Generative AI and the EU Copyright Law’s Blind Spots

Legislatures are notoriously inefficient in legislating for future technologies. These technologies are sometimes developing at the same time as legislatures are debating new laws, resulting in the legislatures missing, by a small margin, opportunities to design the laws for the new technologies. This was the situation, for example, with the negotiations of the TRIPS Agreement, the Agreement on Trade-Related Aspects of Intellectual Property Rights, which was negotiated in the late 1980s and early 1990s, just as the new HTML code was being developed and publicly released. Even though the TRIPS negotiations and the HTML developments at CERN (The European Organization for Nuclear Research) were taking place in parallel only about six miles from each other, the TRIPS Agreement did not address the new intellectual property law issues that were generated by the world wide web that HTML made possible.

Future commentators might similarly view as a missed legislative opportunity the parallel development of EU copyright legislation and generative AI. Although supporters of generative AI tend to suggest otherwise, the text and data mining (“TDM”) exceptions in the EU’s 2019 Digital Single Market (“DSM”) Directive were not drafted with generative AI in the awareness of the legislators. In an article published in November 2025, Dr. Petra Žikovská meticulously reviewed the legislative history of the DSM Directive and confirmed what Professor Tim Dornis had already suggested in his earlier article: according to the accessible documents, EU legislators did not have generative AI on their minds when they drafted and adopted what became Articles 3 and 4 of the DSM Directive on the TDM exceptions to copyright. In fact, at the time of the discussions of the proposal for the Directive, some of the participants in the legislative process were perplexed by the degree of interest in and emphasis on the exceptions that some interest groups were promoting.

In hindsight, the legislators’ lack of appreciation of the rapidly-developing AI technologies, including their impending generative AI abilities, appears as surprising now as the TRIPS Agreement negotiators’ lack of knowledge of the imminently ubiquitous world wide web. But it is not difficult to forget that the proposal for the Directive was launched into the legislative process in September 2016, a year and a half before OpenAI released GPT-1 in June 2018. By June 2018, the text of the Directive proposal had deeply progressed into the legislative process; it had already been agreed upon in COREPER, the Permanent Representatives Committee, and had moved to the first reading in the Council and the European Parliament. Although back and forth followed between the Council and the European Parliament, by the time the proposal was adopted in the Spring of 2019 and eventually signed in April 2019, none of the text of the Directive’s provisions on the TDM exceptions had changed significantly.

In today’s context, the texts of the legislative history that preceded the adoption of the DSM Directive read as evidence of the early naïveté about AI. The preamble to the Directive refers to TDM as a method that enables the “processing of large amounts of information with a view to gaining new knowledge and discovering new trends,” which definition was used in a footnote of a 2015 Commission staff working document that accompanied the Commission’s DSM Strategy. The DSM Strategy itself referred to TDM three times without mentioning the use of TDM to train generative AI—perhaps not surprisingly given the lack of knowledge of the technology among the general public at that time; the DSM Strategy relied on the European Commission’s 2014 expert group report “Standardisation in the Area of Innovation and Technological Development, Notably in the Field of Text and Data Mining,” which also did not discuss generative AI. The report included examples of uses of TDM such as for an analysis of “vast quantities of video [for] research in meteorology and police forensics,” of text for research on “the incidence and meaning of the word ‘digital’ in the work of the EU,” and of data for the study of “shoppers’ spending patterns [and] their leisure time and health.” Linguists were mentioned as significant TDM users with 25–30% of computational linguistics projects relying on TDM. If anything was referred to as “generated” in the text of the report, it was positive externalities, consumer surplus, new innovative actors, and also “deadweight loss for society” if a TDM exception were not enacted.

Future commentators might be curious as well about the fact that while the Directive introduced two TDM exceptions, it did not explain the relationship between the exceptions. Article 3 requires that member states adopt an exception for TDM “by research organisations and cultural heritage institutions [conducted] for the purposes of scientific research”; Article 4 requires an exception for TDM conducted by other entities and/or for other purposes. The Directive does not address what happens if TDM is conducted by a research organization but the extracted data are later used by a commercial entity to train and deploy AI; alternatively, the research organization could also train AI, and the commercial entity would then deploy tools based on the AI.

It might be surprising that the EU legislators did not foresee the problem of TDM being conducted for one purpose and later benefitting another purpose. The EU has been promoting research–industry partnerships and the creation of spin-offs; for example, in as early as 2016, the European Commission noted in its “Europe’s Next Leaders: The Start-up and Scale-up Initiative” that the EU and its member states could support start-ups by connecting them with partners, such  as universities and research centers.

The omission in the DSM Directive might one day seem particularly jarring given that a prominent example exists of a transition in the AI world from a not-for-profit research entity into a for-profit business: OpenAI was founded as a non-profit organization in December 2015, about nine months before the European Commission published the proposal for the DSM Directive in September 2016, which at that stage included only one TDM exception—the exception for “research organisations.” When OpenAI announced the creation of its for-profit OpenAI LP on March 11, 2019, the announcement was made only 15 days before the European Parliament adopted the DSM Directive on March 26, 2019. Of course, to what extent an entity such as OpenAI could benefit from the research organization exception of Article 3 might be debated; recital 12 of the Directive explains that the exception does not apply to “organisations upon which commercial undertakings have a decisive influence allowing such undertakings to exercise control because of structural situations, such as through their quality of shareholder or member, which could result in preferential access to the results of the research.”

It should be recognized that EU legislators did not have generative AI in mind when they designed and adopted the TDM exceptions in the DSM Directive. Sometimes legislators manage to adopt laws that happen to be suitable for future technologies, but it is questionable whether this fortuity exists for the TDM exceptions in the DSM Directive and generative AI. The TDM exceptions deserve clarification as to their applicability, including in instances when the TDM process begins in a research organization and moves to and/or its benefits are enjoyed by a for-profit business.

Image by mono-webdesign from Pixabay