South Korea's AI system for finding child sexual abuse material lifts staff throughput nearly tenfold
The state digital sex-crime victim centre says the system, in use since April and trained on about 40,000 posts and images, raised daily cases per worker from 20.8 to 202.1 and helped remove 96,210 items from April to September; staff still make the final call, and 28.5% of removal requests were ignored last year.
Why it matters
Most AI-safety coverage is about what models might do; this is a government reporting measured results from using AI against a specific online harm. Removals supported in six months, 96,210, were more than double the 45,438 handled manually in all of last year, and the system blurs images to reduce staff exposure. The limits are on the platform side: removal requests are increasingly ignored. Watch the planned expansion from 74 to 139 monitored platforms and the move to detecting grooming from conversation context.
Key facts
- In use since April 2026
- Trained on about 40,000 posts and images
- Daily cases per worker: 20.8 to 202.1
- 96,210 removals supported April–September
- 28.5% of removal requests ignored last year
- Platforms monitored to rise from 74 to 139