Study shows: AI image generation increases carbon footprint

Carnegie Mellon University and researchers from the machine learning community website Hugging Face report that if you are one of the more than 10 million users who use machine learning models every day, you may still be contributing to climate change.

In their first systematic comparison of costs related to machine learning models, the researchers found that the energy required to generate an image using AI models is equivalent to the energy required to charge a smartphone.

"People think that AI has no environmental impact, that it is such an abstract technological entity that exists in the 'cloud'," said team leader Alexandra Luccioni. "But every time we query an AI model, it has a certain cost to the Earth, and calculating that is very important."

Her team tested 30 datasets using 88 models and found variations in energy usage between different tasks. They measured the amount of carbon dioxide emissions for each task.

The most energy-consuming task was Stability AI's Stable Diffusion XL, an image generator. It produced nearly 1600 grams of carbon dioxide in one session. Luccioni said this is roughly equivalent to driving four miles in a gasoline car.

On the lower end of energy consumption, a basic text generation task consumed energy equivalent to a car driving only 3/500 miles.

Other categories of machine learning tasks include image and text classification, image captioning, summarization, and question answering.

The researchers noted that generative tasks, such as image generation and summarization, consume more energy and carbon compared to discriminative tasks, such as movie ranking.

They also observed that using multi-purpose models for discriminative tasks consumes more energy compared to using task-specific models for the same tasks, which is significant given the recent trend of model usage.

"Given the shift from small models tailored to specific tasks to models that perform multiple tasks simultaneously and are deployed to respond to a large number of user queries in real-time, we believe this finding is the most striking conclusion of our research," the report said.

According to Luccioni, "If you are doing a specific application, like searching emails... do you really need these large models that can do anything? I would say no."

While the carbon dioxide usage for these tasks may seem small, when multiplied by the millions of users who rely on AI-generated programs every day and frequently make multiple requests, the total shows a significant impact on efforts to control environmental waste.

"I think, for overall generative AI, we should be aware of where and how we use it, comparing its costs and benefits," Luccioni said.