Which approach best summarises how a generative AI system's outputs should be evaluated before wide release?
Select an answer to reveal the explanation.
Short Explanation
Systematic evaluation against named criteria beats gut feeling every time. You need a check, not a hope, before wide release.
Full Explanation
The correct approach is systematic testing of outputs for groundedness, relevance, harmful content, and policy compliance, applying each criterion where it is relevant. A defined set of evaluation criteria applied systematically is what turns 'we think it's fine' into a checkable, repeatable judgement about whether an assistant is safe and trustworthy to release. Relying on the development team's confidence that the model performs well based on informal testing during development is incorrect because informal developer confidence is not a substitute for structured evaluation against defined criteria. Waiting for user complaints after release and addressing issues only as they are reported is incorrect because it means real users encounter harmful or false answers before anyone catches the problem. Evaluating only response speed and uptime is incorrect because speed and uptime are operational metrics that say nothing about whether the content produced is grounded, relevant, safe, or compliant. For the exam, describe pre-release evaluation as systematic and criteria-based, not informal or purely operational.