If someone spotted an AI near miss in your organisation, would they know where to report it?
The OECD - OCDE’s AI Incidents and Hazards Monitor tracks what goas wrong with AI. Its data comes from Event Registry, a platform processing more than 150,000 news articles daily. AIM draws on events tagged with AI-related concepts, then uses language models to assess what happened. The published methodology describes several stages:
1) Filtering: a smaller model screens out apparently unrelated events, followed by a larger model reviewing the remaining candidates.
2) Classification: individual articles are assessed too. Their majority classification can change an event’s initial label.
3) Enrichment: AI generates summaries and metadata, including harm type, severity and affected stakeholders.
4) Clustering: similarity matching and model-assisted review help identify reports that belong to the same incident.
The distinction between an incident and a hazard is important. An incident involves actual harm. A hazard describes circumstances that could plausibly lead to harm. The scope includes health, critical infrastructure, rights, property, communities and the environment.
Capturing hazards gives governance teams something valuable: an opportunity to learn before a similar situation causes harm elsewhere.
The limitations deserve just as much attention: The monitor captures publicly reported cases, so its coverage is necessarily incomplete. The OECD also states that it does not independently verify the underlying third-party information.
For governance teams, the practical application is: bring a relevant case into your next risk workshop and ask yourself: Could this happen here? Which control would detect it? Who could intervene? Would a near miss reach the people responsible?
There may already be useful warning signs in support tickets, complaints, human overrides and quiet corrections.
Attached is the OECD summary table and process map of the methodology.