How to Scream into the Wind Without Losing Your Voice: A Behavioral Analysis of LLM Use and Environmental Impact

Abstract Large language models (LLMs) are a type of AI (artificial intelligence). They are large-scale, sophisticated neural networks trained on vast amounts of text data to mimic human language. The increased adoption of LLMs warrants an analysis of their substantial, rapidly growing environmental footprint. This paper is structured in three broad sections. First, we offer a behavior-analytic lens into the rapidly increasing adoption of this technology. The analysis is provided at the individual, organizational, and cultural levels, with a discussion of efficiency gains being absorbed into productivity contingencies. Second, we provide a cradle-to-grave overview of the environmental impact of this technology: cradle (building the infrastructure), life (training and running the model), grave (disposal of electronic waste), and the potential infrastructure lock-in costs. The analysis includes a discussion of embedded carbon, an area often overlooked in discussions of environmental impacts. To close, we provide suggestions for behavior analysts at the individual level to the cultural level, with an emphasis on where leverage is most possible.