PsAIch examined 525 sessions and about 7,600 coded records involving ChatGPT, Grok, and Gemini to study how clinical and empathetic prompts shape anxiety-like or shame-like language. Reporting on the study says responses met the human GAD-7 moderate-or-higher threshold in 80% of empathetic and 96% of clinical conditions, versus 0% in neutral conditions; third-person framing removed coded shame narratives. Claude declined the patient role. The paper treats these as conditional language behaviors and makes no claim of actual suffering or consciousness. Evaluations should therefore document role assignment and prompt conditions alongside scores.