Abstract Foundation models are essential building blocks for developing AI technologies. When these models are released openly, they enable broad access, foster innovation, and create substantial public value. Training such models typically involves reproducing vast quantities of copyright-protected material, raising legal questions under EU copyright law. This paper argues that the EU’s text and data mining exception for scientific research, set out in Art. 3 of the Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market, can provide a legal basis for training open foundation models. Scientific collaboration and openness are inherently linked. This connection is reflected in the Directive and implicitly taken up by the case law. This paper examines the key requirements of the exception and interprets them in light of the LAION decision by the Hamburg Regional Court. It proposes a functional and permissive reading of ‘scientific research’ and ‘research organization’ that reflects the collaborative and infrastructural nature of open model development. This framework not only enables open foundation model development under existing EU law but also allows for downstream licensing markets to emerge at the commercial fine-tuning stage, which in turn can subsidize the initial, open development.
Arne Radeisen (Wed,) studied this question.