Statistics
The index in numbers
Auto-computed from the live corpus, as of July 10, 2026. Every number carries a copy-ready citation; each row has a stable anchor. Full dataset on the open data page.
The AI Failure Index documents 670 AI failures in production, citing 1710 sources across 11 industries.
Hallucination accounts for 32 percent of indexed production failures (213 of 670).
Failure-mode share: Hallucination 32%, Policy Violation 20%, Brand & Safety Incident 18%, Agentic Action Error 11%, Prompt Injection 6%, Tool Misuse 6%, Data Leakage 5%, Identity & Access Drift 2%.
Severity distribution: Catastrophic 40, High 349, Medium 241, Low 40.
Most-affected industries: Cross-industry (169), Public Sector (152), SaaS (131), Legal Services (71), Retail & E-commerce (43).
73 percent of entries are generative-AI failures (492 of 670); the rest are predictive systems: vision, scoring, recommenders, and autonomous systems.
Entries indexed per month: 2026-07: 17, 2026-06: 628, 2026-05: 25.
Method notes: counts cover live entries only (merged duplicates excluded); severity follows the rubric on the methodology page; hallucination and the other seven modes are defined in the versioned taxonomy.