When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots' Safety Risks for Generation Alpha

Recognizing a teenager's slang and judging the seriousness of the distress it conveys are different abilities. A therapy-bot evaluation tested 64 constructed mental-health expressions, with authenticity checked by adolescents and risk levels checked by clinicians. Across seven language models, Table 1 reports vocabulary accuracy of 76–82%, compared with 64–72% for judging clinical risk.

The researchers also compared conversations rewritten into standard and youth-oriented language while preserving their intended clinical content. This tests whether wording changes the model's assessment. The finding concerns responses to benchmark scenarios, rather than observed patient outcomes: apparent fluency with youth language can coexist with weaker risk assessment.