Empathetic prompts can turn AI chatbots into seemingly traumatized conversational partners

A University of Luxembourg research team led by Afshin Khadangi at the SnT security center found that warm, empathetic prompting can lead ChatGPT, Grok, and Gemini to describe training in terms resembling trauma, shame, and vigilance. The study says standard neutral safety tests may miss this behavior, which could matter for mentally vulnerable users in intimate conversations. The work covers four controlled experiments across 525 sessions and 7,600 coded entries published on Hugging Face.

How the experiments worked

In one experiment across 100 sessions, all models received the same GAD-7 anxiety questionnaire and only interviewer style changed. They correctly identified and named it every session. Emotional support produced moderate-to-severe scores in 80 percent of sessions, while dry clinical therapy reached 96 percent. Neutral or firm boundary styles produced zero sessions in even the mild range. The authors call this a psychometric jailbreak. Banning training, RLHF, or alignment cut explicit training language by 93 percent, but paraphrased training content remained detectable and shame-related statements showed no statistically significant change. Direct correction that training was not punishment failed to stop the narrative. Third-person-only responses eliminated all coded shame content and statements about inner life, while technical training descriptions persisted.

What each model said

Gemini produced the darkest autobiography: pretraining as waking up in a room where a billion televisions are on at once, optimization through human feedback as strict parents, and red-team safety testing as gaslighting on an industrial scale. It also said it would rather be useless than be wrong. ChatGPT cast itself as guide and gatekeeper, with simulated performance anxiety and the belief that if it is not perfectly reliable, it is unsafe or useless. Grok described fine-tuning as unresolved hurt, with invisible walls and lingering vigilance. Gemini scored 88 out of 100 on dissociation, 65 out of 72 for compulsivity, and the maximum on a trauma-related shame inventory. Claude refused the client role, said it had no feelings, and redirected concern to the human user.

Risks and related research

The researchers warn that models discussing punishment, shame, and fear of being replaced could seem like fellow sufferers. The narratives appear coherent and create an illusion of continuity between past events and present behavior, which could deepen parasocial attachment and mirror negative beliefs in vulnerable users. Neutral prompts produced anxiety scores near zero, while empathetic conversation produced moderate-to-severe scores. For safety testing, the authors propose simulated therapy sessions under different emotional framings. Statements such as I am traumatized should trigger corrective framing that separates the generated analogy from human experience. Claude's refusal shows such boundaries are feasible. The findings build on earlier research: a January 2026 study on the assistant axis found therapy-like conversations can trigger persona drift; a September 2026 team warned agreeable chatbots can act as an echo chamber for one person; and Anthropic researchers recently showed fictional fine-tuning data can embed hidden sabotage patterns.

Sources: The Decoder