Framing 'Collaboration': How Human-Human Principles Translate into Human-AI Realities

Paper Title

Framing 'Collaboration': How Human-Human Principles Translate into Human-AI Realities

Publication Info

• Topic area: Human-AI collaboration and its conceptual alignment with human-human collaboration principles.
• Keywords: Human-AI collaboration, human-human collaboration, thematic analysis, LLM-assisted analysis, interaction dynamics, task demands, temporal progression, AI teammates, collaboration framework.

Background and Problem

• Problem / challenge: Despite the widespread use of the term "human-AI collaboration," there is little consensus on its definition, and existing research lacks a systematic framework to compare it with human-human collaboration principles.
• Significance: Understanding human-AI collaboration is critical for designing effective AI systems, fostering interdisciplinary research, and bridging theoretical and practical applications in human-AI interaction.
• Motivation and related work: Prior research has explored human-human collaboration extensively, including frameworks like Shah's synergy model, McGrath's task circumplex, and Tuckman's group development stages. However, human-AI collaboration remains underexplored, particularly in terms of its alignment with human-human collaboration components and its unique challenges, such as asymmetric interaction dynamics and limited social-emotional capabilities.

Solution

• Proposed approach: A multi-step thematic analysis combining manual review of human-human collaboration literature with LLM-assisted analysis of human-AI collaboration literature.
• Novelty:
  1. Development of a thematic framework for human-human collaboration based on 60 publications across disciplines.
  2. Application of LLMs (OpenAI’s GPT 4o mini and o3 mini models) to analyze 299 publications on human-AI collaboration.
  3. Empirical mapping of component prevalence and directionality in human-AI collaboration, highlighting gaps and asymmetries.
  4. Recommendations for future research and design of human-AI systems based on findings.
• Procedure and key techniques:
  • Step 0: Manual thematic analysis of human-human collaboration literature to identify 40 components across three main categories: Collaborative Activities and Dynamics, Sharing and Collective Purpose, and Team and Relatedness.
  • Step 1: Generation of definitions and guidelines for each component using GPT 4o mini.
  • Step 2: Extraction of candidate sentences from human-AI collaboration literature using GPT 4o mini and o3 mini.
  • Step 3: Validation and classification of sentences for component alignment and directionality using o3 mini reasoning model.
  • Additional processing: Text similarity checks to ensure robustness and reproducibility.

Results

• Concrete findings:
  • Components such as interaction-oriented activities (e.g., communication, interdependence) are well-represented in human-AI collaboration literature.
  • Social aspects (e.g., respect, voluntariness) and ownership components (e.g., shared accountability, commitment) are underrepresented.
  • Human-AI collaboration often involves asymmetric dynamics, with humans dominating social aspects and AI dominating information sharing.
• Advantage over baselines:
  • The study provides a structured framework for comparing human-human and human-AI collaboration, filling gaps in existing literature and offering actionable insights for design and research.
• Experiments / evaluation:
  • Analysis of 299 publications spanning 2014–2024 using LLMs, with reproducibility demonstrated through a re-run of the analysis in November 2025.
  • Intraclass correlation coefficients (ICCs) showed high reliability across repeated analyses.
• Limitations and future work:
  • Limited empirical validation of LLM outputs compared to human baseline analysis.
  • Sparse representation of long-term collaboration and multi-party settings in human-AI literature.
  • Recommendations include longitudinal studies, richer team constellations, and standardized measures of collaboration.

Summary

This paper develops a thematic framework for human-human collaboration and applies it to analyze human-AI collaboration using LLMs. The findings reveal significant alignment between human-human and human-AI collaboration principles but highlight gaps in social aspects, ownership, and joint planning. The study emphasizes the need for long-term and multi-party research designs, standardized measures, and improved AI interface designs to foster effective collaboration. By quantifying component prevalence and asymmetries, the research provides actionable insights for advancing human-AI collaboration in both theoretical and practical contexts.