EU인공지능법 오픈소스 면제조항의 WTO 합치성 검토: 내국민 대우 위반 가능성과 일반적 예외 적용을 중심으로

The EU Artificial Intelligence Act, finally approved in May 2024, represents the world's first comprehensive AI regulatory framework that classifies AI systems through a risk-based approach and applies differential regulation. This study focuses on the fact that the EU AI Act provides regulatory exemptions for models released under open-source licenses, even when they perform the same functions as proprietary AI models. This is significant because European companies in the global AI market currently show a strong tendency to adopt open-source strategies, while leading U.S. companies predominantly compete through proprietary models, raising concerns that such exemption provisions may create unintended trade-restrictive effects. This research aims to legally analyze the possibility that the open-source exemption provisions of the EU AI Act may violate the National Treatment (NT) principle under WTO agreements, and to derive policy implications for establishing a trade-friendly AI governance framework. To this end, an empirical analysis was first conducted on the current state of open-source adoption among major AI companies as of 2025. The analysis revealed a significant gap, with European companies showing exemption rates of 80%, while U.S. companies only reached approximately 40%, representing nearly a two-fold difference. This constitutes a typical case of de facto discrimination that, while formally neutral in criteria, substantially demonstrates regional bias. Regarding the controversy over the legal nature of AI models, this study acknowledges the dual nature of goods and services and adopts a situational approach, conducting analysis from both GATT and GATS perspectives. Potential violations of the NT treatment principle were identified under both GATT Article III when AI models are classified as goods and GATS Article XVII when classified as services. Open-source AI models and proprietary AI models fundamentally perform the same functions and are in a direct substitutable relationship from the end consumer's perspective, making them likely to be classified as like products or like services. The 40 percentage point gap in exemption rates is evaluated as substantial discrimination that far exceeds the de minimis level tolerated by the WTO. The possibility of the EU invoking general exception provisions under GATT or GATS based on public order or protection of human life and health was also examined, but the justification potential was deemed limited. Not only is it difficult to prove a direct connection between open-source exemptions and public order protection, but applying the necessity test criteria established through existing WTO dispute settlement cases suggests that less trade-restrictive alternatives may exist to achieve the same policy objectives. Additionally, there are concerns about arbitrary discrimination under the chapeau requirements, making the possibility of exception recognition low. Consequently, it is highly likely to constitute a violation of NT treatment. Based on these analytical results, improvement measures for establishing trade-friendly AI governance were proposed. In the short term, it was suggested to redesign the open-source exemption criteria of the EU AI Act, introducing a graduated regulatory relaxation system based on transparency levels, explainability, and risk management capabilities instead of the current binary structure. It was also analyzed that international economic law compatibility could be enhanced through the introduction of mutual recognition provisions and strengthening approaches based on international standards. In the long term, the study emphasized the need for regulatory harmonization efforts among major AI powers and the establishment of a comprehensive and binding multilateral AI governance framework similar to the International Atomic Energy Agency in the nuclear energy field. This research empirically demonstrates that international economic law compatibility must be considered from the AI regulation design stage and reaffirms the importance of technology-neutral and outcome-oriented regulatory principles. It is also expected to contribute to the development of digital trade norms by specifically examining the applicability of WTO agreements in the emerging field of AI technology.