Domestic AI: Who's Keeping Score?

The Slogan Arrived First

Read the government announcements and industry reports these days, and 'sovereign AI' and 'AI top three' appear in nearly every sentence. Yet walk into ten different conference rooms and ask what those words mean, and you will get ten different answers. For some, sovereign AI means a homegrown large language model that handles Korean well. For others, it means keeping data centers and chips on domestic soil. For still others, it is a security question about not handing public data over to foreign cloud providers. All of these are plausible — and all are different.

This is where the first misunderstanding takes root. The public and many policymakers understand sovereign AI as a question of ownership: 'let's get a chatbot of our own.' They see it as a race — build a native-language model, plant a flag, check where you rank on the global leaderboard. This reduces AI to a national-scale version of a personal productivity tool. A fight over who has the better hammer.

The essence of sovereign AI is not ownership but controllability. More precisely, it is the capacity of a society to examine, interrogate, and halt the workings of technologies that shape its own fate.

The Pipeline Analogy

To understand AI sovereignty, it helps to think of a model as a pipeline. What matters about a pipeline is not just who owns the pipe. The real questions are: where does the oil come in, what gets mixed in along the way, who can turn the valve, and who is held responsible if it leaks.

AI is no different. Into the model-as-pipe flows training data, and that data carries someone's biases and omissions. The valve is the authority to stop or repair the model; accountability is the question of who you can call when a flawed output harms a citizen. Laying a domestic pipe does not automatically resolve any of these four questions. Conversely, even with a foreign pipe, securing the valve and inspection rights preserves a substantial portion of sovereignty.

That is why the claim 'we've made it into the AI top three' is hollow without measurement criteria. Top three by what? Benchmark scores on model performance? Data center power capacity? Quality of native-language processing? Or the institutional maturity that allows citizens to challenge AI-driven decisions? Rank countries by that last criterion and the leaderboard looks nothing like the current one — and that leaderboard is, in fact, a far better proxy for a nation's governing capacity.

Why We Must Go Beyond Individual Use

The number of individuals who use AI skillfully is growing. They draft reports faster, write code, translate foreign languages. That is a good thing. But the ability of an individual to handle a tool proficiently and the ability of a society to understand that tool and embed it within institutions are entirely different muscles.

The latter is societal literacy: the ability to ask what went into the training data; the ability to know where to go when an automated decision is wrong; the ability, when told 'AI made that judgment,' to ask back — then who is responsible? Personal productivity cannot answer these questions. It is the difference between being good at swinging a hammer and being able to write building safety codes.

This is where the sovereign AI discourse becomes dangerous. If we convince ourselves that sovereignty has been purchased with a handful of domestic models and some data center capacity, the very literacy that society needs to cultivate gets pushed to the back of the line. It is like laying the pipe and never hiring an inspector.

The Blanks Korea Has Yet to Fill

Let us name the specific gaps. First, public data. Proposals in the vein of a 'K-ontology' — training domestic models on administrative, medical, and educational data — are attractive, but the language governing citizens' rights over that data, the scope of consent, and procedures for deletion remains rough. The will to aggregate data is strong; the grammar for returning it to citizens as something that belongs to them is weak.

Second, education. Coding curricula and prompt-engineering workshops have arrived quickly, but 'how to question an AI's judgment' has not entered the standard educational vocabulary. The nation has not even managed to agree on a list of the minimum AI literacy every citizen should have.

Third, accountability. As AI enters public administration, automated judgment will intervene in complaint handling, welfare screening, and risk prediction. When that judgment is wrong, is there a legally guaranteed procedure for citizens to demand an explanation and overturn the decision? This gap is the largest of all.

Here we must be wary of the ghost of the five-year plan. State-led mega-projects have an institutional inertia toward reporting outcomes in countable metrics — target years, budget outlays, number of models. Busan is a case in point. When talk turns to attracting data centers and establishing digital hubs, the figures for power, land, and jobs are crisp — but what rights citizens will hold over that infrastructure is blurry. The countable crowds out the uncountable.

The Counterargument, and a Response

One can push back this way: even if definitions are fuzzy, don't you have to plant the flag and commit resources to at least stay in the chase? Waiting for a perfect definition means missing the train. That is a fair point. In the technology race, speed genuinely matters.

But when terminology outpaces policy, resources flow only toward what can be measured. Without definitions there is no evaluation, and without evaluation budgets gravitate toward the countable — chips and buildings. So even for the sake of speed, benchmarks must be set first. If you sprint without agreeing on what sovereignty means, five years from now you will not even be able to measure what you have achieved.

Many countries will become skilled users of AI. But skilled use and genuine understanding are different things. A country that truly understands AI knows what it can and cannot control, and writes those boundaries in language its citizens can read. Whether sovereign AI remains a slogan or becomes a strategy will not be decided by how a model ranks — it will be decided by whether we can build the language to fill these blanks.