AI nearly tenfolds South Korea's digital sex crime caseload per worker

Cases processed per worker at South Korea's National Center for Digital Sexual Crime Response jumped from 20.8 to 202.1 a day after AI adoption, a roughly 9.7-fold rise the center attributed to automation at a Tuesday demonstration in Seoul. The system now handles each takedown request in about a minute or less, and the center processed 96,210 cases between April and September — more than twice the 45,438 it handled in all of 2025.

Behind those numbers sits a harder problem. A compliment under a teenager's video looks harmless. A question about the poster's age looks harmless too. But strung together, they can mark the start of online grooming, and that pattern is what the center wants its AI to catch next.

How does the system work?

The current AI collects posts around the clock using about 220 keywords drawn from frontline monitoring. It analyzes text and images together, then presents its assessment and the reasons behind it to staff for review. A human staffer approves the report. Then the system assembles the information the relevant platform or authority requires, cutting the submission down to roughly a minute.

"From detecting whether material has been circulated to submitting a takedown request, the system handles the process mostly automatically, without requiring human intervention at every step," said Kim Mi-soon, the center's director.

The automation does more than speed things up. It shields the people doing the work. Images are now blurred by default, and staff can reveal them only when necessary to make a reporting decision. That matters for a workforce repeatedly exposed to disturbing material, a burden the center says takes a toll on mental health.

Kim said the gains go beyond throughput. "AI has not only accelerated the process but is also detecting deepfakes that are difficult to identify with the naked eye," she said.

Why does this matter?

The stakes are visible in the center's own victim data. It supported 10,637 victims in 2025, up from 4,973 in 2020. Teenagers and people in their 20s made up 77.6 percent of those supported last year. Digital sex crimes in South Korea skew heavily young, and the center's monitoring capacity has not historically kept pace with the volume of material circulating online.

The automation push has let staff shift toward work that requires human judgment, while the machine handles collection, analysis and paperwork. The center plans to expand the system's coverage from 14 platforms to 139 by 2027 — a tenfold widening of its watch across the services where the material actually spreads.

What comes next: grooming detection?

The center's next target is conversation, not just content. An official walked reporters through the grooming pattern during Tuesday's demonstration.

"A teenager posts a short video and receives comments such as 'You're pretty' or 'How old are you?' These may seem like harmless compliments or signs of interest. But when they are used to build trust and steer the child toward private contact, they can mark the beginning of online grooming," the official said.

The technical challenge is that no single word gives the game away. "Because the individual words are not inherently suspicious, keyword searches alone cannot detect the pattern. That is why we're now developing technology that looks beyond isolated words to examine how an exchange unfolds and identify attempts to draw young people into sexual exploitation," the official said.

Alongside the contextual grooming detection, the center is building an AI chatbot. Its job: automatically send vulnerable young people information about support services, reaching them where they are rather than waiting for them to seek help.

What still stands in the way?

Faster detection does not guarantee removal. In 2025, 28.5 percent of the center's takedown requests went unfulfilled. Among websites whose server locations the center could identify, at least 95.5 percent were hosted overseas, which complicates enforcement beyond South Korea's borders.

Officials said technical improvements must be paired with international cooperation, stronger platform accountability and efforts to disrupt the revenue streams that sustain illegal sites. The AI can find the material and file the paperwork. It cannot force a foreign host to take a page down.

Who keeps the human in the loop?

Park Kwang-soo, head of the Women's Human Rights Institute of Korea, which oversees the center, framed the technological gains as a product of the people behind them. The advances, he said, reflected the expertise and dedication of staff members determined to eradicate digital sex crimes.

His warning cut against any reading of the numbers as a story about machines replacing people. "Whatever AI or technology we use, victim support must begin and end with people," Park said. "And people are needed at countless points in between."

That principle is built into the current system's design: every assessment the AI produces lands on a human screen, with the supporting reasons attached, before any report goes out. The blurring by default exists for the same reason — the machine absorbs the exposure so the person doesn't have to, until judgment is genuinely required.

The center's 2027 roadmap, with 139 platforms under watch, contextual grooming detection in development and a support chatbot in the works, points to a model where AI does the scanning and people do the deciding. Whether the model holds depends on the piece automation cannot fix: getting overseas hosts and platforms to actually remove what the system finds.