- AI's environmental impact is largely unknown, but a new paper points gives some insight into it.
- Training GPT-3 requires water to stave off the heat produced during the computational process.
- Every 20 to 50 questions, ChatGPT servers need to "drink" the equivalent of a 16.9 oz water bottle.
As the general public rushes to generative AI tools like ChatGPT, the environmental impact of the new technology is beginning to come to light.
While there's still very little data on AI and sustainability, a recent study from researchers at the University of California, Riverside and the University of Texas, Arlington points to the water footprint of AI models like OpenAI's GPT-3 and GPT-4.
While training GPT-3 in its data centers, Microsoft was estimated to have used 700,000 liters — or about 185,000 gallons — of fresh water. That's enough water to fill a nuclear reactor's cooling tower, per Gizmodo, and the same amount that is used to produce 370 BMW cars or 320 Tesla vehicles, per the study.
Using these numbers, it was determined that ChatGPT would require 500 ml of water, or a standard 16.9 oz water bottle, for every 20 to 50 questions answered.
"While a 500 ml bottle of water might not seem too much, the total combined water footprint for inference is still extremely large" due to ChatGPT's large user base, the study's authors wrote.
Microsoft and OpenAI did not respond to Insider's requests for comment.
AI models like GPT-3 and GPT-4 are hosted in data centers, which are physical warehouses that store large swaths of computational servers. These servers identify patterns and linkages across massive datasets, which in turn, utilizes energy, whether that's electricity, coal, nuclear power, or natural gas.
Significant expenditures of energy are used in the training process, which is then converted into heat. Water is then used on-site to keep temperatures in check across the entire infrastructure. Fresh water is required for proper humidity control and because saltwater can lead to "corrosion, clogged water pipes, and bacterial growth," per the study.
Moving forward, these figures could "increase by multiple times for the newly-launched GPT-4 that has a significantly larger model size," the researchers said.
Using their own methodology that computes on-site and off-site water usage effectiveness (WUE), in addition to energy usage, the team of researchers also developed water footprint estimates for Google's large language model known as LaMDA.
Ultimately, though, a lack of transparency surrounding the water consumption numbers involved with AI training makes it difficult to pinpoint the actual footprint. When asked about LaMDA's water usage, Google pointed to a November 2022 report that published 2021 data on the broad consumption of water across data centers.
"While it is impossible to know the actual water footprint without detailed information from Google, our estimate shows that the total water footprint of training LaMDA is in the order of million liters," the researchers wrote.
While the carbon footprint involved with generative AI is beginning to ring alarm bells, researchers said their paper seeks to "highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI."