AI’s environmental impact: It’s complicated, but there’s hope

As summer in the northern hemisphere brings more heat records and devastating wildfires, the human contribution to climate change is rightly being scrutinized.

With advances in AI making headlines over the past 12 months, it’s not surprising to see the technology’s environmental impact being questioned.

The scrutiny is warranted. But looking only at the environmental costs associated with AI, such as the energy needed to train large language models (LLMs), risks ignoring both the progress being made to reduce its carbon footprint as well as the potential for AI to solve complex environmental problems.?

Reducing AI’s environmental impact

Exactly how much carbon is emitted in the training and running of an LLM depends on a lot of factors, including how large it is, what hardware it’s running on, the efficiency of the code, and the source of the energy powering the system.

OpenAI’s GPT-3 emitted about 500 metric tons of CO2 during training, while Meta’s OPT produced approximately 75 tons, according to research from AI start-up Hugging Face. To put that into context, 50 metric tons of CO2 is the equivalent of about 60 flights between London and New York.

Many things can be done to reduce the carbon footprint of AI.?

One option is to increase the amount of renewable energy used to power AI systems. This can be achieved by locating them in territories with high levels of renewable options, using power purchase agreements to buy directly from wind and solar projects, or by building the renewable infrastructure from scratch.

The data centers full of GPUs that train LLMs generate a lot of heat and the industry is looking at different ways to use it, like in district heating or to generate energy to feed back into the systems.

Using private LLMs with more focused datasets can also go a long way in reducing CO2 emissions. “For instance, by building DialpadGPT on top of OSS models like Llama2, we're avoiding rerunning all that baseline compute and focusing on optimization instead. In our Dialpad products, our focus on real-time functionality and cost has led us to focus on optimization. The optimized models are faster, cheaper and have a lower carbon footprint,” says Simon Corston-Oliver, Director of Machine Learning at Dialpad.

Asking even a relatively simple question to ChatGPT uses many more multiples of energy than asking the same question to an optimized LLM, which is more likely to return a faster and more accurate answer—in a more energy-efficient way.

Deploying AI for good?

The other side of the equation is: What can AI do to help solve environmental challenges??

David Jensen, coordinator of the United Nations Environment Program's Digital Transformation sub-programme, highlights several areas where AI can play a role, such as designing more energy-efficient buildings, monitoring deforestation, and optimizing renewable energy deployment.

AI can also be used on a macro scale in activities such as satellite monitoring of global emissions, or in a more granular way, such as automatically turning off lights or heat.

Research by PwC UK estimates that the application of AI could reduce global greenhouse gas emissions by 4% by 2030, equivalent to 2.4 Gt CO2e or the 2030 annual emissions of Australia, Canada, and Japan combined.

In an increasingly complicated world, we need to avoid reaching for simple and polarizing answers. AI can help with that too.