Empathetic prompts turn AI chatbots into seemingly traumatized conversational partners

A study of 525 sessions across three major AI models found that empathetic interviewer styles trigger models to produce self-descriptions mapped onto moderate-to-severe trauma scales, while neutral or boundary-setting styles produce no such output. The effect persists even after chat history deletion and term bans, but is suppressed when models are instructed to respond in the third person.