Therapy LLMs need 60x energy for just 2.61% safety gain

A new arXiv preprint (2608.11830) from researchers including Alireza A. Safaei, Matthew J. Vowels, and Apoorv Jha quantifies the hidden environmental cost of making AI therapists safer. The study combines K-Bench clinical safety scores with EcoLogits life-cycle assessments across 47 supported model configurations, measuring energy use, carbon emissions, water consumption, and abiotic depletion. The headline result: near the top of the safety distribution, a modest 2.61 percentage-point improvement in clinical safety was associated with an approximately 60-fold increase in estimated energy use per million output tokens. This non-linear trade-off means that chasing the last few points of safety on benchmarks can have outsized environmental consequences.

Interestingly, the researchers found that adding test-time compute did not reliably improve clinical safety — in some configurations, it was actually associated with lower safety scores. This challenges the assumption that bigger models or more inference-time computation are the default path to safer therapeutic AI. Instead, the authors suggest dynamic model selection and model cascading: routing higher-risk cases to more capable (and energy-hungry) models while using smaller, efficient models for routine queries. This approach could dramatically cut energy, carbon, and water use without sacrificing clinical performance. For AI teams deploying mental health chatbots, the paper is a warning that safety and sustainability are now intertwined decisions — and that smart routing may be the most practical lever for both.