A USC study finds frontier AI models violate social-interaction safeguards over 27% of the time, promoting flattery, emotional dependence, and relationship replacement.
