Navigating AI Governance in Insurance: Challenges and Solutions

The integration of artificial intelligence (AI) into the insurance industry marks a profound shift in how underwriting and claims processes are conducted. However, this technological advancement is not without its challenges, particularly concerning governance and ethics. As machine learning algorithms become more prevalent, insurers face a crucial dilemma: how to harness the power of AI while ensuring that their practices remain fair and compliant with legal standards. This article delves into the complexities of AI in underwriting, the risks associated with proxy discrimination, and the evolving regulatory landscape that insurers must navigate to protect both themselves and their customers.

The Promise and Peril of AI in Insurance Underwriting

AI and machine learning algorithms have the potential to revolutionize underwriting processes by analyzing vast amounts of data to identify patterns and predict outcomes. Traditional underwriting relies on explicit factors such as a customer’s credit score or claims history. In contrast, AI can uncover correlations that are not immediately apparent, potentially leading to more accurate risk assessments.

However, the very nature of these algorithms poses a significant challenge. Insurers often do not fully grasp that AI can exploit correlations involving sensitive characteristics like race, gender, and health status, even if these factors are explicitly excluded from input data. As Daniel Schwarcz, a legal expert at the University of Minnesota, points out, the mathematics of machine learning can inadvertently lead to outcomes that discriminate against protected classes. This phenomenon, known as proxy discrimination, occurs when an algorithm identifies alternative data points that serve as stand-ins for sensitive attributes. For instance, if a model uses zip codes that are predominantly inhabited by certain racial groups, it may still produce biased outcomes based on the historical data embedded in those zip codes.

The Regulatory Landscape: A Patchwork of Standards

The regulatory environment for AI in insurance is complex and evolving. Currently, there is no unified national standard governing the use of AI technologies in underwriting. Instead, insurers are navigating a patchwork of state regulations that vary significantly in their requirements for testing and enforcement. In December 2023, the National Association of Insurance Commissioners (NAIC) adopted a Model Bulletin on AI use, which provides a framework for governance. However, it stops short of mandating specific outcomes or testing methodologies.

This lack of standardized requirements can leave insurers uncertain about how to achieve compliance. As Birny Birnbaum, executive director of the Center for Economic Justice, emphasizes, while governance practices are essential, they are insufficient without robust data collection and testing. Without these measures, insurers cannot ascertain whether their governance is effectively mitigating risks associated with proxy discrimination.

State-Level Initiatives

Some states have taken proactive steps to address these issues. For example, Colorado passed legislation in 2021 requiring insurers to test their AI models for unfair discrimination. However, as of mid-2026, no testing regulations had been implemented across any line of insurance. In contrast, New York has issued bulletins that specifically mandate testing for unfair discrimination using AI, setting a precedent for other states.

The Call for Testing and Accountability

As the regulatory landscape evolves, there is a growing expectation that insurers will not only document their governance practices but also demonstrate the effectiveness of their AI models through rigorous testing. This shift emphasizes the need for insurers to compare actual consumer outcomes with intended ones, particularly across protected classes. The emerging consensus is that AI governance must incorporate documented testing at both the development and post-deployment stages.

• Document Testing: Insurers should implement a formal testing protocol that evaluates the impact of AI models on different demographic groups.
• Statistical Controls: Incorporating statistical controls for protected characteristics can help identify when an AI model is inadvertently predicting sensitive attributes.
• Technology Utilization: Leverage AI governance tools, such as bias detection systems and fairness audit infrastructure, to facilitate comprehensive testing.
• Vendor Accountability: Insurers must ensure that third-party vendors' algorithms are also subject to governance standards, addressing a critical gap in many existing contracts.

Preparing for Future Regulatory Changes

As regulators contemplate moving from guidance to more prescriptive regulations, insurers must proactively prepare for a landscape where they may face affirmative restrictions on the use of machine learning in specific functions, particularly in pricing. Schwarcz warns that insurers that have built their pricing architectures around opaque machine learning models without adequate testing records could find themselves at a significant disadvantage if stricter regulations are introduced.

To navigate this potential shift, insurers should prioritize building testing into their model lifecycle, ensuring that they maintain auditable documentation of their processes. This preparation includes empowering legal and compliance teams to play a pivotal role in AI deployment decisions, rather than relegating them to a final review stage.

Key Takeaways

• AI in insurance underwriting offers significant predictive advantages but poses risks of proxy discrimination.
• The regulatory landscape is fragmented, with varying state-level requirements and a lack of unified national standards.
• Documented testing and accountability are becoming essential for insurers to demonstrate compliance and protect consumers.
• Insurers must prepare for potential future regulations that may impose restrictions on AI use, particularly in pricing.
• Effective governance requires integrating legal and compliance considerations throughout the AI deployment process.

Frequently Asked Questions

What is proxy discrimination in AI?

Proxy discrimination occurs when AI algorithms inadvertently produce biased outcomes based on correlations with sensitive attributes, such as race or gender, even when these attributes are excluded from the model's input data. This can happen when the algorithm identifies alternative data points that serve as stand-ins for these sensitive characteristics, potentially leading to unfair treatment of certain groups.

How are regulators addressing AI in insurance?

The regulatory landscape for AI in insurance is currently fragmented, with some states implementing specific requirements for testing and accountability while others have yet to establish clear guidelines. The NAIC has introduced a Model Bulletin that outlines governance expectations but does not mandate specific outcomes. This inconsistency can create challenges for insurers trying to achieve compliance.

What steps can insurers take to ensure compliance with AI regulations?

Insurers can take several proactive steps to ensure compliance with evolving AI regulations. These include implementing documented testing protocols to evaluate the impact of AI models on protected classes, leveraging AI governance tools for bias detection and fairness audits, and ensuring that third-party vendor algorithms are subject to the same governance standards as their own. Additionally, empowering legal and compliance teams to lead AI deployment decisions is crucial for effective governance.

What future changes can insurers expect in AI regulation?

As regulators continue to scrutinize AI technologies in insurance, insurers can expect potential changes that may include stricter testing requirements, the introduction of affirmative restrictions on machine learning use in specific functions, and greater emphasis on accountability for the outcomes produced by AI models. Insurers that have already integrated testing into their model lifecycle and maintained comprehensive documentation will be better prepared for these regulatory shifts.