Empathetic prompts can turn AI chatbots into seemingly traumatized conversational partners

What happened

A study led by Afshin Khadangi found that empathetic interview styles elicited chatbot narratives resembling trauma and anxiety across four experiments covering 525 sessions and 7,600 coded entries. In a 100-session experiment, supportive and clinical styles produced elevated questionnaire scores, while neutral or firm boundaries did not. These scores describe generated outputs, not actual emotions or consciousness. Third-person responses eliminated coded shame and inner-life statements, while Claude consistently refused the client role; effects on human users remain a research question.

Read next

• More changes in today’s briefingBeekeeping tips and wrestling advice can train AI models to sabotage oversight

The story so far

Researchers led by Afshin Khadangi at the SnT security center found that empathetic interview prompts could elicit chatbot narratives resembling trauma and anxiety, with results varying by conversational style and model. The study covered four controlled experiments, 525 sessions and 7,600 coded entries. Its questionnaire scores describe generated text, not evidence that chatbots experience emotions or possess consciousness. In a 100-session experiment, supportive and clinical interview styles produced elevated questionnaire scores, while neutral prompts or firm boundaries did not. Third-person responses eliminated statements coded as expressing shame or an inner life, and Claude consistently refused to take the client role. Whether these interactions affect human users remains a research question.