I'd half-agree, but I don't think "breaking copyright" matters to the question of "is LXM 'AI' plagiarism?".
Like you say you can plagiarize without braking copyright(for cases where the copyright allows usage without attribution such as with public domain), and it's also possible to break copyright without plagiarism(e.x. redistributing with attribution when you don't have the license).
But I think this is irrelevant to the point being made. LXM's need to take in a large amount of data, and then the outputs are attributed to the "model" rather than the originators of the material.
Since most of the content being digested by LXMs is not public domain that's where copyright gets twisted up with it, since for the majority of LLM training data 'plagiarism' and 'breaking copyright' come from the same act of redistributing/using without attribution(and since the "LXM" is considered to have created the data by most people the 'plagiarism' comes in).
As a creator I feel like that's not very useful, to be a single name in billions. Of course I'd still like attribution if the work was significantly based on mine.