4 ms·
Is it possible that it "remembers" that url from it's training and the content tied to it?
Is it possible that it "remembers" that url from it's training and the content tied to it?
Assuming the earlier statements from ChatGPT were correct, the feature did not exist before the training data cutoff, so that's impossible.
I don't think they prefix each document with the URL in the training set, but there might be URLs in the clear text.