MEP Axel Voss’s own-initiative report on copyright and generative AI will be put to a vote in the JURI Committee on 28 January. Although framed as a response to legitimate questions around AI and copyright, the proposal suffers from serious shortcomings, most notably its disregard for the existing legal framework. Despite some more constructive compromise amendments, the report as a whole creates more uncertainty than it resolves. We therefore recommend its rejection.
Applicability of the TDM exceptions to AI training
MEP Voss’s original draft asserted that the training of generative AI systems falls outside the existing copyright framework. This interpretation directly contradicted the approach taken in the AI Act and introduced unnecessary legal uncertainty. That uncertainty was felt most acutely by actors seeking to comply with the law, such as researchers and developers of public-interest AI systems.
The compromise text retreats from this position, instead referring to alleged “ambiguities” in the application of the text and data mining (TDM) exceptions to the training of generative AI models. While this represents a clear improvement over the original draft, it remains problematic. Earlier amendments went so far as to delete the relevant sections entirely—an outcome that, in our view, would have been preferable.
Transparency and presumption of use
Another major area of concern relates to transparency obligations and the introduction of an “irrebuttable” presumption of use where such obligations are not met. While the current text still contains a presumption of use, this has now been made rebuttable, which is a more reasonable demand. However, the latest compromise continues to call for “full transparency.” The report considers this essential to restoring the “absolute right of creators”, displaying a distorted view of the nature of exclusive rights.
The report calls for full transparency both for training-time and inference-time access and scraping activities. More specifically the report calls for “an itemised list identifying each copyright-protected content used for training” and the maintenance of a “detailed record of their crawling activities.”
COMMUNIA supports transparency standards that effectively complement the opt-out framework for commercial AI training, but we consider the proposed changes to be unimplementable. Large-scale models are trained on vast, heterogeneous datasets, making the creation of a complete, itemised list of every individual copyright-protected work practically impossible. Requiring detailed record-keeping for dynamic processes, such as inference and retrieval-augmented generation, does not provide a clear public benefit and would raise privacy concerns, while at the same time generating significant technical complexity, legal uncertainty, and compliance costs.
Users’ rights concerns
We welcome the introduction of important safeguards for lawful uses. The revised text urges the Commission to explore measures to counter copyright infringements arising from generative AI outputs, provided that such measures do not prevent the generation of non-infringing content. This explicitly includes content covered by exceptions and limitations such as private use, quotation, criticism, review, caricature, parody and pastiche, as well as incidental inclusion.
Another improvement concerns the framing of the proposed assessment on the existing EU acquis around the needs of all stakeholders, including researchers, universities, libraries and cultural organisations. This broader perspective helps move this discussion beyond the narrow interests of rightholders and AI companies.
Creating a new exclusive right for press and news media
The most problematic addition to the text envisages significant changes to the use of press publishers’ and news media content for the training of generative AI systems, as well as for downstream uses such as inferencing and retrieval-augmented generation. It calls on the Commission to explore extending the press publishers’ right and the broadcasting right in order to grant press and news media rightsholders “full control” over the use of their content for such uses. According to the proposal, these uses would require “explicit consent,” effectively removing press and news media content from the scope of the TDM exceptions.
We consider the proposed language unacceptable and urge the Committee to draw a clear red line here. There is no justification for creating new exclusive rights, elevating press publishers’ and media content to a privileged status above other copyrighted works.
The JURI Committee should recognise that the TDM exceptions are a cornerstone of European research and innovation. They are clearly applicable to AI training and play a vital role in enabling the development of AI systems, including those serving the public interest. On Wednesday, the Committee should thus reject the report, or, at the very least, correct its most glaring mistakes.
