Isn't it also because lowering hallucinations requires repeated training with the same fact/data, which makes the final response closer to the training source itself and might lead to more direct charges of plagiarism (which may not be economical)?