Regulating inputs, underestimating outcomes: The EU AI Act’s intellectual property blind spot
The EU AI Act rests on a critical assumption in the field of intellectual property: that regulating how data is used in AI systems is likely sufficient to control the legal risks they create. It is described as meta-regulation in the field of copyright, imposing preventive and risk-based compliance obligations on AI providers rather than directly governing infringements or creating new rights. However, this assumption is increasingly untenable in generative AI. Lawful inputs do not necessarily guarantee lawful outputs, and even where outputs are lawful, uncertainty remains regarding ownership, entitlement, and who may legitimately benefit from AI-generated content.
As AI systems become more capable of producing complex and unpredictable results, the gap between input regulation and output consequences becomes increasingly visible. These systems remarkably act as independent generators of content that may diverge significantly from user prompts and expectations. In practice, almost every creative output or innovative approach is likely to raise intellectual property (IP) related concerns (Lyon, Watkins & Iwahashi, 2017). Questions surrounding the IP status of AI-generated outputs are now practical rather than theoretical. Although this analysis focuses on the AI Act, contractual arrangements and private agreements may ultimately determine ownership, attribution of rights, and legal responsibility for such outputs.
Input regulation and copyright compliance in AI systems
At the level of inputs, the EU legislator has adopted a relatively structured approach based on compliance perspective. Recitals 105 and 106 of the AI Act emphasise the importance of respecting copyright in the context of text and data mining, while Article 53 establishes obligations for providers of general-purpose AI models to comply with copyright law and related rights. By referring explicitly to Directive (EU) 2019/790, the Act implies that its primary concern lies in ensuring the lawful use of data during the training phase. Advanced generative AI systems, particularly those based on Large Language Models (LLMs), rely on extensive data mining processes that extract vast amounts of information from online sources. This large-scale extraction is not merely a technical feature but a structural characteristic of contemporary AI systems. As a result, it raises significant IP concerns, and challenges the adequacy of existing legal frameworks (Novelli et al., 2024).
In this context, it is important to revisit the concept of creativity, which has traditionally been understood through an anthropocentric perspective in copyright law. This approach assumes that originality and creativity arise directly and intentionally from human activity. By contrast, large language models introduce a form of machinic creativity. Although grounded in human-generated knowledge, these systems process vast amounts of data through exploratory and observational methods. In doing so, they may reveal hidden potentials within existing information and recombine patterns in ways that generate novel outputs, creating new forms from previously unexploited informational structures. ( Boden, Margaret A, 2004).
From lawful inputs to unlawful outputs: a regulatory gap
The AI Act does not directly clarify the legal status of outputs, as it primarily focuses on ensuring compliance at the level of inputs. It does not address fundamental questions such as who, if anyone, should be considered the author of AI-generated content, how ownership should be allocated, or under what conditions liability for infringement should arise. This silence reflects the difficulty of reconciling traditional copyright concepts with the realities of generative AI. Recent legal disputes, such as GEMA v. OpenAI (2025), offer compelling evidence of this inefficiency. In an ongoing dispute, the German “Society for musical performing and mechanical reproduction rights” (GEMA) alleged that AI-generated outputs reproduced protected song lyrics without authorisation. The Munich Regional Court largely upheld the claims against OpenAI, finding that both the memorisation of copyrighted lyrics in language models and their reproduction in outputs may constitute copyright infringement. The court rejected reliance on text and data mining exceptions and attributed responsibility to the model provider rather than users. Importantly, although the dispute falls within the broader scope of the AI Act, the court relied solely on national copyright law, highlighting the absence of a dedicated legal framework within the AI Act for addressing AI-generated outputs. Although the judgment is not yet final, it highlights how AI-generated outputs can raise infringement issues even where training practices are considered lawful. It also highlights the lack of clarity regarding who should bear responsibility for such outputs within the current legal framework; especially in the terms of unpredictable products.
