Gen AI and copyright in Europe – Munich Regional Court seeks to determine the rules of the playing field

In the recent Getty Images v Stability AI decision, the UK High Court rejected Getty’s claim for secondary infringement of UK copyright on the basis that (i) an “infringing copy”, for the purposes of secondary infringement, must be “a copy” and (ii) Stability’s AI model, Stable Diffusion, had never contained or stored copies of any of Getty’s copyright works and was not therefore an “infringing copy” (see blog). However, perhaps highlighting the lack of technical and international consensus, on 11 November the Munich Regional Court, in its judgment in GEMA v OpenAI, found that a model may contain copies of a work where that work is “memorized” by the model. Of course, the models in question and the arguments before the courts were different, and neither of these decisions is final. But, they serve as an important reminder that the rules of the playing field are still being determined. 

The GEMA decision also warns against widespread reliance by commercial AI developers on exceptions to copyright infringement for text and data mining (TDM), a message supported by a recent study report on Generative AI and Copyright commissioned by the European Parliament’s Committee on Legal Affairs (the Study). 

In this blog, we focus on the German court’s findings in GEMA and how these feed into wider conversations around (i) memorisation and (ii) the scope of the EU TDM exception set out in Article 4 of the 2019 Directive on Copyright in the Digital Single Market (TDM Exception). 

GEMA v OpenAI

GEMA is a German music collecting society. Towards the end of last year, it issued a claim against OpenAI alleging that ChatGPT infringed copyright in a number of its members’ song lyrics by reproducing those lyrics in the model itself (through memorisation) and at the output stage in response to simple user prompts. 

OpenAI naturally disagreed, arguing that the lyrics were not “memorised” and “stored” in the model and, in any event, copying was permitted under the German implementation of the TDM Exception. It also denied responsibility for any outputs, arguing instead that the individual users were responsible. 

The Munich Regional Court largely sided with GEMA, finding that:

• The song lyrics were “memorised” and “reproducibly contained” in ChatGPT, infringing copyright.
• The TDM Exception was not available to OpenAI other than in respect of the initial preparatory work in compiling the training data. The memorisation and reproduction of lyrics in the model weights went beyond text and data mining and interfered with the rightsholders’ legitimate interests.
• Similarly to the UK High Court in Getty Images (albeit that was in a trade mark infringement context), the Munich Court rejected the argument that individual users should be responsible for the output of their prompts, finding that, where simple prompts were used, the outputs in question were significantly influenced by the architecture of the model and the training data chosen and used by OpenAI.

The wider conversation

It’s fair to say that tension has increased within the last year, both in legal and technical circles, over both: (i) whether AI models should be treated as containing “copies” of their training data for copyright law purposes; and (ii) whether the TDM Exception should apply to AI training. 

As noted above, the courts in Getty and GEMA reached different conclusions on whether the models in question stored copies of their training data. However, caution should be exercised in drawing any sweeping conclusions from either of these decisions. They are both first instance decisions (which may be appealed), the models in question were different, the issues before the court were different and the cases were pleaded differently (with Getty, for example, not seeking to argue that Stable Diffusion is a copy of, or contains copies of, Getty’s copyright works – despite the court’s findings on this point).

As for whether the TDM Exception is applicable to AI training, the Munich Court’s decision appears to align with the Study’s conclusions, i.e. that the TDM Exception is aimed at the extraction of patterns from data for analytical purposes, while the training and use of generative AI is concerned with the synthetic reassembly of data to produce new content. If followed, this suggests that the TDM Exception might not be the solution AI developers were hoping for. It will also be interesting to see what, if any, impact these developments have on the outcome of the UK’s consultation on copyright and AI and its proposals to extend the UK copyright exception for TDM. 

We will continue to watch, and report on, how this landscape develops over the coming months, with further guidance expected from both national courts and the CJEU in cases that are testing similar questions, including GEMA v Suno AI (also in Germany), SNE v Meta (France) and Like Company v Google Ireland (CJEU).