European AI copyright litigation has a new case, and its legal architecture is worth studying carefully.
What’s been filed
Penguin Random House filed a copyright infringement complaint against OpenAI in the Munich Regional Court, targeting Ingo Siegner’s children’s book catalog, per the Mishcon de Reya AI Litigation Tracker. The filing date is approximately late March 2026, the tracker acknowledges uncertainty on the precise date. This isn’t breaking news of a brand-new filing; it’s a case that’s been advancing through European courts and is now drawing scrutiny from compliance teams tracking cross-jurisdictional copyright exposure.
The dual-liability theory
What makes this case structurally notable is the complaint’s two-point liability architecture. Most AI copyright litigation focuses on one exposure: either the training stage (ingesting copyrighted works without license) or the output stage (generating text that resembles protected expression). This complaint, as characterized by the Mishcon de Reya tracker, alleges both simultaneously.
That’s a materially different legal claim. Training liability and output liability have different evidentiary requirements, different defenses, and potentially different damage calculations. A complaint structured to advance both requires OpenAI to defend on two fronts, not one. Whether the Munich Regional Court accepts that theory as legally cognizable is the question compliance teams should watch, not the existence of the claim, but how the court frames its jurisdiction over each stage.
AI Copyright Liability Framework
The dual-liability framing tracks with arguments being developed in U.S. copyright cases over the same period, though the European copyright framework operates differently. The EU’s text and data mining exemptions under the Digital Single Market Directive provide some defense space at the training stage, but those exemptions are narrower than what U.S. fair use doctrine offers, and output liability doesn’t fall under them at all.
Jurisdictional significance
Germany’s courts have historically been active in intellectual property enforcement, and the Munich Regional Court has handled significant technology IP disputes. A finding, even a procedural one, that output liability can be asserted alongside training liability in German courts would have implications beyond this case. LLM developers distributing products in the EU would need to assess exposure at both pipeline stages, not just the ingestion stage where most legal attention has focused.
That said: this is a pending matter. The complaint alleges. No court has found. Don’t characterize this as settled law or an adverse ruling.
Unanswered Questions
- Does the Munich Regional Court accept dual-liability (training + output) as simultaneously cognizable?
- How do EU TDM exemptions apply when a model trained on licensed vs. unlicensed data generates allegedly infringing output?
- If output liability is established in German courts, does this create an audit trail obligation for inference-stage logging?
What to watch
First hearings and the court’s initial framing of the dual-liability theory. Watch also for whether other European publishers join or file parallel suits, the cross-jurisdictional copyright convergence pattern suggests this case won’t remain isolated.
TJS synthesis
U.S.-focused compliance teams tend to frame AI copyright risk as a training-stage problem, addressed by licensing strategies or fair use arguments. The Munich complaint suggests European courts may be asked to evaluate a connected chain where training liability and output liability reinforce each other. If that theory gains traction in a German court, the legal risk calculus for EU market access changes: output monitoring becomes a compliance function, not just a product quality question.