SERBIA Law and Practice Contributed by: Stevan Pajović, Tamara Simić Pajović, Medo Zornić and Jelena Stoljiljković, T-S Legal
7.3 Training Data, Model Development and Infringement Risk Current Status of Serbian Legislation Serbian copyright law has not yet been aligned with the EU Directive on Copyright in the Digital Single Mar- ket (the “DSM Directive”). Consequently, the statutory exceptions for text and data mining (TDM) are current- ly not provided for in the domestic legal framework. This lack of harmonisation creates a more restrictive environment for automated data extraction compared to the EU, as any reproduction of protected works for the purpose of data mining generally requires explicit authorisation from the rights holders. Distinction Between TDM and Generative AI Even upon the eventual implementation of TDM excep- tions into Serbian law, there are substantive reasons why such provisions should not automatically apply to the training of generative AI (GenAI) systems. TDM and GenAI serve fundamentally different purposes; while TDM focuses on analysing existing information to uncover patterns and correlations for knowledge extraction, GenAI utilises massive datasets to create entirely new synthetic content. TDM is essentially a non-transformative analytical process, whereas GenAI adopts and “memorises” the styles and characteris- tics of original works to generate competing outputs. Functional and Legal Implications The core distinction between these technologies lies in their functional outcome and market impact. TDM seeks patterns in existing data for the purpose of analysis and scientific insight, whereas GenAI utilises those same patterns for synthetic expression, often resulting in content that competes in the same market as the original works used for training. Therefore, the legal interpretation of TDM exceptions should remain narrow and should not be extended to cover the com- mercial training of AI models, which represents a sig- nificant shift from data analysis to content creation. Enforcement and Remedies in Intellectual Property In the event of an infringement or a serious threat to copyright and related rights, the legal landscape in Serbia provides for comprehensive judicial protec- tion, allowing rights holders to seek declaratory relief, permanent injunctions against infringing acts under the threat of monetary penalties, and the recovery of
reflect the author’s personality, spiritual expression, and creative intellect. Since artificial intelligence lacks a personality and self-awareness, the content it gener- ates cannot currently meet the statutory test of “origi- nality” or “spiritual creation” required for authorship. Legal Status of AI as an Author Current Serbian regulations do not recognise artificial intelligence as a legal or natural person, thus exclud- ing it from being considered an author. Even in the case of complex audiovisual works or scripts, such as those generated by models like Sora or Benjamin AI, the law maintains that only a human can express The inability of AI to hold authorship leads to two pri- mary legal possibilities: the work falls into the public domain or is attributed to a natural person. Attribu- tion to a human, such as the programmer or the user, depends on the extent of their creative influence. For a user to be recognised as an author, they must provide a real and essential creative contribution that goes beyond simple data selection or prompts. If the final output is an expression of the user’s creative spirit and key creative choices, it may qualify for protection, though any parts resulting solely from autonomous AI choices remain exempt. a state of spirit in a regulated form. Ownership of AI-Generated Works Corporate Frameworks for AI-Assisted Ownership In the absence of specific statutory provisions, busi- nesses in Serbia primarily structure ownership of AI- assisted content through precise contractual frame- works. Companies incorporate broad intellectual property assignment and “work-for-hire” clauses into employment and service agreements to ensure that all outputs generated by staff or contractors automati- cally vest in the employer. To prevent such works from falling into the public domain, businesses focus on documenting the human creative contribution, such as the formulation of specific prompts, iterative refine- ment, and final editorial selection. Where the original- ity of AI-generated content is legally uncertain, firms increasingly rely on the Law on the Protection of Busi- ness Secrets to protect algorithms, prompt libraries, and resulting datasets as confidential information that provides a competitive advantage.
252 CHAMBERS.COM
Powered by FlippingBook