Exploring the Implications of AI for Climate

Share

Exploring the Implications of AI for Climate

Feature Story| August 2, 2024

By Sara Frueh

Artificial intelligence is being harnessed to research and respond to climate change - for example, by advancing weather and climate modeling and optimizing the operation of some electrical grids. But AI itself uses a lot of energy, raising worries that the technology's growing use across society will extend our dependence on fossil fuels.

"We have this kind of Janus-faced situation," said David Goldston of the Massachusetts Institute of Technology, who led a discussion of AI and its potentially opposing impacts during a session at the National Academies' recent Climate Crossroads Summit.

Goldston was joined by three other panelists - Uyi Stewart of data.org, Topaz Mukulu of the Patrick McGovern J. Foundation, and Priya Donti of MIT. All of them pointed to ways AI is aiding climate-related work, as well as the challenges AI raises for climate, which stretch beyond its energy use to touch on issues such as equity.

"As a practitioner in artificial intelligence for the last two decades, I've seen tremendous progress in the ability for this technology to help generate great insights from data - whether it's in diagnostics, whether it's in optimization, or even prediction," said Stewart. "This technology, and the online data, is becoming more accurate. That's what makes me hopeful."

Donti, too, noted ways that AI is helping make headway on climate. "There are lots of ways that AI is starting to mature in order to enable climate-change-related work, from helping us better predict the weather, better optimize power grids and heating and cooling systems, and also do things like accelerate scientific discovery," she said.

Mukulu offered an example, describing an effort in Rajasthan, India, to use AI to help incorporate solar energy into the electrical grid.

"When you're trying to incorporate solar and renewable energy, it's really hard to predict and to measure how much solar energy is available," she said. "So, they've figured out ways to use algorithms and models to test and predict when there's going to be more solar energy. There are obviously complexities related to that - there's cloud cover, there's rain - and so they're able to have a high percentage of accuracy and get those details down, and support grid operators in India in incorporating some of the solar energy into [the grid along with] coal and wind."

Despite its beneficial applications for climate, AI is also being used to advance industries and systems that contribute to climate change, Donti explained.

"AI is being used to accelerate oil and gas exploration and extraction. It's also a big driver of things like targeted advertising that change how we consume," she said. "AI is an accelerator of many of our systems across our society, and if we don't align that aspect of it as well - if we don't align business-as-usual with climate as well - then we're actually going to exacerbate the problem."

Equity is another concern, in terms of who is developing AI tools and data, the speakers noted. "There is a digital divide between the so-called global north and the global south, and we need to start to pay attention to those who are marginalized and vulnerable," said Stewart, adding that AI is widening this divide.

Donti agreed. "The community that is often bringing forth AI tools is often a very privileged community, often centered in the global north," she said. "If we truly want AI to enable global and equitable climate action, that means we really have to focus on ground-up education and capacity building across many different contexts, and across the world, in order to ensure that the way we develop AI for climate is coming from those places of experience, those contexts, those communities - rather than being foisted top-down from a small set of people onto the rest of the world."

Mukulu explained that the need for equitable AI development also includes data collection. People often say of AI, "garbage in, garbage out" - if the data you feed into AI models isn't good, then your outputs won't be reliable either. But the settings in which data is collected and used also matter.

"If you have 'good' data, data that's complete and that you might consider comprehensive, [but] it's coming from one specific region - I'm thinking of the global north - and being used in a different context that is unrepresentative of the data, then that's also garbage," said Mukulu. "It's 'good data in, garbage out.'" Data is most impactful when it's context-specific and relevant, she said, stressing that communities should be involved in data-collection efforts.

The speakers also discussed whether the intensive energy demands of AI will perpetuate dependence on fossil fuels.

Donti stressed that reliance on fossil fuels to meet AI's energy needs is not inevitable. "We have a lot of choices in how we actually develop and use AI, and employ AI across society," she said. "And every sector of society - including the information and communications technology sector - has to go to net-zero. That includes AI."

A key step toward achieving this is by being realistic about the increases in energy use that may happen because of AI, incorporating that into clean-energy planning, and ensuring we're serving that demand through clean energy, Donti said.

"We need to shift a lot of these data centers to a reliance on clean energy," Mukulu agreed. She also noted that some large private-sector organizations have figured out how to run their data centers on more renewable energy and run their models in a more energy-efficient way, but that data is not public; pushing for more visibility with this information could help expand those efforts.

Mukulu also recommended more transparency about energy use by AI applications, likening it to when a consumer buys an airplane ticket and is shown the carbon footprint associated with the trip. "I think a bit more of that visibility as well in the AI space would be useful - so as you're creating models and developing all of these innovations, being able to indicate how much that fed into emissions would help incentivize us to move in a more sustainable direction."

The speakers pointed to multiple ways to improve AI applications for climate going forward.

Mukulu said that one factor in making AI applications more equitable is open data. "One of the things I think is lacking in the climate space is open data," she said. "So, government policies that push for more open data - not just for nonprofits but also with private-sector actors - that's really critical, so that folks that aren't able to afford some of the initiatives that are needed to implement some of the ideas can pull from an ecosystem and a wealth of knowledge."

Stewart explained that right now many climate models are not at the spatial scale to be useful on the ground. Models need to be downscaled in order to become hyperlocal in terms of the parameters they are using, to inform decisions at a local level.

Donti also stressed the need to make AI development for climate more specific, creating applications that match particular needs and requirements on the ground.

"For example, if we think about AI in the context of power grids - on power grids, you're dealing with physical signals, you're dealing with an engineering system that has robustness constraints - you have to make sure your methods satisfy various notions of safety," she said. "And that actually has to be baked into the way you design those methods in the first place."