Two weeks ago, I wrote a column about how AI training licenses for music might work under the assumption that the development of such deals is a matter of when, rather than if. I said that I’d return to the subject, to focus on how rightsholders might divide money from those deals, but I was interrupted by the Trump Administration’s attack on a late-night talk show, because 2025. Now there’s more relevant news, as an AI-related lawsuit brought by German PRO GEMA goes to court in Munich, offering a look at why I think licensing for AI is something of an inevitability.  

In the U.S., it’s not quite. The major labels’ lawsuit against Suno and Udio involves fair use, so it’s possible — but, I think, unlikely — that a court could determine that Suno and Udio don’t need a license in order to copy music for the purposes of training their AI algorithms. (The two sides are said to be in settlement negotiations.) The European Union doesn’t have fair use, though — and the first of the two big cases that could clarify EU law on AI and music goes to court on Monday (Sept. 29). 

That morning, a Munich district court will hear GEMA’s lawsuit against OpenAI for copying lyrics it controls. (The second case is GEMA’s suit against Suno for copying recordings as well as its output of songs that sound like the originals.) From the U.S. point of view, this might seem unusual — why is a collecting society dealing with lyrics and how important are they by themselves, anyway? European collecting societies control different rights, including mechanicals, which in some cases include lyrics — which are copyrighted in the same way songs are.

From a legal point of view, lyrics present the same issues as recordings. “It’s the same principle as in the Suno case,” GEMA general counsel Kai Welp tells me. Just as in that case, the first question – arguably the most important one — is whether an AI company needs a license to copy a work to ingest it for training purposes. That depends on whether a rights-holder has “opted-out” — signaled to technology companies that a work is not freely available for training purposes — which GEMA has done on behalf of its songwriters and publishers, although OpenAI will argue that it hasn’t done so on every online source of lyrics. Unlike the Suno case, “there’s no dispute about territoriality,” Welp says. That means EU copyright law applies. 

Like the Suno lawsuit, this case also involves issues of “output” — the results generated by giving the algorithm different prompts. In some cases, OpenAI returned all the lyrics to a song, just as Suno generated music that sounds almost exactly like the original recordings. So GEMA is also suing for violating the “making available” right — which is basically akin to public display when it involves lyrics. (There is also what Welp calls “a hallucination in the output” — a change in lyrics — so in one case GEMA is also suing for a violation of the right to create derivative works.)

There is also a more technical question that could become very important: Does the AI generate a new copy of the work in between input and output? Theoretically, a generative AI algorithm would ingest a work, analyze it and then use that analysis and others to create new works. If the algorithm simply outputs an original work, however, it’s presumably copying it in order to do so. However that is resolved, cases like this will almost certainly lead AI engineers to program algorithms to avoid returning original works as output.  

Although German copyright cases can take time to resolve, this court hearing should only take a few hours. Within a few months, the court should either reach a verdict — which either side could appeal to a higher court — or refer a legal question or questions to the European Court of Justice in Luxembourg. That could make an eventual verdict more important, since it would clarify the law, or at least certain issues involved in it, throughout the European Union. 

From a financial standpoint, this case may not be such a big deal — copyright infringement in the EU doesn’t come with the high statutory damages it does in the U.S. That’s not the point, though. “We want to put pressure on the companies to acquire licenses,” Welp says. For that, this case offers a few ways to win, as does the Suno case, although some would be more decisive than others. How cases like this shape the law will determine the negotiating leverage that both sides have, but it is gradually becoming harder to imagine how AI music could become a serious business without some kind of licensing in place. It will take some time, and perhaps some more court cases, to figure out how it will work.