Automatically moderate user-generated ad copy

WHAT THIS TESTS: It probes whether you can build a practical moderation pipeline and reason honestly about the limits of fixed rules versus learned models in a compliance setting.

A GOOD ANSWER COVERS: Use a layered design. Stage one is a cheap, fast rules layer: regex and curated keyword or blocklists that catch unambiguous profanity and known banned phrases instantly and explainably. Stage two is a machine-learning text classifier trained to flag policy violations, profanity, hate, or non-compliant claims, capturing context and obfuscation that lists miss. Combine confidence scores: clearly clean copy auto-approves, clearly violating copy is blocked with a reason, and a mid-confidence band routes to human reviewers whose decisions feed back as training data. Keep an audit log of every decision and let the policy set update without redeploying the whole service.

THE TRADEOFFS: Regex and keyword lists are fast, cheap, deterministic, and easy to explain and update, but they are brittle, trivially evaded by spacing or character substitution, blind to context, and prone to false positives like the Scunthorpe problem where a clean word contains a banned substring. An ML classifier understands context and catches novel or obfuscated violations, but it needs labeled training data, is harder to interpret, can drift as language changes, and adds latency and serving cost.

LIKELY FOLLOW-UPS: How do you handle the Scunthorpe false-positive problem in practice. How do you keep policies current as ad networks change rules. How do you measure precision versus recall and which do you favor for compliance. How does the human-review loop feed back into model retraining.

ONE CONCRETE EXAMPLE: A user submits copy with a non-compliant 'guaranteed weight loss' claim. The regex layer, looking only for banned words, misses it entirely, while the ML classifier flags it as a prohibited health claim and blocks it with a clear reason. A borderline rephrase scoring mid-confidence is sent to a human reviewer, whose verdict is logged and used to train the next model version, steadily improving recall.