🤖 AI Summary
This study investigates the trade-off between clinical safety and environmental impact in therapeutic large language models. By integrating K-Bench clinical safety scores with EcoLogits life cycle assessment, it performs a fine-grained analysis of 47 model configurations across four environmental dimensions: energy consumption, carbon emissions, water use, and abiotic resource depletion. The work reveals, for the first time, that within high-safety regimes, marginal gains in clinical safety incur nonlinear surges in environmental costs—specifically, a mere 2.61-point increase in safety score corresponds to approximately a 60-fold rise in energy consumption, with additional inference compute not necessarily yielding further safety benefits. To address this, the study proposes dynamic model selection strategies, such as model cascading, which can maintain performance in high-risk clinical scenarios while substantially reducing environmental footprint.
📝 Abstract
The deployment of large language models (LLMs) in mental health contexts raises questions about the relationship between clinical safety and environmental cost. In this paper, we examine this relationship by combining K-Bench clinical safety scores with EcoLogits life-cycle assessment estimates across 47 supported model configurations. We evaluate model performance and environmental impact across four dimensions: energy use, carbon emissions, water consumption, and abiotic depletion. The results indicate a non-linear trade-off at the upper end of the safety distribution: a 2.61 percentage-point increase in clinical safety score corresponded to an approximately 60-fold increase in estimated energy use per million output tokens. Row-level analyses further suggest that additional test-time compute did not consistently improve clinical safety and, in some configurations, was associated with lower clinical safety scores. These findings suggest that relying solely on larger models or additional inference-time computation may be an inefficient strategy for improving safety in therapeutic AI systems. We discuss the implications for sustainable deployment and highlight dynamic model selection, including model cascading, as a potential approach for reducing environmental impact while preserving clinical performance in higher-risk cases.