New AI System Spots Bad Ads It Has Never Seen Before

Today's content moderation usually works like a single overworked employee: one giant AI looks at a post's image, video or audio and immediately decides pass or fail. When a platform changes its rules — say, new limits on AI-generated political ads — the whole system often needs retraining with thousands of fresh examples. Worse, unlike text, you can't easily create fake versions of a video or photo to practice on. And when the AI blocks something, nobody can really explain why.

This paper splits that job in two. A "Content Model" watches the post and writes a plain-text description: what's shown, what's claimed, what the tone is. A separate, text-only "Policy Model" then reads that description alongside the rulebook and makes the call. Think of a detective writing up a report, and a separate reviewer applying the law to it. Because the second half only deals with text, researchers can cheaply generate thousands of fake summaries — including tricky borderline ones — to stress-test a rule before launch. The content model is also nudged in an automatic loop to write summaries that actually matter for policy.

On misleading-advertising detection, the system was about 24% better at correctly clearing ads that aren't misleading compared with a standard approach. The standout result: a version trained on zero real violating examples — every bad example was synthetic — matched the full-data model almost exactly (within 0.2%). In plain terms, a brand-new policy could go live on day one without waiting to collect real violations.

The catch is that everything hinges on the written summary. If it misses sarcasm, cultural context or a subtle threat, the judge downstream can't catch what it never sees. Errors in step one quietly become wrong decisions in step two. It's also a research result on ads, not a shipped product — so treat it as a promising direction, not a fix for moderation problems overnight.