GEMA v. OpenAI

The claims

Copyright infringement under German and EU law: reproduction of lyrics through memorization inside the model and communication to the public through chatbot outputs; claims for injunction, damages and information about use and revenues.

What has happened

GEMA sued OpenAI in November 2024 over the lyrics of nine well-known German songs, including “Atemlos”, “Männer”, “Über den Wolken”, “Bochum” and “36 Grad”. On November 11, 2025 the Munich Regional Court I largely upheld the claims. It held that memorization of lyrics in model parameters is itself a reproduction under Section 16 of the German Copyright Act, that the EU text-and-data-mining exception covers only the initial analytical phase of training and not copies retained in the model, and that chatbot outputs reproducing the lyrics are communication to the public attributable to OpenAI rather than to prompting users. The court ordered OpenAI to cease, pay damages, and disclose the scope of use and revenue. OpenAI announced it would appeal; commentary through mid-2026 places the case on track for the Munich Higher Regional Court, and no appellate ruling has been reported.

Key developments

• 2024-11 — GEMA files what it describes as the first collecting-society suit against OpenAI, over nine German song lyrics.
• 2025-11-11 — Munich Regional Court I rules for GEMA: memorization is reproduction, the TDM exception does not cover it, and OpenAI is directly liable for outputs; injunction, damages and disclosure ordered.
• 2025-11 — OpenAI states it will appeal; the judgment is not final while review by the Munich Higher Regional Court remains open.

Why it matters for training data

This is the first European judgment holding an AI developer directly liable for training-derived outputs, and its reasoning is structural: memorization equals reproduction, and the TDM exception stops at the analysis phase. If that framing survives appeal, EU operations cannot rely on TDM opt-out mechanics to cover models that retain protected works verbatim. For suppliers, it converts memorization from a model-quality bug into a per-work infringement. For buyers, EU exposure turns on what a model can emit, not only on how the training set was gathered — a different diligence test than US fair use, and one a collecting society can litigate at repertoire scale on members’ behalf.

Sources

• EUIPO case-law note on the judgment
• Bird & Bird analysis of the ruling
• CMS Germany legal update
• William Fry on appeal posture and EU direction
• Norton Rose Fulbright commentary

Deeper analysis

• Training-data copyright lawsuits: the landscape

← All case files Weekly lawsuit news