Generative Artificial Intelligence under German Copyright Law (Part 1)

Abstract

This short series of two articles examines copyright issues relating to Generative Artificial Intelligence (GenAI), from acquiring training data for GenAI models, through the training process and operation of the models, to using their output. Part 1: A detailed look at the technical functioning shows that GenAI models are constructed in a way that often technically prevents or limits the possibility of copyright infringements. The text and data mining exception in § 44b of the German Copyright Act (UrhG) generally allows for legally compliant AI training. However, acquiring the training data and using the AI output can still infringe 3rd party rights, either by the GenAI developer or user. Part 2: With adequate human instruction and oversight, intellectual property rights to AI-generated results belong to the user; otherwise they are in the public domain. The article concludes with a schema for assessing the copyright status of GenAI outputs. Developers of generative AI must observe third-party copyrights when procuring and using training texts. The text and data mining exception in § 44b UrhG often allows legally compliant AI training. After training is complete, the finished GenAI model contains solely numerical values about statistical patterns between words or word parts. The original training data is never stored in the finalized GenAI model, because it is not usable for its operation. As a rule, developers even delete the texts originally used for training once the AI is fully trained. However, deleting the original data does not "cure" any earlier rights infringement. New training data is not added during ongoing operations, because for technical and liability reasons the GenAI model is "frozen" until release of the next version. These reasons suggest that once training has been completed, the developers can generally hardly commit any further copyright infringements on the original data.