AI as an Agent and Collaborative Space: Exploring the role of generative AI in small group synchronous and asynchronous collaborative dynamics

Synthesis: AI as an Agent and Collaborative Space. A qualitative study by Haowei Xu, Ahmed Kharrufa, Ellis Solaiman & Vasilis Vlachokyriakos (2026) examining how generative AI shapes small-group collaboration through observations and interviews with 27 higher-education students. Through a distributed cognition lens, the study finds a stark contrast: in synchronous settings with a single shared GenAI interface, teams maintain shared awareness, co-construct prompts, treat the chat as shared external memory, and openly negotiate AI's role — whereas in asynchronous teamwork GenAI is used individually, with outputs selectively rewritten or "de-labeled" before being shared, making AI's influence on decisions opaque and harder to trace. The paper proposes a GenAI-Supported Cooperative Work (GSCW) perspective, framing GenAI as both a configurable agent (from individual assistant to team member) and an interactive collaborative space.

Overview

GenAI systems are designed for single-user interaction, yet they are increasingly integrated into collaborative work. This study asks how introducing GenAI influences collaborative dynamics — information flow, role negotiation, and decision-making — in small-group teamwork, and what new collaboration challenges it raises. The distributed cognition lens reframes GenAI as a cognitive participant distributed across people, tools, and environment.

Method

• Observation study: 18 students in six groups of three, working synchronously (3 in-person, 3 online via Microsoft Teams) on designing a cultural educational game with a single shared ChatGPT-4 interface over ~45 minutes, followed by reflection.
• Interviews: 9 additional semi-structured interviews (~45 min) capturing asynchronous, everyday GenAI use in teamwork.
• Analysis: Qualitative thematic coding across three collaborative-dynamics dimensions — information flow, role negotiation, decision-making — plus GenAI as a collaborative space.

Key findings

• Access configuration shapes information flow. With a shared, visible interface, teams co-constructed "collective prompts" and ran a recurring surface–evaluate–embed cycle, sustaining common ground. In asynchronous settings, private, fragmented prompting and output de-labeling produced opaque, privatised information flows that raised the cost of maintaining a shared cognitive model.
• Awareness must extend to the AI. GenAI is cognitively involved but contextually unaware of the team's shifting focus — a paradox that forces teams to constantly repair drift between the AI's limited awareness and the actual collaborative trajectory. Future systems need AI aware of group state and, with multiple agents, aware of each other.
• Roles are negotiated, not fixed. GenAI's role shifted from subordinate assistant to contested teammate depending on context and proactivity expectations; teams debated and constrained AI influence to ensure alignment and accountability.
• Design direction — permeable boundaries with directional context. Team-level AI should pass shared context into personal AI spaces, but insights generated privately should not auto-flow back to the group; humans stay in control of what becomes part of the shared record.

What this means for practice

• Instructors. Require groups to surface AI inputs for collective negotiation. In synchronous shared-interface work teams ran a surface–evaluate–embed cycle that sustained common ground, whereas asynchronous members prompted privately and often rewrote or "de-labeled" outputs before sharing, making AI's influence on decisions opaque.
• Instructors. Configure access deliberately: restricting the group to a single shared interface forced prompt negotiation, because teams could not work in parallel on separate devices.
• Designers. Support both synchronous and asynchronous modes, and design permeable boundaries with directional context — pass shared team context into personal AI spaces, but do not let privately generated insights auto-flow back into the group's shared record.
• Designers. Let teams see and adjust the AI's role as collaboration unfolds rather than fixing one role in advance, and require quick group agreement before the AI contributes to sensitive decisions.
• Researchers. Extend the notion of awareness to the AI itself and study group-state-aware agents, including multiple agents aware of one another, since GenAI in this study was cognitively involved but contextually unaware of the team's shifting focus.

Limitations

• The observation covered 18 students in six groups of three working for roughly 45 minutes with one shared ChatGPT-4 interface, and the interview sample was only nine participants, limiting the diversity of collaboration practices and organizational contexts captured.
• Combining observation with interviews provided complementary perspectives but limited direct comparison across settings, and the findings are shaped by university students and a single text-based AI system.
• Video recording and the centralized-input, shared-display setup may have shaped participant behavior despite efforts to keep conditions naturalistic, and differences in GenAI familiarity among team members may have affected how roles were distributed.
• The participant pool had a skewed gender ratio, the analysis emphasized information flow over social and organizational concerns such as trust or accountability, and the authors state their design implications remain speculative.

Citation

Xu, H., Kharrufa, A., Solaiman, E., & Vlachokyriakos, V. (2026). AI as an Agent and Collaborative Space: Exploring the role of generative AI in small group synchronous and asynchronous collaborative dynamics. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '26).