As long as you transform copyrighted data into something else then it appears so.

Which will be also interesting for prompts used by the users of these AI companies, because even if those prompts contain copyrighted data, they can be used for training because training is transforming these prompts into something else.