During the single winter season of 2020–2021, 30 million chickens were culled. It was because of highly pathogenic avian influenza (HPAI). The culling led to compensation payments and pushed up chicken and egg prices. Higher prices for eggs, a key ingredient, feed into higher prices for processed foods and dining services.
During the winter season (2025–2026), there were also 62 HPAI cases in Korea. Global warming is changing migratory bird routes. Outbreaks used to be frequent mainly along the west coast, but they are now spreading inland. The virus itself is mutating. If left as is, annual quarantine expense in the hundreds of billions of won, as well as blows to the domestic market, will be unavoidable.
Couldn't this problem be addressed by using artificial intelligence (AI) to craft effective countermeasures? Hong Seung-gil, Head of Team for the Plant and Animal Big Data Team at the Animal and Plant Quarantine Agency, and Gu Reum, CEO of the data tech corporations Bigvalue, built an AI-based risk prediction system together. They organized the limits, tasks, and solutions they experienced during development with public AX (AI transformation) and published National Intelligence. On the 2nd, we met the two authors to hear their explanation of the HPAI risk prediction model and asked about the direction of public AX.
CEO Gu emphasized AI-tailored data production and management. Gu said, "Korea's level of public data openness ranks near the top among OECD countries. There are tens of thousands of datasets registered on the data portal," adding, "But the share of data that is actually usable is very low." Gu particularly noted, "Data that AI can read and use is only a tiny fraction of the whole," and pointed out, "There is no interface to link data released by one ministry for use with other ministries."
Head of Team Hong said it is important to decide "how to view artificial intelligence." Hong said, "Do not see AI as a mere tool; see it as a colleague," adding, "You need to carefully design the process structure so that you decide what to entrust to AI and from where to take back full authority and handle it directly." Hong retired from public office on the 30th of last month.
The following is a Q&A with the two authors.
─There are many animal and plant epidemics. Why did you develop the HPAI risk prediction model first among them?
Hong Seung-gil (hereafter Hong) "Among livestock products, eggs are distinctive. Most countries produce only as much as they demand. Many countries produce more pork or beef and export it, but eggs are produced domestically and, when production plunges due to HPAI and the like, they are imported only then. In Korea, HPAI occurred frequently starting in the 2020s, and domestic egg supply and demand faced difficulties. The need grew for a system that could predict this and respond quickly."
─What was the project's goal?
Hong "Until now, when HPAI occurred, we moved according to the response manual. We decided to break that mold. Rather than post-incident response, we aimed to build a system that alerts of risk beforehand and allows preparation."
─Specifically, did the method or structure of culling change with the introduction of the risk prediction model?
Hong "Until now, culling has been used as a means to stop spread. But now we first run in a virtual space the questions of 'An infection has occurred—where is the virus heading? If we block key nodes on the route, how much can we stop? What combination yields the maximum blocking effect with minimal intervention?' and then make the (culling) decision."
─What is the effect of this?
Gu Reum (hereafter Gu) "First, the number of farms culled solely because they are nearby decreases. It is also positive for the national budget. Reducing unnecessary preemptive culling lowers compensation expenditure."
─How much of a reduction in culling compensation do you expect?
Gu "Our analysis shows more than a 20% reduction is expected."
─There are concerns that climate change could alter migratory bird routes or even the types of epidemics.
Gu "Due to global warming, diseases are occurring even in the United States, which used to be a clean zone for avian influenza. Volatility has increased."
Hong "The impact of climate change is especially pronounced in plant quarantine. From the plant's perspective, pests are akin to viruses, and they are moving rapidly between countries on rising air currents. Pests that did not come before are entering Korea, and the problem is serious."
─When situations shift so quickly, doesn't the lack of existing data make prediction harder?
Gu "It depends on how you do the modeling. When you run model experiments to identify causal relationships, you end up digging into the root causes of the variables you assumed as causes. As you study migratory routes of birds, which are the cause of avian influenza, you come to understand the weather conditions along those routes and then incorporate that as a variable."
─What is the accuracy level of the avian influenza risk prediction model you have built?
Gu "Other countries have built models, but ours is far more accurate. The performance of a risk model is measured by how many percent of actual outbreak farms fall into the top 10% of high-risk farms selected. Our model recorded 56.8% accuracy. That means we selected only 10% of all farms, yet more than half of the actual outbreak farms were in that group. If we broaden the group to the top 20%, accuracy rises to 75%. On a regional basis, 95.5% are included in the top 20%. The European Food Safety Authority's model captures 73% when selecting the top 33% of regions."
─In the book, you mentioned difficulties in securing data during AI model development. In particular, many datasets that seemed meaningful were excluded.
Hong "We thought that if we could secure data on traditional markets where live poultry is traded, we could improve the accuracy of the risk prediction model. But there was no system to collect that data. Some markets recorded by hand, some used Excel files, and some had no data at all. The inconsistency problem emerged. There was no guarantee the data would continue to be collected in the future. We considered sustainability and ultimately excluded it. We took into account that if you adopt data without guaranteed sustainability as a key variable, the model itself could be shaken if the data stream later breaks."
─What data did you ultimately use?
Hong "Out of thousands of datasets, we actually used only about 60 as inputs. Through epidemiological investigation reports, AI identified the causal relationships of avian influenza. We then organized four pillars—environmental factors, quarantine factors, transmission factors, and seasonal factors—as variables."
─Public AX is a hot topic, but there is a limitation that goals vary depending on which ministry holds the reins.
Gu "Of course. In quarantine, migratory birds are subjects of control, but from an environmental perspective, they are to be protected. Disrupting migratory birds or damaging stopover sites is considered ecosystem destruction. If each ministry develops AI that suits its own purposes, they will reach conflicting judgments. If it is an integrated system, the entire system will descend into confusion."
─How do you prevent that?
Gu "The top decision-maker must set priorities and decide. Depending on the situation, you have no choice but to weigh quarantine and ecosystem protection differently. This is ultimately a political process."
─Any tips for making national AX run smoothly?
Hong "Do not see AI as a tool; see it as a colleague. You need to design the process structure to decide what to entrust to AI and from where to take back full authority and handle it directly."
─You also propose exporting AI systems in the book.
Gu "Once public AX is completed with national funds, we can export it and create new value. If our system becomes the global standard, it can also help enhance the national image. There is no international standard yet for livestock quarantine AI systems. If Korea is the first to present a validated system on the international stage, that will effectively become the standard."
☞ National Intelligence
"National Intelligence" is a public AX empirical record that, centered on the case of the highly pathogenic avian influenza (HPAI) risk prediction model built jointly by Bigvalue and the Animal and Plant Quarantine Agency, addresses the conditions for Korean administration to move beyond "AI adoption" to the "AX (AI transformation)" stage. It unpacks the methodology of AX—which goes beyond simple automation (RPA) to change the very decision-making structure and operating method of administration—through real operational cases. The two authors' research results and insights can be heard on the 9th at the Administrative AX Forum hosted by ChosunBiz.