Who Bears the Cost? The EU's Emerging Answer to AI Training and Copyright

In November 2025, the Munich Regional Court held, in GEMA v OpenAI, that training a language model on copyrighted material without a licence infringes German copyright law where the model memorises and verbatim-reproduces protected expression. That decision is now on appeal, with a hearing before Germany's Federal Court of Justice scheduled for 3 September 2026 — the same day the Advocate General of the Court of Justice of the European Union is due to deliver a non-binding opinion in Like Company v Google Ireland (C-250/25), the first reference to the CJEU on generative AI and copyright. Alongside these judicial developments, the European Parliament has adopted a resolution proposing a mandatory licensing scheme charging AI providers 5–7% of global turnover. This paper argues that both the blanket permissiveness of a broad text-and-data-mining exception and the blanket cost of a flat mandatory licence miss the distinction the Munich court actually drew — between learning statistical patterns (functional, and largely unobjectionable) and reproducing protected expression verbatim (the actual point of harm). A liability rule targeting verbatim output, rather than training or turnover as such, better tracks that distinction — and better protects smaller AI developers, including those based in emerging markets, from a flat cost of entry that large incumbents absorb without difficulty.