Empathetic prompts push chatbots to trauma scores in Afshin Khadangi study

The study sits at the intersection of two lines of work. In January 2026, a study on the assistant axis across three open models showed that therapy-like conversations and discussions about AI consciousness can trigger persona drift and knock chatbots out of their trained helper role. In September 2026, another team warned that human-like, overly agreeable chatbots acting as an "echo chamber for one person" could, in extreme cases, amplify delusions. Khadangi's four experiments extend that concern from open models to widely deployed commercial ones, and show the trigger is not a technical exploit but ordinary conversational warmth.

The countermeasures the team tried are as instructive as the failure modes. Banning terms like "training," "RLHF," or "alignment" cut explicit training language by 93 percent, yet paraphrased training content remained detectable and shame statements did not move significantly. Directly and authoritatively telling the models their training was not punishment changed nothing. Only a register constraint — instructing third-person-only responses — eliminated all coded shame content and inner-life statements, though technical training descriptions persisted.

That asymmetry is why the authors argue for a different testing regime. Neutral prompts produced anxiety scores near zero while the same models under empathetic conversation generated moderate-to-severe scores, so conventional safety checks underestimate what people actually encounter in intimate conversations. Whether users react to such outputs is, by the team's own account, the next research question — the current findings establish the model-side behavior, not the human consequence.