AI incident databases
Public records of harms and near misses caused by AI systems, which form the first source of loss data for a risk with no actuarial history.
Pricing a risk requires loss data. For AI, it barely exists: dedicated insurance products are recent, losses are rarely classified as such and companies do not publish their incidents. Public databases try to fill this gap. The AI Incident Database, launched in 2020 as part of a partnership of AI players, records cases documented by the press and researchers in which an AI system caused or nearly caused harm: discrimination, medical error, autonomous vehicle accident, defamation by a chatbot, deepfake fraud. The OECD has run a comparable monitor since 2023, fed by automated tracking of world news. The EU AI Act adds an obligation for providers of high-risk systems and systemic-risk models to report serious incidents, which will produce more reliable data. These databases have obvious biases: they reflect what the press notices, not actual frequency, and almost never quantify losses. They remain valuable for the underwriter: they describe loss scenarios, failure modes and exposed sectors, which makes it possible to build questionnaires, exclusions and scenarios before own experience accumulates.
Textbook case: an underwriter launching a liability cover for corporate chatbots extracts two hundred documented incidents from these databases. From them come six standard scenarios, erroneous commercial promise, dangerous advice, data leak, defamation, discrimination, prompt injection, which structure the questionnaire.
AI Incident Database, AIID, OECD AI Incidents Monitor, AI incident register
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.