The rapid development of generative artificial intelligence has made copyright protection dependent on distinctions that traditional doctrine did not need to articulate with the same precision: the status of human contributions to an output, the legal character of a prompt, and the permissibility of using existing works for model training. This article examines these questions through a comparative legal analysis of Kazakhstan, the United States and the European Union, with particular attention to Article 23 of Kazakhstan’s Law on Artificial Intelligence of 17 November 2025. The methodology combines doctrinal interpretation, functional comparison of legal instruments and analysis of official guidance and judicial decisions. The research finds that the jurisdictions share an emphasis on human creativity in the protection of outputs but differ in their treatment of creative prompts, training-related uses and rightsholder reservations. Kazakhstan expressly addresses both human contribution and machine-readable prohibitions on training; United States practice differentiates human-authored elements from autonomously generated content; European Union legislation combines text-and-data-mining rules with transparency and copyright-compliance obligations for general-purpose AI providers. The article proposes a three-part compliance framework covering proof of human creative choices, provenance of training material and interoperable recording of rights reservations. It argues that these questions must be assessed separately to avoid conflating copyright ownership in outputs with permission to use inputs.