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

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.