According to AI startups Hugging Face Every time you use AI to generate an image, write an email, or ask a question to a chatbot, you’re taking a toll on the planet, according to a new study by researchers at UC Davis and Carnegie Mellon University.
In fact, generating one image using a powerful AI model consumes as much energy as fully charging a cell phone, the study found.firstThe carbon emissions generated by using AI models for different tasks were calculated. However, they found that the energy consumption of using AI models to generate text was significantly lower. Generating text 1,000 times only consumed 16% of energy, which is equivalent to charging a mobile phone.
Source Note: The image is generated by AI, and the image is authorized by Midjourney
The study, which has not yet been peer-reviewed, is the work of researchersfirstSasha Luccioni, an AI researcher at Hugging Face, led a team to calculate the carbon emissions generated by using AI models for different tasks. She hopes that by understanding these emissions, she can help us make smart decisions when using AI and be more environmentally friendly.
Luccioni and her team studied the emissions of 10 popular AI tasks on the Hugging Face platform, including question answering, text generation, image classification, captioning, and image generation. They conducted experiments on 88 different models. For each task, such as text generation, Luccioni ran 1,000 prompts and measured the energy consumed using the Code Carbon tool she developed. The team also calculated the emissions generated by performing these tasks using 8 generative models that were trained to complete different tasks.
Generating images is an AI-based task that uses energy and carbon emissionsmaximumThe carbon dioxide emissions from using a powerful AI model to generate 1,000 images were roughly equivalent to driving 4.1 miles in an average gasoline car. In comparison, the least carbon-intensive text generation model in their study produced carbon dioxide emissions equivalent to driving 0.0006 miles in a similar vehicle. Stability AI, the company responsible for Stable Diffusion XL, did not respond to a request for comment.
The study provides concrete data on AI's carbon footprint and shows some worrying upward trends, said Lynn Kaack, an assistant professor of computer science and public policy at the Hertie School in Germany, who leads work related to AI and climate change and was not involved in the study. These emissions can add up quickly. The rise of generative AI has allowed large tech companies to incorporate powerful AI models into many different products, from email to word processing. These generative AI models are now used millions, if not billions, of times a day.
The team found that using large generative models to generate outputs consumes more energy than using smaller AI models that are tailored to specific tasks. For example, using a generative model to classify movie reviews based on whether the review is positive or negative consumes about 30 times more energy than using a tuned model created specifically for that task. The reason generative AI models consume more energy is that they try to accomplish multiple tasks at once, such as generating, classifying, and summarizing text, rather than just a single task, such as classification.
Luccioni said she hopes the research will encourage people to be more cautious when using generative AI and to choose models that are specialized for specific tasks and use less energy when possible. "If you're just doing a specific application, like searching emails... do you really need these big models that can do everything? I don't think so," Luccioni said.
The energy consumption of AI tools has been a missing factor in understanding their true carbon footprint, said Jesse Dodge, a research scientist at the Allen Institute for Artificial Intelligence who was not involved in the study. It’s also important to compare the carbon emissions of new large-scale generative models with older AI models. “This highlights the idea that the new wave of AI systems are more energy-intensive than the models we had two or five years ago,” he said.
Google once estimated that an average online search uses 0.3 watt-hours of electricity, equivalent to 0.0003 miles of driving a car. Today, that number is likely higher because Google has incorporated generative AI models into its search, said Vijay Gadepally, a research scientist at MIT Lincoln Laboratory who was not involved in the research.
Not only did the researchers find that emissions per task were much higher than they expected, they also found that emissions from everyday use of AI far outweighed emissions from training large models. Luccioni tested different versions of Hugging Face’s multilingual AI model, BLOOM, to see how many times it would take to outweigh the cost of training. It took more than 590 million uses to reach itsmaximumCarbon costs of model training. For very popular models, like ChatGPT, emissions from usage can exceed training emissions within a few weeks, Luccioni said. That’s because large AI models only need to be trained once and then used billions of times. According to some estimates, popular models like ChatGPT have as many as 10 million users per day, many of whom use the model multiple times.
Studies like this make AI-related energy consumption and emissions more concrete and help raise awareness of the carbon footprint of using AI, Gadepally said: "I hope this becomes something consumers start to pay attention to." Dodge said he hopes studies like this will help us hold companies more accountable for their energy use and emissions. "The responsibility here is on the companies that create the models and profit from them," he said.
