The doctrine
Plaintiffs increasingly argue the model's outputs reproduce protected expression — independent of how the model was trained. The theory expanded sharply in 2025.
Output-infringement claims sidestep the training-data fair-use debate by focusing on what the model actually produces. Where a model can be prompted to regurgitate near-verbatim copyrighted text, lyrics, or images — or where it produces summaries dense enough to capture protected expression — courts have begun to find that those outputs are themselves potentially infringing copies.
The Authors Guild v. OpenAI ruling of October 2025 was the first U.S. decision to hold that an LLM's plot-level summaries of plaintiffs' novels may infringe. The Munich Regional Court reached an analogous conclusion in GEMA v. OpenAI, holding ChatGPT's lyrics output reproduced protected expression and that the EU text-and-data-mining exception does not cover memorized output. Retrieval-augmented generation systems like Perplexity face direct output-based exposure: their outputs are explicitly grounded in third-party content surfaced at query time.
The remedies question is fluid. Injunctive relief targeting outputs is technically achievable through filters; statutory damages calculated per output, multiplied across millions of users, present an enormous theoretical exposure that no court has yet fully addressed.
Leading cases
First U.S. ruling that LLM plot summaries may infringe.
First European court to find LLM lyrical output infringing.
Watermark reproduction in outputs grounded trademark / passing-off liability.
Output regurgitation evidence central to discovery posture.
Key holdings
- Outputs can themselves infringe. Even without verbatim reproduction, dense plot summaries may capture protected expression.
- RAG systems most exposed. Retrieval-augmented generation grounds outputs in third-party content, narrowing transformativeness arguments.
- Filters now legally relevant. Output filtering capabilities may bear on injunctive remedies and willfulness.