As the use of artificial intelligence grows and societies ponder how to use the powerful tool to improve our lives, increase productivity and tackle our most pressing challenges, few have considered its effects on the environment.
While some have highlighted the technology’s potential to help tackle environmental challenges, others point out that we must first understand AI’s own carbon footprint.
Proponents of technological advances like cryptocurrency were quick to celebrate its potential to reduce carbon emissions, only to be refuted by wasteful practices like Bitcoin mining due to their enormous energy demands.
But experts largely see AI as a positive development, with the United Nations Environment Program lauding it as a tool that could improve our understanding of our environmental impact and the effects of climate change.
AI can be used to sift through large amounts of data, like satellite images researchers use to monitor climate change, said Sasha Luccioni, who works analyzing AI models for sustainability. With the help of AI, scientists can better model climate patterns, identify trends and make predictions so they can have a clearer understanding of climate change and effective mitigation strategies.
“There are a lot of really cool applications of AI in different sectors of climate change — everything from optimizing electricity grids to tracking biodiversity,” Luccioni said.
But some experts are looking at the carbon footprint of AI itself. For them, companies hoping to deploy AI should be transparent about its environmental impact and how they are addressing it.
What is AI’s carbon footprint and why is it worrying some environmental advocates?
AI’s overall carbon footprint is difficult to measure, but it starts with the computers it uses. The raw materials needed to create computer hardware are mined and “that can be really labor intensive and also environmentally expensive,” Shaolei Ren, associate professor of electrical and computer engineering at the University of California, Riverside, said.
Once developers have the hardware they need, training an AI model can consume a lot of energy. AI companies don’t tend to share how much energy is used, but researchers have taken guesses based on the data available to them. One non-peer-reviewed study, led by Ren and other experts, estimates that training GPT-3, which powers a language model of ChatGPT, could potentially have consumed 700,000 liters of freshwater. The water used to prevent data centers from overheating is usually evaporated, which means it can’t be reused.
There’s also the carbon emissions. Researchers at the University of Massachusetts, Amherst found the training process for a single AI model can emit more than 626,000 pounds of carbon dioxide. That’s about the same amount of greenhouse gas emissions as 62.6 gasoline-powered passenger vehicles driven for a year. Carbon dioxide makes up the vast majority of greenhouse gas emissions, contributing to climate change by trapping heat in the atmosphere.
After consulting these independent estimates, CBS News asked AI language models about the technology’s carbon footprint.
Bard, created by Google, said it was difficult to estimate accurately. ChatGPT, created by OpenAI, stressed that as an AI language model, it doesn’t have a direct carbon footprint, but with an estimated 100 million monthly active users, there’s a footprint connected to the electricity and computing resources needed to run the servers hosting and powering the model. (OpenAI did not respond to requests for comment for this story.)
Microsoft, which has invested billions of dollars into OpenAI, declined to share estimates for the carbon footprint involved in developing AI tools.
“AI will be a powerful tool for advancing sustainability solutions, but we need a plentiful clean energy supply globally to power this new technology, which has increased consumption demands,” a Microsoft spokesperson said. “Microsoft is investing in research to measure the energy use and carbon impact of AI while working on ways to make large systems more efficient, in both training and application.”
Can AI tools be designed in an environmentally-conscious way?
Training, deploying and running AI can be energy intensive, so companies should carefully consider the potential consequences while building the systems, said Junhong Chen, professor of molecular engineering at the Pritzker School of Molecular Engineering and lead water strategist at Argonne National Laboratory.
“When we design these types of systems, we have to be mindful of the potential negative consequences and try to minimize it from the beginning by design,” Chen said.
Research from Google shows that water-cooled data centers emit roughly 10% lower carbon emissions than air-cooled data centers. According to the Energy Department, data centers are one of the most energy-intensive types of building in the U.S., consuming 10 to 50 times the energy per floor space of typical commercial office buildings. They collectively account for about 2% of the total U.S. electricity use.
When new sites are picked for Google data centers —largely decided based on proximity to users— the company will look into reclaimed and nonpotable water resources in the area, Ben Townsend, Google’s head of data center sustainability, said.
“Data centers are very similar to your personal computer. They require space, they require energy and they require cooling,” Townsend said.
There’s also a balance to strike when it comes to energy grids, Ram Rajagopal, who leads the Stanford Sustainable Systems Lab, said. With the goals of decarbonization and resiliency in mind, AI can be used in the electricity system to reduce costs, scale up deployments and determine optimal plans for lowering the amount of greenhouse gas emissions, Rajagopal said.
Still, as AI use becomes more common, the data centers currently handling AI tasks may not be up to snuff.
“As this starts to scale up, you create a bottleneck in terms of the data centers and then you have to expand data centers, so the power consumption expands,” Rajagopal said.
How can AI help?
Scientists are already using AI in many helpful ways. AI models can help researchers find ways to recycle and reuse water by identifying contaminants in water and figuring out the best ways to extract them, according to Chen, the professor of molecular engineering. It can also potentially be used to determine ways to reclaim those contaminants for other uses, he added.
In one recent project, Google, American Airlines and Breakthrough Energy teamed up and used AI to piece together and sift through satellite imagery, weather and flight path data. The AI was used to develop maps to forecast contrails — the thin, white lines sometimes seen behind airplanes. The research can help pilots optimize flight routes so they can cut down on contrails, which account for roughly 35% of the aviation sector’s global warming impact.
Artificial intelligence can also be applied to battery research to optimize lithium batteries, which are used by most electric vehicles, experts say.
Several companies, such as AMP Robotics and MachineX, have developed AI tools to identify and recover recyclables with AI-guided robots. AMP Robotics has more than 300 AI systems deployed globally, a spokesperson said.
The robots can, on average, pick up recycled materials up to two times as fast and with more consistency than humans. According to the company, AMP Robotics technology has helped avoid nearly 1.8 million metric tons of greenhouse gas emissions, an impact equivalent to removing close to 375,000 cars from the road, by optimizing recycling efforts.
Scientists in California are using AI to fight wildfires. AI connected to cameras can identify wildfires and detect smoke before they spread more widely. Cal Fire Battalion Chief David Krussow told CBS Sacramento the information on wildfire prediction is a “game changer.”
At the National Oceanic and Atmospheric Administration, scientists are using AI to improve climate, weather and other earth system models.
The United Nations Environment Program uses AI to help analyze and predict the concentration of carbon dioxide in the atmosphere, along with changes in glacier mass and sea level rise. They hope to use the tool as a type of “mission control” for the planet, David Jensen, a coordinator with the team, has said. One U.N. tool, the International Methane Emissions Observatory, or IMEO, uses AI to monitor and mitigate methane emissions, a potent greenhouse gas which affects the earth’s temperature.
“Reducing the energy sector’s methane emissions is one of the quickest, most feasible, and cost-effective ways to limit the impacts of climate warming, and reliable data-driven action will play a big role in achieving these reductions,” Jensen said in a U.N. post.
But despite the promise, companies and even AI language models recognize the technology’s limitations and the uncertainty around its environmental impact.
“As an AI, my purpose is to assist users like you in accessing information and knowledge about various topics, including those related to the environment, so that you can make informed decisions and take actions that align with your values and goals for a better and sustainable world,” ChatGPT wrote in response to a question by CBS News.
Kate Brandt, Google’s chief sustainability officer, said it was difficult to predict the future growth of energy use and emissions associated with AI.
“Ultimately, the environmental impact of AI models like me will depend on how they are used,” Bard said. “If we use AI to solve environmental problems, then we can have a positive impact on the planet. However, if we use AI to create new environmental problems, then we will have a negative impact.”