Estimating the Increase in Emissions caused by AI-augmented Search

TL;DR

This paper analyzes the energy and CO2 implications of AI-augmented search by benchmarking per-query energy against conventional Google-style queries. It updates conventional-search energy using modern data-center efficiency and decarbonization data, then quantifies per-query energy for AI-generated summaries via BLOOM-like models and GPT-3-scale equivalents. Findings indicate per-query emissions can be on the order of 60–90× higher for ChatGPT and ~75× for BLOOM, with total increases potentially exceeding 60–70× when adopted broadly; training and embodied-carbon factors are secondary but non-negligible. The study highlights the climate relevance of AI-enabled search and argues for aggressive sector-wide decarbonization and efficiency improvements to keep ICT emissions in check through 2040.

Abstract

AI-generated answers to conventional search queries dramatically increase the energy consumption. By our estimates, energy demand increase by 60-70 times. This is a based on an updated estimate of energy consumption for conventional search and recent work on the energy demand of queries to the BLOOM model, a 176B parameter model, and OpenAI's GPT-3, which is of similar complexity.