Google's own AI flags a scam ad its human reviewers approved twice

A developer noticed a YouTube app ad disguised as an iOS system alert — a fake “iPhone Storage is Full” popup with mock “Yes”/“No” buttons that trigger an app install no matter where you tap. He reported it, as did several other users, and each time Google replied that the ad didn’t violate its policies. The obvious tension: deceptive ads that exploit users tend to convert well, which is not much incentive for the platform serving them to take them down.

The sharper point is technical. When the same ad was run through Google’s own Gemini model, it was rejected in seconds, cataloging three distinct policy violations: mimicking system UI dialogs, using non-functional fake interface controls, and deploying fear-based scare tactics to coerce clicks. Gemini even recommended disapproving the creative and flagging the advertiser account for possible suspension. The author leans on Hanlon’s razor — human reviewers simply can’t inspect everything at scale — but the contrast is hard to ignore.

The broader significance is an accountability gap. Google possesses ad-review AI capable of catching exactly this class of abuse instantly, yet its live enforcement pipeline let a textbook deceptive ad through twice and defended it on appeal. It’s a concrete example of a company holding a solution it declines to apply where it would cost revenue.