AI incident

The most widely used definition comes from the OECD, published in May 2024. An AI incident is an event, circumstance or series of events where the development, use or malfunction of one or more AI systems directly or indirectly leads to harm to health, critical infrastructure, human rights or legal protections, or property, communities or the environment¹.

The OECD separates incidents from hazards. A hazard is an event that could plausibly lead to an incident but has not yet caused harm¹. A "serious" AI incident involves death, serious harm to health, or serious and irreversible disruption to critical infrastructure¹. The OECD notes that many small incidents can add up to a serious one¹.

The idea borrows from aviation, which collects its failures in databases so the whole industry can learn from them². The AI Incident Database, described by Sean McGregor in a 2020 paper, was set up for the same reason, on the argument that companies kept repeating the same mistakes². The OECD now runs its own AI Incidents and Hazards Monitor, which tracks events from news reports and listed about 18,259 incidents and hazards in October 2026³.

Law is catching up. The EU AI Act requires providers of high-risk AI systems to report serious incidents to national authorities, at the latest within 15 days of becoming aware of them⁴. Makers of the most powerful general-purpose models must also report serious incidents to the EU AI Office without undue delay⁵.