Apple’s new custom emoji comes with weather costs

You’ll soon be able to create your own emojis and generate custom images with new AI features on Apple devices. I hate to be the bearer of bad news (and I’m also guilty of using too many emojis) – but these images could have the biggest environmental impact of all of Apple’s new AI tools.

Generative AI is notoriously energy-hungry, which means it leads to lots of planet-warming greenhouse gas emissions. Imaging, in particular, is one of the worst offenders. The AI ​​arms race has already started to grow some companies carbon footprints. Now, new varieties of fruit emojis and AI-generated sketches are likely to add to the problem.

Apple announced its new set of AI capabilities during WWDC 2024 yesterday under the “Apple Intelligence” banner. Apple is injecting AI into everyday screen life in a variety of ways, from classifying notifications and helping users compose emails to making Siri smarter. Apple also released a partnership with OpenAI to bring ChatGPT to iOS, macOS and iPadOS “later this year”. Its AI imaging features, Genmoji and Image Playground, are expected to come with iOS 18, iPadOS 18, and macOS Sequoia.

Apple did not immediately respond to questions from On the edge about exactly how much energy you expect these new resources to use or how this all fits into your climate goals. But each of these measures carries energy and climate costs.

The data centers used to train and run AI models have an even more voracious appetite for electricity than other data centers used to send and store digital messages and pictures of the world’s cats. To look for from AI company Hugging Face and Carnegie Mellon University found that large imaging models used much more energy and created more greenhouse gas pollution than others used to classify information or generate text. (The study did not include AI generative video models.) Creating a single image could consume as much electricity as charging a smartphone to 50% capacity, it found. For comparison, a text generation model can use about 9% of a smartphone’s total load for 1,000 inferences.

Imagine charging your cell phone after every custom image you create in a text conversation and you’ll see how quickly energy costs can add up.