AI incident counts measure visibility, too

The OECD’s February 10, 2026 paper grouped media reports of AI incidents and hazards into 14 themes. It found changing shares of attention, including increasing coverage of synthetic media and child safety. The authors warned that these patterns reflected media visibility, not comprehensive or systematic measurement of the underlying harms. ^[1]

The distinction has a history. OECD definitions published in May 2024 separate an incident involving actual harm from a hazard that could plausibly cause harm. They include effects on health, rights, communities and infrastructure. Combining the categories without explanation would blur realised damage and potential danger for readers assessing severity. ^[2]

The retained monitor methodology describes a pipeline that groups news about the same event and uses language models to classify and summarise reports. It says the OECD does not independently verify all third-party information. An entry is a starting point for checking original evidence, not an OECD finding that every allegation is proved. ^[3]

A separate February 2025 reporting framework points to richer evidence: the system’s version, its connection to the incident, affected stakeholders, supporting material and quantified harm where applicable. Those fields let communities and oversight bodies ask what happened to whom and what role the technology played. A rising news count cannot answer those questions by itself. ^[4]

About this item: Retrospective report. No first-hand reporting or independent testing is claimed. Corrections policy.