It's not returning the text from its own weights (which is impossible anyway); but it is returning a direct link to that text, which a clean analogy to the DNS resolver. (Which is the analogy I interpreted the GP comment to be making).
Returning text from the weights is not impossible. Just heavily disincentivized in the contemporary models. Back in ~2023 you absolutely could get copyrighted materials from the training data repeated verbatim (assuming of course you had another copy to verify the integrity after the fact).
Returning text from the weights is not impossible. Just heavily disincentivized in the contemporary models. Back in ~2023 you absolutely could get copyrighted materials from the training data repeated verbatim (assuming of course you had another copy to verify the integrity after the fact).