Environmental impact of large language models in medicine
Oliver Kleinig, Shreyans Sinhal, Rushan Khurram, Christina Gao, Luke Spajic, Andrew C.W. Zannettino, Margaret Schnitzler, Christina Guo, Sarah Zaman, Harry Smallbone, Mana Ittimani, Weng Onn Chan, Brandon Stretton, Harry Godber, Justin Chan, Richard Turner, Leigh R. Warren, Jonathan Clarke, Gopal Sivagangabalan, Matthew Marshall‐Webb, Genevieve Moseley, Simon Driscoll, Pramesh Kovoor, Clara K Chow, Yuchen Luo, Aravinda Thiagalingam, Ammar Zaka, Paul A. Gould, Fabio Ramponi, Aashray Gupta, Joshua G. Kovoor, Stephen Bacchi
Abstract
The environmental impact of large language models (LLMs) in medicine spans carbon emission, water consumption and rare mineral usage. Prior-generation LLMs, such as GPT-3, already have concerning environmental impacts. Next-generation LLMs, such as GPT-4, are more energy intensive and used frequently, posing potentially significant environmental harms. We propose a five-step pathway for clinical researchers to minimise the environmental impact of the natural language algorithms they create.