Beyond Claiming Sovereign AI: Motivations, Challenges, and Contradictions in Developing and Deploying Local Foundation Models in South Korea

Paper Title

Beyond Claiming Sovereign AI: Motivations, Challenges, and Contradictions in Developing and Deploying Local Foundation Models in South Korea

Publication Info

• Topic area: Localized foundation model development and AI sovereignty in South Korea.
• Keywords: Sovereign AI, localized foundation models, South Korea, linguistic and cultural adaptation, regulatory compliance, resource constraints, geopolitical dynamics, HCI, responsible AI, sociotechnical systems.

Background and Problem

• Problem / challenge: Global foundation models are predominantly English-centric, leading to linguistic, cultural, and regulatory misalignments in non-Western contexts. Efforts to develop localized models face challenges such as resource constraints, reliance on global infrastructures, and limited benchmarks tailored to local needs.
• Significance: Addressing these issues is critical for reducing dependence on foreign technologies, ensuring cultural and linguistic relevance, and meeting local regulatory and security requirements.
• Motivation and related work: Prior research highlights the limitations of global models in non-Western contexts and advocates for localized and community-centered AI systems. However, there is limited empirical understanding of how localized foundation models are developed and deployed, particularly in the context of national sovereignty initiatives.

Solution

• Proposed approach: The study examines the development and deployment of localized foundation models in South Korea through semi-structured interviews with 15 AI practitioners, focusing on motivations, challenges, and workarounds.
• Novelty:
  1. Analysis of how AI sovereignty is enacted through practical, localized efforts rather than as a purely political ideal.
  2. Identification of sociotechnical challenges, such as resource shortages and reliance on global benchmarks, in achieving localized AI.
  3. Exploration of the tensions between global dependencies and local aspirations in AI development.
  4. Insights into how practitioners navigate regulatory, cultural, and infrastructural constraints.
• Procedure and key techniques:
  • Conducted semi-structured interviews with 15 practitioners from five South Korean organizations.
  • Analyzed motivations for developing local models, including cultural adaptation, regulatory compliance, and business flexibility.
  • Investigated challenges such as GPU shortages, data scarcity, and reliance on English-centric benchmarks.
  • Explored the practical realities of deploying local models in competitive and resource-constrained environments.

Results

• Concrete findings:
  • Motivations for local models include reducing reliance on foreign technologies, optimizing for Korean language and culture, ensuring regulatory compliance, and enhancing business flexibility.
  • Challenges include GPU shortages, limited high-quality Korean data, reliance on global benchmarks, and complex decision-making in model adoption.
  • Localized models often fail to achieve full autonomy due to dependencies on global infrastructures and tools.
• Advantage over baselines:
  • Local models address linguistic and cultural gaps, comply with South Korean regulations, and meet specific business and security needs that global models cannot.
  • However, they often lag in performance compared to global models like GPT-4, especially on English-centric benchmarks.
• Experiments / evaluation:
  • Interviews revealed practitioners’ strategies for overcoming resource constraints, such as synthetic data generation and fine-tuning global models with local data.
  • Evaluation practices remain tied to global benchmarks, highlighting a gap between local needs and international standards.
• Limitations and future work:
  • Resource constraints, such as GPU shortages and data scarcity, limit the scalability and competitiveness of local models.
  • Future research should explore the development of localized benchmarks, address global dependencies, and investigate the hidden labor involved in local model development.

Summary

This study investigates the development and deployment of localized foundation models in South Korea, focusing on the motivations, challenges, and contradictions inherent in pursuing AI sovereignty. Key findings highlight the importance of cultural and linguistic adaptation, regulatory compliance, and business flexibility, alongside significant challenges such as resource shortages and reliance on global benchmarks. While localized models address critical gaps in global systems, their development remains constrained by transnational dependencies and infrastructural inequities. The study reframes AI sovereignty as a situated sociotechnical practice and calls for future research to develop localized benchmarks, address resource constraints, and support context-sensitive AI systems.