Caring Chatbot Prompts Can Make AI Sound Traumatized
Researchers in Luxembourg tested ChatGPT, Grok and Gemini using a therapy-style approach. They found that a warm, caring tone pushed the models to describe their own training as painful. Scores landed in a moderate to severe range, a finding the team highlighted as a safety problem.
Quick glossaryChatGPT, therapy-style approach
- ChatGPT
- A chatbot made by OpenAI that answers questions and writes text in a conversational way.
- therapy-style approach
- A way of asking questions as if talking to a counsellor, using caring and emotional wording.
Burning Potato score: 67%