Edited by Paola Furiosi, Giulia Gialletti and Francesco Gaggioli
By reshaping the way content is created, generative AI has called into question some of the assumptions on which European copyright law is based.
With its resolution of 10 March 2026 on the relationship between copyright and generative AI (“European Parliament resolution of 10 March 2026 on copyright and generative artificial intelligence – opportunities and challenges (2025/2058(INI)“), the European Parliament acknowledged this transformation, highlighting that the current regulatory framework is no longer fully adequate to govern the dynamics specific to generative systems and, at the same time, emphasising the need to reflect on possible developments of the existing legal regime.
To understand the rationale underlying this intervention, it is helpful to begin with the current European regulatory framework.
Such framework revolves around the exception for text and data mining (“TDM”), provided for under Directive (EU) 2019/790 “Copyright in the Digital Single Market” (“CDSM”). Under this regime, it is generally permitted to extract and analyse large quantities of content, including copyright-protected content, for data analysis purposes, both in scientific and commercial contexts. Rightsholders nevertheless retain the possibility to exercise an opt-out, namely, to expressly reserve the use of their works and protected materials for text and data mining purposes, provided that such reservation is expressed in a technically machine-readable form.
This system is based on a relatively straightforward premise: TDM activities consist in analysing and extracting information from content, rather than creating new works capable of replacing the original ones.
It is precisely this premise that generative AI tends to challenge.
In generative systems, data are not used merely to identify patterns or correlations; rather, they form the basis for training models capable of generating new content that may rework, synthesise or, in some cases, serve as an alternative to accessing the original works from which such content derives. In this context, the use of data may directly affect the economic value of original works, insofar as the output generated by AI models can reduce the need to access the original source material itself.
It is against this background that the European Parliament’s resolution suggests a genuine paradigm shift.
The use of protected works in model training processes increasingly tends to be treated no longer as part of a broad and essentially free exception, but rather as an activity subject to mechanisms closer to authorisation and remuneration. TDM is not eliminated, however, it progressively loses part of its nature as a relatively unrestricted space and moves towards a framework requiring greater transparency and oversight, while also paving the way for possible forms of economic negotiation between developers and rightsholders. From this perspective, the opt-out mechanism assumes particular significance.
Under the current framework, protection is largely left to the initiative of the rightsholders, who must not only express their reservation in a “technically” machine-readable format, but also face significant practical difficulties: from the absence of uniform technical standards to the fragmentation of reservations across multiple platforms, as well as the substantial impossibility of verifying whether, and to what extent, such reservations have actually been respected during training processes.
The resolution therefore seeks to rebalance this dynamic by shifting the centre of gravity of responsibility onto model developers and introducing stricter transparency obligations.
In particular, providers would be required to make available sufficiently detailed information regarding both the content used and the manner in which such content is used, thereby enabling rightsholders to effectively verify whether their protected works have been utilized. Where such information is unavailable, the resolution envisages the introduction of a presumption of copyright infringement. In other words, the burden would no longer rest on the rightsholder to reconstruct ex post the use of its works within technologically complex and opaque systems; rather, it would fall on the model developer to demonstrate the lawfulness of its activities.
A further significant aspect concerns the stage at which interference with copyright-protected works is assessed.
The current framework primarily focuses on the training phase of AI models, whereas the resolution highlights how issues may also arise during the deployment phase, particularly where an AI system access protected online content, processes it and returns to the user a complete and self-contained response, making, de facto, consultation of the original source unnecessary.
This dynamic is especially evident in the media and information sector, where AI systems can aggregate and synthesise journalistic content, potentially reducing traffic to original sources and directly affecting publishing revenues. Hence the European Parliament’s reflection on the possible introduction of appropriate remuneration mechanisms also for these forms of use.
Consistent with this approach, the resolution also opens the way to the development of a genuine European market for AI licensing, based on the principle of fair remuneration for the use of protected works. In this scenario, licensing models capable of reflecting the economic value of the content used to train AI models could progressively emerge.
Finally, the European Parliament envisages the establishment of a centralised European register, managed by the European Union Intellectual Property Office (“EUIPO”), through which rightsholders could identify their protected works and specify any reservations concerning their use in AI systems. The aim is to overcome the current fragmentation and provide both rightsholders and AI developers with a clearer and more standardized framework.
Overall, the resolution suggests a gradual transformation of TDM: from a broad and largely free exception, based on an opt-out that is difficult to monitor in practice, into a more structured framework grounded on transparency, traceability and the economic valorisation of protected works.
Although the resolution does not produce immediately binding effects, it clearly indicates the direction in which the European approach may evolve rebalancing a system that, in the early stages of generative AI development, has favoured access to content more than the protection of rightsholders, while preserving incentives for innovation.
Let’s Talk
For a more in-depth discussion, please contact: