The exceptions permitting automated analysis (mining) for the training of AI models, introduced by the Directive on Copyright in the Digital Single Market back in 2019, are today gaining increasing importance.
The aforementioned Directive introduced two limitations for the purposes of Text and Data Mining (TDM):
- The exception in favour of research organizations and cultural heritage institutions enabled the reproduction and extraction of data carried out by research organizations and cultural heritage institutions for the purpose of the automated processing of text and data for scientific research in works or other subject matter to which they have lawful access; and
- The exception for the reproduction and extraction of works and other subject matter for the purposes of automated text and data processing by all users, provided that the use of works and other subject matter to which lawful access can be obtained is not expressly prohibited by rightsholders in an appropriate manner (the opt-out model).
In this way, the Directive enabled the lawful use of copyrighted works and related rights subject matter for the purposes of automated text and data processing. The exception under which virtually anyone may use copyrighted works and related rights subject matter for the training of artificial intelligence created a legal grey area, the so-called TDM loophole. The absence of mandatory consent from authors and rightsholders, combined with the presumption that works and related rights subject matter may be freely used for automated data processing unless authors and rightsholders exclude their works, has enabled AI companies to use the works of others without payment and without permission.
This raises the question of whether such a limitation is contrary to the essence of copyright, which is not designed as a system of “waiving rights,” nor a system in which the default rule requires rightsholders to “reserve” their rights in order for their content to be protected from use by third parties. Copyright grants its holder a set of preventive rights, meaning that a third party wishing to undertake an act encompassed by a limitation must either obtain permission (a licence) or the act must be prescribed by law.
These TDM limitations have not yet become part of the Serbian legal system, but given that our Law on Copyright and Related Rights strives to harmonize with EU directives, it could be expected that these exceptions will be included in the new law.
The Three-Step Test in the Context of TDM Limitations
The Berne Convention for the Protection of Literary and Artistic Works prescribes possibility of limiting copyright under certain conditions. Thus, a limitation of copyright is permitted in special cases, provided that it does not conflict with the normal exploitation of the work and does not unreasonably prejudice the legitimate interests of the author. This provision constitutes the essence of the so-called Three-Step Test, which every copyright limitation must pass. The Three-Step Test is an obligation not only of legislative bodies when drafting relevant provisions, but also of courts, which must assess it in each concrete case when applying national copyright limitations.
The Three-Step Test requires the cumulative fulfillment of three conditions:
- The limitation applies only in certain special cases, which implies only those limitations that are specific and that have a particular purpose and objective;
- It must not conflict with the normal exploitation of the work, meaning that the acts of use covered by the exception must not enter into „economic competition with the ways in which rightsholders normally extract economic value from their right to the work (i.e., copyright), depriving them of significant or tangible commercial gain“;
- It must not unreasonably prejudice the legitimate interests of the author or rightsholder, meaning that such a limitation must not cause „an unreasonable loss of revenue to the copyright holder“.
Taking these conditions into account, the question arises whether TDM limitations are justified in the context of the Three-Step Test as the fundamental benchmark for any copyright exception or limitation. It is challenging to fit the TDM limitation within the framework of the conditions above, bearing in mind the following: the TDM limitation under which practically anyone may use copyrighted works and other protected subject matter without the author’s and/or rightholder’s consent, which is presumed for the training of artificial intelligence, appears overly broad. Consequently, the first condition—requiring that the limitation be narrowly defined as a special case—is called into question. Due to the broadly formulated limitation, the remaining two conditions are also challenged, raising a major dilemma as to whether such a limitation disrupts the normal exploitation of the work and whether it unreasonably prejudices the legitimate interests of authors, since the burden of protection is shifted from the user to the author. This is particularly evident when TDM is performed for commercial purposes without the explicit authorization of the author or rightholder.
Although from a theoretical standpoint this exception presents numerous challenges, raising questions and dilemmas regarding the need for its revision, EU legal systems are already actively implementing it.
AI Framework Act in Italy
At the end of September this year, Italy adopted a comprehensive law on artificial intelligence—the AI Framework Act—the first of its kind among European Union member states, thereby positioning itself as a pioneer in shaping AI regulation within the European Union.
Although this law sets out principles for the use of artificial intelligence in various sectors such as healthcare, employment, public procurement, professional services, and others, we will here focus on the implications that the Italian AI law has introduced into the regulatory framework in the field of copyright.
In the context of copyright, protected works remain “creations of human intellect“, even if AI tools were used in their creation. Protection applies only if the human author has provided a genuine, creative contribution. Assessing this “sufficient human contribution” as a type of legal standard will undoubtedly be complex, but it is expected that judicial and administrative practice will provide answers to disputes that may arise.
The TDM limitation had previously been implemented in Italian legislation, but with the adoption of the AI Framework Act it has now been extended to AI systems, including generative AI models. Thus, previously prescribed TDM limitations also apply to the use of materials for the training of artificial intelligence that are available online or contained in databases, provided that the user has lawful access to such materials.
It is also noteworthy that the AI Framework Act introduced amendments to criminal legislation specifically in the context of TDM limitations. Unauthorized TDM becomes a criminal offence, elevating what was previously a matter of civil liability to the level of potential criminal liability. Although the penalty is largely symbolic (a fine ranging from 51 to 2,065 euros), anyone who, without the right to do so, for any purpose and in any form, reproduces or extracts text or data from works or other materials available online or in databases, contrary to TDM limitation provisions, including through artificial intelligence systems, may bear criminal responsibility.
Rather than a Conclusion
The limitation for text and data mining (TDM) has become one of the key issues in balancing copyright, related rights, and sui generis rights on the one hand, and innovation on the other. Although the primary aim of this exception was to allow freer access to data for scientific research and innovation, in practice this limitation has created a legal grey area. Particularly problematic is the fact that the TDM exception, in its broad form, permits the use of protected works without the explicit consent of authors or rightsholders, based on the presumption that they must themselves take steps to prohibit such use (the “opt-out” model). This shifts the balance between the interests of creators and technological development in favour of users, calling into question the consistency of this exception with fundamental copyright principles and the Three-Step Test of the Berne Convention.
For this reason, the question of the justification and scope of TDM limitations is becoming increasingly significant. Its broad interpretation may lead to infringements of authors’ rights, especially when their works are used for commercial purposes—such as the training of large language and generative AI models—without adequate licensing, remuneration, or control.
The Italian AI Framework Act of 2025 represents an important step toward clarifying and regulating this area. With this law, Italy has not only confirmed existing TDM exceptions but has also extended them to the training of artificial intelligence, including generative models. At the same time, innovations have been introduced into the criminal law framework – unauthorized text and data mining is now treated as a criminal offence, indicating a tendency to view abuses in the TDM domain more seriously, rather than merely through the lens of civil liability.
For the professional public, the Italian model may serve as a guideline for how EU member states approach the balancing of copyright and technological development in the field of artificial intelligence. For the broader public, the message is clear: while AI systems advance thanks to access to vast quantities of data, the legal system must find a way to ensure that such progress does not come at the expense of creators and holders of copyright, related rights, and sui generis rights. Serbia, which aims to harmonize with the EU acquis, will in the future need to carefully consider how to incorporate TDM exceptions into its legislation without undermining the essence of copyright – the protection of human creative contribution and the preventive nature of the bundle of rights belonging to the author or rightsholder.
This article is to be considered as exclusively informative, with no intention to provide legal advice. If you should need additional information, please contact us directly.
By Bojana Veselinovic, Senior Associate, PR Legal
