Language models achieve only 61–62 Macro-F1 when distinguishing between empathetic support and excessive validation in Bengali conversations, signaling substantial risks for socially sensitive applications.