Algorithmic bias
A systematic error of an AI system that produces results unfairly unfavorable to certain groups, often arising from biased training data.
Obermeyer, Powers, Vogeli and Mullainathan, Science, October 25, 2019: a health algorithm used across the United States predicted costs in place of needs, and at the same score Black patients were markedly sicker than white patients. Algorithmic bias is the tendency of an AI system to produce systematically skewed results, often to the detriment of certain groups of people. The failure has nothing random about it. The deviation is reproducible, and it usually starts upstream: unrepresentative training data, variables correlated with protected characteristics, or the design of the model itself. In insurance, this is a central concern. A biased pricing engine can unfairly penalize a class of policyholders and cross into prohibited algorithmic discrimination. The AI Act answers with data governance and human oversight requirements. Directive (EU) 2024/2853 on liability for defective products, for its part, leaves discrimination and privacy infringements outside its scope, as they give no right to compensation on their own, and leaves that litigation to national law.
Textbook case: a model trained on historical data reflecting past inequalities may reproduce an algorithmic bias by systematically overpricing a neighborhood, which the insurer must detect and correct.
- Directive (EU) 2024/2853 of October 23, 2024 on liability for defective products, Article 6 (compensable damage); CMS, "The new EU Product Liability Directive (EU) 2024/2853"
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.