Why an AI Answer Still Needs Its Sources Checked

A link and a supported claim are different things

An AI answer can contain a real link and still make a claim that the linked page does not support. The link might concern the right subject but a different population, date or condition. It might support only one clause of a longer sentence. Evaluating the answer therefore requires checking the relationship between statement and evidence, not merely confirming that a web page opens.

NIST defines retrieval-augmented generation as combining information retrieval with generated output. Its generative-AI risk profile discusses confabulation, including confidently presented false content. Read together, these sources provide an important boundary: retrieval can supply material to a system, but the presence of supplied material does not independently verify every sentence the system produces. The final answer remains an object that needs evaluation.