A consultant recommends evaluating a generative assistant for groundedness, relevance, harmful content and policy compliance. Which of these should be applied to every generative AI use case regardless of domain?
Select an answer to reveal the explanation.
Short Explanation
You can't use one check as a stand-in for the others. Relevance won't catch harm, and safety won't catch a made-up fact. Apply all four wherever each criterion is relevant to the use case.
Full Explanation
The correct choice is to apply all four criteria—groundedness, relevance, harmful content, and policy compliance—wherever each criterion is relevant to the specific use case, rather than picking only one dimension to check. Each criterion catches a different kind of failure, and a system can pass one while failing another, so all four need to be checked where applicable rather than treating any single one as sufficient.
The option that says to check only harmful content is incorrect because factual inaccuracy and irrelevance are real risks in their own right, not lesser concerns than harmful content.
The option that says to check only policy compliance is incorrect because policy documents describe expectations; they do not themselves verify that generated content meets them.
The option that says to check only relevance is incorrect because an on-topic answer can still be ungrounded, harmful, or non-compliant.
For the exam, evaluate against all applicable criteria, not a single proxy for the rest.