The carbon impact of AI vs search engines

Note from author: the idea for this article came before the news that Microsoft had invested in OpenAI and are looking to integrate generative AI into the Bing search engine. 

Artificial intelligence (AI) has become the hottest topic this year, with the exponential growth of ChatGPT and Google’s unveiling of Bard: but how do they compare to search engines?

While there are ethical and societal concerns, the focus of this article is on the environmental impact of using these services in terms of their estimated emissions. For this, we worked on the premise of questions and answer, searching for an answer to a fairly complex question. 

Search engine 

Google themselves reported that they emit an estimated 0.2g CO2e per search, but when you pair that with the landing page emitting an average of 1.15g per page view (where multiple pages can be visited to find the right answer) then it quickly becomes a much bigger issue. 

We ran a short survey and found out that most people searched multiple times and clicked on multiple pages to find the right answer - averaging 2 searches (even more if using different search engines such as Ecosia, which uses Bing in the background) and 3 page visits means that per answer, a user would emit an average of 3.85g. 

This may not seem like much – especially if many users may only search once or twice a day. However, those searches become more frequent when researching a topic, solving a difficult problem, or looking for the ideal product or service, so the impact adds up rapidly. This is even more of an issue for those working in the digital industry and adding to the impact professionally as well as personally. 

Artificial intelligence 

While figures are varied, and none of them have yet been called official, there have been multiple estimates that measure the emissions of ChatGPT at anywhere between 23kg to nearly 25 tons every single day - even asking it this question leads to even more confusion about its emissions. 

With an article updated less than two weeks ago, a member of ClimateAction.tech wrote a study that estimated the emissions at 25 tons, so using his numbers we can estimate the emission of a single response to be 4.14g. 

Once turned into a conversation – with the goal of obtaining more responses to get to the answer you need – this quickly multiplies. With most conversations consisting of around 5 responses, the estimated total on average rises to around 20.72g. 

Cutting down on carbon 

The use of Artificial Intelligence will only grow over the next few years, with that comes additional usage and additional emissions which are at tipping point and desperately need lowering across the board. 

Minimising this impact comes down to a number of factors – many of which relate to existing issues within the digital industry as a whole. The carbon impact of digital is already massive: the internet consumes more electricity than the entire United Kingdom. Generating this energy should come from sustainable sources; by ensuring that data centres are powered by renewable energy. It’s a solution that would minimise the carbon impact of the entire digital industry – AI included.  

Within the wide range of capabilities that AI offers, writing code is one highly useful feature it boasts. Paradoxically, there is a necessity for higher efficiency within its language model to reduce energy demand. Writing efficient code within the software allows for a lower impact for each search query.  

Although the industry is still getting to grips with the range of solutions offered by AI, there is no better time to understand the carbon impact of the technology. Further research will undoubtedly bring to light the exact carbon impact of AI and how it may help us towards lessening the carbon impact of digital rather than adding further to it. 

As with all current digital solutions, the digital industry still has some work to do around educating users and developers into harnessing the power of digital with a sharp focus on efficiency.