A customer wants a scalable way to evaluate its generative AI outputs. What should that evaluation cover?
Select an answer to reveal the explanation.
Short Explanation
A fluent, fast, wrong answer is still a wrong answer. You need evaluation that checks whether the content itself holds up, including groundedness, relevance, harmful content, and policy compliance wherever they apply.
Full Explanation
The correct choice is the option that says evaluation should cover groundedness, relevance, harmful content, and policy compliance, checked wherever each applies to the use case. Generative AI evaluation exists to catch failures that fluency checks miss entirely: answers that are ungrounded, irrelevant, harmful, or non-compliant with policy. The option focusing only on fluency and grammatical correctness is incorrect because grammatically correct text can still be false, irrelevant, or unsafe. The option focusing only on response latency is incorrect because latency is an operational metric unrelated to whether the content itself is trustworthy. The option focusing only on whether output matches the exact wording used in training examples is incorrect because generative systems are expected to produce novel phrasing, not repeat training text, so wording match is not a meaningful evaluation criterion. For the exam, evaluate generated content on groundedness, relevance, harm, and policy compliance, not on surface qualities.