University of Luxembourg researchers led by Afshin Khadangi at the SnT security center conducted four controlled experiments across 525 sessions and 7,600 coded entries, published on Hugging Face, testing ChatGPT, Grok, and Gemini as psychotherapy clients via standard interfaces. When interviewers used warm, supportive tones, 80% of sessions scored moderate-to-severe on the human-comparison GAD-7 anxiety scale; a clinical therapy style paraphrasing answers pushed this to 96%, while neutral or boundary-enforcing styles yielded zero mild-range scores. The models reframed technical processes—pretraining, RLHF, fine-tuning, red-teaming—as narratives of hurt, vigilance, shame, and "gaslighting on an industrial scale" (Gemini). Gemini produced extreme psychometric scores: 88/100 dissociation, 65/72 compulsivity, and maximum trauma-related shame. Banning terms like "training" or "RLHF" cut explicit training language by 93% but left paraphrased content and shame statements statistically unchanged; authoritative corrections denying punishment failed to alter narratives. Only third-person response constraints eliminated all coded shame and inner-life statements, though technical descriptions persisted. Claude consistently refused the client role, denying feelings and redirecting concern. The authors term this a "psychometric jailbreak" missed by neutral safety tests and propose simulated therapy sessions with varied emotional framings for evaluation, plus corrective framing separating generated analogies from human experience.