Executive Summary

Key Figures

Sources: ZDNet Korea (November 6, 2026), Electronic Times (April 10, 2026)

Three of the four numbers come from the pooling side and one from the service side. The first three count bodies linked up and data registered; the fourth counts how far along the work of turning that data into something citizens can use has got. The units differ, so the sizes cannot be set against one another. What can be compared is what each one counts, and none of the four counts whether a question a citizen asked got an answer. That blank is where this article starts.

The third plenary meeting of the National AI Strategy Committee, held at Seoul Square on November 6, was the first full meeting Ha Jung-woo chaired after taking office as standing vice-chair. About 60 people, government and private-sector members and heads of related agencies, gathered and approved three items: a legislative framework for artificial intelligence in Korea, a plan for forming and running the committee task forces behind the national AI core tasks, and a report on the state and use of the cross-government data pipeline. The third item is the one this article follows.

Bae reported that each ministry is running about 30 AI transition projects built on public data, and that work is under way to connect them so citizens can reach those services through Moduui AI. He added that connecting the individual datasets is no small task. Moduui AI is due in mid-December, and he has described it as a first step toward the agent era rather than a generative chatbot.

In the same room Cho Jun-hee, chairman of the Korea Software Industry Association, took up the other side. He chairs the committee's industrial AX and ecosystem subcommittee. Securing data and developing a service with that data, he said, are separate jobs. For Moduui AI to become something other than an overseas big-tech service, it has to use public data to actually produce better answers, and that requires each ministry to have development capacity, budget and staff of its own.

This is the example Cho raised at the meeting. When a citizen asks whether health insurance will cover another dental cleaning, what does an AI need in order to answer? At first glance the treatment record looks like enough. The date of the last cleaning is written there.

The date, though, is not the answer. It is the material for one. What tells you where to hold that date against is the health insurance benefit rule: how long before the procedure becomes eligible again, whether age or condition changes the terms, and when the clock restarts. None of that is in the treatment record. It is written in the policy document. Both have to be in one place before yes and no can be told apart.

The example is useful because it shows what is missing in one go. A million more treatment records will still not produce the answer so long as the benefit rule is absent. With one line of rule beside them, the records already on hand are enough. The ability to answer a question turns on whether context has been attached, more than on how much has been gathered.

Another remark at the meeting reached the same point from a different angle. Jeong Eun-kyeong, Minister of Health and Welfare, said a standard is needed for linking data across bodies while protecting personal information. Bodies such as the National Health Insurance Service and the Korea Social Security Information Service are still working out how far they will share. The very data behind the dental cleaning example is, in practice, caught on the question of standards.

The gap is not there because the pooling side is thin. The opposite is true. The third agenda item at this meeting sets out the scale. The cross-government data pipeline is a different thing from the Public Data Portal, which opens data to the public. It is a road laid inside government so that one body can find and draw on another body's data, which is why the unit for measuring its scale is the number of bodies served. The Ministry of the Interior and Safety opened a pilot in September, and the pipeline now reaches 848 public bodies. As of last month, 54,768 shared datasets and 2,442,269 metadata items were registered on it. The law behind it is Article 8 of the Act on Promotion of AI and Data-Based Administration.

The way to find those datasets was laid this spring. The National Shared Data Platform, built in April 2026, lets bodies search for and obtain data held elsewhere, and an AI search function shows which body holds what. Training data moved on a similar schedule. The Ministry of Science and ICT and the National Information Society Agency decided to survey every AI training dataset held by public institutions and source 100 with high potential for use, and to spend 6 billion won on an integrated delivery system that will remake AI Hub, which has already opened 903 datasets. That plan was made public in April this year.

On the numbers alone, the preparation looks finished. That is exactly where Cho's point lands. A scale of 848 bodies and 50,000-odd datasets tells you data has been pooled. It does not tell you which questions that data can answer. The two are different kinds of maturity, and the first one growing does not make the second grow with it.

The government is not blind to the gap. The future plans in this agenda item include cleaning and combining the data that many bodies need in common and delivering it as a single dataset. Cross-ministry policy issues such as welfare, youth employment, regional decline and disaster safety are the targets, and quality verification results, worked examples and usage guides are to be supplied with the data. That amounts to recognizing the work of attaching context as a task of its own. The question left over is how large a share that work gets.

The number of bodies linked up, the number of datasets registered, the number of datasets opened are all easy to count. They go into a year-end report as a rise on the previous year, and anyone reading knows what they mean. Which questions the data can now answer is hard to count. The list of questions has to be settled first, and a separate standard is needed for judging whether an answer was right. Budgets and performance targets crowd onto the easy side not because anyone is lazy but because the cost of measuring differs.

This tilt has shown up in numbers once already. In the Ministry of the Interior and Safety's 2025 evaluation of public data provision and operation, 684 bodies scored 89.5 on management systems and 59.2 on openness and use, while 'AI-friendly, high-value data opening', an indicator added that year, came back at 37.8. The same bodies scored close to 90 on management and fell below half on putting data out in a form AI can use. That evaluation is covered in Korea's public data evaluation scores management high and AI-Ready data low.

Two other remarks at the meeting point the same way. Bodies raised a worry that they could end up carrying the blame if something goes wrong after they hand data over, and the view was added that this needs a legislative fix. Lee Hyoung-il, deputy prime minister for the economy and Minister of Economy and Finance, proposed supporting bodies that work hard at sharing data with budget and staff. Behind the proposal is a reading that under the present structure a body putting data out carries the risk and receives none of the reward.

How the committee runs changed at this meeting too. It moves from subcommittees to projects, with a deliverable and a deadline set for each task, and seven new core-task task forces. Three report directly to the standing vice-chair: youth advancement, the national response to superintelligence, and strategic global AI cooperation. Four are tied to subcommittees: strengthening national cyber resilience, AI innovation in cultural content, building a data flywheel, and building a K-AI full-stack ecosystem. Delivering results that citizens and companies can feel was given as the reason for the shift. Whether something is felt is judged, in the end, by whether a question got an answer.

The place the dental cleaning example points at exists outside government data too. The scene we run into most often, in the work of getting data into a usable state, has this same shape. The warehouse holds plenty of records, and the rules that would turn those records into a judgment are scattered in someone's head, in an approval document, or in an old spreadsheet. Changing the model does not move that spot.

This is what AI-Ready Data points at. The test is not how much data has been gathered but whether the standards, definitions and history needed to answer a question are there beside it. Generalized, the relation between one column of treatment dates and one line of benefit rule says this: a value and the rule that interprets it have to be in the same place. We have followed the same distinction twice before: over research data in opening data is not the same as making it usable, and over the terms of use on opened public data in the license change behind Korea's 100 public datasets.

Three questions give a rough fix on where the data your organization has pooled stands.

• Can you write down ten questions you would actually put to your data? Write them out and about half of them, usually, get no answer from the data you have now.
• For each question that gets no answer, can you tell whether what is missing is a record or a rule? If a record is missing, gather more. If a rule is missing, gathering more will not help.
• Does a system know where that rule is written, rather than one person? A rule that disappears when the owner changes jobs is not attached to the data.

None of the three asks how much data you hold. They ask what you can judge with what you hold. While performance targets count only the first, the second goes uncounted, and attaching AI in that state only lengthens the list of questions that go unanswered.

Thank you for reading this far. The meeting is written up in ZDNet Korea's report from the floor. If your organization has a question that stalled because the data was pooled and the answer still did not come, we would like to hear whether what was missing was the record or the rule.