As AI embeds itself into all of our lives and climate change becomes more prominent, the climate impacts of AI have become an unavoidable topic. Like any other new technology, AI has a significant impact on society and the environment. This article is part of a series in which we dive deeper into the impacts of AI, big and small. This article is specifically about the impact of AI on carbon emissions and climate change.

As you may have heard, the climate impacts of AI are anything but insignificant. However, in order to fully understand how creating AI models impacts climate change, you’ll need to understand what goes behind creating something like ChatGPT.

Before a model like ChatGPT can do anything, it needs to be trained. Training a model is where a majority of the direct cost shows up. AI models need to be trained on millions or billions of data points to learn the skills they have. This applies to any deep learning-based model, whether it’s ChatGPT, or medical vision bots that can detect tumors better than doctors<link to article>. The amount of data that AI models are fed allows them to have a strong understanding of the problem they are meant to solve. This process is extremely resource and energy intensive. And this isn’t where the direct climate cost of AI ends, these models continue to use energy every time they are run, even after training, albeit much lower. Every time someone asks ChatGPT a question, it uses some energy, and although it is much lower than the energy used during training, that small amount adds up when billions of people ask it questions.

Another important concept is model size. Deep learning models have parameters that encode their knowledge during training. This is important because all of these parameters are used during training and when a model is run after training; as you might expect, larger models use more energy. This is particularly concerning as, over the past few years, deep learning models, have grown exponentially. We have seen major advances in AI as a result, but it must be understood that the carbon footprint of AI has also majorly increased.

We can’t know the exact impact of training or running every AI model, as that data isn’t public. However, researchers can estimate these impacts based on parameter size and training methodology. AI leader Hugging Face has publicized a report on the emissions of its own LLM (large language model), which helps contextualize other LLMs like OpenAI’s GPT. As a disclaimer, Hugging Face’s model was trained on a supercomputer that was mostly powered by nuclear energy and therefore had a significantly smaller carbon footprint than similarly sized models like GPT. The actual impact of training Hugging Face’s BLOOM model was 50 metric tons of carbon while the impact of each day of use was 19kg.

For comparison, researchers from Google and UC Berkeley estimated the training of GPT-3 to be a whopping 500 metric tons.

Below are some visualizations of how this compares to other things that are considered to be carbon-intensive.

One might note that the data shows that training an AI model, by itself, has a significant climate impact but can be smaller compared to other things considered to be serious threats in terms of climate change. For example, the entire training of GPT-3 emitted as much carbon as the creation of just 8 automobiles. However, millions of people use GPT, while only a handful of people can utilize 8 cars. Given the amount of users AI has, the carbon impact of AI doesn’t seem so bad per capita. So the question is, what is all of the climate concern about? Does this really matter when there are much more significant contributors to climate change that have already been widely adopted and societally accepted?

The answer is a resounding yes for a number of reasons. First, anything we can do to avert the most catastrophic of scenarios regarding global warming is significant. The tech sector has a responsibility to keep emissions low, and as Hugging Face showed, reducing the emissions of AI models is doable. Even if GPT’s per capita emissions are lower than an airplane’s, that doesn’t mean they are low enough.  Secondly, the only reason GPT’s emissions per capita are so low is because it has so many users. There are hundreds of massive AI models being created, from universities to tech companies, and many of them have limited use outside of as thought experiments. The energy usage of these models are massive, and reducing their emissions is of paramount importance. And finally, as you might have understood from the title, there are further impacts of AI with regard to climate change that could be even worse than those of creating and using a model.

One of the most popular applications of AI is in advertising and recommendations. Have you ever wondered how a relevant ad or video found you? The answer lies in data and AI. Through AI, companies are able to more than ever push you towards their products. This is where AI can create a much bigger impact on climate change and environmental destruction.

Now, an increase in consumption is something that generally happens with the introduction of every technology. But the degree to which technology has increased consumption has only grown recently. And AI is increasing consumption in ways like this more than anyone can really estimate. Recommendation systems are simply the most egregious examples, but AI is driving consumerism in more ways than can be listed. Compound this with the already significant impact of AI on emissions just through existing, and you’ll see exactly why there is so much concern regarding this issue.

Researchers estimate that this is where AI can have a truly massive impact on emissions. They create a reinforcement cycle of consumption. Whether it be recommendation algorithms on TikTok or Youtube that keep you hooked onto them, or Amazon’s AI that encourages you to buy more and more, AI has had a undeniable impact on consumption. This is where being careful to recognize AI’s influences on our consumption habits can make an impact.

The climate impacts of AI are a pressing concern, but there is reason to be hopeful. The case of Hugging Face's BLOOM model demonstrates that it is possible to reduce emissions from AI models. By employing sustainable practices and technologies, we can mitigate the carbon footprint associated with training and using AI. Additionally, recognizing the broader implications of AI on consumption patterns and actively working towards sustainable solutions will help us navigate the challenges posed by AI and climate change. With collective efforts and a commitment to environmentally friendly practices, we can pave the way for a better and more sustainable future.

‍

For more on this topic and the impacts of things around us on people and the planet follow @joinbeaker on Instagram.

‍

This article is an extended piece for this Instagram post.