Moderation

WildGuard: the refusal label was the right call, the GPT-4 comparison is weaker than it looks

WildGuard was the first open moderation model to label prompt harm, response harm and refusal separately, and that part holds up. Its margins over GPT-4 sit inside base-model noise, GPT-4 helped grade the test set, and later work shows a refusal sentence in front of a harmful reply makes the guard call it safe.

Nov 8, 2026 · 10 min