That's a strawman. What people are concerned about are the extreme number of examples of LLMs reproducing text and code verbatim from the training set, and it's inherent to the way the models are trained and designed.
If LLMs actually produced "homeopathic plagiarism", far less people would be concerned about it.
https://ai.stanford.edu/blog/verbatim-memorization/
https://arxiv.org/abs/2510.25941
https://aclanthology.org/2025.emnlp-main.1789.pdf