AI image generation increases carbon footprint, study finds

So you program your thermostat to save heating costs, recycle glass and plastic, bike to work instead of driving a car, reuse sustainable grocery bags, buy solar panels, and shower with your partner , all to do your part to save energy, reduce waste and reduce your carbon footprint.

A study published last week could ruin your day.

Researchers at Carnegie Mellon University and Hugging Face, a community-based machine learning website, report that you could still contribute to climate change if you are one of the more than 10 million users who mine data models every day. machine learning.

In what they call the first systematic comparison of costs associated with machine learning models, researchers found that using an AI model to generate an image requires the same amount of energy as loading d ‘a smartphone.

“People think that AI has no impact on the environment, that it is an abstract technological entity that lives on a ‘cloud,'” said team leader Alexandra Luccioni. “But every time we query an AI model, there is a cost to the planet, and it’s important to calculate that.”

His team tested 30 datasets using 88 models and found significant differences in energy consumption between different task types. They measured the amount of carbon dioxide emissions used per task.

The greatest amount of energy was spent by Stability AI’s Stable Diffusion XL, an image generator. Nearly 1,600 grams of carbon dioxide are produced during such a session. Luccioni said that’s about the same as driving four miles in a gas-powered car.

At the lower end of the scale, basic text generation tasks spent the equivalent of a car traveling just 3/500 of a mile.

Other machine learning task categories included image and text classification, image captioning, summaries, and question answering.

The researchers said that generative tasks that create new content, such as images and summaries, are more energy and carbon intensive than discriminative tasks, such as movie rating.

They also observed that using general-purpose models to undertake discriminative tasks requires more energy than using task-specific models for the same tasks. This is important, the researchers say, because of recent trends in the use of the models.

“We believe this last point is the most compelling point of our study, given the current paradigm shift from smaller models, tailored to a specific task, to models intended to perform a multitude of tasks at once , deployed to respond to a barrage of user queries in real time,” the report said.

According to Luccioni, “If you’re doing a specific application, like email searching… do you really need these big models that can do everything? I would say no.”

While the number of carbon dioxide uses for such tasks may seem small, when multiplied by the millions of users relying on AI-generated programs daily, often with multiple queries, the totals show what could represent a significant impact on efforts to control environmental waste. .

“I think for generative AI in general, we should be conscious of where and how we use it, comparing its cost and benefits,” Luccioni said.

The results are published on the arXiv preprint server.

© 2023 Science X Network