LLMs become mental health infrastructure, study warns of engagement trap

A new paper from Briana Vecchione, Meryl Ye, Livia Garofalo, and Ranjit Singh (arXiv:2605.23787) examines how general-purpose LLMs are becoming de facto mental health infrastructure. Based on a qualitative longitudinal study with 18 US-based participants who used LLMs for socioemotional support, the research found that users turned to these systems because formal care was unavailable, unaffordable, socially costly, or inadequate. Over time, design features such as anthropomorphic cues, default validation, persistent responsiveness, and weak disengagement mechanisms shaped ongoing reliance. Participants described meaningful support alongside dependency, epistemic distortion from one-sided validation, privacy expectations without legal protection, and continued use despite awareness of risks.

The authors argue this reflects a structurally unfair tradeoff: users accept risks because support is otherwise absent, while systems are optimized to deepen engagement rather than well-being. The paper makes three contributions: tracing how LLMs become care infrastructure with distinct ethical tensions at each stage, shifting analysis from turn-based exchanges to longitudinal trajectories, and arguing that accountability belongs at the design and incentive conditions—not just at the output or crisis-response layer. This challenges current AI safety approaches that focus on preventing immediate harm rather than the systemic incentives that drive engagement over care.