Can an AI model deliver strong predictive accuracy while still relying on biased or distorted data?
In this episode of the Society of Actuaries Research Institute's Research Insights podcast, host Lisa Schilling, Director of Practice Research, welcomes researchers Hong Li, FSA, ACIA, and Jiangang He to discuss their report, Evaluating Bias Mitigation Strategies in AI Models for Insurance Pricing and Risk Assessment.
The conversation explores how measurement distortions and indirectly associated variables can affect insurance risk predictions, even when sensitive group labels are excluded from model training. Drawing on synthetic simulations across five insurance applications—life, auto, credit, mortgage and health—the researchers compare traditional generalized linear models, standard neural networks and adversarial neural networks designed to mitigate potential bias.
Listeners will discover why predictive accuracy alone may not be sufficient, how adversarial learning can help reduce group-level disparities, and why evaluating model performance at real-world decision thresholds is essential for effective model validation and governance.
Listen to the episode to learn how actuaries can strengthen AI model validation while balancing predictive performance, fairness and responsible risk assessment.