Verletzungen des Urheber- und Persönlichkeitsrechts durch die Ausgaben großer Sprachmodelle

This interdisciplinary paper examines the legal risks involved in generating and using textual
outputs using large language models (LLMs). The authors assume the perspective of an end
user who attempts to prompt the LLM to generate copyright-protected
content. An experiment
designed to provoke such model outputs resulted in five different types of output: 1)
reproductions without context, 2) reproductions in a broader context, e. g. quotes, 3) hallucinations,
4) summaries, and 5) mixes of reproductions and hallucinations. The paper evaluates
each category first from a copyright perspective, discussing the issues of reproduction,
adaptation, and applicable exceptions and limitations. Second, it points to issues regarding
personality rights. While the end user can generate and use model outputs of all types for
purely private purposes, their use in other contexts (e. g. on social media) may come with
risks with regard to both copyright and personality rights.