In these situations, legal actors are compelled to fall back on general copyright regimes, including domestic laws (e.g., UrhG for Germany) and existing EU directives. However, these frameworks are fundamentally rooted in an anthropocentric understanding of creativity. The consequence is a structural mismatch between law and technology. Where no human author can be meaningfully identified, copyright protection may not arise at all under AI act or domestic copyrights’ regulation umbrellas. Conversely, where infringement occurs, assigning responsibility becomes legally and practically challenging, irrespective of the extent to which such matters may be addressed through contractual arrangements commonly used in practice. As Guadamuz (2025) notes, existing legal frameworks remain ill-equipped to address the complexities introduced by AI-generated content. Similarly, Rosati (2025) argues that the AI Act does not provide an adequate basis for addressing liability in relation to such outputs, leaving significant gaps in enforcement and protection.
Conclusion
This situation reflects the difficulty of effectively regulating output-related copyright issues and unpredictable AI-generated content. On the one hand, a growing number of AI-generated products may remain outside the scope of copyright protection, creating uncertainty for users, developers, and markets alike. On the other hand, potential infringements may not be effectively addressed, weakening the deterrent function of copyright law. The result is a system that is both underinclusive and insufficiently responsive to new forms of technological creativity. On the other hand, the limitations of the AI Act in this regard also reflect a broader issue in legislative design. The author believes that the AI act creates a preventive copyright compliance system but lacks effective solutions for consequences and its protection capability for outputs remains highly risk-taking.
Unpredictability is emerging as a central concept in generative AI and must be distinguished from mere chance. Chance refers to accidental similarities or errors, such as independent creation, which may even be considered a “bug” in computer science and cannot form the basis of legal rights. By contrast, unpredictability in generative AI results from structured, data-driven processes that produce outputs diverging from user expectations while the system still operates as intended. This transforms unpredictability from an isolated coincidence into a systemic feature of AI-generated content. In this sense, unpredictability may be understood as an aspect of machinic creativity, although it remains largely undefined in legal terms. The absence of such a framework further reinforces the inadequacy of existing copyright regimes when applied to AI-generated products. Similar tensions can be observed in other areas of intellectual property. For instance, the DABUS cases before the European Patent Office reveal comparable difficulties in attributing inventorship to AI-generated outcomes. Patent applications were filed for inventions generated by an AI system (DABUS), with the AI designated as the inventor. The applications were refused on the grounds that, under European patent law, an inventor must be a natural person (Article 81 EPC & Rule 19 EPC). These decisions illustrate how existing legal frameworks, built around human inventorship, struggle to accommodate AI-generated outcomes that lack a clearly identifiable human creator.
If the European Union aims to maintain a coherent and future-proof intellectual property system, this regulatory blind spot must be addressed through a more comprehensive framework and a specific annex dealing with AI-generated outputs. Such an approach would not remove human actors from rights and responsibility, nor grant legal personality to AI systems. Rather, it would move beyond the strict requirement of direct and identifiable human creativity as a condition for protection. From this perspective, AI-generated unpredictable outputs may be understood as the result of indirect human creativity. Recognising this indirect relationship would allow rights and responsibilities to be allocated more coherently across the AI development chain through causal connections between outputs and the human actors involved.
It may be possible to draw on the approach adopted in database protection to address AI-generated unpredictable outputs and their related intellectual property rights. Similar to the sui generis database right, which protects investment in time, effort, and financial resources without relying on human creativity, a comparable mechanism could be considered for AI-generated outputs. Such an approach could move beyond the rigid anthropocentric framework of current intellectual property law, while allocating rights and responsibilities across the chain of human actors involved in AI development. At the same time, it would avoid attributing legal personality to AI systems and preserve the core principles of European intellectual property law with minimal structural disruption.
In conclusion, this approach proposes recognising rights within the development chain through a sui generis-like mechanism, while attributing responsibility and rights to human actors based on causal contribution. Without such reform, the current framework risks losing normative coherence and practical effectiveness in the face of rapid technological change.