A recent report by researchers from UC Riverside and UT-Arlington has highlighted the undisclosed and significant water footprint of large AI models like ChatGPT on the environment. The study raises concerns for the future of the Middle East and North Africa (MENA), the world’s most water-scarce region.
The report reveals a large amount of water required to cool computational processes of advanced AI-powered language learning models (LLMs) like GPT-3 and GPT-4. The water usage is “kept as a secret,” and it is estimated that OpenAI’s ChatGPT requires a 500-milliliter bottle of water for every 20 to 50 questions and answers exchanged. More advanced models like GPT-3 and GPT-4 consume even more.
On a corporate level, training GPT-3 in Microsoft’s US data center can consume 700,000 liters (185,000 gallons) of water, which is the same amount needed to cool machines producing 370 BMW cars or 320 Tesla electric vehicles.
Read more: DEWA world’s first utility provider to enrich its services with ChatGPT
The MENA region is the most water-scarce globally, with hot climates and low water resources. The region is already a hotspot of unsustainable water use, with only 18% of wastewater recycled. Many countries in the region rely on large-scale air conditioning systems and desalination plants for freshwater. The use of AI models like ChatGPT on a large scale could exacerbate the already precarious situation.
The World Bank predicts that the MENA region will experience economic losses from climate-related water scarcity equivalent to 6-14% of gross domestic product (GDP) by 2050. The study highlights the need for corporations to disclose their water usage publicly and for countries to address unsustainable water use to mitigate the impact of AI models like ChatGPT and to secure the future of the MENA region.
For more on ChatGPT, click here.
