AI's Environmental Impact

Part 1: The Real Cost of AI

If you asked ChatGPT what the environmental costs of speaking to it were, it might tell you something like “The environmental cost of using ChatGPT involves significant energy consumption associated carbon emissions”.

However, not only does this fail to conclusively assess what the actual cost of AI is, but it also only considers the electricity cost of training and deployment, disregarding so many other environmental impacts. For example, when crunching the numbers you can find that the average conversation with ChatGPT consumes around 0.0875-kilowatt hours of electricity and also 500ml of fresh water.

This environmental footprint arises from the usage of data centres, which are huge warehouse-like structures which contain an immense amount of servers and processing power to host and train artificial intelligence models. Unsurprisingly, the amount of electricity required to run and maintain these centres is very high and most of this energy will not be coming from renewable sources, meaning that carbon emissions are inevitable. In fact, the cost of running data centres collectively accounts for about 1-2% of global electricity usage. What is less clear, however, is how ‘thirsty’ the centres are in their water consumption.

The water cost can be broken down into two categories. Firstly, there is the indirect consumption of water through the manufacturing of the data centres themselves. Annie Marshall, who supports Microsoft's environmental communications, breaks it down by explaining "the costs come from the construction of data centres and the associated hardware components such as semiconductors, servers, and racks."

In addition to this, it is often seen as more efficient to mass-produce cheaper parts and replace them when they inevitably break. For example, lead-acid batteries are often used and these are inefficient and temperamental, but also cheap and replaceable. As a result, there are over 2 million tonnes of e-waste each year from data centres. This is wasted energy, resources and water.

However, there is also a direct water cost of running a data centre. Shaolei Ren, an expert in AI's water footprint, explained, “Data centres use an enormous amount of water for both on-site cooling and off-site electricity generation”. As electricity flows between the many racks of servers in a data centre, it begins to heat up the servers. As these temperatures climb, so does the risk of damaging the complex, and often delicate, components within the servers. To avoid this, these servers have to be continually cooled down. Whilst many home computing systems may rely on fans to do this, this method can actually be quite inefficient and slow on a bigger scale. Marshall shared "Air cooling is a common method, but it’s not as effective as liquid cooling". This is also commonly referred to as water cooling, which works by running cold water past the parts to take the heat and transport it elsewhere, before repeating this process.

Whilst this may seem as though the water is being reused, it can only be cycled between 3 and 10 times before a build-up of salt and minerals which can cause issues such as corrosion and biological growth. This is a problem as fresh water is a scarce resource in many areas of the world, leading to data centres competing with the local populations’ water needs. Marshall explained, "The growth of AI could stress local electric grids and water resources in some regions where data centre development is expanding to meet the needs of AI". A full map of data centres shows which areas might be most affected by this. In fact, issues have already arisen across the world as a result of this, ranging from Arizona to Chile to Spain.

At the same time, these considerations do not seem to be the top priority for those hosting AI technology. Marshall shared how when choosing data centre sites, Microsoft "considers a range of criteria, including availability of skilled labour, ample and reliable resources, and multiple high-capacity network connections to determine the long-term viability of each site."

Without greater consideration for the environmental impact of AI, the emissions and effects of the climate crisis will only get worse. For example, Microsoft’s carbon emissions have grown by 30% since 2020 and they said that this was a result of their investment in AI. This is on top of the fact that companies are “trying to play all types of tricks to make them look better”, according to Ren.

He explained how companies typically show market-based carbon emissions instead of their actual emissions based on their physical locations. In other words, the figures they provide are often the numbers after they have applied carbon offsetting, a process by which companies can compensate for greater emissions by investing in projects that reduce or remove emissions elsewhere.

Whilst still useful, this ultimately avoids responsibility for their own emissions and is used as a shield against criticism. It also allows companies to continue to grow emissions as long as they equally invest in offset. Instead, we should focus on more concrete ways to increase the sustainability of AI and reduce its environmental impact.

Part 2: Finding Sustainability

What can we do to make AI less environmentally harmful?

The lack of consistent data highlights one of the key changes that needs to be made if AI is to be more sustainable, a call for more transparency from the big tech sector. How can we understand and counteract the impact of this technology if we do not have access to basic figures about it?

But transparency alone is unlikely to bring a green revolution within the technology sector, for these companies, the chance to win the AI race is too important and the environment is often a secondary concern. It would be all too easy to assume that AI will naturally become more sustainable, but the planet cannot afford to wait.

Matthew Pye, director of the climate academy, emphasised how the fallacy of the Kuznets Curve law has a role to play here. The curve suggests that as an economy develops, inequality is first increased before ultimately decreasing. In terms of the environment, this would suggest that the problem might eventually fix itself.

Initially, emissions start low as a new technology begins production and development, as there is little investment and deployment.

Then, as development and funding increases emissions peak because the technology becomes more widely deployed and used.

Finally, as the technology is widely deployed and used there is interest in reducing its footprint. Furthermore, supply chains are established and the cost of production is lower and efficiency is higher, reducing emissions.

This is the basic theory behind the Kuznet's curve when applied to an emerging technology's environmental footprint.

