Contrary to popular belief, AI has a minimal effect on greenhouse gas emissions

Artificial intelligence is working its way into nearly every corner of the economy, and the pace keeps accelerating. In the United States, the share of companies using AI climbed to about 5.4 percent by early 2024, rising more than a point and a half in just five months. Analysts expect that number to tick up again within half a year. Adoption is happening overseas even faster. Some industries in Germany already report rates that top 27 percent.

This rapid growth reflects a larger hope. Businesses of all sizes believe tools such as computer vision and large language models will make work easier and more productive. At the same time, the excitement comes with a quiet concern. These systems use a lot of electricity. Running and cooling data centers requires constant power. Training the largest models can burn through as much energy as several hundred households do in one year. Even after training ends, day-to-day operation consumes far more. Some experts warn that data center electricity use could triple over the next decade.

Companies have worked hard to build more efficient hardware, but those gains have not kept up with the needs of newer and more powerful models. Much of the public debate focuses on this direct energy use. A new study steps back and asks a broader question that matters just as much. What happens to national energy demand when AI boosts productivity and helps the economy grow?

Researchers from the University of Waterloo and the Georgia Institute of Technology created what they describe as the first detailed estimate of AI’s wider energy and emissions impact on the U.S. economy. Their approach starts at the smallest scale, looking at how AI interacts with individual tasks, then expands to entire industries and their energy needs.

They examined 19,265 tasks scored by experts for their exposure to AI. A score of 1 meant a task was highly exposed. The team labeled tasks with scores above 0.5 as exposed and built two alternate ranges so they could check the strength of their results. They averaged these scores to estimate exposure within each occupation and then linked occupations to industries using wage data from the Bureau of Labor Statistics. This allowed them to tie the level of AI exposure to specific lines of work across the economy.

Once they calculated which industries stood to be most affected, they pulled in data from the World Input–Output Database. This provided information on economic output, energy use, and carbon emissions. By connecting thousands of tasks to broad economic sectors, they built a picture of how AI-driven productivity might influence energy demand.

To translate exposure into real-world change, the team used something called the Cost Savings Factor. This number captures how many tasks can be automated and how much labor cost can be saved when automation happens. Earlier studies suggest that AI can cut labor costs by about 27 percent in the tasks it replaces. They also estimate that about 23 percent of exposed tasks could be automated over ten years, which works out to about 2.3 percent per year. When combined, these numbers give a Cost Savings Factor of 0.0621.

The researchers used this single value to estimate how productivity improves across each industry. When productivity grows, output usually rises as well. That extra output then translates into higher energy use depending on how energy-intensive the industry already is.

The differences across industries are stark. Education, publishing, and motor vehicle retailing each generate roughly the same amount of output, but their energy needs per dollar could not be more different. Education requires 16 terajoules per million dollars of output.

Publishing needs only 0.005. Even with similar gains in productivity, education ends up with an additional 12 petajoules of energy use while publishing sees only a tiny fraction of that. Their emissions also diverge because publishing burns dirtier energy sources on average. These contrasts show why sweeping claims about AI’s energy footprint often miss the real story.

After combining the results across 55 industries, the researchers found that AI adoption would raise U.S. energy use by about 28 petajoules per year. That equals around 7.8 terawatt-hours of electricity. For perspective, one estimate suggests that running ChatGPT requires about 0.2 terawatt-hours each year. The study’s figure is nearly 40 times that amount, yet it still represents a very small share of national electricity use.

They also found that AI would raise carbon emissions by about 897 kilotons of CO2. That is less than 1 percent of emissions from U.S. manufacturing and construction and only a tiny fraction of national totals.

Even when the team tested extreme assumptions that pushed automation to its limits, the increases stayed small. The study suggests that the national impact of AI on energy use is modest, even though specific communities near data centers may feel much stronger local effects.

The authors are open about the limits of their work. Their model looks only at labor cost savings and does not include new tasks created by AI or the role of capital investment. That likely means the estimates undercount the true scale of economic change.

The energy data they used also ends in 2014. When they projected their results forward to 2023, the numbers shifted only slightly, which helped confirm their findings. Still, better data would sharpen future estimates.

Despite these limits, the takeaway is steady. AI will raise energy demand, but not enough to shift the national picture. At the same time, AI could help other sectors cut waste and use cleaner technology.

Dr. Juan Moreno-Cruz of Waterloo says the effects of AI will be felt most in the places where data centers draw their power. “Some places could see double the amount of electricity output and emissions,” he said. But at the national level, “AI’s use of energy won’t be noticeable.”

Moreno-Cruz believes the results should ease fears. “For people who believe that the use of AI will be a major problem for the climate and think we should avoid it, we're offering a different perspective,” he said. “The effects on climate are not that significant, and we can use AI to develop green technologies or to improve existing ones.”

He and coauthor Dr. Anthony Harding plan to repeat their analysis for other countries to see how AI adoption shapes energy use around the world.

These findings show that AI’s national energy and emissions footprint is small enough for leaders to manage with smart planning. The research suggests that local communities hosting data centers may see sharper increases in electricity demand, so regional planning and cleaner grids will matter.

The study also points toward a future where AI could help create more efficient systems, reduce waste, and support the development of greener technology. For policymakers, the message is clear.

AI can grow the economy without creating a major strain on national energy supplies as long as the country continues to expand low-carbon power sources.

Research findings are available online in the journal Environmental Research Letters.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.