Exclusive: Foreign Models Dominate Government AI Platform, Contradicting Sovereign AI Policy

In an exclusive report published on October 5, Chosun Ilbo said South Korea's public-sector AI platform is dominated by foreign models, contradicting the government's stated Sovereign AI policy. The platform, called the "Pan-Governmental Super-Sized AI Common Infrastructure," is meant to let central and local agencies share AI models and GPU resources and streamline administrative work.

Citing data submitted to Rep. Kim Jang-gyeom of the People Power Party by the Ministry of the Interior and Safety and the Ministry of Science and ICT, the report said only 12 of the 31 models introduced to the platform were domestic as of last month, while 19 were foreign. Among the 11 models already deployed, seven were foreign. The remaining models can be activated at any time, it added.

Samsung SDS's "Fabrix" and Naver Cloud's "CLOVA Studio for Gov" are the two platforms currently operational. The report frames the composition of the model list as a direct contradiction of the government's push to reduce reliance on overseas AI.

Rep. Kim said the government launched a homegrown AI foundation model project citing risks of technological and economic security dependence, and concentrated government GPU support. Yet after Naver Cloud was disqualified in January's first evaluation for using Alibaba's model weights, deemed "lacking originality," the common AI platform for public servants showed no preference for domestic models.

Security is a central concern. The list of 31 models includes "Bge-M3," developed by the Beijing Academy of Artificial Intelligence (BAAI), which the U.S. Department of Commerce added to its Entity List in March last year. The U.S. cited BAAI's unauthorized acquisition of controlled items, including high-performance AI chips, and its development of large-scale AI models and advanced computing chips for defense purposes.

The Ministry of Science and ICT responded that Bge-M3 is an embedding model that converts text into vectors rather than a generative AI model, and that it operates within a physically isolated, secure administrative network to prevent data leaks. The report notes, however, that embedding models are typically used in the retrieval-augmented generation (RAG) stage to find relevant laws or official documents, and that if manipulated they could distort search results, while verifying weights for tampering is practically impossible.

Cost is another question. The largest model on Samsung SDS's platform is Upstage's "Solar Open2 250B," while Naver Cloud uses NVIDIA's "Nemotron 550B." Rep. Kim's office said Nemotron requires extensive computational resources for lengthy reasoning, potentially increasing operational costs, and noted that the ministry has not disclosed token-based pricing, flat-rate fees, or cost calculation criteria.

The ministry admitted there were no criteria to evaluate the country or institution behind AI models during the platform's development, saying models were selected by platform providers based on universality, search performance, multilingual support, and usability, and that the government does not restrict models by country.

"The ministry said nothing about the model's developer being on the U.S. sanctions list," Rep. Kim said. "If they claim to have comprehensively reviewed security, we must ask: What exactly was reviewed?" The episode highlights the gap between sovereign-AI ambition and procurement practice, and puts pressure on Seoul to define how model origin and supply chain are vetted.