved" resources do not simply disappear. Instead, the lower cost stimulates new demand that previously would have been too expensive to justify.
Dustin Wright and Raghavendra Selvan describe this in Efficiency is Not Enough: A Critical Perspective of Environmentally Sustainable AI as a socio-technical feedback loop. In AI, this manifests in three primary ways:
- Scaling Up: If training becomes 2x more efficient, researchers may choose to train a model that is 2x larger or uses 2x more data to achieve higher accuracy, rather than pocketing the energy savings.
- Increased Accessibility: As Radosvet Desislavov and Jose Hernandez-Orallo note in Compute and Energy Consumption Trends in Deep Learning Inference, efficiency makes AI viable for billions of IoT and mobile devices, potentially increasing the global volume of computations enough to offset per-unit gains.
- Task Proliferation: When content creation becomes nearly effortless through generative models, the sheer volume of output can negate per-task energy savings, a point emphasized by Bill Tomlinson in Reconciling the contrasting narratives on the environmental impact of large language models.
The Embodied Carbon Trap
The focus on "operational efficiency"—how much power a chip uses while running—often ignores the "embodied carbon" required to manufacture the hardware. Shvetank Prakash and Matthew Stewart highlight this in their analysis of TinyML in Is TinyML Sustainable?.
While a single microcontroller uses negligible power, the rebound effect suggests that their extreme efficiency will lead to the deployment of billions of devices. If the carbon footprint of manufacturing these billions of sensors exceeds the efficiency gains they provide in sectors like agriculture or buildings, the technology results in a net environmental loss. This "sensor-blind" approach to sustainability, as argued in Materiality and Risk in the Age of Pervasive AI Sensors, fails to account for the total lifecycle of the hardware.
The Informatization Paradox
The rebound effect is not limited to the AI lab; it propagates through the entire digital economy. In Spatiotemporal dynamics and driving mechanisms of total factor carbon productivity in China's provincial transportation sector, Yanming Sun and Jiale Liu identify an "informatization paradox." While digital tools optimize logistics and reduce emissions per mile, they simultaneously fuel a surge in e-commerce and delivery demands. This increased consumption scales faster than technological efficiency can mitigate it, leading to a net decrease in carbon productivity.
Similarly, Felix Creutzig and Lynn H. Kaack argue in Digitalization and the Anthropocene that digitalization reduces the friction of services, which triggers rebound effects that concentrate political power and drive up total resource use, complicating the narrative of digital "dematerialization."
Shifting from Efficiency to Redirection
If efficiency alone is insufficient, how can the research community achieve true sustainability? Several researchers propose moving beyond narrow technical optimization:
- Ecological Redirection: Valentin Girard and Antoine Martin advocate for a framework that moves from "Green IT" toward the active renunciation of unsustainable digital practices in A Protocol to Address Ecological Redirection for Digital Practices in Organizations.
- Carbon-Awareness: Instead of just using less energy, systems should align computation with the availability of carbon-free energy. Benjamin C. Lee and Udit Gupta propose Carbon Connect, a research agenda that integrates demand response and circular economy principles to tackle both embodied and operational carbon.
- Transparency and Policy: Alexandra Sasha Luccioni and Meredith Whittaker emphasize in Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI that technical gains must be paired with usage-aware sustainability strategies and democratic oversight to prevent efficiency from simply fueling further excess.
Ultimately, the research suggests that while efficiency is a necessary tool, it is not a goal in itself. Without accounting for the economic and behavioral shifts that follow optimization, the "Green AI" of today may inadvertently become the high-carbon infrastructure of tomorrow.