ChatGPT’s carbon tab

I’ll admit – I’ve lost count of how many times I’ve asked ChatGPT for help planning my weekly meal preps or double-checking my emails. Like so many others, I’ve grown reliant on having a digital assistant at my fingertips, ready to solve my problems in seconds. But while we marvel at the convenience of our (relatively new) AI-powered friend, it’s worth asking: how much is the Earth paying for all its instant information?

Generative AI models, like ChatGPT, are deep learning systems designed to generate new content by predicting sequences based on patterns in massive datasets. These models require enormous amounts of computational training, and that power doesn’t come cheap – at least not for the environment.

Research from Cornell University estimates that just training GPT-3 alone produced 552 tons of carbon dioxide (the equivalent of what 120 gasoline-powered cars would emit in a year). And that’s just for getting the model up and running. With millions of users accessing ChatGPT daily, energy consumption skyrockets.

Each message sent to ChatGPT produces approximately 4.32 grams of CO₂, according to ML CO₂ Impact, an emissions calculator. For reference, one Google search emits only 4% of that, 0.2 grams. Keeping this in mind, let’s do some oversimplified math: sending 15 messages produces as much CO₂ as watching one hour of Panopto lectures; sending 50 messages equals the CO₂ produced by boiling a kettle for your daily cup of coffee or tea; sending 150 messages is the same as taking a bus from Imperial’s South Kensington campus to Hammersmith; and sending 100,000 messages generates as much CO₂ as taking a flight from London to New York.

At first glance, these numbers might seem small. I mean, who is asking 100,000 questions every day anyway? But when you consider that ChatGPT handles an estimated 50 million unique visits daily, the numbers add up.

As these AI models continue to grow, the challenge isn’t just cutting down on our usage but making that usage smarter. Researchers and tech companies are already exploring ways to reduce AI’s carbon footprint, from more energy-efficient algorithms to powering data centres with renewable energy.

As users, we can think more about when we really need AI’s help – sometimes, a good old-fashioned Google search or a bit of human problem-solving might be the greener option.

Ultimately, if both AI’s creators and consumers limit its environmental impact, we can continue benefiting from AI without leaving such a heavy footprint on the
planet.