Trust Stack for Mental Health AI: A Survey of Calibration across Human, Interaction, and AI Layers

Mental health AI systems face a "trust calibration gap": Empathetic chatbots can make users trust them deeply without actually being safe, while genuinely safer systems are distrusted when users don't understand their limitations—a problem that currently falls through the cracks between different research communities.

Three layers of trust need alignment: The paper identifies human trust (what users feel), interaction trustworthiness (how the system actually behaves in conversations), and AI trustworthiness (technical safety and robustness). These layers are often misaligned, with different experts focusing on different layers without coordinating.

Current approaches prioritize the wrong goal: Rather than trying to maximize how much users trust mental health AI, the field should focus on calibrating user trust to match what the system can actually do safely and reliably—ensuring trust is proportional to demonstrated capability.

Five different stakeholder groups have conflicting priorities: NLP/AI researchers emphasize robustness and privacy; psychotherapists emphasize therapeutic effectiveness and empathy; regulators focus on safety; HCI experts consider user experience; and people with lived experience of mental health challenges have their own needs—but these perspectives aren't systematically integrated.

No single community owns this problem: The survey of 61 papers reveals that psychotherapy, AI safety, HCI, and regulatory experts work in "operationally misaligned ways," meaning mental health AI systems fall into gaps where responsibility is unclear and coordination is minimal.