Google recommends enabling search grounding whenever a model may need to know obscure or recent facts. What problem does that address?

The principle — Fluency and accuracy are produced by the same machinery, so one is no evidence of the other. Recent and obscure facts are where the gap opens widest.

Why the key is correct — Google states that grounding with search connects the model to real-time web content and should be enabled whenever it may need to know obscure or recent facts, and frames tool use as the way to avoid hallucinations where a model might otherwise produce incorrect responses.

Why the others are wrong — Models do not reliably refuse; they answer confidently. Grounding is about the question rather than the product. And self-checking cannot supply information the model never had.

Remember this — Confidence is a style. For anything recent or obscure, give it a source or use one.

Sources — Google, Prompt design strategies.