Sign up for CNN Life, but an eco-friendly newsletter. Our exclusive newsletter series guides you on how to minimize your personal role in the climate crisis and reduce environmental anxiety.
Whether you respond to work emails or draft wedding vows, generative artificial intelligence tools have become trusted co-pilots in many people's lives. However, an ever-growing number of research shows that hidden environmental costs are rising for all the problems AI solves.
Each word in the AI prompt is divided into clusters of numbers called “token IDs” and sent to a large data center (larger than a soccer field) equipped with coal or natural gas plants. There, a large stack of computers generates responses through dozens of rapid calculations.
According to frequently cited estimates by the Electric Power Institute, the entire process can complete up to 10 times the energy of a regular Google search.
So, what damage does it do to AI at each prompt? To investigate, German researchers tested 14 leading language model (LLM) AI systems by asking both free responses and multiple choice questions. Complex questions produced up to six times more carbon footprint than questions with a brief answer.
Furthermore, to answer the same question, the “smartier” LLM, which has more inference capabilities to generate up to 50 times the carbon emissions of a simpler system, reported the study.
“This illustrates the trade-off between energy consumption and accuracy of models' performance,” said Maximilian Dauner, a doctoral student at Hochschule München University of Apply Sciences, the first author of the frontier of communication studies published Wednesday.
Typically, these smarter, more energy-intensive LLMs have bias (more parameters) used to process token IDs than small, concise models.
“You can think of it like a neural network in the brain. The more neurons you have, the more you can do to answer the question,” says Dauner.
What you can do to reduce your carbon footprint
According to Dauner, complex questions require more energy for the lengthy explanations that many AI models are trained to provide. He said that when you ask an AI chatbot to solve an algebraic question, you may take the necessary steps to find the answer.
“AI carefully spends a lot of energy, especially when users say “please” and “thank you,” Dauner explained. “But this will make their responses longer and consume more energy to produce each word.”
For this reason, Dauner suggests that users will be easier when communicating with the AI model. Specify the length of the answer you want, limit it to one or two sentences, or you don't need any explanation at all.
Most importantly, Dauner's research highlights that not all AI models are created equal, while Sasha Lucciioni, the AI company's climate lead, emphasizes her embrace in her face in emails. Users who want to reduce their carbon footprint may be more intentional about the model they choose in any task.
“Task-specific models are often much smaller, more efficient and better for context-specific tasks,” explained Lucciioni.
If you are a software engineer who solves complex coding problems every day, you may need an AI model that is right for your coding. But for the average high school student seeking homework help, relying on powerful AI tools is like using a nuclear-powered digital calculator.
Even within the same AI company, offering different models can have different inference power, so we investigate the best capabilities for our needs, Dauner said.
If possible, Lucciioni recommends going back to the basic sources of online encyclopedias and telephone calculators to accomplish simple tasks.
Why it's difficult to measure the environmental impact of AI
It turns out to be challenging to put numbers on the environmental impact of AI.
This study noted that energy consumption could differ based on the user's proximity to the local energy grid and the hardware used to run the AI model.
That's part of why researchers chose to represent carbon emissions within their scope, Dauner said.
Additionally, many AI companies do not share information about energy consumption — or details such as server size and optimization techniques that help researchers estimate energy consumption, said Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, studying AI water consumption.
“I can't say that on average AI consumes so much energy and water. That just makes no sense. You need to look at individual models and then (find what you use) in each task,” says Ren.
One way AI companies could become more transparent is to disclose the carbon emissions associated with each prompt, Dauner suggested.
“In general, if people are informed of more about the average (environmental) cost of generating responses, people say, “Is it really necessary to turn it into an action figure just because they're bored?” or, “Do you need to tell ChatGpt jokes because they have nothing to do?” Dauner said.
Furthermore, as more companies push to add generative AI tools to their systems, people may not be able to choose how or when to use the technology, Lucciioni said.
“We don't need the generation AI for web searches. No one was asked for it on AI chatbots (messaging apps) or social media,” says Lucciioni. “This race, packing them into all existing technology, is really infuriating because it brings real results to our planet.”
With less information on AI resource use, consumers have fewer options, Ren said, adding that regulatory pressure on more transparency is unlikely in the US soon. Instead, the best hope for more energy-efficient AI could be the cost effectiveness of using less energy.
“Overall, I'm still positive (about the future) and many software engineers are working hard to improve the efficiency of their resources,” Ren said. “Other industries also consume a lot of energy, but this is not a reason to suggest that the environmental impact of AI is not an issue. We definitely should be paying attention.”
Create an account at cnn.com for other CNN news and newsletters
