Axel Voss’ Good Intention – The European Parliament adopts report to recalibrate EU copyright for Generative AI

On 10 March the European Parliament adopted a report on copyright and generative artificial intelligence, tabled by the Committee on Legal Affairs. The report urges the European Commission to initiate legislation tackling the ‘opportunities and challenges’ of generative artificial intelligence (GenAI) and to make Europe a ‘world leader in AI’ by enabling 'technological development [that] supports sustainable economic growth, competitiveness and innovation’. The preambulatory part of the report sets the tone for its programmatic orientation: the fundamental right to (intellectual) property must be respected in the development and deployment of GenAI. Accordingly, the proposed way forward is to supplement the existing rules with legal and non-legal initiatives that ensure that copyrighted material can be efficiently licensed.

The report is a tour de force through the shortcomings of existing copyright rules, most notably art. 4 of the 2019 CDSM Directive, but a reference to art. 17 CDSM Directive also makes an appearance. It seeks to facilitate and centralize licensing solutions and infrastructure for GenAI developers and to strengthen the position for rightholders through elevated transparency obligations and procedural presumptions. Whereas the general approach as well as most of the proposed measures are laudable, the core of the proposal – a voluntary licensing approach for GenAI – requires reflection.

The substantive proposals are prefaced by a commitment to free use of training material for scientific research and educational purposes or non-commercial innovation. Here, the report’s intention not to suffocate innovation becomes apparent while it also underlines that subsequent commercialization requires adequate authorization – and remuneration.

Exemptions for press and news media

The broad licensing sweep (see below) would not apply to the press and news media sector, due to its fundamental importance in safeguarding democracy and democratic structures. This sectoral exemption assumes that aggregated information enabled by GenAI will have detrimental effects on diverse service offerings in this sector. Therefore, producers of press and news media content (the report refers to press publishers, journalists and news editors) should have stronger and more exclusive control over the use of their material. Interestingly, these sectors have already developed more refined licensing models for GenAI, as evidenced by a recent EUIPO study.

The document refers to competing products, including AI applications that, beyond mere training, use inferencing and retrieval-augmented generation (RAG) to challenge the business models of original content creators by, for example diverting traffic from their offers. For such uses, express consent should be required, which necessitates the extension of existing related rights to cover such uses. For that purpose, and to ensure compliance with the principles of media pluralism and diversity of information, on which the European Media Freedom Act is based, enforcement measures must be strengthened and comply with provisions of the Digital Markets Act on self-preferencing practices (art. 6(5)).

Sectoral collective licensing for GenAI training

Key to realizing that authors receive fair remuneration, the report calls upon the Commission to facilitate voluntary sectoral licensing solutions. Mandatory collective licensing or de facto statutory licenses through remunerated exceptions do not feature in the text. Flanking the licensing approach are measures that enable effective opt-outs under art. 4(3) CDSM Directive. To create legal certainty, the number of machine-readable opt-out standards should be limited and the EUIPO should function as ‘trusted intermediary’. The institutionalisation of licensing solutions and supporting mechanisms would also enable reactive approaches to technological developments, however, also bears the risk of an over-bureaucratization of copyright management. Still, charging the EUIPO with developing and managing standards, and with facilitating licensing solutions has the potential to establish European – and possible global – standards.

Enhanced Transparency

Essential for a functioning licensing framework for protected material is access to information of training data of (largely) GenAI model providers. The report suggests tasking the ‘trusted intermediary’, together with stakeholders, with refining existing templates to record training data. For that purpose, the General-Purpose AI Code of Practice and its related templates and guidelines should be understood as ‘living documents’ which will be continuously refined and updated under the auspices of the AI Office. Furthermore, for continuous training and the use of RAG, crawlers should keep detailed records of crawling activities, also with a view to protect trade secrets and confidential business information. Transparency can further be safeguarded through the continuous development and research on mechanisms such as digital watermarking and other technological standards.

Extraterritorial application of copyright

In line with the (controversial) approach of art. 53(1)(c) of the Artificial Intelligence Act, which requires GenAI model providers to put in place policies to comply with EU copyright law, especially with reservations declared pursuant to art. 4(3) CDSM Directive, the report supports an understanding of the territoriality principle that requires GenAI systems and models to comply with EU copyright rules when put on the market within the Union. The jurisdiction in which systems or models are developed or trained would not be relevant if the respective model or system is made available to users in the EU. This would avoid jurisdictional forum shopping for AI development and training and ensure that European creators would benefit from the proposed licensing approach irrespective of the place of production of a GenAI model or system that uses their creations for training purposes.

