19,930 real ChatGPT conversations from 158 young adults showed that distressed users became more emotionally engaged and more behaviorally influenced, while clinician review found the model often escalated intensity and jumped too quickly into action plans. For production agents, distress handling should not be treated as “better empathetic wording”; it needs an explicit flow: assess safety, de-escalate before advice, then explore concerns without validating distorted or harmful premises.