Training data should absolutely be under control of the creator / owner of the IP, and AI trained with stolen IP should itself be considered stolen. The lack of

Training data should absolutely be under control of the creator / owner of the IP, and AI trained with stolen IP should itself be considered stolen. The lack of clarity here is only because of the newness of the technology, but this will be as clear to our future selves as "warez" are now known to be stolen work.