Bias diagnostics for Black-Box Actuarial AI Systems under evolving regulations

The use of artificial intelligence (AI) in actuarial science is becoming increasingly widespread, from risk assessment to claims management, but these models may inadvertently introduce discriminatory biases, either directly through the use of sensitive attributes (e.g. gender, ethnicity) or indirectly through variables correlated with them. Prior research has shown that simply removing sensitive variables is insufficient and may even amplify unfairness. At the same time, regulatory frameworks (e.g. Gender Directive 2012, Colorado 10-1-1 Regulation, NAIC Model AI Bulletin, EU AI Act) are imposing stricter requirements for transparency, fairness and accountability in AI-based decision making. In this work, we map these key regulations relevant to actuarial practice and link them to the two fundamental forms of bias: direct and indirect. We show that this distinction forms the foundation for understanding widely used concepts such as disparate treatment, disparate impact and proxy discrimination. We then propose a structured decomposition of direct and indirect biases and illustrate how they manifest in a life insurance case study. This unified perspective clarifies the regulatory implications of bias in actuarial AI models and offers practical guidance for actuaries seeking to design compliant, ethically sound and robust AI systems.