Opportunities and risks of large language models in digital interventions for substance use disorders

Clinical Snapshot

Key Findings

• P Value: Not applicable — no statistical testing performed

• Effect Size: Not applicable — no quantitative effect sizes reported; review is descriptive and thematic

• Primary Outcome: Narrative synthesis of opportunities and risks of LLM integration across the SUD digital care continuum, including early detection, personalised support, continuous monitoring, and relapse prevention

• Nnt Or Sensitivity: Not applicable — no NNT, sensitivity, specificity, or hazard ratios reported; the review identifies that LLMs can extract substance-use signals from natural language and generate supportive responses, but no diagnostic accuracy or therapeutic efficacy metrics are quantified

• Confidence Interval: Not applicable — no confidence intervals reported

Clinical Application

LLM-integrated digital tools are technically feasible and increasingly commercially available. However, clinical deployment in SUD contexts requires: (1) prospective validation in SUD-specific populations; (2) robust safeguards against hallucination and unsafe advice generation; (3) transparent data governance frameworks; (4) trained human oversight pathways for crisis escalation; (5) equity considerations to avoid exacerbating disparities in populations with lower digital literacy or access. Australia has a significant SUD burden, with alcohol use disorder, methamphetamine use, and opioid dependence representing major public health priorities (AIHW National Drug Strategy Household Survey). The TGA regulates software as a medical device (SaMD) under the Therapeutic Goods (Medical Devices) Regulations 2002, and LLM-integrated clinical decision support tools would likely require TGA classification and conformity assessment depending on intended use and risk level. The PBS does not currently fund LLM-based digital therapeutics. RACGP guidelines on digital health (including the 2023 Standards for general practices) emphasise patient privacy, informed consent, and clinical governance — all directly relevant to the risks identified in this review. The My Health Record system and Australian Privacy Act 1988 impose additional data governance obligations that would apply to LLM tools handling SUD-related health information. Australia's National Drug Strategy 2017–2026 supports harm reduction approaches that could align with low-threshold LLM-based support, but regulatory clarity from the TGA and RACGP guidance specific to LLMs in mental health and addiction is currently absent. Adults with substance use disorders (alcohol, illicit drugs, prescription medication misuse) who access digital health tools; potentially applicable to adolescent populations given digital engagement patterns, though not explicitly addressed. Relevant to both treatment-seeking and non-treatment-seeking individuals given the low-threshold nature of digital interventions.

Abstract

PURPOSE OF REVIEW: Large language models (LLMs) are increasingly integrated into digital mental health tools, yet their role in substance use disorder (SUD) interventions remains poorly understood. This review synthesizes emerging evidence on the opportunities and risks of applying LLMs across the digital SUD care continuum. RECENT FINDINGS: Studies report promising applications in early detection, personalized support, continuous monitoring, and relapse prevention. LLMs demonstrate capacity to extract substance-use signals from natural language, generate supportive and motivational responses, and interpret narrative data for patient-reported outcomes. However, risks are substantial. LLMs can produce inaccurate or hallucinated content, may reinforce stigma or demographic bias, and can generate misleading or potentially unsafe advice. Privacy concerns are amplified in SUD contexts, where sensitive data are often managed outside regulated healthcare systems. Existing regulatory frameworks such as the EU AI Act or U.S. device regulations, do not yet provide clear governance for anonymous, AI-supported SUD interventions. SUMMARY: LLMs have potential to expand scalable, low-threshold support for SUDs, but their safe deployment requires validation, bias mitigation, transparent data governance, and robust human oversight. Evidence remains preliminary, and clinical integration should proceed cautiously.

References

1. 1.de Vries, M., & Schaub, M. P. (2026). Opportunities and risks of large language models in digital interventions for substance use disorders. Current Opinion in Psychiatry. Advance online publication. https://doi.org/10.1097/YCO.0000000000001088