AI Boosts Efficiency but Increases Carbon Emissions? The Modern Replay of the Jevons Paradox

Why AI Efficiency Gains May Actually Increase Carbon Emissions

When we talk about artificial intelligence, a common optimistic narrative is that AI will help humanity tackle climate change by optimizing energy systems and improving production efficiency. However, a recent study based on a global energy-economy model has reached a counterintuitive conclusion—the productivity gains brought by AI may actually drive a net increase in global CO₂ emissions.

This finding challenges the widely held "AI equals green" optimism across the industry, and reveals the complex and nuanced relationship between technological progress and environmental costs.

The Core of the Paradox: A Modern Interpretation of the Jevons Effect

What Is the Jevons Effect?

To understand why AI might exacerbate carbon emissions, we first need to revisit a classic phenomenon in economics—the Jevons Paradox. In the 19th century, British economist William Jevons observed that improvements in steam engine efficiency didn't reduce coal consumption. Instead, because the cost of using coal decreased and its applications expanded, total coal consumption actually rose dramatically.

The essence of this paradox is: when the efficiency of using a resource improves and its unit cost decreases, people tend to use it on a much larger scale, and ultimately total consumption exceeds the savings from the efficiency improvement. This is what the "rebound effect" looks like in its extreme form.

The Jevons Paradox is not merely a historical curiosity—it has rigorous theoretical support in economics. At the micro level, the cost reduction from efficiency improvements produces two effects: the substitution effect (replacing other inputs with the now-cheaper resource) and the income effect (funds freed by cost savings are used to expand production or consumption). When the income effect exceeds the substitution effect, the rebound effect surpasses 100%, resulting in what's called "backfire"—total consumption rises rather than falls. Since the 20th century, this phenomenon has been repeatedly verified across multiple domains: LED lighting is several times more efficient than incandescent bulbs, yet global total lighting energy consumption hasn't significantly decreased, because cheap lighting has spawned countless new lighting applications.

How AI Replays This History

The role AI plays in the energy-economy system is remarkably similar to that of the steam engine. AI improves productivity across industries and reduces the marginal cost of products and services. When economic activity becomes more efficient and cheaper, the overall economic scale expands accordingly—more production, more consumption, more logistics and transportation—all of which require energy to sustain.

From an economics perspective, AI's impact on productivity manifests as an increase in Total Factor Productivity (TFP). TFP measures the additional output growth attributable to technological progress and organizational optimization when labor and capital inputs remain constant. Many economists believe AI's impact on TFP could rival the spread of electricity and the internet—it's not a technological upgrade for a single industry, but a General Purpose Technology capable of penetrating virtually every economic sector. Historical data shows that every leap in TFP (such as the Industrial Revolution and the Information Revolution) has been accompanied by dramatic GDP growth and a simultaneous climb in energy consumption. Unless the energy structure itself undergoes a fundamental transformation (such as complete decarbonization), the coupling between economic growth and carbon emissions is extremely difficult to break.

In other words, every kilowatt-hour and every unit of computing power that AI saves may be "consumed" by the economic expansion that follows—and even exceeded. The higher the productivity, the larger the economic volume, and the greater the corresponding energy demand and carbon emissions.

Systemic Risks Revealed by Global Energy-Economy Models

What makes this study noteworthy is that it employed a global integrated energy-economy model for its projections, rather than analyzing AI's carbon emissions from a partial or single-industry perspective.

Global integrated energy-economy models (such as GCAM, MESSAGE, REMIND, etc.) are core tools in climate change research. These models typically cover multiple subsystems including energy supply, energy conversion, end-use energy, economic growth, and land use, coupling them through price signals, technology diffusion curves, and policy scenarios. Unlike partial life cycle assessments, these models can capture cross-industry and cross-regional feedback loops—for example, after AI improves manufacturing efficiency, it lowers consumer product prices, stimulates consumer demand, and in turn drives energy consumption in upstream segments like logistics, warehousing, and raw material extraction. Such cascading effects can only be fully quantified within a system dynamics framework.

