Algorithmic discrimination
Unfavorable and unjustified treatment of individuals by an automated system, on the basis of protected characteristics, directly or through proxy variables.
ProPublica, "Machine Bias," May 23, 2016: among Broward County defendants who did not reoffend, the COMPAS tool rated 44.9% of Black defendants high risk against 23.5% of white defendants. Algorithmic discrimination arises when an automated system treats individuals unfavorably and unjustifiably on the basis of protected characteristics such as origin, sex or health status. It may be direct, where the characteristic is used as a criterion, or indirect, where an apparently neutral variable, a postcode for instance, serves as a proxy for a protected characteristic. Distinct from mere algorithmic bias, which is often its technical cause, discrimination is a matter of legal characterization. In insurance, it runs up against the principles of actuarial fairness and non-discrimination already tested by case law. This matters all the more because the Product Liability Directive (EU) 2024/2853, in its Article 6, compensates only death, personal injury, damage to property and loss of data: an infringement of fundamental rights is left to national laws and the GDPR.
Textbook case: an algorithm that more often refuses cover to residents of certain neighborhoods may constitute indirect algorithmic discrimination if the geographic criterion masks a protected characteristic.
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.