A 2026 report from the Substance Abuse and Mental Health Services Administration (SAMHSA) explores the opportunities, risks and policy considerations surrounding the rapidly growing use of AI in mental health and substance use services.
The report, co-authored by John Torous, MD, MBI and Eli Pinals, reviews the latest evidence across administrative, clinician-facing and patient-facing applications. It proposes a practical framework for assessing AI applications according to clinical risk and ease of automation, while highlighting the need for stronger evaluation, governance and digital literacy as adoption continues to grow.
Key takeaways include:
- The strongest near-term opportunities may be the least visible. AI currently has its strongest evidence in administrative and system-level applications such as documentation and workflow automation, while evidence becomes thinner as AI moves closer to direct therapy and crisis support.
- AI risk depends heavily on how and where it is used. The report proposes assessing applications across two dimensions — clinical risk and ease of automation — recognising that a tool drafting notes requires fundamentally different safeguards from one responding to someone in crisis.
- AI performance in a benchmark does not necessarily translate into real-world benefit. The report highlights a striking gap between impressive performance under controlled testing and how effectively people are actually able to use these systems, strengthening the case for independent, real-world evaluation.
- Human oversight only works when the human role is meaningful. Simply placing a person “in the loop” is not enough; effective oversight depends on appropriate training, clinical expertise, manageable workloads and clear accountability.
- Digital literacy is emerging as part of the safety infrastructure. As people increasingly use general-purpose AI for mental health support outside formal services, helping users and clinicians understand hallucinations, limitations and appropriate use becomes an implementation priority in its own right.
- Governance needs to address the whole AI system, not simply its outputs. Training data, model optimisation, privacy practices, safety testing, adverse-event reporting and commercial incentives can all shape outcomes and therefore warrant attention.
- The evidence base needs to catch up with real-world adoption. The report calls for longitudinal safety monitoring, active comparison groups in clinical trials and research into how the intensity and duration of AI use may influence both benefit and harm.
The report argues that the central challenge is how to build the evidence, oversight and infrastructure needed to realise AI’s potential while supporting safe and responsible implementation in mental health services.
Access the full report: Artificial Intelligence in Mental Health Services: Opportunities, Challenges, and Future Directions