Looking at a single data center, AI optimization can indeed reduce energy consumption; looking at a single production line, automation can indeed reduce waste. But when the perspective is elevated to the global system level, these localized efficiency gains are amplified and transformed through the transmission mechanisms of economic growth, ultimately presenting as a net increase in total carbon emissions.

This reminds us that assessing AI's environmental impact cannot be limited to "direct accounting" (such as the electricity consumed by training large models or data center cooling). We must also do "systems accounting"—how AI changes the way the entire economy operates and its scale.

Direct Energy Consumption Is Just the Tip of the Iceberg

Currently, public discussion focuses heavily on AI's direct energy consumption, such as the enormous computing power required to train GPT-class large language models and the continuously rising electricity demand of data centers. These are indeed real sources of carbon emissions.

Data from the International Energy Agency (IEA) shows that global data centers consumed approximately 460 terawatt-hours of electricity in 2022, accounting for about 2% of global electricity consumption. With the explosion of generative AI, this figure is climbing rapidly—training a GPT-4-class large language model is estimated to consume tens of millions of kilowatt-hours of electricity, and the continuous operation during the inference phase represents an ongoing energy burden. Goldman Sachs predicts that by 2030, global data center electricity demand could increase by 160% compared to 2023. Yet even so, these direct energy consumption figures remain a limited share of total global emissions. What truly has systemic impact is the economy-wide energy demand growth that AI drives as a productivity engine.

But this study points out that indirect, systemic emissions growth may be the greater hidden concern. By improving total factor productivity, AI leverages the expansion of the entire macroeconomy, and its carbon footprint is far more massive and concealed than the energy consumption of data centers themselves.

Practical Implications for Policy and Industry

Technological Optimism Needs Tempering

For a long time, the tech industry has tended to respond to external criticism of its energy consumption with the narrative that "AI will save the planet." This study reminds us that such a narrative may be overly simplistic. Improvements in technological efficiency do not necessarily lead to environmental improvement—what matters is how institutions and policies direct where efficiency gains go.

Complementary Carbon Reduction Mechanisms Are Needed

If AI-driven efficiency gains automatically translate into economic expansion and carbon emissions growth, then relying solely on technological progress cannot solve the climate problem. This means strong external constraint mechanisms must be in place—such as carbon pricing, carbon taxes, and emissions caps—to ensure that efficiency gains truly translate into emissions reductions rather than being swallowed by growth.

Carbon pricing is a policy tool that internalizes the external costs of greenhouse gas emissions, primarily taking two forms: carbon taxes and Emissions Trading Systems (ETS). A carbon tax raises the cost of fossil fuel use by imposing a fixed fee per ton of CO₂ emitted, while an ETS achieves reduction targets by setting an overall emissions cap and allowing companies to trade emission allowances. As of 2024, over 70 carbon pricing mechanisms are operating globally, covering approximately 23% of global greenhouse gas emissions. However, most carbon prices remain far below the level needed to achieve the Paris Agreement goals (estimated at $75-150 per ton). In a scenario where AI accelerates economic growth, the intensity and coverage of carbon pricing need to increase correspondingly; otherwise, emissions growth from efficiency gains will easily outpace decarbonization efforts.

In other words, we cannot expect AI to "automatically" turn green—we need to proactively design an institutional framework that makes AI serve emissions reduction goals.

Conclusion: Beware of Technological Solutionism

The value of this study lies not in being pessimistic about AI, but in providing a more clear-eyed and comprehensive perspective. AI is undoubtedly a powerful productivity tool, but its impact on climate depends on how we use it and how we manage the accompanying economic growth.

History has already taught us this lesson through the Jevons Paradox—efficiency improvements don't necessarily mean resource conservation. In the AI era, we may be repeating the same story. Only by confronting this paradox head-on and embedding technological progress within sound policy and institutional frameworks can AI truly become an ally in addressing climate change rather than an invisible accelerant.

It should be noted that this study is currently based on model projections, and its specific assumptions, parameter settings, and conclusions still await broader academic discussion and real-world verification. But the questions it raises undoubtedly deserve serious attention from the entire tech community and policymakers alike.