GPT-6 Astra, unveiled recently by OpenAI, stunned the world when it was used to crack a mathematical problem that had gone unsolved for years. Days later, an internal OpenAI system far more powerful than Astra produced a solution to the Navier-Stokes Millennium Prize problem, which had stood unsolved for roughly 90 years. Yet the change that deserves attention lies elsewhere: AI is no longer merely a tool that answers questions. It searches the internet on its own, operates computers and carries out complex, multi-step tasks. It has begun to substitute for human intelligence in search, software development, scientific research and professional work.
The South Korean government is moving fast as well. In August, the Ministry of the Interior and Safety announced its Strategy for Realizing an AI Democratic Government. It will roll out On AI, an intelligent work management system, to 47 central ministries by year-end and use AI for reviewing legislation and drafting reports. The plan also envisions analyzing public opinion with AI to propose policies and compare alternatives, and standardizing government data for retrieval-augmented generation (RAG). In September, it went a step further, pledging to expand the AI Government Lab, where civil servants build AI services themselves, and to train 20,000 "AI champions" by 2030.
What warrants close attention here is the "public GitLab." GitLab is originally a development collaboration platform for storing source code and development documents and letting multiple people edit and manage them together. The government has brought that approach into the public sector. Civil servants are required to register the documents, source code, prompts, data specifications and test results of projects they build at the AI Government Lab in a public repository of development outputs — the public GitLab. The idea is to let other civil servants see, use and improve ideas that would otherwise sit dormant on personal computers. It is already spreading quickly. Two months after the pilot began on July 3, the number of registered civil servants had grown to 2,073 and registered projects to 770. Of those, 376 have been made open for other civil servants to use.
The advantages are clear. A small work tool built by a deputy director at one ministry can spread rapidly to other ministries and local governments. There is no need for each agency to build the same program from scratch. The government plans to scale up outstanding projects across the entire government after verifying security, personal data protection and quality.
But that very advantage can become a risk. In a structure where good ideas spread quickly, insufficiently verified reasoning methods and errors are replicated at the same speed. In an AI government, development speed is not the only thing that matters. What gets verified at which stage, how far something is shared, and when to stop — those matter.
There is a particular reason for the South Korean government to be careful. It has a track record of success with e-government. It built systems quickly, spread them from the central government down to local governments, and set standards applied nationwide. This was the strength of Korean public administration. But the grammar of that success does not carry over intact to AI government.
E-government handled decisions made by people more quickly and conveniently. It issued civil documents online, connected information across agencies and turned administrative procedures digital. Judgment remained a human responsibility. AI is different. It finds information, filters materials, summarizes content and compares competing arguments to produce draft reports. It intervenes in what people see before they make a judgment. If e-government "processed" decisions, AI "intervenes" in judgment. That is where the problem begins.
When a civil servant reviews a company or an industry, in an AI government they will ask AI first, before reading dozens of articles and reports themselves. Summarize this company's major controversies. Compare overseas regulatory cases. Suggest three policy alternatives. The final decision is of course made by a person. But the information AI retrieves becomes the starting point of judgment, and the first summary sets the frame through which the issue is viewed.
This shift poses new homework for companies as well. Until now, corporate external communication has grappled with how people would read it. Companies polished messages, searched for persuasive numbers and picked examples that were easy to understand. Now one more question has been added: what does AI read when it searches our company? Companies need to check whether the numbers on their website differ from those in their sustainability reports, whether positions they have already changed still stand in old press releases, and whether dozens of articles surface from the time of an accident while the subsequent investigation findings and corrective measures are hard to find. This is not about domesticating AI to a company's advantage. It is the opposite. It means companies must manage the accuracy and consistency of the information they release to the world far more rigorously than they do now.
The risk of AI does not lie only in "hallucination," the fabrication of facts that do not exist. The thornier problem is the "selection of facts." An accident five years ago is a fact, and this year's corrective measures are also a fact. But select only the former and an entirely different picture of a company emerges. The moment a 100-page report is cut to five lines, important conditions and exceptions can disappear. Even when each number is correct, placing numbers with different reference dates and standards side by side can yield a wrong conclusion. Companies must therefore produce information that AI finds hard to misread.
The greater responsibility lies with the government. It is dangerous if a government accustomed to the success of e-government starts competing over adoption speed, user counts and the number of development projects on the grounds that "we are first again this time." In a structure like the public GitLab that replicates and spreads ideas quickly, the moment verification speed fails to keep pace with the speed of diffusion, one agency's small error can become a government-wide error. In the e-government era, system errors generally ended in administrative inconvenience. In an AI government, the wrong selection and interpretation of information changes policy judgment. Decisions on regulation or permits can turn out differently. Administrative action or investigations could be launched on the basis of wrong information. The direction of regulation and the targets of support and procurement could change.
An AI error at a private company may end in a bad recommendation, but an AI error at the government becomes an error of state power. The truly dangerous moment is not when AI's answer is absurd but when it is plausible. The prose is smooth, the logic is orderly, and sources are even attached. People stop checking from the beginning. This is what is known as automation bias.
The government is not unaware of these risks. It has said it will conduct additional verification and confirm security, personal data protection and quality before applying AI projects to actual administration or converting them into public-facing services. Public AI ethics standards and impact assessments are also underway. The direction is right. But it must be stricter. If AI has been used in an administrative judgment, it must at minimum be possible to answer three questions. What materials formed the basis of the answer? How did the official in charge verify it? And if wrong information became the basis of a judgment, how can citizens and companies correct it? At the very least, the black box of the AI judgment process must be eliminated from the policy process.
The South Korean government succeeded in e-government through speed. But the formula for success in AI government is different. In the e-government era, connecting more was innovation. In the AI government era, stopping and checking is sometimes innovation. How many civil servants use AI and how many services have been built must not become the measure of achievement. How quickly an error is detected when AI gets something wrong, and how decisively a flawed judgment can be halted before it spreads to other agencies — that is the true competitiveness of an AI government. A good AI government is not the one that uses AI the most. It is the one that best knows how to doubt the plausible answers AI produces.
