But if it produces the whole article verbatim, it should be relatively straightforward to match its output to the training set and give attribution, no?
Doesn't make legal what? Copying or publishing the articles? The model output could as easily be modified to output link to the article instead of quoting it verbatim.
