A content moderation LLM outputs a refusal for a legitimate customer request about fireplace safety. A PM argues the guardrails are "too sensitive." What is the technical term for this failure mode?
- A.B. False positive (over-refusal)
- B.A. Underfitting
- C.C. Model inversion
- D.D. Gradient leakage
Why A is correct
Refusing a benign request is a false positive in the content moderation context - also called over-refusal. It represents a safety-utility tradeoff failure. Underfitting is a training generalization concept. Model inversion and gradient leakage are privacy attacks on training data. The tension between recall (catching real violations) and precision (not blocking legitimate use) is central to guardrail calibration.
Know someone studying for AI Security Fundamentals? Send them this one.