image: “Understanding AI risk” by Prof. Dr. Martin Eling of the University of St. Gallen, published in China Finance Review International
Credit: China Finance Review International Editorial Office
As artificial intelligence becomes embedded in financial services, healthcare, transportation and other industries, an important question is emerging: can insurers protect businesses against the risks of AI failures? A new study published in China Finance Review International argues that AI risk cannot simply be treated as another form of cyber risk.
In “Understanding AI risk,” Martin Eling of the Institute of Insurance Economics at the University of St. Gallen develops a framework for classifying AI-related risks and evaluating their insurability. The paper draws on academic and policy literature as well as descriptive evidence from four AI incident and litigation databases. It also considers what the early development of cyber insurance can teach insurers about this new category of exposure.
The study finds that AI and cyber risks overlap but have important differences. Cyber incidents often involve external attacks, such as hacking or ransomware. AI-related harm can arise even when a system has not been compromised: inaccurate predictions, biased automated decisions, model drift and weak oversight may generate operational, financial and legal losses. Because many organizations depend on the same models and providers, a single failure could affect multiple firms and sectors at once.
The descriptive evidence illustrates both the scale of reported incidents and the difficulty of measuring AI risk. Across four databases, reported AI incidents and litigation cases rose sharply over the period examined. Estimated annual growth rates range from 29.7% to 62.2%, depending on the database. The author cautions that these figures reflect reporting patterns, growing AI adoption and changes in database coverage as well as underlying risk, and should not be read as actuarial loss estimates.
Analysis of the OECD AI incidents and hazards data also identifies governance concerns. Accountability and transparency issues appear in 95.5% and 82.1% of documented incidents, respectively. In the same database, the rate of physical harm is more than twice as high in systems without human oversight as in fully supervised systems. These descriptive patterns underscore the importance of effective AI governance and monitoring.
From an insurance perspective, some AI-related exposures appear manageable. Operational errors with clear causes and limited losses, certain liability claims and AI-enabled cyber incidents may be covered under carefully designed policies. By contrast, systemic failures affecting many organizations simultaneously, losses with unclear responsibility and extreme scenarios with little historical data are much more difficult to insure.
The paper recommends narrowly defined coverage, conservative policy limits, scenario-based pricing and underwriting that examines model testing, human oversight and governance standards. It also calls attention to the role of clearer liability rules and transparency requirements in helping insurance markets develop. As AI technology and incident data evolve, the study suggests that specialized insurance and complementary capital-market risk-transfer tools could become increasingly important.
The research provides a foundation for insurers, businesses and policymakers seeking to distinguish insurable AI-related losses from risks that may require broader risk-sharing or regulatory responses.