What's the carbon footprint of AI?

October 4, 2023 6:30 AM

Johan Mathe, co-founder of Atmo AI, left, demonstrates an Atmo AI supercomputer at the Uganda National Meteorological Authority (UNMA) headquarters in Kampala, Uganda, on Tuesday, Feb. 8, 2022. Artificial intelligence, in theory, can deliver on-par weather forecasts with less computing. Photographer: Esther Ruth Mbabazi/Bloomberg — Johan Mathe, co-founder of Atmo AI, left, demonstrates an Atmo AI supercomputer at the Uganda National Meteorological Authority headquarters in Kampala, Uganda, on Feb. 8, 2022.

The energy required to store and process data associated with artificial intelligence is a growing concern. As insurers move legacy systems to the cloud and continue to implement more AI into their systems, it may have an impact on their carbon emissions.

Digital Insurance received written responses from Joseph B. Keller, a visiting fellow at the Brookings Institution in Foreign Policy, affiliated with the Strobe Talbott Center for Security, Strategy, and Technology and Artificial Intelligence and Emerging Technology Initiative, about the need to quantify the potential carbon footprint of AI.

How does the energy consumption of data storage and cloud computing contribute to carbon emissions?

Cloud service providers are becoming the primary hosts of AI activities. Because of the rapid growth in reliance on the cloud, people are starting to pay more attention to the sector's sustainability practices.

The global information and communications technology (ICT) sector, which includes AI-related activities, emits more than 1.5% of all global carbon. This number may be growing but it's difficult to determine how much is driven by machine learning and commercial AI applications. Some of the biggest algorithms are large language models like ChatGPT, which emitted more than 500 tons of carbon over its lifecycle. In many parts of the world, fossil fuels are still used to power ICT and train AI models. They remain a major culprit in rising emissions.

People often consider the "cloud" and AI to be an intangible thing. However, many are waking up to the fact that digital communication, processing, and storage require a very tangible and physical footprint. AI systems are certainly things you can touch.

Are there initiatives or strategies being developed to reduce the carbon footprint of AI-related data storage and cloud computing?

We need more innovative strategies in the public and private sectors alike, particularly at the global level. Because AI threatens to widen the gap between the wealthy countries and lower/middle-income nations, equity and AI sustainability should go hand in hand.

The "FAIR Forward — Artificial Intelligence for All" initiative, supported by UNESCO, enables a data-sharing community including countries located in sub-Saharan Africa. It promotes sustainable design of AI-related activities and makes that information more accessible to those often left out of discussions typically dominated by industry and advanced economies. Inclusive coordination between stakeholders is necessary to have an impact — you must try to bring everyone to the table.

While training machine learning models takes a lot of energy, technological innovation can help realize some efficiencies as well. Some models can increase efficiency and optimize operations to exploit economies of scale.

What are the trade-offs between AI advancement and its environmental consequences, particularly in terms of energy consumption and climate impact?

Balancing the consumption of energy and natural resources to meet AI demand remains a challenge. Data centers, for example, spur competition for resources and land usage encroaches upon local communities. Many are not thrilled to have them as neighbors, and it is causing some backlash. This juxtaposition highlights a glaring strain between the industry's need to grow and the realities of climate change in local communities.

We're hearing about plans for new tech infrastructure in water-stressed regions suffering from extreme heat. At the same time demand is rising for computers and new data centers. It's worthwhile considering how these two trends, worsening climate and expanding tech, will confront one another.

Are there geographical variations in the energy efficiency of data storage and cloud computing for AI applications?

Quantifying the carbon footprint for AI, data storage, and cloud computing has led to mixed results. Some believe carbon emissions will rise exponentially with AI, while others see emissions as something that can be neutralized in the ICT sector. There's a serious need for better data on global ICT carbon emissions.

We know less about how distinct regions contribute to the overall efficiency of the AI network at a global scale. However, there's some research proposing new approaches to this assessment. They're trying to partition the network and see if the individual components of the ICT network offer efficiencies that can be capitalized upon. This methodology could lead to a reorganization in how we selectively choose geographic interventions that have a chance to yield broader benefits.

What challenges exist in accurately measuring and assessing the carbon footprint of AI-related activities?

There are many challenges, but two stand out. The sector still struggles with a dearth of accessible and meaningful environmental information related to AI. There's a lack of transparency from companies on data inputs and emissions. Companies also may not track and capture important data or are not inclined to make it public. Understandably, the motivations of private industry and the public sector don't always align.

In addition, it's difficult to tell things apart. It's hard to tease out the individual contributions from different digital activities and AI-specific processing and storage. How much is specifically attributable to AI? Evidence suggests that AI activities, in particular, need more energy and resources than other digital services. At the moment, however, we're limited to making rough estimates. It's tough to look at the system and distinguish AI from other digital activities that use shared infrastructure.

How can interdisciplinary collaboration between AI researchers, environmental scientists, and policymakers help address the climate impact of AI?

Policymakers need accurate and reliable measures and improved data transparency. Without it, effective governance and regulation are out of reach. Industry will want to reduce energy in order to reduce costs — optimizing networks to achieve efficiency gains is in everyone's best interest. Collaboration is essential.

Improving carbon emissions and water usage reporting are important steps. Some countries are making demands of industry stakeholders while acknowledging the tension between AI development, environmental impacts, and climate change. The European Commission is reportedly inquiring about data center metrics and water usage, and the UK also has plans to request water consumption data from tech companies.

Many suggest an overhaul to the energy grid by decreasing fossil fuel usage and completely electrifying power exclusively from sustainable and renewable energy sources. These calls are commonly made by wealthier and western countries, offering little accommodation for nations that may not possess the financial resources necessary to make a swift transition.

But it's not just the energy usage that's worthwhile scrutinizing, it's the intersection between AI, the environment, and the climate. The increasing frequency of extreme heatwaves, devastating natural disasters, and persistent droughts are our new reality. Meanwhile, digital expansion continues to accelerate. These issues should be on the docket at the next UN Climate Change Conference in Dubai. Innovation is necessary to help manage these problems, but it's not sufficient.