The environmental impact of generative AI
Everyone’s been talking about the untapped potential and general downsides of generative AI this past year. Far less time has been spent discussing its environmental impact. From what we know, it’s not great. And according to experts, that negative effect is set to worsen, especially with major tech giants rushing to integrate AI into search engines.
Why generative AI severely impacts the environment
It can be easy to forget the amount of energy needed for computing. According to the International Energy Agency, data centers account for around 1% of the world’s electricity demand and greenhouse gas emissions, while the worldwide tech sector is responsible for approximately 1.8% to 3.9% of global emissions. Regular indexing and searching Internet content already requires quite hefty resources, but AI computing needs a lot more. Alan Woodward, professor of cybersecurity at the University of Surrey in the UK, told Wired:
“It requires processing power as well as storage and efficient search. Every time we see a step change in online processing, we see significant increases in the power and cooling resources required by large processing centers.”
In recent years, several studies examining how much computing power and carbon emissions are generated by AI confirm this. In 2019 researchers at the University of Massachusetts, Amherst, found that the process of training several large AI models can emit more than 626,000 pounds of carbon dioxide equivalent. For context, that is more than five times the emissions the average American car will emit in a lifetime, from manufacturing to usage. The same study found that training its own Natural Language Processing model generated 284 tons of carbon dioxide. To put this number in perspective, the average human generates about 5 tons of carbon dioxide annually.
MIT Technology Review has reported that during training, OpenAI’s GPT-3 and Meta’s OPT are estimated to have emitted more than 500 and 75 metric tons of carbon dioxide, respectively. However, the precise impact is difficult to calculate because there’s no standardized way of measuring carbon dioxide emissions yet. These numbers are based on external estimates and limited data released by the companies.
How the impact is set to worsen
With generative AI’s recent surge in popularity, it should come as no surprise that tech companies want to double down on their efforts to get their piece of the pie. Microsoft has already integrated ChatGPT into its Edge browser and Bing search engine, and Google has launched its own AI chatbot, Bard, which it has described as “a complement to search.” Chinese search engine Baidu recently announced it would also be launching its own AI chatbot.
According to Wired, this ramping up of AI-related tools could lead to a fivefold increase in necessary computing power and carbon emissions. It’s estimated that ChatGPT has around 13 million users per day, while Bing handles around 500 million daily searches. Integrating the two will undoubtedly come with a hefty increase in power usage, particularly since ChatGPT’s model is only trained to understand the world up until late 2021. As Martin Bouchard, the co-founder of Canadian data center company QScale, told Wired, “If they’re going to retrain the model often and add more parameters and stuff, it’s a totally different scale of things.”
How tech companies could reduce the impact
There are numerous ways tech companies can reduce the strain these models can have on the environment. Building data centers that run on cleaner energy sources is a significant one. Towards Data Science suggests 17 solutions, from running computational procedures in geographic locations powered by more renewable energy sources and only training models at times of day when the energy is cleaner. The Machine Learning Emissions Calculator is a great resource for estimating the carbon impact of various regions, cloud providers, and types of hardware.
