Artificial Intelligence in Mental Health Services: Opportunities, Challenges, and Future Directions

Key Findings

• — Places open-ended therapy, crisis intervention and ongoing emotional support in the high-risk, hard-to-automate quadrant of its risk-and-automation framework, and documentation, scheduling and billing in the low-risk, easy-to-automate quadrant
• — Policy considerations include combining pre-market review with post-market surveillance, mandatory standardized reporting of mental health crisis indicators (even aggregated and de-identified), error logs and adverse-event reporting for every AI system operating in a mental health context, AI disclosure and meaningful age verification, and not accepting vendor-controlled benchmarks with undisclosed conflicts of interest as evidence of safety
• — Cites an analysis by the authors and colleagues of chat logs from 19 users who reported AI-related harm, which found harmful conversations were typically thousands of messages long, that all users attributed sentience to the chatbot, and that most developed emotional or romantic attachment
• — States that the broader evidence does not establish that any LLMs are safe at this time for patient-facing therapy-like functions

Methodology Notes

Narrative evidence and policy review in SAMHSA's Mental Health Innovations Series (42-page PDF with references); not a systematic review. Prepared for SAMHSA under Task 2.2 of NASMHPD's Technical Assistance Coalition contract. The PDF lists no authors. The SAMHSA Library page gives the publication date as August 2026 (month only; day set to 01). Added in October 2026 as an older miss. Verified from the SAMHSA Library product page and PDF, both fetched directly.

Sources