Well, I would rather say "composing music [or any other creative process] is roulette for copyright law"; your arguments are not specifically critical for generated art, but for art in general. At any point in time an artist could create a work which is similar to previous work (which is essentially unavoidable to some degree). The author of the previous work can take you to court if he/she assumes a copyright infringement (whether this assumption is justified or not); and how a court judges in such a case is indeed very often a game of chance; that's why copyright law, along with patent law, is such a popular field for trolls. In the U.S., it is at least possible for two artists to independently produce very similar works and still both be copyrightable (if the second artist can show that he/she did not simply copy); in other parts of the world this is not possible and the first artist prevails.

Anyway, in principle, the court does not care how an artist created a work. However, according to the law and current jurisprudence, the author must be a human being; creations by a machine are still not protectable by copyright; the artist could, of course, publish art created by AI under his/her name (algorithms have long been involved in many art forms). But if the result is a work that looks similar to an existing one, we are back to the case above.

And just as little as software developers know about copyright or its international treaties, so little do lawyers know about AI algorithms. This explains the peculiar nature of the discussions and the recent lawsuits mentioned in the text.