AI Copyright: Training-Data Lawsuits and Licences — 2026-09-02

Top developments

Sony and Warner Music Sue Anthropic Over "Blatant Theft" of Songs

On August 31, 2026, the music publishing branches of Sony Music and Warner Music Group filed a lawsuit against Anthropic in California federal court. The plaintiffs allege that Anthropic misused copyrighted song lyrics from artists including The Beatles, Taylor Swift, and Michael Jackson to train its Claude models. The complaint describes the actions as "blatant theft" and seeks damages for the unauthorized ingestion of "tens of thousands" of songs, marking a significant escalation in the legal battle between major music labels and generative AI developers.

German Publishers Sue OpenAI Over "NEINhorn" Children's Book

In a notable European development, the German publisher Carlsen Verlag, along with author Marc-Uwe Kling, has initiated legal proceedings against OpenAI. The lawsuit alleges that ChatGPT has memorized and reproduced content from the popular children's book series "Das NEINhorn," raising questions about whether the AI model's training data included these works without permission. This case highlights the growing trend of individual publishers challenging large language models for specific content infringement rather than broad class-action claims.

Japanese Government Proposes Mandatory Disclosure of AI Training Data

The Japanese government is preparing new intellectual property protection principles that would require AI companies to disclose their training data collection methods and provide information to rights holders. This policy shift aims to increase transparency in how generative AI models are trained, addressing concerns from creators about unauthorized use of their works. While penalties are being debated, the move signals a regulatory tightening that could impact international AI firms operating in Japan.

Local view

Japan: Major Japanese media outlets, including Nikkei and Yomiuri Shimbun, have extensively covered the Sony/Warner vs. Anthropic lawsuit, noting it as a significant precedent for music copyright in the AI era. Additionally, local coverage focuses on the Japanese government's upcoming guidelines which will require AI operators to disclose training data sources, a move seen as balancing innovation with creator rights.

Germany: Legal analysts in Germany are closely watching the Federal Court of Justice (BGH) hearing scheduled for September 3, 2026. The case, involving the creation of datasets for AI training, is expected to set a critical precedent for whether copying copyrighted works solely for dataset creation constitutes copyright infringement under German law, potentially influencing future licensing deals across the EU.

Context & numbers

The legal landscape for AI training data is becoming increasingly defined by high-stakes settlements and licensing benchmarks. The recent settlement in the Bartz class action against Anthropic, approved in mid-2026, established a de facto market reference price of approximately $3,000 per work across more than 482,000 books. This figure is now being used as a benchmark for unlicensed ingestion in subsequent negotiations and lawsuits.

Furthermore, the volume of AI content licensing deals continues to rise, with projections suggesting 36 major deals in 2026 compared to just 12 in 2023. The largest disclosed deal remains OpenAI's reported $250 million five-year agreement with News Corp, though newer agreements are increasingly focusing on specific niche datasets and music catalogs.

On the radar

• September 3, 2026: The German Federal Court of Justice (BGH) will hold an oral hearing in case I ZR 281/25, deciding on the legality of duplicating copyrighted works for the purpose of creating AI training datasets.
• UK Policy Review: Following the March 2026 report that ruled out a broad text-and-data-mining exception, UK policymakers are expected to release further guidance on voluntary licensing frameworks later this year.