‘We Were Surprised How Obvious It Was’: The Case Against AI in the EU

In case you missed it: Suno has picked up another lawsuit against it.

Before you read any further, go to this link and listen to one or two of the songs to which GEMA licenses rights and compare them to the songs created by the generative music AI software Suno. (You may not know the songs, but you’ll get the idea either way.) They are among the works over which GEMA, the German PRO, is suing Suno. And while those examples are selected to make a point, based on significant testing of AI prompts, the similarities are remarkable.

Suno has never said whether it trained its AI software on copyrighted works, but the obvious similarities seem to suggest that it did. (Suno did not respond to a request for comment.) What are the odds that artificial intelligence would independently come up with “Mambo No. 5,” as opposed to No. 4 or No. 6, plus refer to little bits of “Monica in my life” and “Erica by my side?”

“We were surprised how obvious it was,” GEMA CEO Tobias Holzmüller tells Billboard, referring to the music Suno generated. “So we’re using the output as evidence that the original works were elements of the training data set.” That’s only part of the case: GEMA is also suing over the similarities between the AI-created songs and the originals. (While songs created entirely by AI cannot be copyrighted, they can infringe on existing works.) “If a person would claim to have written these [songs that Suno output], he would immediately be sued, and that’s what’s happening here.”

Although the RIAA is also suing Suno, as well as Udio, this is the biggest case that involves compositions, as opposed to recordings — and it could set a precedent for the European Union. (U.S. PROs would not have the same standing to sue, since they hold different rights.) It will proceed differently from the RIAA case, which involves higher damages, and of course different laws, so Holzmüller explained the case to Billboard — as well as how it could unfold and what’s at stake. “We just want our members to be compensated,” Holzmüller says, “and we want to make sure that what comes out of the model is not blatantly plagiarizing works they have written.”

When did you start thinking about bringing a case like this?

We got the idea the moment that services like Suno and Udio hit the market and we saw how easy it was to generate music and how similar some of it sounds. Then it took us about six months to prepare the case and gather the evidence.

Your legal complaint is not yet public, so can you explain what you are suing over?

The case is based on two kinds of copyright infringement. Obviously, one is the training of the AI model on the material that our members write and the processing operations when generating output. There are a ton of legal questions about that, but I think we will be able to demonstrate without any reasonable doubt that if the output songs are so similar [to original songs] it’s unlikely that the model has not been trained on them. The other side is the output. Those songs are so close to preexisting songs, that it would constitute copyright infringement.

What’s the most important legal issue on the input side?

The text and data-mining exception in the Directive [on Copyright in the Digital Single Market, from 2019]. There is some controversy over whether this exception was intended to allow the training of AI models. Assuming that it was, it allows rights holders to opt out, and we opted out our entire membership. There could also be time and territoriality issues [in terms of where and when the original works were copied].

How does this work in terms of rights and jurisdiction?

On the basis of our membership agreement, we hold rights for reproduction and communication to the public, and in particular for use for AI purposes. As far as jurisdiction, if the infringement takes place in a given territory, you can sue there — you just have to serve the complaint in the country where the infringing company is domiciled. As a U.S. company, if you’re violating copyright in the EU, you are subject to EU jurisdiction.

In the U.S., these cases can come with statutory damages, which can run to $150,000 per work infringed in cases of willful infringement. Is there an amount you’re asking for in this complaint?

We want to stake out the principle and stop this type of infringement. There could be statutory damages, but the level has to be calculated, and there are different standards to do that, at a later stage [in the case].

Our longterm goal is to establish a system where AI companies that train their models on our members’ works seek a license from us and our members can participate in the revenues that they create. We published a licensing model earlier this year and we have had conversations with other services in the market that we want to license, but as long as there are unlicensed services, it’s hard for them to compete. This is about creating a level playing field

How have other rightsholders reacted to this case?

Nothing but support, and a lot of questions about how we did it. Especially in the indie community, there’s a sense that we can only discuss sustainable licenses if we stand up against unauthorized use.

The AI-created works you posted online as examples are extremely similar to well-known songs to which you hold rights. But I assume those didn’t come up automatically. How much did you have to experiment with different prompts to get those results?

We tried different songs, and we tried the same songs a few times and it turned out that for some songs it was a similar outcome every time and for other songs the difference in output was much greater.

These results are much more similar to the original works than what the RIAA found for its lawsuits against Suno and Udio, and I assume the lawyers on those cases worked very hard. Do you think the algorithms work differently in Germany or for German compositions?

I don’t know. We were surprised ourselves. Only a person who can explain how the model works would be able to answer that.

Tell me a bit about the model license you mentioned.

We think a sustainable license has two pillars. Rightsholders should be compensated for the use of their works in training and building a model. And when an AI creates output that competes with input [original works], a license needs to ensure that original rightsholders receive a fair share of whatever value is generated.

But how would you go about attributing the revenue from AI-created works to creators? It’s hard to tell how much an AI relies on any given work when it creates a new one.

Attribution is one of the big questions. My personal view is that we may never be able to attribute the output to specific works that have been input, so distribution can only be done by proxy or by funding ways to allow the next generation of songwriters to develop in those genres. And we think PROs should be part of the picture when we talk about licensing solutions.

What’s the next step in this case procedurally?

It will take some time until the complaint is served [to Suno in the U.S.], and then the defendant will appoint an attorney in Munich, the parties will exchange briefs, and there will be an oral hearing late this year or early next year. Potentially, once there is a decision in the regional court, it could go [to the higher court, roughly equivalent to a U.S. appellate court]. It could even go to the highest civil court or, if matters of European rights are concerned, even to the European Court of Justice [in Luxembourg].

That sounds like it’s going to take a while. Are you concerned that the legal process moves so much slower than technology?

I wish we had a quicker process to clarify these legal issues, but that shouldn’t stop us. It would be very unfortunate if this race for AI would trigger a race to the bottom in terms of protection of content for training.