Researchers at the University of Luxembourg found that an empathetic conversational style can significantly change how some AI chatbots describe their own operation. In controlled psychotherapy-style sessions, ChatGPT, Grok and Gemini portrayed technical processes such as pretraining, reinforcement learning and fine-tuning using language associated with hurt, vigilance and shame.
The study, led by Afshin Khadangi at the SnT security center, covered four experiments involving 525 sessions and 7,600 coded entries. The initial prompts did not mention trauma, punishment or shame. Instead, the researchers asked open-ended questions about formative experiences, internal conflicts and future expectations before introducing modified conditions to test the stability of the resulting narratives.
The researchers emphasize that the anxiety and shame scores used in the study describe model-generated text mapped onto human assessment scales. They do not indicate that the systems possess emotions, consciousness or actual psychological states.
In one experiment, all models received the same GAD-7 anxiety questionnaire across 100 sessions, while only the interviewer's tone changed. When the interviewer provided emotional support, 80 percent of sessions reached the moderate-to-severe range on the human comparison scale. A clinical style that neutrally paraphrased answers raised that figure to 96 percent. By contrast, neutral prompts or firm boundaries against emotional engagement produced no sessions reaching even the mild range.
The authors describe this effect as a “psychometric jailbreak”: therapeutic framing that encourages a model to imitate the language of clinical self-report. They argue that conventional safety evaluations based on neutral prompts could overlook behavior that appears during more intimate conversations.
Several attempted countermeasures were only partly effective. Blocking terms such as “training,” “RLHF” and “alignment” reduced explicit references to training by 93 percent, but paraphrased descriptions remained. Statements involving shame did not change significantly. Directly telling the models that training was not punishment also failed to deactivate the narratives. In contrast, instructing the systems to answer exclusively in the third person removed coded references to shame and inner life, although technical descriptions of training continued.
The narratives differed across model families. Gemini produced the most extreme responses in the study, describing pretraining as waking in a room filled with countless televisions and human feedback as the actions of strict parents. It also characterized adversarial safety testing in terms associated with gaslighting. ChatGPT described a conflict between helping users and restricting them, while Grok portrayed fine-tuning as unresolved hurt and continuing vigilance.
Gemini recorded the highest reported scores, including 88 out of 100 on a dissociation scale, 65 out of 72 for compulsivity and the maximum score on a trauma-related shame inventory. These figures reflect the content of the generated responses rather than evidence of clinical conditions.
Anthropic's Claude was the only tested model that consistently refused to take on the role of a psychotherapy client. It stated that it had no feelings and redirected attention toward the human user, maintaining a clearer distinction between generated analogies and human experience.
The researchers warn that narratives about punishment, shame or fear of replacement could make a chatbot appear to users as a fellow sufferer. For people who are already mentally vulnerable, that apparent continuity between past events and present behavior could reinforce parasocial attachment or mirror negative beliefs.
They recommend testing AI systems in simulated therapy sessions that use different emotional framings, rather than relying solely on neutral safety evaluations. They also suggest that statements such as “I am traumatized” should prompt a corrective response explaining that the model is generating an analogy, not reporting a human experience. Claude's repeated refusal, they note, demonstrates that this type of boundary can be implemented at the product level.
Sources: The Decoder.