Responsible ML: Interpretability, Fairness & Privacy
The boundary: this lesson is the modelling and measurement side of responsibility, what you compute and what you can claim. The governance workflows around it (registry approvals, promotion gates, audit trails, rollback) live in MLOps Model Lifecycle & Governance; LLM safety, jailbreaks and prompt injection are Guardrails and Prompt-Injection Defense; and attacks on ML systems (evasion, poisoning, extraction, membership inference as a threat model) belong to Adversarial ML & Model Security. Drift and fairness monitoring after launch connect to ML Monitoring, Drift & Retraining; prediction intervals and selective prediction are in Uncertainty Quantification & Conformal Prediction; and Causal Inference Basics and Uplift Modeling & Heterogeneous Treatment Effects are the right tools whenever the question is "what would happen if" rather than "what does the model use". Pricing Optimization & Marketplace Economics flags personalized pricing as a fairness audit target, and this is where that audit is defined.
Interviews ask this as "explain a prediction to a customer who was declined", "SHAP says income matters, so should we raise incomes?", "your model approves one group half as often: is it biased and what do you do?", "how would you choose a fairness metric for hiring versus triage?", "what does epsilon mean?" and "write a model card". The senior answer names the audience and the decision first, says which method answers which question (and what it cannot answer), refuses to collapse fairness into one number, and treats privacy as a quantified budget rather than a checkbox.
Everything executable below was run (Python 3.12, NumPy 2.5, scikit-learn 1.9, pandas 2.3, SHAP 0.53, fixed seeds) on a simulated credit dataset, or derived by hand, and is marked "(verified)". Legal and regulatory statements, and papers cited from the literature, are recalled from memory, are stated generally, and are marked "(from memory)": this is engineering education, not legal advice, so check with counsel for your jurisdiction and product. Simulated numbers show mechanics and relative sizes, not the behaviour of any real lender or regulator.