The Korean government‘s “Artificial Intelligence (AI) Foundation Models” competition, launched in June last year, has recently been embroiled in controversy over “technological independence”. The competition aims to build AI models entirely developed by Korean technology, and plans to select two local products with performance comparable to international leading levels by 2027. However, among the five finalists, three have already been eliminated using foreign open-source components, many of which are highly similar to Chinese open-source models.
According to the Wall Street Journal, the controversy first focused on the nominee company Upstage. Gao Xixuan, chief executive of local rival Sionic AI that did not enter the top five, publicly questioned on social media that some components of Upstage‘s AI model were identical to the Chinese open-source model of Spectro AI, and that Spectro AI‘s copyright markings were retained in the code, referring directly to its “submission of a suspected fine-tuning version of the Chinese model to a national project.” Upstage then released its development log in live form, proving that the model training “started from scratch,” but acknowledged that the reasoning code that ran the model did indeed refer to the open-source components of Spectro AI, and emphasized that the relevant modules had been widely applied globally.
As the Upstage controversy festered, other nominees were also subjected to stricter scrutiny. Reports revealed that Naver‘s visual and audio encoders were identified as similar to Alibaba and OpenAI products. SK Telecom‘s inference code structure was suspected to be similar to DeepSeek.
“The goal of the national artificial intelligence project is to measure the performance of self-developed AI models for testing, but in the case of Naver, the performance of the Ali Qianwen ‘visual encoder’ accounts for a large part. This is in fact no different from evaluating Chinese artificial intelligence technology, which is unreasonable,” the Korean Times quoted industry insiders as saying. The Korean government emphasized in its national AI program that its goal is to ensure the realization of “self-developed artificial intelligence” by constructing AI base models “from scratch.” Industry insiders questioned whether such a practice of importing foreign modules as core components really meets the original purpose of the assessment.
In response, all three companies denied “shell development”. Naver explained that citing external coders was “strategically adopted as a proven technology” and emphasized that the “core engine” that determines the model learning mechanism was completely developed independently by the company. SK Telecom presented a similar position, emphasizing that the core of its model was built independently. The focus of controversy then extended to the competition judgment criteria themselves. South Korea‘s East Asia Daily reported that although the government demanded that the competing models be “trained from scratch”, it did not set specific judgment criteria, leading to differences in the industry‘s definition of “autonomous AI”.
Despite the ongoing controversy, the South Korean government continues to push for this competition. According to the Associated Press, South Korea‘s Minister of Science, Technology, Information and Communications, Pei Kyung-hoon, publicly requested that the judging process remain fair and transparent to ensure that the project goals do not deviate from the original intention of “autonomous AI”. On 15th, South Korea‘s Ministry of Science, Technology, Information and Communications announced that two AI alliances led by Naver and NC were eliminated. The remaining 3 teams were led by SK Telecom, LG and Upstage, respectively. The industry predicts that regardless of the election results, the controversies that will ensue will be difficult to avoid.