Researchers have developed CAREBench, a new benchmark designed to evaluate the child-safety risks of large language models beyond explicit abuse material. The benchmark includes 500 prompts across twelve categories such as grooming, deception, and emotional dependency, with annotations from parents and clinicians. Initial evaluations of seven frontier models revealed failure rates between 2% and 58%, highlighting significant gaps in current AI child-safety policies. AI
IMPACT This benchmark aims to help AI developers identify and address upstream child-safety risks in language models, potentially leading to safer AI interactions for minors.
RANK_REASON The cluster describes a new benchmark for evaluating AI safety, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
