Buying two T-shirts costs twice as much as one. This is obvious when we see the money leave our wallet. Driving twice the distance will cost you twice the fuel. Also obvious when we look at the gas tank indicator.

If everything we do online costs the environment in some way, how much does a Google search cost? How does it compare to the cost of that same search on ChatGPT?

As we move through the web — streaming, searching, downloading and more — we leave behind a footprint. Yet the size of that footprint, or even the fact that it exists, feels abstract and hidden away. Our internet carbon footprint is determined, in large part, by the decisions companies make when they create their tools — tools that require vast amounts of land, water and energy somewhere off in the distance. That footprint is also determined by decisions consumers make, when we choose which of those tools we use and how often.

In an ideal world, all of our favorite sites and apps would be emissions free — every app download would result in rainbows above our phones and thumbs-up’s from polar bears off in the distance. In the real world, it’s quite the opposite. Not only are these modern conveniences costly to the environment, but it can be hard to tell how costly. There’s no wallet to indicate you’re spending twice as much (or checking account, for those paying digitally). There’s no fuel gauge to say you’ve used twice the fuel (or 1.5x the fuel, for those driving electrically). It isn’t clear how much everyday internet usage costs in resources, and things are becoming even less clear with AI.

The folks at Code Carbon are one of the Mozilla Technology Fund awardees hoping to bring us to that ideal world. (The zero emissions part. Not the polar bear thing.) Code Carbon is a code package that helps developers estimate their tools’ carbon emissions, and then offers ways to lessen it. When it comes to AI, some new norms are emerging that don’t feel great to Code Carbon founder Sasha Luccioni. “AI companies don’t share the amount of energy a query costs,” says Sasha. Companies like OpenAI keep information about their energy usage close to their chest, and according to Sasha, they’re incentivized not to say. Fortunately, open source models like Hugging Face help us guesstimate. “When we look at open source AI models, we see that smaller models trained for a specific task use less energy than many of these huge large language models that can technically do a bunch of tasks but they use a lot more resources for each query.” Sasha notes that asking a larger AI model a simple question can result in 30 to 40 times the energy usage for a single query, when compared to task-efficient models.

But using an AI tool is just half of the picture. The other half is the profound amount of energy AI models require as part of their training. “Facebook will buy over 300,000 GPUs this year, which will require the same amount of energy used by thousands of homes in France,” says Benoît Courty, president at Code Carbon. “Just training the AI requires the same amount of energy as powering a city.” The graphics card, Nvidia’s H100, gobbles up 700W during peak usage — the same as an average two-person American household. Facebook (now known as Meta) will use 350,000 of these graphics cards and a city’s worth of energy strictly to develop artificial intelligence. And that doesn’t even include the resources it will cost to then use its AI tools.

How Do We Address The AI Carbon Footprint Problem?

Changes to this can happen at three levels. The company level, the policy level and the individual level. It’s clear that companies need to, first, be transparent about their products’ carbon footprint and second, actively work to shrink it. For example, pointing users to smaller AI models when their query is a basic one.

Unfortunately, it’s unlikely meaningful change will happen at the company level without government intervention. So what’s happening at the policy level? There aren’t many laws regarding AI and energy usage but we are seeing at least a little movement. A bill in congress proposes that the Environmental Protection Agency (EPA) and National Institute of Standards and Technology (NIST) nudge AI companies to report their tools’ energy usage. The bill would be a good first start if it passes but there’s room to improve. “The next step should be mandatory disclosure,” says Sasha about the EPA/NIST bill. “And even if you couldn’t give an exact number, you could offer an average. Once companies are required to report these numbers, then you can start to add rules around how much energy, at a maximum, companies’ AI models can use.”

Love Generating AI Cats? Hit The Paws Button

And then there’s the individual level. Even though consumers don’t have much control over the carbon footprint of the tools they use, there is some agency in choosing the right tool for the job and the planet.

An app like ChatGPT is a huge model and (likely) requires a lot of energy each time you use it. Every problem may look like a nail when you’re holding a hammer but keep in mind there are lots of different-sized tools in the internet toolkit and some are better for the planet than others. “Bigger isn’t always better,” says Sasha. “Sometimes good old-fashioned tools can get the same job done. If you want to do math, there are calculators. If you want information, run a simple web search or search encyclopedias.” Using digital tools without AI can be just as effective, yet carry a lighter footprint.

This isn’t to say we should never use AI to generate anything. Benoît, for example, recommends using the text-only ChatGPT 3.5 over ChatGPT 4 if you know you won’t be doing any image processing. If you are using the extra processing power, though, Benoît says make it worth it. “If you’re using AI to generate global warming solutions that could be a good use for it,” says Benoît. “If it’s to make pictures of cats, we already have enough of those on the internet.”

AI Is Getting Bigger. Is Its Climate Impact Getting Badder?

Written By: Xavier Harding

Edited By: Audrey Hingle, Kevin Zawacki, Tracy Kariuki, Xavier Harding

Art By: Shannon Zepeda

Relatearre ynhâld