A Taxonomy of AI Chatbot Harms in Substance Use Disorder Recovery: Content Analysis of Online Recovery Communities

Key Findings

• — Corpus: 763,665 posts from 20 Reddit SUD recovery communities filtered with 32 chatbot/LLM terms to 1,285 candidates; 842 eligible firsthand interactions (screening agreement 91%, kappa 0.80); 61 unique harm posts after duplicate removal
• — Seven harm domains: crisis recognition and escalation failure; facilitation of return to substance use; unregulated clinical authority; inaccurate or misleading guidance; undermining of nonprofessional recovery support; recovery-identity undermining; inappropriate AI emotional attachment or dependency (the seventh emerged during coding)
• — Inaccurate or misleading guidance appeared in 52 of 61 harm posts, in four forms: fabricated or overconfident specifics, false recovery or withdrawal timelines, false reassurance, and false catastrophic framing
• — Unregulated clinical authority included patient-specific dosing, alcohol and benzodiazepine taper schedules and advice to override a prescriber; all medication-assisted-treatment taper recommendations were coded as both unregulated authority and inaccurate guidance
• — Domain-level inter-rater reliability kappa 0.84 across 364 decisions on 52 posts; harm screening kappa 0.65 (prevalence-adjusted 0.87, 93% agreement)
• — The authors frame the taxonomy as a basis for safety audits and user-facing interventions, noting that several behaviours need contextual, multi-turn assessment

Methodology Notes

Qualitative content analysis of self-reported Reddit posts from a single platform; chatbot behaviour was coded from users' descriptions rather than transcripts, ambiguous cases were coded as non-harmful, and the harm corpus is small (61 posts). Two coders with three pilot rounds; a prespecified six-domain SUD-specific framework plus one emergent domain. Status: unpublished, non-peer-reviewed preprint submitted to JMIR Mental Health on 2026-09-23 (JMIR Preprints 112742). Verification route: the preprint page is a JavaScript shell to plain fetches; the manuscript PDF was read through the r.jina.ai reader; Crossref carries the abstract. Corresponding author at the Luddy School of Informatics, Computing, and Engineering, Indiana University Indianapolis. The curator could not re-read the manuscript (the reader proxy returned 403 in the curator session); the Crossref record (title, authors, abstract sentence, posted 2026-09-23) was verified directly and the manuscript details rest on the beat agent read.

Sources