The example from the article doesn't show that LLM is trained on copyrighted data - it's just Bing fetching the source article, providing it to GPT, and GPT rephrasing the article. An agent trained on entirely copyright-free data would provide exactly the same output.