Safety & Alignment

Content Moderation

Content moderation checks text for harmful content. Moderation uses classifiers to identify problematic outputs. Moderation can be rule-based (keyword lists) or learned (ML classifiers). Moderation requires defining what's inappropriate. Moderation tradeoff: strictness vs. utility. Moderation is computationally efficient as a filtering layer. Moderation is widely used for safety. Moderation is imperfect.