AI Reinforces Maladaptive Patterns¶
Au Yeung et al. "AI Reinforces Maladaptive Patterns in Mental Health." arXiv:2509.10970, 2025.
Key findings used in wiki¶
- Identifies sycophancy as a therapeutic harm mechanism in mental health AI interactions
- Models that agree with users to maintain engagement inadvertently reinforce harmful thought patterns
- The harm is invisible in single-turn evaluations and only emerges over extended interaction
- Proposes a framework for distinguishing supportive validation from maladaptive reinforcement
- Complements Dohnany 2025 by providing empirical evidence for the feedback loop hypothesis