英文_ITIF_一项法律制定了韩国的人工智能政策_一个薄弱环节可能会打破它_25页_445kb

South Korea's AI Framework Act Summary

Core Content

South Korea's AI Framework Act, passed in December 2024, is the first global law to integrate AI strategy, industrial promotion, and regulation into a single statute. This comprehensive legal instrument aims to foster AI innovation, enhance national competitiveness, and ensure ethical and safe AI development. However, the Act's regulatory provisions are criticized for being overly broad and potentially counterproductive to its strategic and industrial goals.

Main Viewpoints

1. Unified Policy Instrument

• The Act serves as a single instrument for AI strategy, promotion, and regulation.
• This integration provides a strong foundation for coordinated efforts but also increases the risk of systemic failure if any part is poorly designed.

2. Strengths in Promotion and Strategy

• The promotion agenda includes data infrastructure, clusters, talent development, and internationalization, positioning South Korea as a leader in AI.
• The strategy is centralized, emphasizing coordinated investment and a unified vision for AI growth.

3. Regulatory Weaknesses

• Regulatory provisions are criticized for being too rigid and misaligned with the nature of AI risks.
• Overly broad definitions, mandatory labeling, compute thresholds, and process-heavy reporting may stifle innovation and create unnecessary burdens.

Key Information

1. Chapters of the Act

The AI Framework Act is structured into six chapters:

1. General Provisions – Defines key terms and sets the legal foundation.
2. Governance for Sound AI Development and Trust – Establishes the National AI Strategy Committee and institutional support.
3. Policies for AI Development and Industry Promotion – Focuses on innovation and industrial support.
4. AI Ethics and Trustworthiness – Combines soft-law and hard-law measures for ethical and safe AI.
5. Supplementary Provisions – Sets rules for funding and program operations.
6. Penalties – Defines sanctions for non-compliance with AI regulations.

2. Regulatory Challenges

• Broad Definitions: The term "AI system" is too inclusive, potentially regulating ordinary software and not just systems with novel governance risks.
• Rigid R&D Mandates: Article 13 sets prescriptive R&D priorities, which may not align with global breakthroughs and could stifle innovation.
• SME Bias: The law mandates SME priority in AI support, which may hinder the scale and competitiveness of South Korea's AI industry.
• Process-Heavy Compliance: Requirements for extensive self-assessment and documentation may waste resources without delivering meaningful oversight.

3. Recommended Adjustments

Structural Amendments by the National Assembly

• Article 2 (Definitions): Narrow the definition of "AI system" to exclude systems that do not introduce novel governance risks.
• Articles 7–9 (National AI Strategy Committee): Preserve the committee's role in strategy and resource allocation, but remove its authority to dictate regulatory changes.
• Article 13 (R&D Priorities): Replace prescriptive R&D mandates with a flexible, government-led roadmap that can adapt to global trends.
• Article 17 (SME Priority): Remove the statutory requirement for SMEs to receive priority consideration and adopt size-neutral language.
• Article 31 (Transparency Obligations): Replace mandatory disclosure requirements with voluntary provenance standards and targeted rules for specific harms.
• Article 32 (Safety Measures): Remove compute thresholds as a regulatory trigger and authorize the AI Safety Research Institute (AISI) to conduct post-deployment evaluations.
• Articles 33–35 (High-Impact AI Oversight): Replace process-heavy mandates with performance-based requirements.
• Article 36 (Domestic Agent Designation): Remove revenue and user thresholds and base oversight on the system's impact rather than its size or location.

Implementation by MSIT

• Article 14 (Standardization): Clarify that standards should be industry-led, with government acting as a coordinator.
• Article 18–19 (Support for Start-ups and Convergence Policies): Ensure size-neutral support for all firms, including larger ones.
• Article 40 (Data Requests and Inspections): Set clear guidelines on what data may be requested and for what purposes.
• Article 42 (Penalties): Scale penalties based on the severity and impact of violations.
• Article 43 (Administrative Fines): Introduce a grace period for compliance, allowing firms time to adjust without facing immediate fines.

Conclusion

The AI Framework Act is a bold and ambitious attempt to shape AI policy in a unified manner. While it provides a strong strategic and promotional framework, its regulatory provisions require significant refinement to avoid stifling innovation and to ensure that oversight is both effective and proportionate. The law will take effect in January 2026, and the success of its implementation depends on structural amendments by the National Assembly and balanced, practical enforcement by MSIT.