AI Incidents and Marginalized Populations: A Descriptive Analysis of the AI Incident Database (1983–2026) from a Social Work Perspective

Author Note (August 18, 2026): This preprint is currently undergoing systematic review and revision. An updated version will be posted upon completion. Readers are encouraged to consult the latest version of the manuscript.Purpose: Artificial intelligence (AI) systems are increasingly deployed in the institutional core of social work practice, such as law enforcement, healthcare, education, and public administration. Despite this implementation, the social work profession lacks a population-level empirical baseline of documented AI-related harm. This study characterizes the landscape of documented AI incidents, focusing on marginalized populations and the sectors in which social workers practice.Method: We conducted a cross-sectional descriptive analysis of the AI Incident Database (AIID; snapshot as of June 29, 2026), a curated repository of publicly reported AI incidents. The record comprises 1,549 incidents (1983–2026), with three independent post hoc classification systems that provide harm coding for overlapping subsets of these incidents. We report descriptive population proportions with explicit denominators for every reported statistic.Results: Most incidents (83.8%) were dated 2020 or later, though documented incidents date back to 1983, predating large language models by decades. Among the 214 CSET-coded incidents, a subset weighted toward pre-2020 events, 32.7% involved protected characteristics, most often race or ethnicity. Social-work-relevant sectors, including healthcare, law enforcement, education, and public administration, accounted for 23.4% of these incidents. Incidents in these sectors implicated protected characteristics (50.0% vs. 27.4%) and rights violations (40.0% vs. 4.9%) at markedly higher rates than other sectors, and 80.0% involved the public sector. Among incidents classified by intent and deployment timing, discrimination and toxicity incidents were coded as unintentional in 80.7% of cases, and 97.6% occurred after the AI system was already in active use rather than during development or testing.