GEMA has become the first collecting society to sue an AI company for allegedly infringing copyright through the training of a generative AI model. The German song rights society is suing OpenAI over allegations it used lyrics written by GEMA members without getting permission.
Many AI companies continue to deny they need to get permission to make use of copyright protected works in their training datasets. This means there are “fundamental legal questions that we absolutely must clarify” relating to copyright and AI, says GEMA’s General Counsel Kai Welp.
This lawsuit, the society hopes, will provide legal clarity and allow the development of an AI licensing model. “GEMA wants to establish a licence model”, it insists, that would allow AI companies to make use of its members’ works legitimately, and for songwriters and publishers to be remunerated.
That will be easier to achieve, it adds, once this lawsuit has refuted the “contention that training with and subsequent use of the generated content is free of charge and possible without the rightsholders’ authorisation” that many AI platforms currently base their businesses on.
The ultimate aim of the litigation, the society says, is to “guarantee fair participation for authors in the use of their copyright-protected works by generative AI”, while also hopefully initiating “a public discussion on copyright and AI”.
Both record companies and music publishers have already sued AI companies for copyright infringement, with the labels going after Suno and Udio, and the publishers Anthropic.
This case is interesting not only because it’s the first AI lawsuit involving a collecting society, but also because it’s the first music AI case filed with a European court. Up until now the music industry’s lawsuits targeting AI companies have been filed in the US courts, against US companies, while GEMA has filed its lawsuit against US-based OpenAI in the courts in Munich.
Although music companies in both the US and Europe believe AI companies must secure permission before using existing recordings or songs to train their models, the legal arguments are different in the US and the European Union.
In Europe, to circumvent the need to get permission when copying existing music onto their servers, AI companies have to rely on a specific copyright exception. There is a text and data mining exception in EU copyright law, but copyright owners can opt out of it by ‘reserving their rights’.
In the US, AI companies are relying on the slightly more ambiguous concept of ‘fair use’. Whether or not AI training constitutes fair use is currently being tested in multiple lawsuits in the US.
GEMA explains these legal differences in an FAQ that accompanies the lawsuit. “The legal situation in the USA differs from that in Europe”, it says.
While in Europe the use of copyright protected works for AI training is “expressly and clearly regulated”, meaning AI companies “may not train AI using protected works for which a reservation of rights has been declared”, the copyright obligations of AI companies are “highly controversial in the USA”.
That’s because of the fair use principle, it adds, under which “uses that do not significantly affect the rightsholders are permitted without authorisation and without remuneration”.
While the fair use argument is debated in American courts, GEMA adds, “this potential US loophole also appears to be motivating primarily US-based providers to refuse to pay the licence fee owed in Europe, despite the divergent legal situation”.
Although this is the first music and AI case in Europe, there has already been a copyright and AI case in the German courts. A photographer called Kneschke sued LAION, a German non-profit that maintains the LAION 5B training dataset, an open large-scale dataset for training next generation image-text models.
The photographer said LAINO had infringed his copyright by reproducing his photo when it was ingested into the dataset. LAINO said it was covered by a text and data mining exception under European law, specifically the separate text and data mining exception for “research organisations and cultural heritage institutions”, which does not provide an opt out for copyright owners.
A court in Hamburg agreed that that exception applied to LAINO as a non-profit, even though commercial entities like StabilityAI use its dataset. The case is now being appealed.