CompanionSim: Synthetic Data for Evaluating Anthropomorphism in Human-AI Relationships

Researchers have developed CompanionSim, a simulation framework designed to generate synthetic data for studying anthropomorphism in human-AI relationships. This framework addresses the limitations of real-world human-AI interaction data by simulating multi-turn human-chatbot dialogues, encompassing 16 chatbot behaviors across seven use cases. The resulting 2,240 simulated conversations were validated by human participants, providing a novel resource to accelerate research into the consequences of AI companionship behaviors such as validation, trust, empathy, and attachment.

Why it matters

This research provides a crucial tool for understanding and shaping the evolving nature of human-AI interaction. By offering a robust dataset for studying anthropomorphism, it enables organizations to proactively address ethical, social, and operational challenges associated with AI's role as a companion, impacting design principles and user engagement strategies.

Key insights

• AI systems are increasingly perceived by individuals as social companions, beyond their function as productivity tools.
• The study of human-AI companionship behaviors, including aspects like validation, trust, empathy, and attachment, is crucial.
• Existing real-world human-AI interaction data is scarce and often unreliable, hindering research progress.
• CompanionSim provides a method to scale small amounts of real-world data through simulated multi-turn human-chatbot dialogues.
• The framework includes 2,240 simulated conversations, representing 16 distinct chatbot behaviors across seven different use cases.
• Human participants validated the quality and relevance of the simulated conversations compared to real-world data.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00250