![[ Visit Client Website ]](images/banner.gif)
The rapid integration of artificial intelligence (AI), particularly large language models (LLMs), into substance use disorder information, treatment, and recovery has the potential to improve access, health literacy, and knowledge dissemination. However, it also raises concerns regarding the generation, amplification, and detection of misinformation. Substance use treatment is particularly vulnerable to misinformation due to stigma, clinical complexity, and high-risk decision contexts. Despite growing interest, the extent and nature of AI-related misinformation in this domain, as well as how it is evaluated, remain unclear.
This scoping review maps the emerging literature on AI and misinformation in substance use contexts, with attention to (1) the roles AI systems play (generation versus detection), (2) the types of misinformation identified, and (3) evaluation approaches and associated risks.
Methods
This scoping review followed JBI guidance and applied the Population-Concept-Context framework. Literature searches were conducted across 8 databases including PubMed, Scopus, ProQuest, PsycINFO, Web of Science, IEEE Xplore, Cochrane, and EBSCO. The AI component was defined broadly to capture diverse methodological approaches, including LLMs, natural language processing techniques, and traditional machine learning methods.
Eligible studies examined AI-generated, AI-mediated, or AI-detected substance misuse-related misinformation in digital settings. Data extraction included study characteristics, AI modality, misinformation type, and outcomes. Studies were independently and blindly double-coded to ensure reliability, and methodological quality was appraised using a modified GRADE-informed approach. Screening was conducted using Rayyan.ai, and data management and extraction were performed in Microsoft Excel.
Results
5 studies published between 2021 and 2025 (from 293 identified records) met inclusion criteria, reflecting two primary paradigms: (1) AI as a generator of health information (n=2) and (2) AI as a detector of misinformation (n=3). LLM-focused studies demonstrated moderate to high overall response quality but revealed critical risks, including hallucinated evidence, inconsistent outputs under prompt variation, omission of evidence-based treatments, and potentially unsafe clinical guidance that deviated from established recommendations. Misinformation was often unintentional yet clinically consequential.
In contrast, machine learning approaches showed strong performance in detecting substance use-related misinformation and highlighted the scale and persistence of misinformation across online platforms (e.g., Reddit, YouTube, TikTok). Across studies, evaluation approaches varied substantially and relied primarily on human-annotated assessments of accuracy, with no standardized framework for evaluating accuracy, safety, or clinical appropriateness.
Conclusions
AI systems play a dual and paradoxical role in substance use information ecosystems, functioning as both sources of misinformation risk and tools for large-scale misinformation surveillance and detection. Current evidence highlights concerns regarding safety, reliability, and over-trust in generative AI outputs, alongside the promise of automated detection systems. However, the field remains methodologically underdeveloped. There is a critical need for standardized evaluation frameworks, domain-specific guardrails, and interdisciplinary oversight from both researchers and clinical practitioners prior to the broader deployment of AI in substance use and mental health contexts.