Lately I’ve lost a lot of sleep over climate change. It’s just over five weeks until Christmas, and last weekend in London, it was warm enough to have a pint outside without a coat. As world leaders gather in Egypt for the final week of climate conference COP27 to “blah, blah, blah,” this week I’m focusing on the carbon footprint of AI.
I’ve just published a story about the first attempt to calculate the broader emissions of one of the most popular AI products right now — large language models — and how it could help nudge the tech sector to do more to clean up its act.
AI startup Hugging Face calculated the emissions of its large language model BLOOM, and its researchers found that the training process emitted 25 metric tons of carbon. However, those emissions doubled when they took the wider hardware and infrastructure costs of running the model into account. They published their work in a paper posted on arXiv that’s yet to be peer reviewed.
The finding in itself isn’t hugely surprising, and BLOOM is way “cleaner” than large language models like OpenAI’s GPT-3 and Meta’s OPT, because it was trained on a French supercomputer powered by nuclear energy. Instead, the significance of this work is that it points to a better way to calculate AI models’ climate impact, by going beyond just the training to the way they’re used in the real world.
The training is just the tip of the iceberg, because while it’s very polluting, it only has to happen once. Once released into the wild, AI models power things like tech companies’ recommendation engines, or efforts to classify user comments. The actions involved use much less energy, but they can happen a billion times a day. That adds up.
Tech companies want us to just focus on the emissions from training AI models because it makes them look better, David Rolnick, an assistant professor of computer science at McGill University, who works on AI and climate change, told me.
But the true carbon footprint of artificial intelligence is likely to be bigger than even Hugging Face’s work suggests, Rolnick argues, when you take into account how AI is being used to boost extremely polluting industries — not to mention its broader, societal knock-on effects. For example, recommendation algorithms are often used in advertising, which in turn drives people to buy more things, which causes more emissions.
And while AI may play a part in fighting climate change, it’s also contributing to our planet’s death spiral. It’s estimated that the global tech sector accounts for 1.8% to 3.9% of global greenhouse emissions. Although only a fraction of those emissions are caused by AI and machine learning, AI’s carbon footprint is still very high for a single field within tech.
The Hugging Face paper is a good way to begin addressing that, by trying to provide honest data on the broader emissions attributable to an AI model. Tech companies like Google and Meta, which dominate this sector, do not publish this data. That means we really don’t have a remotely accurate picture of AI’s carbon footprint right now.
