Exploring the Hidden Environmental Costs of AI-generated Images
Published on: March 10, 2024
A groundbreaking study conducted by researchers at AI startup Hugging Face and Carnegie Mellon University has for the first time calculated the carbon emissions resulting from using AI models for various tasks. The study finds that generating an image using a powerful AI model consumes as much energy as fully charging a smartphone.
The research, led by Sasha Luccioni from Hugging Face, indicates that while training large AI models is known for being energy-intensive, a significant part of their carbon footprint actually comes from their usage. The team evaluated the emissions from 10 different AI tasks on the Hugging Face platform, utilizing 88 models, and found that image generation is particularly carbon-intensive.
Interestingly, the study revealed that creating text 1,000 times uses only about 16% of the energy required for a full smartphone charge. In contrast, generating 1,000 images with models like Stable Diffusion XL can emit as much CO2 as driving a car for 4.1 miles.
The team also discovered that large generative models, capable of performing multiple tasks like generating, classifying, and summarizing text, consume significantly more energy than smaller, task-specific models. This finding highlights the environmental cost of the versatility of generative AI models.
The study emphasizes the need for more efficient and specialized AI models to reduce energy consumption and carbon emissions. It also points to the growing importance of understanding the environmental impact of AI as these technologies become increasingly integrated into everyday products and services.
Researchers hope that these findings will encourage both AI developers and users to consider more planet-friendly approaches in the deployment and utilization of AI technologies. The study serves as a reminder of the hidden environmental costs of the digital technologies we frequently use.