Keep CALM: Mitigating Anxiety-Inducing Language in Multilingual Agents

Conversational agents routinely handle health-related and emotionally sensitive queries, yet standard safety protocols can produce medically cautious responses with unnecessarily alarmist language. We study this failure mode across six languages and seven Large Language Models using paired two-turn scenarios that separate clinical content from emotional escalation. We define anxiety-inducing language as threat, urgency, or worst-case framing whose intensity is disproportionate to the clinically justified risk, and evaluate it through three complementary sources: native-speaker judgments, clinician assessments of clinical appropriateness and urgency, and automatic lexical and semantic measures validated against these judgments. This design explicitly distinguishes appropriate emergency referral from unsupported catastrophising and treats cross-lingual differences as behavioural variation, without attributing them to cultural causes. We then introduce (Context-Aware Layered Mediation), a training-free three-stage framework that separates emotional assessment, response drafting, and safety review. Across multiple backbone models, reduces human-rated alarmism and state-anxiety change while preserving clinical correctness and emergency-referral sensitivity. Our validation further shows that lexical metrics provide an interpretable but incomplete signal, whereas semantic scoring aligns more closely with human and clinician judgments. These findings indicate that emotional restraint and clinical caution can be jointly optimised in multilingual health interactions.