Abstract
<jats:p>This article explores the issue of training artificial intelligence (AI) in the context of copyright law and compares it to human learning. The authors emphasize that such comparisons must be approached critically, considering the specific nature of both processes. Not every use of copyrighted works in AI training constitutes an infringement; however, AI’s capacity for large scale memorization and reproduction of protected works necessitates a clear distinction between permissible and impermissible uses. The article proposes a classification of such uses into “negative learning,” “subconscious memorization,” and “verbatim reproduction,” highlighting the legal significance of each category. Special attention is paid to the comparative analysis of regulatory approaches in the United States and the European Union, grounded in dominant theories of intellectual property: the utilitarian and welfare theories in the U.S. versus the personhood theory in the EU. This analysis helps explain the divergence in AI regulation across legal systems. Given Ukraine’s legal tradition rooted in the continental model, the country is encouraged to continue aligning its regulatory path with EU law. The authors argue that there is no need to develop new intellectual property theories to address the challenges posed by AI. Instead, existing frameworks remain effective if applied complementarily: welfare theory should guide early-stage model development, while personhood theory should inform fair compensation mechanisms for creators whose works are used during training. The article also supports the adoption of an output-based remuneration model as a viable solution to balance the interests of authors and AI developers.</jats:p>