Environmental impact of generative AI – 20 stats & facts

Generative AI is a branch of AI that specializes in creating new content such as text, images, videos, or sound. Generative AI creates seemingly original content by being trained on large datasets.

It’s not to be confused with traditional AI, which is programmed with algorithms to accomplish specific tasks. Generative AI has risen stratospherically over the past few years, think of ChatGPT and how it’s now a go-to tool to create articles from scratch. Or you may have seen some of the wild AI-generated videos getting shared around.

As sustainable marketers here at Akepa, we’re keen to move beyond being impressed (or not) and better understand AI’s environmental impact. We’re aware that traditional AI has the potential to positively impact the planet with tools that can predict the weather, identify pollution, improve waste management, and clean up marine plastic. A study by the Boston Consulting Group even stated that if AI is used wisely, it could help mitigate 5 to 10% of GHG emissions by 2030. But generative AI is a different proposition as it requires higher amounts of computational resources and is being used by masses of people at the same time.

So we’ve compiled 20 statistics on generative AI’s environmental impact and how it could be contributing to the climate crisis – mainly through energy use, carbon emissions, and water use.

1. In a study of 88 different AI models, it was found that a single AI-generated image can use as much energy as half a smartphone charge, using the least efficient model. Although, there is a large variation between image generational models. (ACM Digital Library, 2024)
2. The most carbon-intensive image generation model generates the amount of carbon equivalent to 4.1 miles driven by an average gasoline-powered passenger vehicle for 1,000 inferences (a prediction or response to a query). (ACM Digital Library, 2024)
3. But the least carbon-intensive image generation model generates 6,833 times less carbon, equivalent to 0.0006 miles driven by a similar vehicle. (ACM Digital Library, 2024)
4. AI-generated text requires significantly less energy than AI-generated images. Using the most efficient text generation model studied, creating text 1,000 times can use as much energy as 9% of a full smartphone charge. (ACM Digital Library, 2024)
5. Generating images is by far the most energy- and carbon-intensive AI-based task. (ACM Digital Library, 2024)
6. General, multi-purpose AI models are orders of magnitude more energy-intensive than task-specific models. (ACM Digital Library, 2024)
7. Training the bigger, more popular AI models like GPT-3 produced 626,000 pounds of carbon dioxide, equivalent to approximately 300 round-trip flights between New York and San Francisco—nearly five times the lifetime emissions of an average car. (University of Massachusetts, 2019)
8. Yet, AI’s predictions and responses to our queries (interference) impact the environment just as much or more than training AI models because interference happens far more frequently than model training. While training AI models remains more energy-intensive than inference, it could take just a couple of weeks or months for usage emissions to exceed training emissions for popular models like ChatGPT. (ACM Digital Library, 2024)
9. When comparing electricity demand, a single Google search takes 0.3 watt-hours of electricity, while OpenAI’s ChatGPT takes 2.9 watt-hours of electricity. That’s nearly 10 times as much electricity needed. (IEA, 2024)
10. If ChatGPT replaced the 9 billion Google searches daily, the electricity demand would require almost 10 terawatt-hours yearly. (IEA, 2024) That’s equivalent to the annual electricity consumption of 1.5 million EU citizens. (Our World in Data, 2020)
11. Data centers are responsible for powering generative AI. By 2028, Goldman Sachs analysts expect AI to represent about 19% of data center power demand. They also estimate that the AI revolution will cause data center power demand to grow by 160% by 2030. (Goldman Sachs, 2024)
12. But another study already states that roughly 25% of data center workloads are related to machine learning. (Nature Climate Change, 2022)
13. By 2030, data centers are predicted to emit triple the amount of CO2 annually than it would have without the boom in AI development. The amount of GHG emissions predicted, 2.5 billion tonnes, equates to roughly 40% of the U.S’s current annual emissions. (Morgan Stanley, 2024)
14. The computing power required for AI is doubling every 100 days and is projected to increase by more than a million times over the next 5 years. (SPJ, 2023)
15. A short conversation of 20-50 questions and answers with ChatGPT (GPT-3) costs half a liter of fresh water. (arXiv, 2023)
16. Training GPT-3 in Microsoft’s U.S. data centers can directly evaporate 700,000 liters of clean fresh water. That’s enough water to produce 370 BMW cars or 320 Tesla electric vehicles. Although, the “when” and “where” of training a large AI model can significantly affect the water footprint. (arXiv, 2023)
17. Google’s data centers in The Dalles, Oregon consume more than a quarter of all the water used in the city. (Oregon Live, 2023)
18. All told, a mid-sized data center consumes around 300,000 gallons of water a day, or about as much as 1,000 U.S. households, says Shehabi of Lawrence Berkeley National Laboratory. Their direct, on-site consumption ranks data centers among the top 10 water users in America’s industrial and commercial sectors. (NPR, 2022)
19. Although, some studies say the proportion of data center water consumption (1.7 billion liters/day) is small compared to total water consumption (1218 billion liters/day), at least in the U.S. (Nature, 2021)
20. Lastly, another area in which generative AI can harm the environment is through e-waste, where one study found that the e-waste generation of generative AI will grow at a rapid pace – 16 million tons of cumulative waste by 2030. (Research Square, 2024). Meanwhile, e-waste is one of the fastest-growing waste streams in the world.

These are all pretty convincing stats, but still, you – like us – might wonder how much negative impact this causes in the real world. Practically speaking, what does this look like? Is using ChatGPT worse than eating animal products? Taking a flight a few times a year? Throwing away food regularly? Googling?

We don’t have all the answers but we should remind ourselves that because of the generative AI boom, the seemingly-small emissions could be adding up quickly.

So, should we cancel AI and revert to regular search? It’s a bit more complicated than that. Technological advancements could decrease AI’s energy intensity. Big tech companies are working on reducing electricity and water consumption in their data centers. If they’re successful, we can harness the climate-saving benefits of AI without harming the climate. The International Organization for Standardization (ISO) is also developing sustainable AI standards to be rolled out by late 2024, aiming to lessen AI’s environmental footprint and empower users to make informed choices about AI usage.

And having this dialogue is a good first start. Model builders need to implement sustainability into their design choices and the more this issue is brought up, the likelier it is to be addressed.

Maybe one thing we should be doing though, is being more judicious with its use.