Understanding and curbing the environmental cost of AI

A lot of people say artificial intelligence (AI) can help solve many of the world’s greatest challenges. This includes environmental crises, such as climate change, biodiversity loss and pollution.

There is much truth in this line of thinking – right now, we are still at the nascent phase of leveraging AI and there are already countless examples of the technology being used to make our world cleaner and greener.

At one end of the business spectrum, tech giant Google has been using AI to change the flight paths of aeroplanes and reduce the line-shaped clouds known as contrails that streak behind them and which account for a third of the aviation industry’s climate impact. At the other end of the scale, there are start-ups such as Greyparrot using AI to analyse waste facilities and increase the amount of items recycled.

However, there is also a direct environmental cost of AI because of the technology's huge power consumption requirements. At the most basic level, a ChatGPT query is estimated to require 10 times as much electricity as a simple Google search. Generative AI systems are also believed to use 33 times more energy than machines running task-specific software. In addition, data centres (not accounting for those related to cryptocurrencies) could account for as much as three per cent of the world’s electricity demand by next year, according to the International Energy Agency – double the levels observed in 2022.

It seems clear, therefore, that the current corporate focus on AI is starting to have an adverse impact on the environment. For example, Google reported a 48 per cent increase in its greenhouse gas emissions between 2019 and 2023, concluding that the rise was caused by the higher energy demands of its AI initiatives.

AI consumes so much power that the big players, aware of their net zero pledges, have literally gone nuclear to start addressing the demand, as well as related infrastructure challenges. Nuclear energy is almost completely carbon-free and Google recently signed a deal to use small nuclear reactors to help power its data centres. This follows in the footsteps of Microsoft, which has recommissioned a closed nuclear power plant that was the site of a partial meltdown in 1979 and Amazon, which paid $650m for the privilege of putting a data centre campus adjacent to a nuclear plant.

Regardless of where you stand on the use of nuclear energy, most businesses simply don’t have a spare $650m to relocate next to an existing nuclear facility. So, what steps can the average company take to offset the environmental impact of their AI usage?

Informed decision-making

First, it’s important to be aware of the simple fact that increased AI usage will impact a company’s energy consumption and use that knowledge to inform decision-making in all areas of the business.

Capturing quantitative carbon data associated with AI usage will form part of existing governance and regulatory reporting responsibilities for many organisations. However, for others, it may be a new capability that needs developing to support the company’s ethical AI policies and frameworks.

Companies also need to work with their internal teams – including those in governance, sourcing, operations and technology – as well as external suppliers to measure, track and forecast the carbon footprint of their AI usage, just as they would with any other system they use. Again, this is likely to form an extension of their existing carbon accounting for many, while it may be a new muscle requiring development for others.

Factors to consider when designing solutions 

The second step is to make carbon impact part of the criteria when architecting platforms and solutions. Just as not all AI models perform equally, they also consume power differently too, as have been evidenced by the cases of IBM and DeepSeek. When designing solutions that use AI, companies should make conscious decisions by considering the following:

• Whether AI is really needed for this particular problem and whether another technology might be better suited.
• What the right model is in this instance. Smaller and more efficient AI models, for example, require fewer resources but can yield similar accuracy in the right circumstances, thus lowering the system’s carbon footprint. 
• The AI model’s refresh rate. With new models released frequently, consider the longevity of the model in question, as a large proportion of  energy usage comes in the initial training phase.
• Tools. To measure and monitor the performance of models throughout their lifecycle, it’s important to put the right tools in place.

The big picture

The third step is to understand why all this matters. After all, ESG looks to be taking a back seat in boardrooms following its 2021 peak. We’ve also had a succession of underwhelming COP summits and that is putting it politely, with populist politicians rising in influence. From all this, a sense that sustainability is too expensive has started to permeate into the mainstream. While being green was generally seen as a positive across the political spectrum a few years ago, that is no longer the case today.

However, evidence suggests that consumers still want to see companies caring for the environment and that many will pay a so-called ‘green premium’. For example, EU research has found that the environmental impact of a product is “very important” to 73 per cent of customers when making a purchase decision. Even in the face of inflationary pressures, PwC has also suggested that consumers are willing to spend an average of 10 per cent more on sustainably produced products.

Companies should keep these trends and findings in mind when deliberating over the adoption of AI programmes. For all its transformative power, at some point leaders will have to consider the implications for the environment. If they don’t, their customers may well do it for them.