German lawsuit tests copyright limits for AI music training

According to the ‌verdict, Suno will have to pay damages that have yet to be quantified. Suno has disagreed with the ruling and said it would evaluate all available options, including an appeal.

"This is a verdict of global significance," Gema CEO Tobias Holzmueller told Reuters .

To view this video please enable JavaScript, and consider upgrading to a web browser that supports HTML5 video

"Our goal is not to clear Suno from the face of the earth, but obviously to get into licensing negotiations on an eye-to-eye level, which was not possible with Suno so far," explains GEMA CEO Tobias Holzmüller, who describes the company as "by far the largest and commercially most successful of all AI music models." He believes Suno pursues a much more aggressive business strategy than its competitors.

The outcome of this case is also linked to another central question: How should creators be compensated when their music is used to train generative AI systems?

OpenAI was ordered to cease and desist from reproducing the works in the model, to disclose information and pay damages. But as the company pursues an appeal, the ruling remains non-final.

In contrast to that decision, the current lawsuit does not focus on song lyrics, but on entire music pieces.

One of the aspects that makes this lawsuit important is that it has not been initiated by a private company or record label, but rather by a music collecting society, notes Martin Senftleben, professor of intellectual property law at the Amsterdam Law School.

Such initiatives "are particularly important because, by definition, you could say that they're more inclusive," Senftleben tells DW. "You can become a GEMA member even if you're not a very well-known high-profile artist or composer. And that means there's a chance of bringing remuneration for AI users to a much broader group of rights holders."

GEMA represents more than 95,000 composers, songwriters and music publishers in Germany, along with over 2 million rights holders worldwide.

Suno's music creation platform generates songs through simple text prompts. Users can describe the mood, style or genre they desire and the system synthesizes a structured track, complete with lyrics, vocals and instrumentation, in under a minute.

During the court proceedings, GEMA demonstrated this by prompting Suno AI to create tracks with the original lyrics of different popular songs in its repertoire, including "Forever Young" by Alphaville, "Mambo No. 5" by Lou Bega, "Daddy Cool" by Boney M., "Cheri Cheri Lady" by Modern Talking and "Atemlos durch die Nacht," written by Kristina Bach and popularized by Helene Fischer .

Along with the lyrics, the plaintiffs prompted the desired musical style and the title of the work, without specifying any melody, rhythm, harmony or arrangement.

Yet the resulting AI songs are very similar to the original works.

Suno admits that the original songs at the center of GEMA's lawsuit are among the numerous pieces that were used to train the AI model.

AI companies such as Suno claim that the ingestion of data for AI training can be done freely on the basis of what is known as "fair use" in the US and the "text and data mining exception" rule in the EU.

Anyone who wants to become a good musician needs to listen to and learn from existing music.

Similarly, the AI industry claims this analogy applies to how their models are trained, by absorbing countless creative works in order to generate new ones.

But from the rights holders' perspective, AI models are rather copying and commercially exploiting their works.

This argument that an AI learns from music and is inspired by it like a human being is "completely ridiculous," notes Tobias Holzmüller. "These are concepts that come from the human brain and have been designed for the human brain, and not for a machine that is able to digest millions of sound files in split seconds. I mean, just the amount of sheer data, they say it's data; for us, it's music, that these machines have swallowed up and then digested and are now spitting out, this is completely incomparable to what's happening in a human brain.

"To humanize these machines by using these concepts is not logical, and it's certainly not moral."

Across Suno's platform globally, users collectively generate roughly 7 million songs per day. Around 75,000 of them end up on streaming services daily, adds Holzmüller. "This is definitely where it's not sustainable anymore, but really toxic for the music ecosystem."

"If you want to create a sustainable model where you accept that AI music is here to stay, but at the same time, that it shouldn't push human-created music out of the market, the only way to tackle this is to put a price tag on the output," says the CEO of GEMA. "Because otherwise you will have free-of-charge AI sounds that will outcompete human music in a lot of circumstances."

For example, music in movie scores or social media has until recently allowed various songwriters to make a living. For Holzmüller, if licences lead to an increase in costs in AI-created sounds to a point that they are as expensive as human-created songs, it could provide a level playing field. "Then, it's my firm belief, people will continue to use human music, because it has this additional element."