AI image uses as much energy as a phone charge — are ChatGPT & co. harmful to the climate?

Generative AI is a powerful technology used in various fields, from product development to medical research. However, its use has a significant carbon footprint. According to a recent study by researchers at Hugging Face and Carnegie Mellon University, generating a single image using AI consumes as much energy as fully charging a smartphone.

The study, led by Sasha Luccioni from Hugging Face, analyzed the impact of 10 prompts on 88 different (cloud-based) AI models. The tool Code Carbon was used to measure energy consumption. The analysis revealed that creating images was the most carbon and energy-intensive activity.

1,000 images are equivalent to a 4.1-mile drive

Using a powerful AI model like Stable Diffusion XL to create 1,000 images generates approximately the same amount of carbon dioxide as driving an average petrol-powered car for 4.1 miles, or about 1.1 kilograms of CO₂. However, text generation models have a significantly lower carbon footprint.

The lowest-carbon model analyzed produced only as much CO₂ as a 0.0006 mile journey in a similar vehicle, or around 0.002 kilograms of CO₂. However, if you consider NPUs that are used locally, for example, AI applications can be operated in a much more energy-efficient manner.

Large AI models contribute to increased emissions

One might assume that the high CO₂ emissions of AI imagery are due to the considerable amount of energy required for training, as AI models are trained with huge datasets of images that are processed on supercomputers. However, the majority of emissions result from actual use rather than from training large models.

Luccioni estimates that the energy used to train large language models such as ChatGPT is exceeded after just a few weeks of use. This is because the popular chatbot has around 10 million daily users. Studies also show that running large generative models is significantly more energy-intensive than using more specific models that are only needed for specific tasks.

If you’re doing a specific application, like searching through email … do you really need these big models that are capable of anything? I would say no.

- Sasha Luccioni  

Luccioni hopes that the results will encourage the conscious consumption of generative AI and the selection of more energy-efficient models whenever possible. The researchers aim to raise awareness of this topic and encourage companies to take more responsibility for their energy footprint.

The responsibility here lies with a company that is creating the models and is earning a profit off of them.

- Jesse Dodge, a research scientist at the Allen Institute for AI