Test Underwriting Algorithms for Unfair Bias

State rules, including Colorado's requirements for life insurers and New York DFS guidance on AI in underwriting and pricing, expect carriers to test models for unfairly discriminatory outcomes and document their methods. PointGuard AI generates bias-oriented probes tailored to each carrier's lines and requirements, runs them continuously, and preserves evidence as part of a broader fairness and model risk program.

Fairness testing must be repeatable, documented, and current as models change.

• Underwriting models can encode bias from training data or proxies.
• Model and data changes can alter outcomes after approval.
• Regulators expect documented methodology, assumptions, and results.
• Testing is often manual and difficult to repeat.

AI Red Teaming, Agent Mission Control, and AI Governance help carriers test models and enforce approved boundaries.

1. Define fairness scenarios. Create probes reflecting life, auto, and health underwriting tasks and protected characteristics, with criteria set by the carrier's actuarial and compliance teams.

2. Test and retest. Run probes on a schedule and after model or data changes. Compare results across versions to detect regressions.

3. Enforce boundaries. Route sensitive underwriting actions for human review and restrict agents from acting on unapproved recommendations.

4. Document for examiners. Record methodology, results, and remediation in a reviewable format for internal governance and examiner requests.

Adversarial testing complements statistical fairness analysis and legal review; it does not by itself establish compliance with any state rule.

• OWASP Top 10 for LLMs: LLM09 Misinformation and LLM06 Excessive Agency, addressing misleading outputs and acting on them without appropriate limits.
• OWASP Top 10 for Agentic Applications: ASI09 Human-Agent Trust Exploitation, covering overreliance on agent recommendations.
• NIST AI RMF: GOVERN defines accountability; MAP considers affected policyholders; MEASURE evaluates harmful bias; MANAGE prioritizes remediation.