Pye does not believe the curve is accurate. He said “This Kuznets ‘law’ is not mechanically true: Just as economic growth leads to poverty reduction only when the state acts proactively to redistribute, raising living standards does not automatically lead to greater respect for the environment, as shown by the US model, which is extremely predatory. Correcting the negative environmental effects of increased growth and consumption requires regulatory intervention.” In other words, the emissions will not lower themselves automatically over time, unless there is active work in this area.

Whilst the environmental costs of AI are high and changes are needed to increase the sustainability of the technology, Dr Ramit Debnath, a researcher at the University of Cambridge’s Artificial Intelligence to the Study of Environmental Risks team (AI4ER), believes that “As with every new technology, AI’s environmental cost is huge now, but it can be used to power major climate mitigation and adaptation plans globally.” In other words, there are ways in which this machine learning technology can already be used for the benefit of the environment, and these use cases will only become more frequent.

Furthermore, Ren explores the idea of considering the cost of not using AI. Using the transport industry as an example, which has a huge environmental cost, if self-driving technology can decrease emissions and a smart grid can allow for smoother and sustainable powering of electric vehicles, all of a sudden the cost of the data centres might be quite negligible compared to the global impact this technology could have.

Part 3: Environmental Potential

With proper regulation, how can AI be used to help protect our environment and bring about positive change?

AI technology is already occasionally being deployed to measure, record and mitigate the effects of the climate crisis. Dr Debnath highlighted that “AI is standing apart with better weather prediction, enabling preparedness and reducing damages to extreme weather events.” There are projects which track the melting of icebergs, improve recycling and predict weather disasters.

But the potential of AI has not yet been reached, Debnath added “AI has tremendous potential in managing energy use of cities, estimating carbon storage potential of forests and find the most optimal location for planting trees that can offer wider co-benefits to people and society.” From this, it is clear that the potential for AI is immense, even if the environmental cost is high. To better balance the cost and potential, Dr Debnath emphasised the importance of “actively working towards greener and sustainable AI.”

Ren suggests that one change that could be made to increase this sustainability is to work on where the code itself is deployed. Titled Geographical Load Bearing, this solution uses predictive data to distribute user requests to an AI (such as ChatGPT) to various data centres across the globe with a focus on minimising environmental harm, instead of prioritising faster response times.

Ren explained how “different regions have different water efficiencies. For example, Arizona has a much worse water efficiency compared to Iowa”. As a result, Iowa might be better positioned to handle more user requests. But there are many more factors to consider.

Unfortunately, this requires more processing power and data than humans can handle. Yet, we can utilise AI technology to help with this issue. A specifically trained AI program could be trained to consider the availability of fresh water in an area, the outside temperatures, the cost and sustainability of electricity (and more) to dynamically spread requests across the globe, thereby minimising the environmental impact. To play around with what this might look like, you can use this interactive tool I created.

AI projects can also benefit the environment more indirectly. For example, AI has the potential to revolutionise the way in which our economy and supply and demand are organised, reducing waste and overconsumption. Leo Schlichter, an academic at the Berlin School of Law and Economics, decided to explore how the degrowth movement can be combined with AI technology to facilitate a planned economy. Leo said:

"My hope is that when consequences of the climate crisis and environmental destruction become clearer and increasingly affect people in wealthier nations, people will realize that a growth-based system cannot be sustained much longer and that there is interconnectedness of ecological issues with other important issues like war, migration, political instability, poverty and more."

Simply put, degrowth seeks to challenge the common assumption that growth is always good in an economy, and argues that it is a leading cause of overproduction, poverty and emissions. To enact these ideas in society, Leo explores how AI can help to predict demand and prevent overproduction, facilitate greater knowledge sharing and increase collaboration, and empower democratic insitutions to ensure society is developed with the interests of everyone in mind. Interestingly, the roadblock here might not be the technology as "mega-corporations like Amazon or Walmart already use AI for internal planning", so the framework already exists. Instead, Leo believes that "The main obstacles are political. It will be incredibly difficult to radically transform an economic system."

Perhaps the most widespread use of AI is currently chatbots and image generation. This technology has widespread potential across different sectors, and it would be a shame to limit the scope of its uses to only things which benefit the environment. In fact, elements of this article have been supported by various AI technologies as various graphics were generated using Adobe Illustrator and Microsoft's Copilot helped to discover key papers and sources. At the same time, users of this technology must be transparent about how and where they use it and companies must fairly and consistently obtain permission to use other people's work as training data and compensate them accordingly.

The point is not to shame and shut down non-environmental uses of AI, but to highlight the importance of cleaning up the way in which AI is deployed and utilised in order to increase the sustainability of the technology. Furthermore, a greater awareness of the costs of using AI will allow users to make informed judgements about when and how they use it.

It is of the vital that we ensure that the companies which are spearheading the development of AI technology are held to high ethical and environmental standards and pressured to explore the positive environmental potential of their technology. If this can be achieved, the power of AI could have transformative benefits for our planet. However, if ignored, its harmful environmental impacts could spell disaster for future generations.