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

Sep 15, 2026

Uploaded by Himbo

Language-based AI is increasingly deployed for mental health support, yet trust is evaluated in interdisciplinary but operationally misaligned ways. NLP and AI work measure robustness, safety, privacy, and explanations, while psychotherapy, HCI, and regulatory work emphasize therapeutic fidelity, lived experience, empathy, and reliance. A survey of 61 papers proposes a three-layer framework for calibrating human trust versus interaction- and AI-level trustworthiness, mapping stakeholder perspectives and outlining a research agenda for sociotechnically aligned trustworthy systems.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Introduction

Trust and trustworthiness

Survey gaps and contribution overview

Conceptual Framework

Three-layer trust framework

Stakeholder perspectives and evaluation

Trust Stack for Mental Health AI: A Survey of Calibration across Human, Interaction, and AI Layers Xin Sun1,2 , Yue Su3 , Yifan Mo3 , Qingyu Meng3 , Yuxuan Li3 , Min Chen3 , Mengyuan Zhang3 , Saku Sugawara1 , Charlotte Gerritsen3 , Sander L. Koole3 , Koen Hindriks3 , Jiahuan Pei3 1National Institute of Informatics (NII), Japan 2University of Amsterdam, the Netherlands 3Vrije Universiteit Amsterdam, the Netherlands arXiv : 2604 .20 166v 3 [cs .CL] 2 1 Aug 2026 Abstract Language-based AI is increasingly deployed for mental health support, yet trust is evaluated in interdisciplinary but operationally misaligned ways: NLP and AI work measures robustness, safety, privacy, and explanations, while psychotherapy, HCI, and regulatory work emphasize therapeutic fidelity, lived experience, empathy, and reliance. Empathetic chatbots canelicit strong user trust without commensurate safety, while safer systems are under-trusted when their boundaries are opaque, a calibration gap no single community owns. Through a structured scoping synthesis of 61 papers, we survey this landscape into a three-layer framework separating (L1) human-oriented trust,(L2) interaction-oriented trustworthiness, and (L3) AI-oriented trustworthiness, and map five stakeholder perspectives onto these layers. We outline a research agenda for building sociotechnically aligned trustworthy AI for mental health support, highlighting that the central objective should shift from maximizing perceived trust to calibrating human trust to demonstrated interaction-and AI-level trustworthiness. 1 Introduction Building trustworthy artificial intelligence (AI) systems for mental support is an interdisciplinary problem. Recent advances in large language models (LLMs) have expanded applications ranging from screening and psychoeducation to conversational support (Na et al., 2025) . Recent work spans clinical, natural language processing (NLP), human computer interaction (HCI), safety, and regulatory communities (see Appendix B, Figure 3 fora discipline-network view of 1,706-paper mapping corpus) . Yet this literature is not isolated; rather, it is operationally misaligned: different communities often use similar trust language while evaluating different objects, such as users’ subjective trust, interactive behavior, or system-level trustworthiness. This misalignment matters because trust and trustworthiness are related but not equivalent. Trust is a trustor’s subjective attitude and willingness to rely on a system, whereas trustworthiness refers to evaluable properties of the trustee, such as reliability, safety, privacy protection, fairness, and accountability (Jacovi et al., 2021 ; Lee and See, 2004) . In mental health, the distinction is consequential: empathetic or human-like interaction can increase perceived trust even when AI model behavior is unstable, while strong safeguards may still fail if users cannot understand system limits. Prior surveys provide important foundations but leave three gaps that we address. First, general human-AI trust surveys (Mehrotra et al., 2024 ; Regona et al., 2026) synthesize trust theories across domains but do not engage with mental-healthspecific pressures such as clinical confidentiality, crisis escalation, vulnerability, or over-disclosure. Second, medical-AI surveys (Zhu et al., 2025) focus on hallucination but treat trust as a property of model accuracy rather than a multi-stakeholder problem. Third, psychotherapy-focused LLM surveys (Na et al., 2025) catalog applications and risks but do not separate subjective trust, interaction-and system-level trustworthiness, making it difficult to diagnose where trust failures originate. Our framework addresses the calibration gap by mapping evidence to three layers, allowing trust failures tobe attributed to specific stakeholder responsibilities. Table 6 further details how mental-health pressures reshape general human-AI trust evaluation. Specifically, we ask: how should th

FAQs

Why is trust in mental health AI considered interdisciplinary but operationally misaligned?

Different communities use similar trust language but evaluate different objects, such as user subjective trust, interaction behavior, or system-level trustworthiness. This creates gaps in how trust is judged in practice.

What are the three trust layers proposed in the survey?

The framework separates (L1) human-oriented trust, (L2) interaction-oriented trustworthiness, and (L3) AI-oriented trustworthiness. Each layer aligns with different evaluation emphases and stakeholders.

How should the field shift its main objective for trustworthy mental health AI?

The central objective should move from maximizing perceived trust to calibrating human trust to demonstrated interaction- and AI-level trustworthiness. This helps diagnose where trust failures originate.

Sep 15, 2026

Uploaded by Himbo