Why should generative AI evaluation include a check for harmful content in addition to groundedness and relevance?
Select an answer to reveal the explanation.
Short Explanation
Grounded and on-topic just means the facts check out and the answer stayed on subject - neither one screens for tone or safety. You still need a harmful-content check because safe, respectful output is its own requirement.
Full Explanation
The correct answer says an answer can be factually grounded and on topic while still containing offensive, dangerous or otherwise harmful material. Groundedness and relevance evaluate factual basis and topicality; neither is designed to detect offensive, dangerous or otherwise harmful phrasing, so a separate check is needed. The option claiming harmful content checks are a legal requirement in every jurisdiction without exception is incorrect because legal requirements vary by jurisdiction and are not the reason this check is included in evaluation design. The option claiming groundedness and relevance checks automatically flag harmful content as a side effect is incorrect because those checks do not inherently screen for harm; they test different properties entirely. The option claiming harmful content only appears in ungrounded answers, so checking groundedness would otherwise miss it, is incorrect because harmful content can appear in answers that are otherwise well grounded and accurate. For the exam, treat harmful content as its own evaluation dimension, independent of groundedness and relevance.