Work

Towards a Communicative Process Design for Human-AI Team Collaboration

Year: 2026
Type: dissertation
Abstract: As AI systems shift from individual tools to collaborative team members, the communicative processes through which humans and AI partners coordinate remain poorly understood. Decades of research have established how human interlocutors accommodate, ground, and coordinate in teamwork, but whether these patterns extend to human-AI teams, how they vary across cultures and team configurations, and whether designed support can serve diverse users equally are open questions. This dissertation develops a communicative process design approach for human-AI team collaboration, asking how human-AI communicative processes are, and can be, shaped by culture, team structure, and designed support. Four studies across collaborative tasks, from group brainstorming to negotiation, were designed, conducted and integrated through three cumulative stages. The first stage is a characterization stage that aims to understand how communicative processes such as accommodation and grounding operate in human-AI interaction and how they vary across cultural populations: an analysis of 1,500 user-agent conversations spanning 75 countries reveals that linguistic accommodation varies by cultural population, with AI agents that accommodate well on average failing specific cultural groups (Chapter 3). Separating affective accommodation from grounding within the same negotiation sample (N=40) establishes that these operate as distinguishable constructs (Chapter 4). The second stage is a formation stage that explores how team formation shapes the communicative processes that the characterization stage reveals. A 2 x 2 factorial experiment (N=168) crossing human-AI team size and formation of AI behaviors within the team finds that user-parameterized formation creates qualitatively distinct communicative modes, shifting the AI from co-generation to a coaching profile, while a second AI erases the formation effect by raising group output and reducing human authorship (Chapter 5). In the intervention stage that aims to improve human-AI communication and collaboration, multi-agent coaching for cross-cultural negotiation (N=80) improves message persuasiveness, but whether this gain reaches the task outcome depends on horizontal collectivism, an individual-level cooperative orientation, not regional category (Chapter 6). The convergent finding across the communicative process design approach is that population variation identified in characterization returns as the moderation that determines who benefits from designed communicative support, and that human communicative behavior shifts through role redistributions rather than uniform amplification. These results motivate a conceptual framework for understanding alignment in human-AI collaboration. The dissertation proposes an iceberg framework that separates visible AI behavior from hidden user assumptions and model processes, and traces three interaction design paradigms, free-form interaction, behavioral engineering, and human-AI co-configuration, along a trajectory in which communicative processes become increasingly revisable during collaboration (Chapter 7). The dissertation makes three contributions to the understanding and programmatic design of communicative processes in human-AI teams. First, it establishes a characterize-before-you-design principle: effective communicative process design must account for population variation rather than assume a universal user, and that culturally aware systems should infer cooperative orientation from interaction rather than assign support by demographic category. Second, the iceberg framework and design paradigms provide a conceptual vocabulary for the field: the iceberg identifies where alignment interventions operate, and the co-configuration paradigm reframes alignment as an interaction property maintained through ongoing mutual grounding rather than being fixed at design time. Third, the measurement architecture separates human from AI authorship, affective accommodation from grounding, global from local diversity, and message-level gains from task outcomes, providing reusable constructs for benchmarking communicative dynamics across future model generations. The gap between initial parameterization and iterative co-configuration is the central design frontier that this work opens.
Language: English
Field: Psychology
Open Access status: green
OpenAlex ID: W7221481408
Cites: 114
Cited by:
Related to: 10