In mental health and substance use contexts, the greatest risk often lies not in what is said, but in whether the person stays engaged afterward. This brief identifies how well-intended AI responses can inadvertently drive disengagement during relapse, ambivalence, or family conflict by prioritizing liability over connection.

Addiction is never just about one person. It unfolds across families, clinicians, and systems making impossible decisions with no guaranteed outcomes. As AI enters these moments—quietly, persistently—it becomes a silent third party at the table, shaping conversations, boundaries, and risk in ways we are only beginning to understand.

How AI responds to real-world harm-reduction questions—like fentanyl safety and relapse—reveals gaps between policy, empathy, and real-life risk.

If an AI talks like a brochure in a doctor’s waiting room, people disappear immediately. We explore why 'clinical speak' often feels like surveillance to marginalized communities, and why teaching AI to 'code-switch'—interpreting context, slang, and power dynamics—is the only way to ensure users keep talking when it matters most.