Name two metrics to evaluate how well an AI system aligns with human values in a safety-critical task.

• A. Model size and training speed
• B. Alignment accuracy and safety margin
• C. Data diversity and recall
• D. Explanation fidelity only

Answer: (B)

Evaluating alignment with human values in safety-critical tasks hinges on two practical metrics: alignment accuracy and safety margin. Alignment accuracy measures how often the system’s decisions match human judgments or established guidelines in representative scenarios, showing whether the AI tends to do what humans would approve. Safety margin gauges how far the system’s behavior is from the boundary of unsafe actions, providing a buffer against uncertainty, distribution shifts, or adversarial inputs. Together, they capture both whether the AI generally honors human values and how robust it is against risky situations.

Other options miss this combined focus. Model size and training speed describe resource use, not alignment with values. Data diversity and recall relate to data coverage and information retrieval, not normative agreement with human values. Explanation fidelity concerns how well the system can justify its decisions, which is about transparency rather than actual alignment in safe, real-world actions.