Unpacking Trust Dynamics in the LLM Supply Chain: An Empirical Exploration to Foster Trustworthy LLM Production & Use

Research Background and Issues

• Identified Problems or Challenges:
  This study highlights that existing research on artificial intelligence (AI) trust often focuses on limited trust relationships (e.g., between end-users and AI systems) and lacks empirical exploration of complex trust dynamics in real-world contexts, particularly within the AI supply chain. These supply chains involve multiple technical artifacts (e.g., training data, foundational models) and decision nodes (e.g., model development, deployment, and usage), where intricate trust dynamics influence the production and adoption of AI systems.

• Importance of the Issue:
  In the supply chain of large language models (LLMs), trust can profoundly impact the accountability of their production and usage, as well as the trustworthiness of their technological outputs. Poorly managed trust may lead to inaccurate or miscalibrated trust, increasing the risks of blind spots in usage or misuse of technology.

• Research Motivation and Related Work:
  Inspired by organizational psychology, this study broadens the scope of trust considerations, focusing on the complex relationships between various trust dynamics within the AI supply chain. While previous research has primarily concentrated on interactions within single organizations or between users and AI systems, this study seeks to address the gap in understanding trust relationships within the AI supply chain.

Solution

• Proposed Methods or Solutions:
  The authors conducted semi-structured interviews with 71 practitioners involved in the LLM supply chain (totaling over 3,600 minutes) to explore their collaborative practices through the lens of trust, aiming to uncover diverse trust relationships and interaction dynamics.

• Innovative Contributions:

  • Systematically studying cross-organizational and interpersonal trust dynamics within the AI supply chain for the first time.
  • Examining not only technical artifacts but also the trust relationships between organizational participants.
  • Highlighting the interdependence and complexity among multiple trust objects, including technical artifacts, organizations, and individuals.

• Implementation Steps and Key Techniques:

  1. Interview Design and Data Collection: Semi-structured interviews and questionnaires were used to understand trust behaviors and decision-making patterns of various roles within the supply chain.
  2. Data Analysis: Reflexive thematic analysis techniques were employed to code interview transcripts, which were interpreted using trust frameworks (e.g., competence, benevolence, and integrity models) and organizational psychology theories.
  3. Theme Extraction: Three core themes were identified: trust in technical artifacts, interpersonal trust relationships, and their interactions at supply chain nodes.

Research Findings

• Specific Findings:

  • Revealed the diversity of trust relationships within the LLM supply chain, including personal trust, inter-organizational trust, and broader technological trust (e.g., trust in AI technologies and academic research).
  • Clarified how key factors of trust within the supply chain (e.g., technical competence, organizational integrity, and transparency) influence trust relationships.
  • Found that the dynamics and complexity of the supply chain can lead to phenomena such as "blind trust" or "distrust," for instance, where blurred responsibility boundaries within the supply chain exacerbate trust misjudgments.

• Comparison with Existing Solutions:

  • Unlike studies that focus solely on user-AI system trust, this research broadens the perspective by treating the socio-technical AI supply chain as the research context.
  • Emphasizes the need to integrate trust dynamics into the design and governance of the AI lifecycle.

• Experimental or Evaluation Results:
  Empirical interviews revealed variations in trust relationships across different nodes of the supply chain. For example, upstream organizations in the supply chain often rely on feedback from downstream users, while downstream consumers heavily depend on transparency tools provided by model suppliers.

• Limitations and Future Directions:

  1. Limitations:

     • The study sample is concentrated in the private sector, lacking data from other industries (e.g., public sector).
     • Interview subjects may not fully represent all relevant roles within the supply chain (e.g., foundational model developers).
     • Due to the complexity of supply chains, trust dynamics may vary depending on specific scenarios.

  2. Future Directions:

     • Expand the scope of research to include more industries and supply chain participants to validate the universality of trust dynamics.
     • Explore the relationship between trust behaviors and power dynamics between supply and demand sides.
     • Design transparency tools and methods to support trust calibration within the supply chain.

Conclusion

This study broadens the scope of trust research by uncovering various trust relationships within the LLM supply chain and their impact on AI production and usage. The authors highlight the potential risks of trust misjudgments and propose new possibilities for fostering accountability and calibrated trust. Future research should deepen the understanding of supply chain complexity and trust compensation mechanisms while providing empirical evidence for AI policymaking and supply chain governance.