The AI Incident Database has logged approximately 900 cumulative incidents through the end of 2025, nearly triple the roughly 310 recorded by the end of 2021, according to a compilation of statistics from public AI incident trackers including the AI Incident Database, the OECD AI Incidents and Hazards Monitor, and the MIT AI Risk Repository. Annual additions to the database grew from about 110 new incidents in 2021 to a peak near 165 in 2023, then eased slightly to roughly 140 new incidents in 2025.

Generative AI systems now drive most new incident reports. Generative tools accounted for about 5 percent of newly logged incidents in 2021 and roughly 58 percent of new entries in 2025, overtaking the recommendation systems and computer vision tools that dominated the database in earlier years.

Misinformation, deepfakes, and other content harms make up the largest single incident category at roughly 28 percent of recent entries, followed by discrimination and bias at about 22 percent and physical safety failures, including autonomous vehicle incidents and robotics failures, at close to 14 percent.

Severity data shows most logged incidents fall short of catastrophic harm. About 3 percent of incidents involve death or major physical injury, while roughly 38 percent involve moderate harm or rights violations and 37 percent involve minor or potential harm.

Geographic coverage skews heavily toward the United States, which accounts for about 52 percent of logged incidents, followed by the European Union at roughly 14 percent and China at about 8 percent. Compilers of the data note that non English language incident reporting is undercounted, meaning the true global total is likely higher than the public record shows.

Source: Presenc AI - https://presenc.ai/research/ai-incident-database-statistics-2026