"The Conduit by which Change Happens": Processes, Barriers, and Support for Interpersonal Learning about Responsible AI

Research Background and Issues

• Main Issues and Challenges:
  In artificial intelligence (AI) development, the practice of Responsible AI (RAI) is gradually gaining traction, with tools and resources such as datasheets and model fairness evaluation tools being developed to mitigate AI's potential societal inequities. However, in real-world applications, practitioners often hesitate to address AI's potential negative impacts due to factors such as social environments, internal team power dynamics, and organizational incentives. Furthermore, while research indicates that AI practitioners primarily learn about RAI through "learning by doing" in their work, there is currently little understanding of the core interpersonal learning processes involved in RAI.

• Significance:
  RAI requires not only reliable technical tools but also a learning environment that emphasizes teamwork and discussion to help practitioners collaboratively identify and address the societal impacts of AI projects. Failing to adequately support interpersonal learning in RAI could significantly undermine practitioners' ability to develop responsible AI, potentially leading to broader systemic inequities or societal harm.

• Research Motivation:
  The authors aim to bridge the gap between learning theory and RAI practice by deeply exploring the interpersonal learning processes, barriers, and support mechanisms within the RAI domain. The study seeks to help practitioners systematically understand and implement RAI. The three research questions include:

  1. How do AI practitioners learn RAI through interpersonal interactions in the workplace?
  2. To what extent do RAI learning resources support interpersonal learning?
  3. What factors influence interpersonal learning, and how do they impact practitioners' engagement with RAI?

Solutions

• Core Methods and Research Design:
  The authors conducted 10 workshops (each lasting 90 minutes) with 21 RAI educators to explore the interpersonal learning processes in RAI. These educators had experience designing or implementing RAI learning materials, with many involved in informal RAI education efforts. The workshops included focus group discussions, mind map creation, and intervention design activities to gather educators' insights and experiences regarding interpersonal learning in RAI.

• Innovative Contributions:
  This is the first systematic study on interpersonal learning within the RAI domain, providing a detailed analysis of interactive learning scenarios and potential challenges in team collaboration. It highlights the critical role of socio-emotional skills and interpersonal relationships in RAI learning.

• Implementation Steps and Techniques:

  1. Data Collection:
     • Pre-surveys to understand participants' backgrounds and their focus on interpersonal learning in RAI resource design.
     • Focus group discussions to share participants' experiences and perspectives on RAI learning.
     • Mind map creation to help participants explore ideal and negative scenarios in RAI systematically.
     • Intervention design to guide participants in developing educational interventions for interpersonal learning in RAI.
  2. Data Analysis:
     • Iterative analysis to identify themes, focusing on the emotional skill requirements, power dynamics, and organizational culture influences in specific learning design cases.

Research Findings

• Key Findings:

  1. Identified key processes of interpersonal learning in RAI work, including:

     • Critical Reflection: Teams collaboratively examining the ethical implications of AI development on society.
     • Collective Meaning-Making: Teams jointly exploring how to translate high-level RAI principles into concrete practices.
     • Emphasizing the importance of interpersonal relationships in driving change within RAI, such as building alliances to promote RAI transformation and fostering supportive environments for mutual assistance and learning among colleagues.

  2. Found that RAI education design often prioritizes individual learning while neglecting interpersonal collaboration. This results in AI practitioners relying primarily on informal interpersonal interactions for learning, creating a disconnect between learning resources and actual needs.

  3. Identified key barriers to interpersonal learning in RAI:

     • Power Dynamics: Unequal power structures within teams and organizations can suppress members from raising concerns.
     • Value Tensions and Knowledge Gaps: Differences in team members' understanding of AI ethics concepts hinder collaboration and consensus-building.
     • Adversarial Stances: Conflicts between RAI teams and AI product teams create challenges for collaboration and knowledge sharing.

• Experimental and Evaluation Results:

  • Clear Steps: Through qualitative feedback from participants, the authors identified factors contributing to successful interpersonal learning in RAI, such as socio-emotional skills (e.g., active listening, respecting differing opinions) and a safe team culture (e.g., building trust-based relationships).
  • Practical Implications: The study suggests embedding RAI across entire teams rather than assigning it to specific roles and enhancing interpersonal skills through team learning to significantly improve RAI efforts.

• Limitations and Future Directions:

  1. Study Limitations:
     • Participants were primarily from large U.S.-based companies, with a majority being women, which may limit the study's generalizability.
     • Data relied heavily on participants' self-reports (e.g., surveys and discussions), which may not fully capture the dynamic realities of RAI learning scenarios.
  2. Future Research Directions:
     • Conduct more diverse cross-disciplinary and global studies to explore the impact of power dynamics and cultural differences on RAI team collaboration.
     • Develop and evaluate RAI educational interventions that support the development of socio-emotional skills and interpersonal relationships.
     • Expand RAI learning to smaller-scale or startup environments and explore how to establish robust RAI learning systems under resource-constrained conditions.

Conclusion

This study fills a critical gap in the field of interpersonal learning for RAI, elucidating the core processes and challenges involved in fostering responsible AI development. The research not only provides an educational practice model for the RAI field but also highlights the significant impact of socio-emotional skills and communication culture on RAI success, offering systematic guidance for future RAI education and intervention design.