AI Ethics, Governance, and Risk
Core concepts
Ethics and governance address how AI is built and used responsibly. Know the recurring principles: fairness, transparency/explainability, accountability, privacy, safety, and human oversight. Understand bias sources-historical data, sampling, labeling, proxy variables-and that a model can be technically accurate yet discriminatory. Recognize privacy risks: personal data in training sets, memorization, and re-identification from supposedly anonymized data.
Governance
Candidates should understand risk-based regulation (higher-risk uses demand stricter controls), impact assessments, documentation (model cards, datasheets), audit trails, and human-in-the-loop review for consequential decisions. Know transparency as disclosure that AI is in use and explainability as justification of specific outputs-related but distinct. Understand that governance spans the lifecycle: design, data, training, deployment, monitoring, retirement.
Traps
Questions often present a single "ethics checklist" as sufficient; durable governance needs ongoing monitoring. Watch for conflating explainability with accuracy-an accurate model can be opaque. Another trap: assuming removing a protected attribute eliminates bias, when proxy variables (e.g., ZIP code) can reproduce it. Also distinguish accountability (a named responsible party) from mere disclosure.
Rule of thumb
Apply the harm-and-reversibility test: the more consequential and irreversible a decision (hiring, medical, credit), the more you need human review, documented rationale, and appeal paths.
Worked example
A résumé screener trained on past hires learns historical bias. Removing names is insufficient because correlated features remain. Mitigation: audit outcomes across groups, test for proxies, add human review, and document limitations.