Part 6: How do LLMs impact society?

Quick Overview

The overall environmental impact of LLMs is projected to be much worse across energy consumption, water footprint, and carbon emissions based on current growth scenarios, despite the hype suggesting otherwise, while experts and developers consistently overestimate the productivity gains from AI tools in real-world tasks.

Key Points: The environmental impact of LLMs is projected to be much worse by 2030 across energy consumption (up to 444.9 TWh in the highest case), water footprint (up to 1124.5 million m³), and carbon emissions (up to 80 Mt), according to projections from a recent study. Data centers are increasingly located in water-stressed areas, with large data centers consuming up to 5 million gallons of water per day, equivalent to the water use of a town of 10,000 to 50,000 people. A survey of 48,000 people across 47 countries shows that while 73% use general-purpose generative AI tools, 48% of employees report uploading company information into public AI tools, raising liability concerns. Actual measurements from a randomized controlled trial with developers show that AI usage slows down task completion time by about 20% (contrary to the 30-40% speedup often estimated by experts and developers). The work impact is characterized by two claim types: 'It helps me do something I couldn't otherwise do' (more beneficial for novices) versus 'It has the potential to revolutionize something' (often marketing hype). Learning to use LLMs ('prompt engineering') is insufficient for experts, who show no strong evidence of improvement over doing the work themselves, and may hinder skill development for novices by creating dependency on subpar AI output. The current AI business model risks 'enshittification' and is propelled by circular investment, where companies buy from each other hoping not to be the last one holding the bag when the bubble bursts or prices rise significantly.

Context: This presentation, identified as Part 6 of an 'All about AI' series focusing on societal impacts, transitions into an analysis of the economic and environmental consequences of Large Language Models (LLMs) and AI. The speaker moves from discussing job impacts (where LLMs might eliminate jobs or hinder skill development) to economic fears of an AI bubble, evidenced by massive capital expenditures not matched by current revenues, and finally addresses the significant environmental footprint of data centers, particularly concerning water consumption and carbon emissions.