Colorado Tackles AI Bias in Insurance, Raises Questions on Defining Unfair Discrimination

Insurance is increasingly relying on sophisticated computer algorithms to make key decisions affecting consumers. This shift has spurred regulators to grapple with the challenge of preventing discrimination embedded within artificial intelligence (AI) systems. Colorado is at the forefront in addressing algorithmic discrimination within the insurance sector. The state faces complex questions: What counts as unfair discrimination when AI systems make insurance decisions, and how can such discrimination be measured without explicit data on policyholder race?

Regulators in Colorado aim to protect consumers from unfair discrimination under a law passed in 2021. This law focuses on how insurance companies utilize extensive customer data. The law is still in its implementation phase, and an industry group is set to present its proposal for measuring discrimination and inferring race next week.

Traditionally, insurers do not ask applicants for their race. To comply with Colorado’s law, insurers might have to infer the race of their policyholders through methods like Bayesian Improved First Name Surname Geocoding (BIFSG). For instance, BIFSG calculates the probability of an individual’s race based on their first name, last name, and ZIP code. However, the accuracy of this method can vary, posing another challenge for regulators.

The complexities are not confined to technical methods. There is a fundamental question that Colorado must address: what qualifies as unfair discrimination. For example, insurers charge higher premiums based on risk factors such as smoking or youthful driving age, criteria generally accepted as legitimate. However, complications arise when considering factors that correlate strongly with race, such as neighborhood crime rates affecting car insurance premiums for predominantly Black neighborhoods. Disentangling these calculations from historical racial discrimination in housing, education, and job sectors is not straightforward.

According to Jason Lapham, Big Data and AI Policy Director at the Colorado Division of Insurance, stakeholders are still working to define what constitutes an unfairly discriminatory outcome. A consultation process remains ongoing, initially targeting the life insurance industry before extending to auto and health insurance.

For example, Verisk Analytics Inc., a prominent insurance rating bureau, has proposed an alternative to BIFSG, relying solely on geographic data rather than names.

Colorado law scrutinizes the use of external consumer data beyond traditional underwriting factors, such as credit scores and social media history. These data sources can closely correlate with race, inadvertently leading to discriminatory outcomes even when race is not explicitly used in underwriting decisions. Earlier guidelines suggested testing for discrimination in premiums and approval rates between non-White and White applicants. However, industry feedback indicated that these tests might not yield meaningful results.

Next week, the American Council of Life Insurers (ACLI) will present its response to the draft rules proposed by the Colorado Division of Insurance. ACLI’s approach suggests testing the predictive power of external data to ensure it does not act as a proxy for race, but questions remain about what actions should be taken if such tests reveal discriminatory outcomes.

To read more about these ongoing regulatory efforts, visit the full article on Bloomberg Law.