Google study finds LLMs are embedded at every stage of abuse detection

This survey examines the use of large language models at each stage of the content-moderation lifecycle—labeling, detection, review/appeals, and auditing—highlighting gains in contextual reasoning and scale alongside new biases, over-refusal and false positives, unfaithful explanations, demographic disparities, and the impracticality of running large models at full scale; it recommends hybrid architectures combining smaller specialists, retrieval-augmented policies, continuous red-teaming, and sustained human oversight.