Analysis of the Voss Report: Technical, Economic and Legal Concerns for European AI Policy

Cross-posted from C4C’s LinkedIn page – see the original article

German MEP Axel Voss (EPP) shared his draft report on copyright and generative AI (2025/2058(INI)) last week. This own-initiative report of the EU Parliament JURI Committee, while lacking direct legislative force, carries some influence in shaping future EU policy on artificial intelligence and copyright law.

The non-binding nature of this report should not diminish scrutiny of its proposals, as Parliament resolutions often inform subsequent Commission legislative initiatives and Member State positions.

This draft report is scheduled for discussion at the 15 July JURI Committee meeting (16:00-18:30 CEST). MEPs will have until 12 September (TBC) to table amendments to the report.

This analysis examines serious concerns with the report’s approach across technical feasibility, economic impact, legal consistency, and strategic alignment with EU competitiveness goals, arguing that the proposed framework contains fundamental flaws that would undermine European AI development and affect European consumers while failing to achieve meaningful protection for creators.

Technical Impossibility Meets Economic Reality

The Voss report’s core requirement for detailed itemisation of copyrighted works in AI training datasets represents a computational impossibility that would single-handedly eliminate European AI development. Modern frontier models like GPT-4 are trained on 13 trillion tokens from billions of sources, with the largest datasets containing 15+ trillion tokens. The computational overhead of tracking individual source attribution would increase training costs by 10-100x, making European AI development economically unviable.

The mathematics are unforgiving: Training costs already reach $100 million for current models, with next-generation systems projected to cost $1-10 billion. Adding granular content tracking would require additional petabytes of storage and exponentially more computational resources. As Stanford Professors Mark A. Lemley and Bryan Casey explain in the Texas Law Review, “there is no plausible option simply to license all of the underlying photographs, videos, audio files, or texts” on which ML models are trained. If you were surprised at the cost of saying “Thank you” and “Please” to ChatGPT, as claimed by Sam Altman, just imagine the impact of the proposed obligation.

Legal scholars who have closely examined the AI Act’s transparency requirements consistently emphasise the impossibility of detailed content tracking. Dr. João Pedro Quintais of the University of Amsterdam’s Institute for Information Law (IViR) notes in his comprehensive analysis of the AI Act that “the goal cannot be for [General Purpose] AI providers to list all or most of the copyrighted material they are including in their training data sets in an itemized manner with clear identification of rights ownership claims, etc., as this would make this provision impossible to comply with”. Quintais further explains that “the low threshold of originality, the territorial fragmentation of copyright and its ownership, the absence of a registration requirement for works, and in general the poor state of rights ownership metadata demonstrate this impossibility.”

The Voss report’s requirements would require tracking individual tokens across distributed training systems operating on thousands of GPUs, a technical impossibility that reveals the report’s fundamental misunderstanding of both AI architecture and existing legal frameworks.

Copyright Levies: Taxing Innovation to Fund Dysfunction

The numbers are damning: Past studies have shown that it costs 51.2 cents to collect each Euro for EU copyright levies, with only 53 cents of every euro reaching artists in Germany. This represents a catastrophic administrative failure that benefits bureaucrats more than creators. The system operates as what competition law scholars identify as “price-fixing cartels” with monopolistic positions.

Collecting societies exhibit profound structural inadequacies with “complete lack of accountability” and tariff-setting processes that can take up to 10 years. They “usually privilege top-selling artists and are not well suited to accommodate non-commercial music production,” creating massive coverage gaps. The distribution of levy revenues to recording artists amounts to less than €0.01 per album, a figure that exposes the system’s fundamental failure to serve its stated purpose.

The levy variations across EU countries range from €0.02 in Romania to €2.6 in France per capita, variations that cannot be explained by any underlying concept of economic harm. This arbitrary system taxes innovation while failing to compensate creators, representing the worst of both worlds for European competitiveness.

Scope Creep: Data Is Not Copyrighted

The Voss report contains fundamental conceptual errors that violate established legal principles distinguishing raw data from copyrightable expression across the globe. The EU Database Directive explicitly states that protection “does not extend to the data or material itself”, establishing a clear distinction between data compilation and the underlying data elements. This principle aligns with international jurisprudence, including the US Supreme Court’s Feist decision, which definitively established that “raw data does not satisfy the originality requirement” for copyright protection).

The legal foundation is clear: Copyright protects expressions, not ideas, procedures, or methods. The CJEU’s Infopaq decision requires that protected works be “original in the sense that it is the author’s own intellectual creation”. The Voss report’s conflation of factual data extraction with copyright infringement represents a worrisome expansion of copyright beyond its traditional boundaries.

The report’s approach creates confusing regulatory contradictions with the AI Act, which requires transparency about training data without presuming copyright infringement. The AI Act’s measured approach to copyright compliance directly conflicts with Voss’s presumption of infringement, creating legal uncertainty that violates due process principles and undermines the rule of law.

Fundamental Rights Versus Regulatory Capture: The Deeper Structural Failure

From a fundamental rights perspective, the Voss report violates multiple rights of the Charter of Fundamental Rights of the European Union: Article 11 (freedom of expression and information), Article 13 (freedom of arts and sciences), and Article 16 (freedom to conduct business).

AI training involves processing publicly available information to extract patterns and correlations. This represents a core aspect of freedom of information protected by the Universal Declaration of Human Rights and the European Convention on Human Rights. When the Voss report restricts AI research and development through licensing requirements and payment obligations, it fundamentally contradicts academic freedom and the established right to use available materials for legitimate research purposes and puts up another (artificial) barrier to the 5th Freedom, the free circulation of knowledge, at a time where the EU claims to promote it.

The economic dimension reveals an even more troubling pattern. MEP Voss’s approach creates what economists recognise as insurmountable transaction costs for licensing millions of individual works, while simultaneously generating windfall profits for rightsholders without corresponding benefits to consumers or broader social welfare. The mathematical reality of AI training datasets exposes the fundamental impossibility of meaningful individual creator compensation. As technology law expert Andrés Guadamuz observes, “for there to be a traditional licensing system, there has to be a value to each individual work that is worth the transaction costs, but if what you are going to get is 0.00000001% of an already low transaction cost, then most creators would not expect to see any money at all”. He aptly describes this as “the Spotify remuneration model, but a thousand times less money going to creators.” Basically, when training datasets contain billions of works, individual contributions become economically meaningless while administrative costs consume any potential distribution.

This represents a textbook case of regulatory capture, where legacy interests successfully manipulate the policy process to extract rents from innovation at the expense of European competitiveness and fundamental rights.

Conclusion: Defending European Digital Sovereignty

The Voss report represents regulatory capture by legacy interests seeking to extract rents from European innovation. Its technical requirements are impossible, its economic impact is destructive, its legal foundations are unsound, and its strategic implications are catastrophic for European competitiveness.

A proper policy response must reject the report’s fundamental premises, defend fundamental rights to information and innovation, and align copyright policy with European strategic autonomy goals. The EU’s choice is simple: lead the global AI revolution or become its casualty. The Voss report ensures the latter outcome while serving narrow interests at the expense of European innovation, competitiveness, and fundamental rights. The choice lies with European policymakers, but the consequences will impact European citizens, researchers, and businesses for decades to come.