Path-Dependent Triage Exposes Chatbot Blind Spot

The problem is not simply that chatbots lack information; they can receive the right facts and still fail because their conversational trajectory locks them into an early, incorrect assessment.

• –Multi-turn evaluation exposed failures that single-message benchmarks largely miss.
• –Re-presenting all disclosed facts reduced emergency under-triage to 12.1%, while presenting everything upfront produced a 5.6% failure rate.
• –Developers should separate evidence gathering from risk scoring and re-evaluate urgency from scratch after each turn.
• –Stereotyped questioning and conversational perseveration suggest that safety tuning must measure what models ask, not only what they answer.
• –Human escalation remains essential for mental-health crisis workflows; these systems are not reliable standalone triage tools.