A developer in his 20s, identified as A, recently took part in a Korea Advanced Institute of Science and Technology (KAIST) study observing how suspending artificial intelligence (AI) use affects work — and ran into major disruptions. Although the AI-free period lasted only four days, he struggled to gain momentum from the very first stage of grasping the context of a problem. "My work slowed from the early stage of understanding the context of the problem," A said. "After adopting AI for my work, I realized I had come to rely on AI for a significant portion of my working hours." He added, "I felt like I had become a developer who couldn't even catch a simple bug without AI."
As individuals, companies and public institutions ramp up AI use, concern is growing that core operations and public services could become dependent on the performance, outages and access policies of specific foreign AI models, according to the information technology (IT) industry on Aug. 22.
Korea is one of the countries adopting AI fastest in the world. According to the Ministry of Science and ICT's 2025 survey on value-added telecommunications businesses, 78.1% of 2,500 adults aged 19 to 69 had used generative AI in the past three months. Among those in their 20s, the figure reached 92.6%. The problem is that usage is concentrated on specific foreign models. In the same survey, the most-used generative AI was ChatGPT (68.1%), followed by Gemini (13.8%). Domestic models did not make the top ranks.
The concentration is even more pronounced in corporate settings. Over the past two to three years, major Korean companies have accelerated enterprise-wide AI transformation (AX), using generative AI for drafting documents, summarizing meeting minutes, translation, code generation, customer service and internal knowledge search. The problem is that much of this AX is being built on top of foreign models from global Big Tech players such as OpenAI, Google and Anthropic.
Given performance, stability, developer ecosystems, cloud integration and global references, the reality is that companies find it hard to avoid foreign AI. But if access to a specific model is cut off or an outage occurs, productivity can be severely shaken. During last year's ChatGPT access outage, users who relied on AI for writing reports, coding and searching for materials reported work disruptions.
Such concerns have recently spread into the security domain as well. On Aug. 12 (local time), the U.S. government issued export controls fully blocking foreign nationals' access to Anthropic's latest models, "Mithos5" and "Fable5." Mithos is a high-performance model notable for its ability to automatically detect cyber vulnerabilities. In Korea, the Korea Internet & Security Agency (KISA), along with Samsung Electronics, SK hynix and SK Telecom, had joined Anthropic's security cooperation body "Project Glasswing" to secure access, but the controls came down about 10 days after joining, blocking access before they could properly use the model. It has been revealed that access to foreign security AI can be shaken at any time by a single policy variable.
Going forward, if a company designs its vulnerability detection, source-code review and breach-analysis systems on the premise of a high-performance security AI like Mithos and then loses access, the situation could require not just swapping out tools but re-examining security inspection methods, contracts and system integration. However, Anthropic believes the current controls will not last long. Chris Chaurhi, head of Anthropic International, said at a press briefing for the opening of its Seoul office on Aug. 17, "I am confident that the suspended Mithos use will resume within days," adding, "The export controls will not be maintained for an extended period."
Another concern is that while the U.S. controls access to its own models, China is moving quickly to seize the role of an alternative. China's Z.ai has open-sourced its GLM-series models and is focusing on attracting global developers and startups with a low-price strategy.
For these reasons, experts agree that developing domestic models is urgent. The government has set out to foster domestically built foundation models under the banner of "sovereign AI," but in corporate settings, a preference for foreign models remains strong in terms of performance, stability and ecosystem. Choi Ki-young, president of the Korean Society of Semiconductor Engineering, noted, "High-performance models like Mithos5 can be used to hack other countries' systems to steal critical information or cause malfunctions," adding, "One country monopolizing such a model poses a major threat to others." He continued, "Countries with independent AI capabilities — China as well as France and Canada — will invest more in their own development," and stressed, "Korea, too, must accelerate the development of a top-tier independent foundation model."
