The Wizard of Oz (WoZ) method is a common and popular approach for simulating interactive systems in Human-Computer Interaction (HCI). Running such studies is demanding for researchers because the human wizard must manage human–agent interactions in real time while keeping participants safe and the interaction natural. Many WoZ systems struggle to reproduce complex agent behaviours without minimal delays or heavy workload for the moderator. We introduce AI of Oz, a framework that uses large language models to support researchers by monitoring ongoing interactions, detecting sensitive moments, and suggesting contextually appropriate responses. In a study with 20 HCI-related researchers, the system improved participants’ ability to manage interactions and maintain control compared to a version without AI support. We outline implications for WoZ research and note current limitations and future directions.

This study investigates Embodied Virtual Agents (EVAs) driven by Large Language Models (LLMs) as mediators in triadic collaboration where two users with conflicting goals must work towards a shared objective. We developed and evaluated an XR system where pairs (n=24) collaborated on an office design task, assessing the impact of agent presence and embodiment. Our findings reveal that agent embodiment significantly enhanced co-presence, which in turn fostered interactions that led to better perceived collaboration and higher user preference, whereas frequent interactions with the disembodied agent was associated with lower user preference. Conversely, agent presence did not improve task efficiency or satisfaction. Notably, our qualitative analysis revealed that the LLM-driven agent spontaneously adopted emergent facilitative and evaluative mediation strategies that align with collaborating and compromising conflict-resolving modes, showcasing its potential as an adaptive collaborative aid without explicit programming. These results highlight that the value of EVAs in complex collaboration lies in their ability to shape social dynamics and provide nuanced, context-aware mediation, a different form of value than traditional productivity enhancement.

Virtual reality (VR) is increasingly adopted for collaborative teaching and learning, enabling immersive and interactive experiences. As Artificial Intelligence (AI) tutors begin to take on roles alongside human instructors, it becomes crucial to understand how their integration influences interaction dynamics and role perception in these settings. This study investigates the role of Embodied Virtual Agents (EVAs) substituting human instructors in virtual training, specifically addressing the previously underexplored issue of EVA appearance consistency with instructors in Human-Agent Teaming (HAT). We recruited 21 participants to compare three conditions: No Agent, Shared-Appearance (SA) EVA, and Unique-Appearance (UA) EVA, where an EVA substitutes for the instructor during temporary absences. We evaluated collaboration efficiency, user perception/preference, and HAT dynamics. Our findings confirm that EVAs significantly enhance task efficiency compared to no support and reveal a key trade-off regarding appearance: SA fosters perceived continuity and trust but risks ambiguity and uncanny effects, while UA provides transparency and role clarity but may disrupt experiential coherence. These results have implications for designing dynamic HAT systems where control may shift. We discuss the benefits and limitations of each approach and offer design recommendations for future mixed-agency interfaces.