What happened when AI directed the 1996 Gangneung manhunt

Jang Seok-kwang

The author is the general secretary of the Academy of National Intelligence.

AI is emerging as a new arena of national security competition. At the United Nations General Assembly on Sept. 22, U.S. President Donald Trump rejected international controls on AI and stressed U.S. leadership. Two days later, Chinese President Xi Jinping said at his summit with Trump that the two countries should cooperate to prevent AI risks and misuse even as they compete.

What could AI change for intelligence agencies and battlefields? I revisited an incident from 30 years ago.

Before dawn on Sept. 18, 1996, a North Korean submarine ran aground off Gangneung. It belonged to the 22nd Squadron under the Maritime Department of the Reconnaissance Bureau and was carrying 26 people at the time: the department chief and deputy chief, 19 crew members, two guides and a three-man reconnaissance team.

Unable to return by sea, they tried to escape on land. South Korea launched a counter-infiltration operation involving 1.5 million military personnel, reservists and police. Most of the 26 North Koreans aboard the submarine were killed, either by their own comrades after the vessel ran aground or in clashes with South Korean forces; one was captured. South Korean soldiers, police, reservists and civilians were also killed, and dozens were wounded. The 49-day operation ended on Nov. 5, when the last two armed infiltrators were killed in Inje.

Thirty years later, what would AI conclude? I asked two questions: “What did [the North Koreans] come to do?” and “How would the 49-day search differ with today’s technology?” In both experiments, the infiltrators, terrain, weather, forces and intelligence environment remained as they were in 1996. Only South Korea’s available technology changed.

For the first experiment, I gave AI only field evidence and objective facts available by Nov. 5, 1996, excluding government assessments, press releases and expert commentary.

AI found that this was not simply an armed infiltration but a specialized maritime infiltration and intelligence-gathering operation. The plan was to secretly land personnel by submarine, collect military data, transmit the information through communications equipment and recover the team by submarine.

Evidence included the reconnaissance and guide teams; film showing the Yeongdong power plant and Gangneung airfield runway; maps; reconnaissance gear; communications equipment; and code tables. The presence of senior Maritime Department officials was another significant clue. AI concluded that heavy weapons, including RPG-7s, were insufficient to establish that a specific attack or sabotage mission had been planned.

That reminded me of teaching intelligence analysis at a national intelligence university more than 20 years ago. When students grew sleepy, I asked questions to wake them up. Why had all the Gangneung infiltrators undergone appendectomies, but none had been circumcised? Why was one bullet invariably found in the breast pockets of dead infiltrators, and sometimes two?

Those questions were intriguing but relatively simple. My second question for AI was more complex: How would we find them?

In 1996, troops, police and reservists physically combed mountains and valleys. When new traces appeared, investigators analyzed them and searched again. Now, search teams could add satellite imagery, drones and thermal cameras while incorporating terrain and weather data.

AI could connect information arriving from different sources, calculate likely movement routes and continually reorder search priorities as new evidence emerged. The model could shift from “people searching a vast area” to “AI narrowing down the most probable areas, and people verifying them on the ground.”

AI’s conclusion about the infiltration’s purpose did not differ greatly from assessments by expert agencies at the time, including what was then the Agency for National Security Planning. Even without access to their internal intelligence, AI could mimic the judgment of experienced analysts.

The second experiment did not prove that modern technology could reduce a 49-day search to several days. It did, however, demonstrate a different approach: recalculating priority search areas whenever new information arrives.

AI’s strength is not merely processing enormous amounts of data quickly. It can connect scattered information, identify meaning and suggest what humans should examine next.

That change is already visible in war. Ukraine combines Delta, a battlefield information system integrating data from satellite and drone imagery, radar and ground forces with the AI-based Avengers image-analysis platform to detect tanks, artillery and air defense systems. Processes that once took hours or days from target detection to strike can be compressed dramatically.

Russia is also deploying Svod, a tactical situational-awareness system that integrates multiple reconnaissance feeds, builds a battlefield picture and links it to strike systems. AI is widening Russia’s field of vision as well as Ukraine’s.

And my answers to those classroom questions? Appendectomies reduced the risk of acute appendicitis during operations in North Korea, and the absence of circumcision reflected the North’s social and cultural practices. One bullet was for suicide, but two suggested a commander prepared to shoot a subordinate who hesitated. Simple answers, perhaps, but grounded in analysis of many cases.

A single paulownia leaf can signal autumn, and ice forming in a jar can reveal the cold across the land. Intelligence is about reading large changes from small signs. Thirty years ago in Gangneung, we were searching for those traces. AI connects that scattered information, finds meaning and reads changes not yet visible.

Then it asks what we should look at next. That is where AI may become a new tool of intelligence analysis itself.