The important question here isn't who owns the copyright in the output, but rather whether the output also infringes on somebody else's, and if so, who's liable for that: The company that trained the AI, the one that hosts it as a service, or the user that prompted it.