Do AI underwriting models discriminate against protected groups

A Genuine, Documented Risk

Yes, AI underwriting models genuinely can produce discriminatory outcomes against protected groups, and this isn’t merely a theoretical concern — documented research and regulatory findings have identified real cases of concerning disparate impact in some insurance AI models, which is precisely why this has become a significant focus of insurance regulatory attention.

How Historical Data Bias Can Get Learned by AI Models

AI underwriting models trained on historical claims and policyholder data can inadvertently learn and replicate patterns of past discrimination or historical inequities present in that underlying data, since the model is identifying statistical patterns in the data it’s given, without independently distinguishing between patterns that reflect legitimate risk and patterns that merely reflect historical bias or inequitable past practices.

Understanding Proxy Discrimination

A particularly significant, well-documented concern is proxy discrimination, where a factor used in a model — such as a specific geographic location or a certain data pattern — closely correlates with a protected characteristic like race or national origin, effectively producing discriminatory outcomes even though the model doesn’t directly use the protected characteristic itself as an explicit input.

Why This Risk Has Prompted Regulatory Requirements

Given these documented risks, insurance regulators in a growing number of states have introduced or proposed specific requirements for insurers to test AI-based underwriting and pricing models for potential discriminatory impact, including specifically checking for proxy discrimination effects, reflecting a policy recognition that facially neutral models can still produce genuinely discriminatory real-world outcomes.

Why Insurers Have a Strong Incentive to Proactively Test for Bias

Beyond regulatory requirements, insurers face genuine legal and reputational risk from deploying AI underwriting models that produce discriminatory outcomes, giving responsible insurers a strong practical incentive to proactively test their models for bias and proxy discrimination effects before and during deployment, rather than waiting for a regulatory finding or legal challenge to surface the problem.

Why This Remains an Active, Ongoing Area of Concern and Oversight

Given the genuine complexity of detecting subtle proxy discrimination effects, particularly in more sophisticated AI models analyzing many combined data factors, ensuring insurance AI models don’t discriminate against protected groups remains an active, ongoing area of regulatory attention and industry practice, rather than a fully and permanently solved problem.

Bottom Line

AI underwriting models genuinely can produce discriminatory outcomes against protected groups, whether through learning biased patterns from historical data or through proxy discrimination where seemingly neutral factors closely correlate with protected characteristics — a documented, real risk that has prompted regulatory bias-testing requirements and gives insurers strong incentive for proactive testing.

Go deeper

• What Real-World Harms Have Resulted From Biased AI Systems?
• Why Do AI Models Sometimes Produce Biased or Discriminatory Outputs?
• What is adverse impact analysis and why does it matter for ai hiring tools?