I Know Why Lying about AI Water Use is So Easy

Quick Overview

The reason lying about AI water use is easy is that the actual life cycle water use for training models is complex and often hidden, leading to misleading comparisons where AI water use seems small compared to massive agricultural water use, while ignoring the pollution impact of thermal power plants used for AI electricity generation and the fact that AI companies are not transparent about their water sources.

Key Points: Sam Altman's claim that a ChatGPT query uses 1/15th of a teaspoon of water is highly misleading because it ignores the vast water used in the training phase of the AI models. US Corn production alone requires approximately 20 trillion gallons of water annually, which is about 80 times more water than the total estimated global AI server water use (200 billion gallons) when comparing usage per unit of output. Thermal electric power plants, which generate electricity for AI data centers, account for 45% of all fresh water withdrawals in the US, emphasizing that cooling water use is a major factor often omitted in AI water discussions. The water used for cooling AI data centers is typically municipal water, which is often treated to drinking water quality, unlike industrial water used for power generation, which is often self-supplied and returned to local water bodies with thermal pollution. The total environmental impact of AI is complex, involving not just direct water use but also energy consumption (leading to large carbon budgets) and the environmental impact of cooling water discharge that can harm local ecosystems. The water used to train large models like GPT-5 is not directly counted in query estimates; instead, the training process—which is energy-intensive—is often separated from the query process, obscuring the true resource cost.

Context: The video analyzes a claim made by Sam Altman regarding the minimal water usage of an individual ChatGPT query, contrasting this small figure with the much larger, often hidden, water consumption associated with the entire lifecycle of training large language models (LLMs). The speaker critiques the transparency of AI companies like OpenAI and discusses the broader resource implications of massive AI infrastructure, particularly in terms of electricity generation and water withdrawal from local sources.