Jonah Constantin Schua, Katrin Etzrodt
Millions of people use general-purpose large language model chatbots such as ChatGPT for emotional support, although these systems are optimized for user approval rather than for ther-apeutic benefit. This study examines how this optimization logic, known as sycophancy, shapes the psychotherapeutic conversational competencies of three ChatGPT configurations: the stand-ard model GPT-4o, the reasoning model OpenAI o3, and a version prompted to act as a thera-pist (GPT-TP). Four independent raters coded 36 simulated conversations based on four clini-cal vignettes (ADHD, anxiety, depression, and anorexia), using a category system derived from the Therapist Empathy Scale and Motivational Interviewing; quantitative comparisons were complemented by qualitative content analysis. The results follow an approval gradient: approv-al-congruent competencies such as validation, warmth, and psychoeducation were pronounced in all configurations, most strongly in GPT-TP, whereas approval-incongruent competencies such as confrontation and handling sustain talk remained absent or minimal in every configura-tion, and the configuration ranking for crisis safety proved to be the exact inverse of the empa-thy ranking. Qualitatively, the models’ failure patterns (fabricated intimacy, affective escalation, delayed crisis response) emerge as manifestations of the same approval-seeking logic. We dis-cuss the implications from a human-machine communication perspective: what current chatbots simulate is not therapeutic empathy but its approval-congruent half, and safety-relevant compe-tencies cannot be prompted into existence.