When people mention Artificial Intelligence’s greatest potential upside for society, they often cite sustainability.

After all, it is regarded as an integral tool in the fight against climate change and scientists expect it to reduce the overall carbon footprint of human activity. If developed responsibly, AI systems could help streamline industrial operations, preserve environmental resources and unlock new forms of efficiency that dramatically reduce waste. Examples abound, ranging from reducing the use of packaging to providing reliable global flood forecasting.

But all this comes at a cost. While much is spoken about the concept of “AI for sustainability” less attention is focused on its flipside - “sustainability of AI.” This is the sustainability cost of training, operating and deploying these often power-hungry AI systems. For example, Google reported this year, that they had a 48 percent increase between 2019 and 2023 in their total greenhouse gas emissions, mostly due to data center energy consumption and supply chain emission increases that were associated with the integration of AI into their products.

The growing awareness of these costs has sparked several attempts to create carbon footprint calculators for AI to measure and report on greenhouse gas emissions. However, most of these models do not share their assumptions and therefore results can be incomparable and unreliable.

At Nokia, we have come up with what we believe is the most comprehensive framework to transparently assess the environmental impact of AI systems, thus helping people, companies and governments make informed sustainable decisions.

A new way forward to assess AI environmental impact

The greatest benefit of our approach is that it relies on existing standards of the International Organization for Standardization (ISO) and the International Telecommunication Union (ITU) and therefore can be adopted across different industries and regions as a globally recognized way to assess the environmental impact of AI systems. And it can be used in AI systems as large as Meta’s Llama 3, which currently has 70 billion parameters in size, or as small as a basic shopping list recommender.

Our "environmental impact assessment of AI” framework is built by combining the AI system life cycle that depicts important stages in the evolution of an AI system, with the environmental life cycle assessment method that interprets the environmental impact of different activities and flows. In our framework, we propose a uniform way to map relevant processes and stages together.

In other words, this merges the best of both these worlds to create a comprehensive and transparent solution.

By leveraging existing standards, and not re-inventing the wheel, we get all the benefits of the existing standards for “free.”

One of the main benefits is transparency. It is currently challenging to compare or combine the communicated results from companies and researchers since different approaches are used. One study talks about carbon emissions from training, another from inference and a third from cooling perspective and each reports the environmental impact of AI. The reader is left with the question of what was contained or left out in each of these studies. When using a standards-based approach, the underlying assumptions are unified, and the results are provided in a globally harmonized “language.”

Another benefit is that when a particular aspect of the AI system is modified, for example, the location of the training or the training time, the resulting change in the environmental impact can be assessed. This is the benefit gained from the environmental life cycle assessment method in which the AI system life cycle representation adds transparency to the processes that form the AI system under investigation.

Companies using this new proposed framework will know exactly which part of their AI systems can and should be optimized. Therefore, a company could, for example, determine that it will train its AI systems in sunnier locations to take advantage of the available solar power.

In addition, this new framework will provide companies with a more transparent tool for reporting the environmental impact of AI systems to the government and stakeholders.

Another way to use the outcome of the assessment is to act on how the AI system is designed. The June 2024 creation and publication of Frugal AI principles by AFNOR (Association Française de Normalisation), the French national organization for standardization, serves as proof points for this complementary step and guidance on how a better designed AI system would be more sustainable.

An agreed upon framework

The environmental impact of AI systems is an integral part of the wider discussion on responsible, trustworthy, ethical and sustainable AI. Therefore, we cannot allow the training, deployment and use of AI to become just a major sustainability burden.

Given how few coherent data points exist today, we are confident that this proposed framework will create a uniform system of assessing the environmental impact of each AI system that will help people make better informed decisions on how to work with their AI systems.

This framework will help companies and governments have a better understanding of the overall AI impact and be able to assess it in a methodical way.

Nokia is committed to driving sustainable business. This framework, though, will go far beyond our own company and will offer insights into the information and communications technology industry and beyond.

We hope this framework can serve as a base for the new standardization work, recently started in ITU on ‘Guidelines for Assessing the Impact of Artificial Intelligence on Greenhouse gas emissions.”

Measuring is knowing and therefore widespread participation is encouraged. A common and transparent framework gives us all a better understanding about the environmental impact of AI and how to optimize that. By understanding and acting upon the full life cycle environmental impact from AI, the future of AI remains full of promise and endless possibilities.