AI 420 · Year 4 · Semester 2 · 3 credits · AI & Machine Learning

Explainable AI, Fairness & Model Governance in Insurance

From black-box predictive power to defensible actuarial decisions: mastering explainability, algorithmic fairness, and model governance.

Full practice exam · 30 Q · 120 min

Season 1 · 8 episodes

Auditing the Black Box

When Centennial Mutual's flagship auto pricing model faces an unexpected regulatory audit under Colorado SB21-169, newly appointed Model Risk Actuary Maya Lin must dissect, explain, debias, and govern a high-dimensional gradient-boosted machine before the state insurance commissioner revokes their license to underwrite.

Protagonist · Maya Lin, FSA, newly appointed Lead Model Risk Actuary at Centennial Mutual, caught between aggressive data science teams and uncompromising state insurance regulators.

Setting · Centennial Mutual headquarters in Denver, Colorado, spanning the executive boardroom, data science bullpen, and state insurance commission hearing chambers.

Stakes · If Maya fails to prove Centennial's machine learning rating algorithm is explainable, statistically sound, and free from unfair proxy discrimination, the Division of Insurance will reject their statewide rate filing, impose a multi-million-dollar market conduct fine, and freeze underwriting across five Western states.

Pick an episode

Episodes

Season finale: Maya stands before the Colorado Insurance Commissioner with a fully validated, debiased, and mathematically audited filing binder under ASOP 56, proving that Centennial's neural pricing pipeline achieves Pareto-optimal fairness without compromising actuarial solvency or violating statutory anti-discrimination laws.

From the lecturer: Welcome to AI 420, your capstone bridge between advanced machine learning and professional actuarial responsibility. In this course, we demystify complex pricing and reserving models, transforming opaque algorithms into transparent, fair, and legally compliant actuarial systems. To get the most out of these lesson-films, keep a dedicated notebook and a Python REPL open at all times. You will not simply call off-the-shelf interpretability libraries; you will derive Shapley values by hand on three-player cooperative games, compute Accumulated Local Effects across correlated rating territories, and prove the mathematical impossibilities inherent in simultaneous fairness constraints. Although AI 420 does not target a single preliminary multiple-choice exam, it directly prepares you for the highest tier of modern actuarial practice: complying with ASOP No. 56 (Modeling), ASOP No. 12 (Risk Classification), the NAIC Model Bulletin on Artificial Intelligence, Colorado SB21-169 quantitative testing rules, and the EU AI Act. Treat every validation framework and governance memo as an auditable professional work product that you would confidently sign before an insurance commissioner.

Cold open

At six fifteen on a Monday morning, Maya Lin's desk phone rings. It is the Chief Actuary. The Colorado Division of Insurance has just flagged Centennial's pending personal auto rate filing for an emergency investigatory review: their brand-new LightGBM frequency-severity model is quoting thirty percent higher premiums in zip codes with high minority populations, and the rate filing contains zero interpretable rating relativities. Maya has seventy-two hours to explain how the model actually makes its predictions.

preparing lecture…

Transcript

Formulate Partial Dependence Plots (PDP) and pinpoint why feature correlation causes them to evaluate impossible risk profiles.
Construct Accumulated Local Effects (ALE) curves from conditional distributions to eliminate out-of-distribution extrapolation.
Derive Shapley values from cooperative game theory and prove how the four fundamental axioms enable exact additive feature attribution for complex insurance pricing models.