ScholarWorks@경상국립대학교

초록

Harmful language in conversational AI poses distinct safety and governance challenges, as Large Language Model (LLM) chatbots interact in private, one-to-one settings. Understanding the types of harm and their temporal concentration is crucial for responsible deployment and time-aware moderation. This study investigates the types and diurnal dynamics of harmful speech, comparing patterns between play-oriented chit-chat and task-oriented LLM services.We analyze two large-scale, real-world English corpora: a chit-chat service (SimSimi; 8.7 M utterances) and an LLM service (WildChat; 610 K utterances). Using the Perspective API for multi-label classification (Toxicity, Profanity, Insult, Identity Attack, Threat), we estimate the incidence of harm categories and compare their distribution across five dayparts. Our analysis shows that harmful speech is significantly more prevalent in the chit-chat context than in the LLM service. Across both platforms, Toxicity and Profanity are the dominant categories. Temporally, harmful speech concentrates most frequently during the dawn daypart. We contribute an empirical baseline on how harm varies by chatbot modality and time of day, offering practical guidance for designing dynamic, platform-specific moderation policies.

키워드