Environmental Impact of AI-Generated Images Under⁢ Scrutiny

Each week, we⁤ examine potential solutions to address the ongoing climate ⁢and environmental crisis.

Posted: 8:00 p.m.

The recent surge in AI-generated images, fueled ‍by platforms like ChatGPT, has sparked debate about their environmental cost. One reader, Émilie, inquired about the potential⁣ impact of creating these virtual images.

The process typically involves users providing a name, image, and a few defining characteristics. Within seconds, AI algorithms generate a personalized figurine, resembling a toy. These images gained popularity in early April, following a trend of mimicking the style of studio Ghibli, known⁢ for animated films such as My ⁣Neighbor Totoro and Spirited Away.

OpenAI, the company behind ChatGPT, promoted these trends ‍to showcase ⁢the capabilities of its new image generation feature. Brad Lightcap, the company’s number two, stated that over 700 million images were generated by more than 130 million users in just 48 hours in early April.

This rapid proliferation prompted designers ⁣and‍ illustrators ⁢to ‍protest, creating “starter packs” made without AI, using⁢ the hashtag ⁤#starterpacknoai. Prominent figures⁣ also raised concerns about the environmental implications.

French astronaut Thomas Pesquet wrote on social media, “Behind magic, there is a reality that we often⁢ forget to look: the environmental cost of‍ these⁣ technologies.”

Subsequent media reports have highlighted the⁣ water needed to cool servers and the electricity consumed in generating these images, and also the greenhouse ⁣gases emitted.

Though, Sasha Luccioni, climate manager at Hugging Face, a platform promoting accessible AI, notes, “We cannot really estimate the impact of such a trend.”

Estimating the Impact: A Complex challenge

Luccioni published a study last June ⁢estimating⁤ the⁣ energy consumption of AI models that generate text or images. Image generation models were found to be the most energy-intensive.According to the study, the most demanding models required the equivalent of half a cell phone ⁢charge to produce a single image. Generating 1,000 images resulted in 1.6 kg of CO2 equivalent emissions – roughly equivalent to driving a car 7 kilometers.

The ⁣study focused on open-source models accessible to researchers, unlike ChatGPT, which‍ uses a ‍”closed”‍ model, making performance assessment difficult.

“But as it generates rather high quality images, one can imagine that it is indeed rather on the‍ upper part of the spectrum. It is probably a very large model ⁤that is used underlying,” Luccioni said.

The energy required for a single image ⁢generation could be ⁣substantially higher than the open models⁢ studied.

“It is indeed impossible to know it, as they do not say how their model works, on‍ how manny images they train‍ it, etc.,” Luccioni said.

The same uncertainty applies to water consumption and greenhouse gas emissions.

A Slippery Slope?

While the energy consumption per image may seem small, the‍ cumulative effect of hundreds of millions of requests could be substantial.

“And as these companies give absolutely no environmental information, people are not aware of the cost that it represents,” Luccioni said.

The lack of transparency from major tech companies makes it ⁣difficult for individuals to make informed decisions about their AI⁣ usage, unlike choices related to car use or meat consumption, where energy, greenhouse gas, and water consumption estimates are available.

Furthermore, the resources consumed are “far from us,” ⁤making the impact difficult to perceive.

“If we saw the battery of our phone decrease by half when we generate an image,it would help us to⁤ think before doing ⁤so,” Luccioni said.

luccioni advises against wholly abandoning AI tools but encourages mindful consideration before defaulting to AI for every task.

“people sometimes use chatgpt⁣ on a daily basis to have fun and generate images, to look for recipe ideas, to answer questions … or ⁤even as a calculator,” Luccioni said.

“Couldn’t we do otherwise?”