OpenAI loses song lyrics copyright case in German court

Large language models like ChatGPTinfringe on German authors’ rights laws if they use song lyrics in their responses without having paid license fees for them, a Munich court ruled on Tuesday.

Judge Elke Schwager at the Munich District Court I said that OpenAI, the US company that owns ChatGPT, would be charged damages for the unauthorized usage. She did not specify a sum.

The claimant and a German journalists’ trade union both claimed the case could have far-reaching implications for AI or large langugage models and intellectual property and copyright laws.

The verdict can be appealed.

“We do not agree with the verdict and are examining further steps,” OpenAI said in response, adding that it respected intellectual property rights and was in negotiations with relevant organisations around the world.

What was the case about?

The German association GEMA that seeks to defend authors’ rights brought the lawsuit.

Authors’ rights law (or Urheberrecht in German) is separate from and not to be confused with the more commonly understood Anglo-American copyright law. It places more emphasis on the individual artist or author and considers the rights non-transferable, rather than the property of the owner of the content (like a publisher or record label).

GEMA used nine specific songs as examples for the purposes of the case, including titles like “Männer (Men)” by Herbert Grönemeyer, “In der Weihnachtsbäckerei (In the Christmas bakery)” by Reinhard Mey, and “Atemlos (Breathless)” originally by Kristina Bach and popularised more recently by Helene Fischer.

Although this case only concerned German law and usage, one of GEMA’s lawyers claimed that Tuesday’s ruling would prove groundbreaking for Europe as a whole, given that the applicable rules were “harmonised.” He said he anticipated negotiations with companies like OpenAI on suitable licensing fees.

“We are of course extremely pleased that the chamber has ruled so clearly,” GEMA lawyer Kai Welp told journalists. “The goal is not to remove anything from the market, but rather to receive appropriate compensation.”

GEMA made international headlines about a decade ago with its restrictive approach to German music videos on YouTube, though a deal was ultimately reached to permit their publication on the platform.

Judge baffled by oversight from ‘highly intelligent’ defendants

Judge Schwager said while issuing her ruling that she was astonished that OpenAI had not taken heed of what she called a clear legal situation.

“We have highly intelligent defendants who have managed to create the most modern of technologies,” Schwager said.

Anyone who created something and used outside content in so doing had to pay for that content or otherwise obtain permission, she said, finding that the current usage amounted to unlicensed distribution and reproduction.

“Authors’ rights are protected intellectual property,” Klager said. “And so it’s clear that this is out of order.”

On what grounds did OpenAI dispute the allegations?

Neither side disputed during the trial that the songs’ lyrics had been used to “train” the fourth iteration of ChatGPT.

What was at issue was whether or not the lyrics had been actively stored in the large language model’s database for future use.

OpenAI argued that ChatGPT did not store or copy specific training data, but rather reflected in its parameters what it had learned in its entire training dataset. It also argued that “outputs” from ChatGPT answering user questions were generated only in response to user prompts and so if anybody were responsible for their generation it would be users more than OpenAI.

The court found that the coincidental generation of text that happened to match the lyrics either exactly or in large part was not plausible.

“Given the complexity and length of the song text, coincidence can be ruled out as the cause of the reproduction of the song lyrics,” the court wrote in a press release.

German journalists’ union hints at wider implications

Several media organizations had also questioned the legality of large language models’ training processes, with journalism among the sources used.

The chairman of one leading journalists’ trade union, Mika Beuster of the DJV, called Tuesday’s ruling “a partial victory for authors’ rights.”

“The training of AI models is intellectual property theft,” Beuster said, arguing that journalists seeking compensation from companies like OpenAI would now have an improved legal position.