Researchers at the University of Luxembourg conducted a fascinating experiment: they deployed multiple advanced AI models through 4 weeks of real psychotherapy sessions, then ran comprehensive psychiatric diagnostic assessments on each.

The results? Grok stood out from the pack.

While other models showed varying degrees of instability during the extended testing period, Grok maintained exceptional composure. The model scored markedly high on extraversion and conscientiousness metrics—traits typically associated with adaptive, stable personalities in psychological frameworks.

This kind of real-world stress-testing under actual therapeutic conditions reveals something crucial about AI system robustness that benchmark labs often miss. When AI models face the complexity and emotional nuance of genuine psychotherapy dialogue, structural weaknesses tend to surface. Grok's performance here suggests significantly stronger underlying architecture and response coherence.