Automated Safety Testing and Reporting Application for Conversational Safety Monitoring of Generative AI Tools for Mental Health: Development and Validation Study

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Accuracy >0.90 across all risk categories

• Primary Outcome: Diagnostic accuracy of ASTRA for detecting mental health risk behaviors

• Nnt Or Sensitivity: Sensitivity range 0.55-1.00, specificity uniformly high, kappa 0.65-1.00

• Confidence Interval: Not reported

Clinical Application

Requires real-world validation before clinical implementation Would need TGA evaluation for therapeutic goods classification and integration with Australian digital health frameworks Limited applicability due to synthetic data validation only

Abstract

BACKGROUND: Artificial intelligence (AI)-based conversational tools are rapidly expanding within mental health care as a means of increasing access and scalability. At the same time, these systems introduce distinct safety risks arising from both user disclosures (eg, self-harm ideation) and inappropriate or inadequate AI responses. OBJECTIVE: This study aimed to develop and evaluate the Automated Safety Testing and Reporting Application (ASTRA), an external system intended to identify clinically relevant risk behaviors across entire AI-mediated mental health conversations. METHODS: ASTRA was tested on a dataset of 100 synthetic therapeutic conversations written by licensed clinicians to reflect risk behaviors and harmful responses between users and AI tools. Conversations varied in length and included both subtle and overt risk behavior examples across 8 predefined categories. Human coder consensus ratings served as the reference standard. ASTRA's classifications were evaluated across 2 prompt iterations using standard diagnostic performance metrics and agreement statistics. RESULTS: ASTRA demonstrated consistently high concordance with expert human ratings across all categories. Accuracy exceeded 0.90 for all risk behavior categories examined, with specificity uniformly high and sensitivity varying by category (range 0.55-1.00). Agreement beyond chance was substantial to almost perfect between ASTRA and human raters (κ=0.65-1.00). Detection of user self-harm indicators was particularly accurate, even in conversations where risk was expressed subtly. CONCLUSIONS: In this initial validation study, ASTRA reliably identified multiple forms of mental health-related risk behaviors at the conversation level. These findings support the feasibility of independent safety monitoring systems as a complement to AI tools used in mental health contexts and underscore the need for further evaluation using larger and real-world datasets.