Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models

This highly cited research paper provides a methodology for estimating the water footprint of AI models, revealing that training GPT-3 alone can consume millions of liters of freshwater. It quantifies both direct water use (data center cooling) and indirect water use (electricity generation), showing that a single ChatGPT interaction consumes roughly 16.9 milliliters of water. The authors call for greater transparency in reporting AI's water consumption and argue that water footprint must be addressed alongside carbon footprint to achieve truly sustainable AI.