Documented artificial intelligence incidents cluster around a handful of harm categories, with misinformation, deepfakes, and content harms accounting for roughly 28 percent of recent cases, according to incident classification data. Discrimination and bias follow at about 22 percent, and physical safety failures make up around 14 percent.

The scale of tracked incidents varies by source and methodology. The OECD AI Incidents Monitor tracks roughly 5,000 to 7,000 incident reports drawn from news monitoring, capturing broad media coverage. A separate academic tracker classifies more than 1,400 real world reported incidents by risk, cause, harm, and severity, providing a structured view of how AI systems fail.

The composition of harms has shifted alongside the technology. As generative tools spread, content related harms such as fabricated images, synthetic audio, and false information have grown as a share of the total, while traditional algorithmic harms tied to recommendation and vision systems have become relatively smaller.

The breakdown helps organizations prioritize controls. Concentrations in misinformation and bias point to content review, provenance, and fairness testing as focus areas, while the physical safety share underscores the stakes when AI systems control or influence machinery, vehicles, or medical decisions. The data offers a map of where AI risk is concentrating as deployment widens.

Source: OECD.AI - https://oecd.ai/en/incidents