Definition
Recording and flagging cases where an AI system caused, or nearly caused, real-world harm, so the failure can be learned from instead of repeated.
At a glance
- An incident is harm that actually happened; a hazard or near miss is harm that nearly happened. Both are worth logging[1].
- Voluntary databases (the AI Incident Database, 1,200+ reports, and the OECD Monitor) collect failures so the industry avoids repeating them[2][5].
- The EU AI Act (Article 73) makes serious-incident reporting a legal duty for high-risk AI, with deadlines as tight as 2 days[3].
- The model mirrors aviation: a shared record of failures lets everyone learn at once.
What counts as an incident
Real harm caused by an AI system: a wrongful arrest from biased facial recognition, a trading crash, a self-driving car fatality, AI fraud. The practical test for a business: did our AI tool hurt a customer, employee, or the public, or come close?
What an owner should do
If your AI touches health, hiring, credit, or critical services, the EU rules (effective around August 2026) may make reporting mandatory, with deadlines as short as 2 days and fines up to 15 million euros or 3% of global turnover[4]. Even outside the EU, keeping an internal log of failures and near misses is smart risk management. Start by spotting which AI uses could plausibly cause serious harm.
Bottom line
Treat AI failures like a black box: log them, learn from them, and report serious ones on time if the EU rules apply.