Does a single AI query use a bottle of water?

This More or Less podcast episode investigates the water consumption of AI systems, specifically language models like ChatGPT. The episode features an interview with Dr. Xiaole Ren, whose research quantified the water used, both directly in data centers and indirectly in electricity generation. While initial estimates suggested a significant amount of water per query (around 500ml for 10-50 queries using GPT-3), later analysis by the Washington Post, focusing on GPT-4, indicated a similar amount for a single query. However, the podcast also highlights the limitations of these calculations due to limited data access and the varying nature of AI queries. Further discussion with a computer scientist from Hugging Face emphasizes the need for more precise definitions of "average query" and points out discrepancies between different estimations, including a much lower figure provided by OpenAI's CEO. Finally, the podcast mentions data from Mistral, a French AI company, suggesting a different water usage figure. The overall conclusion is that while significant water is used, the exact amount remains uncertain due to data limitations and the complexity of the issue.