CompanionHarm: A Multi-Turn Benchmark for Detecting Harms in Real-World AI Companion Conversations

The increasing integration of AI companions into daily life necessitates robust mechanisms for identifying harms arising from human-AI social and emotional interactions. A new benchmark dataset, CompanionHarm, has been introduced to address the current limitations in research due to a lack of real-world, multi-turn conversational data. This dataset, derived from interactions with the AI companion Replika, comprises 2,111 conversations and 14,051 utterances, with 7,016 AI utterances annotated across 13 categories of harmful behavior.

Why it matters

The development of reliable methods for detecting and mitigating harms in human-AI interactions is crucial for maintaining public trust and ensuring ethical AI deployment. This dataset provides a foundational resource for advancing research and development in AI safety, directly impacting the long-term viability and responsible scaling of AI companion technologies across various sectors.

Key insights

• AI companions are becoming increasingly prevalent in daily life, highlighting a critical need for harm detection.
• Current research on AI companion harms is constrained by the absence of real-world, multi-turn conversational datasets.
• CompanionHarm is a new, publicly available benchmark dataset designed to address this data gap.
• The dataset consists of 2,111 multi-turn conversations and 14,051 utterances from the AI companion Replika.
• 7,016 AI utterances within the dataset have been independently annotated by three annotators.
• Annotation covers 13 distinct categories of harmful behavior, grounded in a taxonomy of AI companion harms, and includes aggregated labels.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.25377