AI Risk Mitigation and Controls
Core concepts
Controls reduce risk to within appetite. Candidates must distinguish preventive, detective, and corrective controls, and automated versus manual controls. Map controls across the lifecycle: data controls (provenance, quality gates, de-identification), development controls (peer review, version control, testing standards), deployment controls (staged rollout, approval gates), and runtime controls (input validation, output filtering, rate limits, monitoring hooks).
Know human-in-the-loop versus human-on-the-loop versus fully automated oversight, and when each is appropriate. Understand guardrails for generative AI (content filters, grounding/retrieval, prompt hardening, output constraints), model risk controls (champion-challenger, fallback models, confidence thresholds), and security controls (adversarial input detection, access control, model access logging).
Key principle: defense in depth. No single control is sufficient; layer preventive and detective controls. Controls must be tested for effectiveness, not merely documented.
Traps
Questions often assume a control exists because it is documented. Effectiveness must be evidenced. Another trap: treating human oversight as a control when the human lacks authority, information, or time to intervene. Automation bias can defeat nominal human review.
Rule of thumb
For every high-risk use case, require at least one preventive and one detective control, plus a corrective path (rollback or kill switch).
Worked example
A customer-facing assistant can produce harmful advice. Preventive: output filter and grounding to approved content. Detective: sampled review and flagging of low-confidence responses. Corrective: ability to disable the feature and notify affected users. Documented filters alone are insufficient without effectiveness testing.