Trusting Autonomous Teammates in Human-AI Teams - A Literature Review

Research Background and Issues

• Identified Problems or Challenges: As artificial intelligence (AI) increasingly integrates into human teams to form "Human-AI Teams" (HATs), trust plays a critical role in the success of these teams. However, existing research lacks consistency in defining and measuring human trust in their AI teammates, which limits theoretical development and hinders the optimization of AI system design in practice. Moreover, most studies focus on simple team setups involving a single human and a single AI teammate, with insufficient exploration of multi-member teams and complex task contexts.

• Significance: AI is transitioning from being a tool to becoming a team member, a novel role that requires systems to be designed with interdependence and social functionality at the team level. A lack of trust may lead to inappropriate reliance on AI or outright rejection of AI interventions, thereby affecting work efficiency and the achievement of team goals.

• Research Motivation and Related Work: Previous studies have emphasized the importance of trust in automation and human team collaboration, but there is no unified understanding of trust in AI as a "teammate." This study aims to fill this research gap by integrating the concept of trust and providing a set of guidelines for future research.

Solution

• Methods and Solutions: The authors conducted a systematic literature review to analyze factors influencing human trust in AI teammates and to study the inconsistencies in the definition, measurement, and operationalization of trust. The main research questions include:

  1. How is trust defined and operationalized in HAT research?
  2. What factors influence trust in HATs, and what is their impact?

• Innovations: This study proposes a continuum framework based on "functional-interpersonal trust," integrating the technical and social capabilities of AI into the definition of trust. Additionally, the authors conducted a meta-analysis to quantify the effect sizes of trust-influencing factors across 57 studies and proposed a dynamic trust measurement framework to capture real-time changes in trust.

• Implementation Steps and Techniques:

  1. Literature search and screening were conducted using PRISMA guidelines, selecting 57 relevant studies from ACM, IEEE, and other databases between 2008 and 2022.
  2. Trust-influencing factors were categorized based on AI agents, human participants, team dynamics, and environmental characteristics.
  3. The concept of a trust continuum was proposed, and the differential impacts of functional (performance-based) and behavioral (behavior-based) factors on trust evaluation were analyzed.

Research Findings

• Specific Findings:

  1. Consolidated existing knowledge on trust in HATs and categorized influencing factors (e.g., reliability, transparency, behavioral style).
  2. Clarified the differences between the concept and measurement of trust, highlighting the misconceptions between "instrumental trust" and "teammate trust" in research.
  3. Revealed, through meta-analysis, several key factors (e.g., reliability) with significant impacts and quantified the overall effect sizes (Cohen's d).

• Advantages Over Existing Solutions:

  • Utilized a multidimensional framework to capture the dynamic changes in trust, providing a theoretical basis for the design of human-AI teams.
  • Offered operational guidelines for trust, promoting consistency in future research.
  • Positioned trust as a core element of team dynamics rather than merely an evaluation of AI performance.

• Experimental or Evaluation Results: Meta-analysis results showed that the strongest factor influencing trust was the reliability of AI teammates (Cohen's d = 1.13, significant impact), followed by transparency (smaller impact, d = 0.31). Behavioral factors such as apologies and responsibility attribution also demonstrated significant effects.

• Limitations and Future Directions:

  1. Current research is largely limited to two-member teams (one human and one AI), lacking exploration of complex teams. Future studies should focus on trust contagion effects and dynamic changes in multi-member teams.
  2. Existing studies predominantly focus on military and emergency task contexts. Future research should expand to creative collaboration and decision-making tasks to explore the trust requirements for different task types.
  3. The study calls for the comprehensive implementation of natural language processing technologies in HATs to explore the impact of interaction capabilities on trust.
  4. Very few studies have been conducted in Global South contexts. Future research should consider cross-cultural perspectives on trust-building.

Conclusion

This study provides a systematic literature review on trust in HATs and, for the first time, proposes a trust continuum framework that comprehensively analyzes how functional and social dimensions of AI teammates influence team dynamics. By integrating existing knowledge, the research offers guidelines and directions for future trust measurement, team design, and dynamic adjustments, while highlighting numerous unexplored areas, particularly the need for research on more complex teams, diverse task types, and cross-cultural adaptability. This has profound implications for the evolving role of AI as a team member.