Efficient voluntary licensing

The report does not suggest a one-size-fits-all licensing approach for GenAI. It recognizes that different systems and models create different vulnerabilities. As one example, the text refers to GenAI integrated in search engines or other services whose outputs unfairly compete with original training data sources. Voluntary licensing would therefore have to have regard to the specificities of the different content sectors and should enable rightholders to effectively bargain for fair remuneration for uses  of protected material by GenAI producers and deployers. Coherent and functioning licensing frameworks should also apply to past uses, developed prior to the anticipated legislative intervention. However, past infringement that occurred in a temporal regulation gap should not be compensated with global flat-rate models, but instead be compensated through fairly negotiated proportionate licensing solutions.

Presumption of use

Supporting the AI Act’s transparency obligations and the proposed enhanced transparency obligation, which would require a much more granular listing of training data, the report introduces a procedural element to ensure rightholders can properly enforce their rights. Similar to the presumptions for non-compliance (art. 3) and causality (art. 4) contained in the failed AI Liability Directive, the text suggests introducing a rebuttable presumption that any AI model placed on the market in the EU was trained on works and other protectable subject matter if the relevant (proposed) transparency requirements are not fulfilled. This would, in any case, require that model developers track training material; either to comply with the transparency requirements ab initio, or to rebut the presumption of use during training. Borrowing from another instrument, the report suggests that successful claimants can recoup reasonable and proportional legal expenses from providers of AI models or systems. This is very similar to the cost-sharing mechanisms under art. 21(5) of the Digital Services Act.

AI outputs and Deepfakes

Finally, Parliament urges the Commission to consider measures to address copyright infringements through GenAI and to address the problem of deepfakes. Presumably triggered through rulings in Germany  (Gema v OpenAI) and domestic initiatives to strengthen the position of individuals in relation to their physical likeness, these initiatives round off the suite of measures the Commission is asked to explore. The output side of GenAI is only addressed in three meagre paragraphs, which also include a general requirement to label purely AI-generated content that goes beyond the watermarking requirements of the AI Act.

An ambitious Agenda, but a feasible model?

The crux of the report of notorious Rapporteur Axel Voss is its veer towards licensing solutions for GenAI, which would make remunerated uses the (more attractive) norm without clearly breaking with the systematic approach of art. 4 CDSM Directive. The tendency is clear: the impossibility to foresee technological developments in the field of GenAI now requires readjustment of text-and-data mining rules and supplementary initiatives for AI outputs to rebalance the economic disruptions caused by the de facto free use of protected material for GenAI training purposes.

The proposals aim at creating efficiency for licensing solutions through a mixture of legal and infrastructural interventions. However, it is unclear whether the suggested measures can address the practical difficulties of an opt-out system. Providing more transparency and facilitating declarations of and respect for opt-outs will certainly help larger rightholders, who will be able to navigate these transactional systems post-opt-out declaration. Whether the value gap in the GenAI ecosystem can be closed, which is how the reference to the ‘rationale of Article 17 of the DSM Directive’ in the Explanatory Statement must be understood, remains uncertain. Statutory remuneration rights, as proposed elsewhere, would create more legal certainty, but it has also not been proven that they would benefit the creative sector. Full transactional autonomy, as proposed by Axel Voss’ drafting party, would only be reserved for a special sector, where the costs of displacement would be disastrous for the structure of European democracies, freedom of expression and the sustainability of the rule of law. What is missing in the proposal is clear evidence that the proposed adjustment would help to share the economic benefits of GenAI and its various applications throughout the value chain – and not only benefit a small group of larger rightholders. The report’s focus on one approach – licensing – therefore seems premature, and other options should at least be explored.

An interesting aspect of the report are the borrowed elements from other legal domains, such as the (eventually abandoned) presumption on infringement in the absence of fully transparent datasets and the (criticized) low-risk rule for bringing successful claims against GenAI developers who have (likely) infringed copyright. They could be harbingers of a stronger alignment of copyright law to other areas of digital regulation. However, the limited scope of the proposal and its relative isolation from other areas of copyright law make this observation somewhat speculative.

The adoption of the proposed measures is still a way off, the own initiative report being the first step in the direction of concrete legislative measures. The legislative path will be hard and long. While this initiative is targeted and would only apply to GenAI, the proposal ends its Explanatory Statement with the afterthought that European copyright generally requires further harmonization. There lurks a certain danger that another painful micro-reform will lessen the appetite of the EU legislator to tackle copyright reform more holistically as opposed to the fragmented, highly sectoral scope of the report. Integrating considerations of fairness, openness, sustainability and innovation in a larger debate on the EU copyright rules could ensure that sectoral interventions do not cause systemic disruptions. Albeit limited in scope, the proposed changes to EU copyright law are bold. An even bolder approach and call to action, fully reflective of the EU’s economic and constitutional order, even if later implemented in stages, could have set the EU on a path for a more comprehensive and desperately needed overhaul of copyright law. 

Image by Kohji Asakawa from Pixabay