The ruling restricts accountability to memorization and reproduction of specific works, not a broad prohibition on using copyrighted materials for AI training.

The decision clarifies that training AI models on copyrighted material is not inherently unlawful, but memorization and reproduction of identifiable protected works can be actionable.

OpenAI was found infringing German copyright law by memorizing and reproducing lyrics from nine songs in GPT-4 and GPT-4o, according to the Munich I Regional Court in GEMA v OpenAI.

The author calls for balancing copyright protection with AI’s potential benefits to innovation, education, and research, and invites consideration of future frameworks that serve both creators and progress.

Storing copyrighted material in numerical form does not remove it from protection, and outputs that reproduce lyrics are not exempt by text and data mining or the scientific research exception given OpenAI’s commercial nature.

OpenAI was held directly liable for infringing outputs and ordered to stop reproductions, account for revenues, and pay damages under the judgment terms.

The court found that the models memorized lyrics during training and that prompts were not merely incidental, resulting in substantial reproductions of protected works.

The Philippines lacks an AI-specific copyright framework, and Philippine law, including fair use, may address similar issues, though the German decision is not binding there.

The court distinguished three AI development phases—training data collection, model training, and output generation—and ruled that memorization occurred during training, not merely as a by-product of prompting.