There’s been a shift in AI. No longer a field reserved just for researchers and data scientists, AI has gone mainstream. Today, practically everyone has read a poem written by AI, or seen what they look like rendered as a cartoon character. These are impressive advancements that indicate the rapid progress being made in AI. But we’re not asking a crucial question: What does the growing popularity of large-scale AI models mean for the planet?

Other industries from fashion brands to manufacturing companies have started to report on their own environmental impact as it’s become clear that we must do everything we can to curb the amount of CO2 we release into the atmosphere. AI is no exception. Across the field, there is a lack of transparency around the carbon emissions generated from AI, and efforts to meaningfully quantify its carbon footprint have just begun.

With AI’s level of carbon output only expected to rise as models become more powerful, and more widely used, it’s critical that we establish transparency today in order to mitigate the potential impact of AI’s carbon footprint tomorrow. Establishing this transparency will allow us to develop actionable tools for researchers that will reduce harmful carbon emissions from AI workloads, and further cement AI as a tool to help combat climate change, not exacerbate it.

UNDERSTANDING AI’S CARBON FOOTPRINT

How do you measure the carbon footprint of AI? Alongside a team of researchers from Microsoft, Carnegie Mellon, and the Hebrew University of Jerusalem, I set off to answer that very question. Using a framework called the software carbon intensity developed by the Green Software Foundation—a nonprofit that aims to reduce the emissions of software—we were able to map emissions from AI models trained on Microsoft Azure back to the supporting energy grid.

AI can generate carbon emissions, and thus have a high software carbon intensity, when the energy powering a data center running an AI model is not clean. To begin understanding the magnitude of this problem, we trained a number of different-size models on Microsoft Azure across various geographies and time frames. The results were staggering. We found that the largest model we trained at 6 billion parameters—which is about a tenth the size of models like GPT-3—could emit more CO2 than the average U.S. home does in a year, totaling 8.3 metric tons. And that model was only trained at 13% of its full capacity, suggesting the actual emissions could be much larger.

That’s just the tip of the iceberg. Today, increasing numbers of AI models are trained on cloud platforms like Microsoft Azure, Google GCP, and AWS, which are housed in large data centers that consume vast amounts of energy. In 2018, researchers estimated that global data center energy use represented close to 1% of global energy usage. Because of this, AI’s carbon footprint is only expected to grow.

THE CASE FOR AI TRANSPARENCY

These findings are some of the first estimates of AI’s total carbon footprint because we, as an industry, are not cooperating to measure our environmental impact. Our team was only able to generate these initial estimates of AI’s emissions because we were working with a cloud provider that provided transparent reporting about carbon intensity—the amount of CO2 emitted per unit of energy—of the data center’s supporting grid. If we want to truly understand the scope of this issue, and generate meaningful change, this type of transparency has to be replicated at scale by the AI industry.

We can take my team’s research as an example for what’s possible when our industry is transparent about reporting carbon emissions. Thanks to the clear reporting from Microsoft Azure, our team was able to begin quantifying AI’s growing environmental impact and identify a number of actions AI practitioners can take today to curb our carbon footprint.

For example, we found that altering where an AI model is run has the biggest impact on its emissions. Data centers are located all across the globe, and certain countries have higher mixes of clean energy in their grids than others. If cloud providers provide an option to select a data center location with more renewable energy sources, AI practitioners may reduce the software carbon intensity of a model by up to 75%. We also found that timing models to run when the carbon intensity of that grid is low could be a viable solution as well. When AI researchers know how much CO2 the supporting grid emits at a given time, we can time a model to run when that carbon intensity is low, lessening our impact.

THE SUSTAINABLE FUTURE OF AI

It can seem daunting to grapple with carbon emissions when the AI industry doesn’t have a true understanding of the magnitude of our contributions today. But there’s hope. My initial research has shown that there are clear actions we can take now to lessen our environmental impact. And as more research into AI’s carbon footprint is published, and new measurement tools are released, we will undoubtedly find even more ways to actively reduce the AI industry’s footprint. An industry-wide commitment to transparency will enable the AI community to arm ourselves with the knowledge necessary to address AI’s role in climate change.

As we consider the emissions of our models, we also need to consider AI’s applications. Not only do we need to think about how we are powering our models, but we also need to think about what we are powering. AI is an accelerant. It can, and already is, directly helping to fight the climate crisis by fast-tracking breakthroughs and spurring Earth-saving innovations. Some of the other amazing work being done at the Allen Institute for AI is doing exactly this, from using AI to protect our oceans or to improve climate modeling.

But it can also exacerbate climate change if we don’t address these challenges head-on today. As we demand industry-wide transparency, we must also make conscious decisions to support AI applications that help save our planet.