Every time a child safety system flags the wrong person, trust in the entire system erodes. A teenager falsely banned from a platform they use to talk to friends. A teacher wrongly suspended from an educational tool. An adult gamer kicked out of a community they've been part of for years. False positives in child safety moderation are not just technical errors. They're injustices that fall disproportionately on specific groups, create legal liability, and undermine the social license that makes any safety system viable long-term. This post is about the false positive problem in child safety AI — what causes it, how different system architectures handle it, and why we at SENTINEL made specific engineering choices around it. Two categories of false positives Child safety AI has two distinct false positive problems that are often conflated: Statistical false positives — the model is wrong on individual cases. Every classifier has a false positive rate.…