Large language models now help write a great deal of code, but they use energy in training and in every request. This paper compares the energy, and so the carbon footprint, of solving programming tasks by hand with solving them with an LLM assistant, using Codeforces problems as a controlled stand-in for everyday development. Averaged across the tasks, LLM-assisted code generation had 32.72 times the carbon footprint of the manual approach, and the gap widened as tasks became more complex. The authors propose ways to reduce the footprint of LLM-assisted development, a question at the heart of GreenCode's own use of AI.
Authors: Kuen Sum Cheung, Mayuri Kaul, Gunel Jahangirova, Mohammad Reza Mousavi, Eric Zie
Contributing partner: King's College London
