Can you build ethical review from blank page to operational process? They want evidence of systematic thinking, not just philosophical awareness.

1. Impact assessment — MAP 5.1: identify all affected communities and assess likelihood/magnitude of potential harms (bias, discrimination, privacy). 2. Fairness evaluation — define fairness metrics appropriate to the context (demographic parity for hiring, equalized odds for lending). 3. Technical review — audit training data, model architecture, and outputs using Fairlearn/AIF360. 4. Stakeholder engagement — consult affected communities and cross-functional teams. 5. Documentation — produce Model Cards, ethics review report, and remediation plan with monitoring KRIs.

This tests both technical and communication competence. Can you use the tools AND explain findings to non-technical stakeholders?

Start with data audit: examine training data for representation gaps and historical bias. Run disaggregated evaluation using Fairlearn across protected attributes (race, gender, age). Apply appropriate fairness metrics — demographic parity for equal opportunity contexts, equalized odds for predictive accuracy contexts. Use SHAP to identify which features drive disparate outcomes. Document findings with quantitative evidence and recommend mitigation: resampling, Exponentiated Gradient Reduction, or ThresholdOptimizer. Set up continuous monitoring for fairness drift post-remediation.

This is the defining challenge. ACM FAccT documents “decoupling” — ethics on paper without infrastructure. They want evidence you can drive adoption, not just write policies.

Frame ethics in business terms: EU AI Act fines (up to €35M or 7% of turnover), brand risk from public bias incidents, customer trust erosion. Propose risk-tiered review — lightweight review for low-risk systems, full review for high-risk, so ethics doesn’t bottleneck every launch. Build champion networks within engineering and product teams who can advocate for ethical review. Offer parallel track options: identify issues early in development so remediation happens during sprint cycles, not at launch gate.

Technical fluency check. The Ethics Officer must understand mathematical fairness definitions, not just philosophical concepts. They want someone who can navigate the impossibility theorem tradeoffs.

Demographic parity: positive outcomes are distributed equally across groups (e.g., equal hiring rates regardless of race). Equalized odds: true positive and false positive rates are equal across groups (e.g., equal accuracy for credit decisions). Calibration: predicted probabilities match actual outcomes across groups (e.g., a 70% risk score means 70% risk for every group). The impossibility theorem (Chouldechova 2017) proves these metrics cannot all be satisfied simultaneously when base rates differ. The Ethics Officer’s job is to choose which fairness definition is most appropriate for each context.

Hands-on tool knowledge matters. They want someone who has used SHAP and LIME in practice, not just read about them.

Two primary tools: SHAP (SHapley Additive exPlanations) provides global and local feature importance based on game theory; best for understanding which features drive model behavior overall. LIME (Local Interpretable Model-Agnostic Explanations) provides local instance-level explanations; best for explaining individual decisions to affected individuals. IBM AI Explainability 360 provides additional methods across tabular, text, image, and time series data. Enterprise platforms like Fiddler provide production-grade explainability dashboards. The key is matching the explainability method to the audience: SHAP for data scientists, plain-language LIME explanations for end users.