arXiv:2606.04867v1 Announce Type: new Abstract: As AI companion platforms such as Replika and Character.AI rapidly grow, concerns about unsafe human-AI interactions have intensified. This study introduces AICompanionBench, to our knowledge the first publicly available benchmark dataset of human-AI companion conversations annotated with fine-grained safety risk categories. The dataset contains 2,123 real-world Replika conversations collected from Reddit and annotated through human-AI collaboration across nine categories: sexual behavior, antisocial behavior, physical aggression, verbal aggressi
Source: arXiv cs.AI — read the full report at the original publisher.