Fairness in Child Safety AI: Why Demographic Parity Audits Are Not Optional

There's a particular failure mode in content moderation AI that the industry doesn't talk about enough: the system works, on average, but it works badly for specific groups. Keyword filters disproportionately flag African-American Vernacular English. Toxicity classifiers flag LGBTQ+ content at higher rates than equivalent heteronormative content. Spam detection penalizes non-native English speakers. These failures are documented, reproducible, and — when they happen in a child safety context — c