M39B: Deciding between Recall vs. Precision for Meta's Anti-Scam Detection

Meta’s anti-scamming product protects billions of users, and the protected ones never know it happened; false-positive victims tell everyone, immediately. Most teams read that asymmetry and chase recall: catch everything. The harmonic mean punishes that instinct as hard as it punishes a missed scam.

NSM: Scam Detection F1 Score, harmonic mean of precision and recall, the industry-standard metric that penalizes both missed scams and false positives simultaneously.

Optimize for recall. A missed scam drains savings, hijacks an account, or manipulates a relationship; a false positive only flags a post. The fix looks obvious: lower the bar, catch more.

Two mechanisms reinforce that instinct. Press coverage is asymmetric: an unprotected scam victim’s story damages Meta more than an over-flagged post’s. Regulatory exposure runs the same way, several markets mandate aggressive detection coverage, and recall-first tuning satisfies all of them at once:

• Lower detection thresholds across all four scam vectors to catch edge cases

• Accept higher false positive rates as the cost of comprehensive protection

• Prioritize recall in engineering reviews and OKRs to signal seriousness about safety

Trap: Most teams run one recall dial for all four vectors. Financial fraud and a fake-news post aren’t equally severe, and a uniform threshold treats them like they are.