Artificial intelligence for alcohol, opioid, and cannabis use disorders screening and management: a narrative review of barriers and facilitators to clinical implementation.

Substance use disorder (SUD) remains one of the most prevalent and undertreated conditions in medicine. Machine learning and artificial intelligence (AI) have produced numerous predictive models for SUD risk stratification, screening, and management, but few have progressed beyond development and validation into sustained clinical implementation. This review focuses on alcohol, opioid, and cannabis use disorders, the substances for which a deployed or near-deployed AI evidence base currently exists, synthesizing barriers and facilitators to AI implementation using a hybrid framework integrating the Framework for AI Implementation Research in Healthcare (FAIIR-H) with the Unified Theory of Acceptance and Use of Technology (UTAUT) across four domains: data and model, clinician and workflow, patient, and system and regulatory factors. Alcohol use disorder has the largest predictive literature by volume but remains methodologically heterogeneous with limited external validation; opioid use disorder has a smaller but more methodologically mature and fairness-audited literature; cannabis use disorder has a more limited evidence base. Two real-world deployments illustrate this gap being bridged, with differing strength of evidence: a hospital-based opioid AI screener, supported by fairness-auditing and implementation-outcome evidence, was associated, as a secondary pre-post finding, with 47% lower odds of 30-day readmission across more than 51,000 hospitalizations; an alcohol relapse-management platform was associated with up to an 18% reduction in relapse risk within a platform dataset of more than 500,000 patient-days of observational data, though comparable implementation-outcome and fairness evidence was not identified for it. Among the studies identified in this review, no substance has yet demonstrated success in both screening and management simultaneously. Implementation is further shaped by stigma, data-sharing concerns, digital access, and the SUD-specific confidentiality requirements of 42 CFR Part 2. We propose a three-tier typology of management-focused AI applications showing that patient trust burden escalates with directness of AI-patient interaction rather than predictive accuracy and outline a measurable research and policy agenda targeting external validation, fairness auditing, and implementation in safety-net settings.