AI risk insurability mapped: 55 threats vs. 26 insurance products

A groundbreaking arXiv paper by Leung, Zhang, Ling, Toyoda, and Loh systematically maps the insurability of AI risks, coding 55 threat classes (from OWASP/MITRE) against 26 insurance products using public carrier materials and exclusion regimes. The result is a four-tier frontier: (1) affirmatively insured perils (e.g., model performance drift at Munich Re, hallucinations via Armilla/Lloyd's, IP/E&O from Tokio Marine Kiln/CFC, autonomous system liability from Apollo ibott, deepfake-enabled cyber response from Coalition); (2) silent-AI exposures in legacy cyber, D&O, EPLI, crime, and media policies where AI is an instrumentality; (3) actively excluded perils; and (4) perils outside conventional private insurance.

The paper's most striking finding: foundation model concentration creates a genuinely novel insurability frontier because upstream model failure can correlate losses across many cedents simultaneously—a systemic risk that defies traditional pooling. The authors stress their coding reflects public positioning, not executed contract wording, but the patterns are clear. Legacy lines still carry silent-AI exposure wherever AI is a tool rather than the legal cause. The market design question is which insurability constraint (e.g., adverse selection, moral hazard, correlated risk) each candidate structure relaxes, not simply which systemic risk template exists. For professionals, this means insurers are actively segmenting AI risk—and coverage gaps are widening fast.