AI Risk Management and Ethics
Frameworks and concepts
Candidates should understand risk management as a continuous cycle: identify, assess, mitigate, monitor, and communicate. Know the harm categories AI systems can create - bias and unfairness, privacy harms, safety incidents, opacity, misuse, and environmental or economic impacts. Be familiar with fairness concepts at a conceptual level: disparate treatment versus disparate impact, and the tension among fairness metrics that cannot all be satisfied simultaneously. Understand transparency, explainability, and interpretability as distinct ideas, and the role of human oversight, contestability, and redress. Know that ethics is not only about intent but about foreseeable outcomes and affected stakeholders.
Traps that appear in questions
A common trap is assuming a technically accurate model is automatically fair or ethical; accuracy and fairness are separate axes. Another is believing a single fairness metric can be optimized universally - tradeoffs are inherent. Questions often confuse explainability (why this output) with transparency (how the system works overall). Watch for scenarios that treat human-in-the-loop as a checkbox rather than meaningful oversight with authority to override, and for risk assessments that omit downstream or third-party users.
Worked example / rule of thumb
Rule of thumb: risk severity = likelihood × impact, judged against the most affected stakeholder, not the deployer. Example: a hiring screener with 95% accuracy but materially lower selection rates for one group carries disparate-impact risk that accuracy alone does not reveal; mitigation may require reweighting, feature review, or human review of borderline cases.