Signal Quantifying the relationship between clinical safety and environmental impact in therapeutic LLMs

Summary

Researchers examined the relationship between clinical safety and environmental cost in large language models used for mental-health support by combining K-Bench clinical safety scores with EcoLogits life-cycle assessment estimates across 47 supported model configurations. They evaluated performance and environmental impact together across energy use, carbon emissions, water consumption and abiotic resource depletion. The results showed a sharp non-linear trade-off at the upper end of the safety distribution, where each 2.61 percentage-point rise in clinical safety score corresponded to roughly a 60-fold increase in energy use per million output tokens. The researchers found that adding test-time compute did not consistently improve clinical safety, and in some configurations safety scores actually fell. This suggests that leaning on larger models or extra inference-time computation alone can be an inefficient way to improve safety in therapeutic AI, while dynamic model selection could cut environmental impact without sacrificing performance.