How can insurers stop AI proxy discrimination?

Removing protected characteristics from an underwriting model does not guarantee fair outcomes.

For years, removing protected characteristics from insurance data has been viewed as a simple solution to algorithmic bias. The assumption is that if a model does not access race, gender, or ethnicity, it cannot discriminate based on them.

Machine learning makes that assumption difficult to defend.

An AI model does not need a field labeled 'race' to pick up racial signals. ZIP codes, occupation, credit behavior, purchasing patterns and location can contain information about a person's race or other protected characteristics. When a model learns from those relationships, it can produce disparate outcomes across groups. That is proxy discrimination.

The Actuaries Institute's research on fairness in data-driven decision-making describes the underlying problem as the 'myth of blindness,' or fairness through unawareness. Removing sensitive attributes does not guarantee fairness, as other variables can serve as proxies. Gender may correlate with vehicle engine size, while postcodes can reflect racial or ethnic distributions. Those relationships can allow ostensibly neutral variables to act as proxies for protected characteristics. The research also notes that indirect discrimination becomes harder to address as models incorporate more information.

More data can make proxy bias harder to see

Auto insurance provides a familiar example. Geographic location can be closely associated with demographic composition.

A MoneyGeek analysis of ZIP codes across 69 U.S. cities found that drivers in less-white ZIP codes could pay as much as $891 more annually than drivers in predominantly white ZIP codes, using a consistent driver profile. The analysis found a strong correlation between the percentage of white residents in a ZIP code and premiums in many of the cities studied.

That correlation does not establish that race itself caused the difference. Geography can capture legitimate differences in insurance risk. The concern arises when geographic and other variables carry demographic information into a model, contributing to outcomes that disproportionately affect protected groups.

The same problem has appeared in healthcare. A widely used algorithm used healthcare spending as a proxy for patients' health needs. Researchers found that the algorithm underestimated the health needs of Black patients because it assumed lower healthcare spending indicated lower need. 

That matters as insurers move beyond traditional actuarial models and use increasingly complex machine learning systems. A model can identify relationships that were not obvious during feature selection or model design. More variables create more opportunities for those relationships to emerge.

Reviewing the feature list is only the starting point. Insurers also need to examine the outcomes those features produce across protected groups.

Regulators are looking beyond the model itself

U.S. insurance regulators are moving in that direction.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers states that decisions affecting consumers and supported by AI must comply with applicable insurance laws, including laws addressing unfair discrimination. The bulletin also sets expectations for governance of AI systems that insurers develop, acquire, or use, including systems supplied by third parties.

Colorado's Regulation 10-1-1 establishes governance and risk-management requirements for insurers that use external consumer data and information sources, including algorithms and predictive models that use such data.

Virginia has taken a similar approach. Its Administrative Letter 2024-01 sets expectations for insurers to govern and manage risks associated with the development, acquisition, and use of AI systems.

The European Commission's guidance on the AI Act classifies AI used for risk assessment and pricing in life and health insurance as a high-risk use case. Requirements for high-risk systems include risk management, data governance, documentation, traceability, and human oversight.

In the U.K., the Financial Conduct Authority's review of insurance firms' outcomes monitoring under the Consumer Duty says firms should monitor the outcomes customers receive and identify whether different groups are experiencing different outcomes.

For insurers, reviewing the variables going into an AI system is only part of the job. The resulting decisions need scrutiny too.

Explainability can reveal bias without removing it

This is where explainability tools such as SHAP (Shapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) come into play.

They can show which features influenced an individual prediction or help teams understand how a model behaves.

They do not, by themselves, establish that an outcome is fair.

A model can be transparent about the factors influencing its decision and still produce disparate outcomes. An explanation shows how features contributed to a prediction. It does not determine whether the resulting decision is fair.

Fairness needs to be addressed during model development and after deployment.

Taking three technical measures can help eliminate bias:

Debias the data before training. Sample weighting, rebalancing, and other preprocessing techniques can reduce the influence of historical disparities before a model learns from them.

Test for proxy relationships. An adversarial model can test whether protected characteristics can be inferred from the information available to the primary model. High inference accuracy can indicate that ostensibly neutral inputs carry demographic information.

Monitor outcomes after deployment. MLOps (machine learning operations) processes can track relevant fairness measures alongside model performance and business metrics. Changes in those measures can trigger further investigation.

The goal is not to reduce predictive accuracy, but to ensure that risk signals are not accompanied by hidden indicators of protected characteristics.

An insurer may remove race from datasets, restrict access to sensitive attributes, and document model features to explain individual decisions. But none of these steps tells the insurer whether the model is producing different outcomes for different groups.

This is where AI fairness in insurance must be evaluated.