Assuring trustworthy AI: An integrative review of hierarchical and relational assurance approaches

• Identifies the need for an integrative review of hierarchical and relational assurance approaches to AI trustworthiness. • Demonstrates—via a directed content analysis—the emphasis currently placed on hierarchical compliance-based approaches with inadequate attention given to relational approaches for increasing confidence in AI. • Develops a forward-looking agenda for enhancing AI assurance through the combined use of hierarchical and relational assurance strategies. • Highlights implications for social science approaches to trust and trustworthiness. • Provides a researchable, practice-based framework for integrating the use of hierarchical and relational assurance strategies within the AI life cycle. Trustworthy AI has recently emerged as a composite construct combining the legal, ethical and technical conditions relevant to mitigating the risks of AI systems. The methods used to assure the trustworthiness of AI systems have typically taken the form of a hierarchical approach and assurance strategies. These aim to provide documentary evidence demonstrating an AI system’s compliance with pre-defined legal, ethical, and technical criteria. Much less attention has been accorded to the use of a relational approach and assurance strategies that aim to increase stakeholders’ confidence in the trustworthiness of AI systems. Drawing on a directed content analysis of 90 research and policy documents, this article presents an integrative review of the hierarchical and relational strategies used to assure AI trustworthiness. Based on the content analysis, a hierarchical-relational classification is presented of the strategies that currently exist to assure the ethical principles, governance, accountability, transparency and explainability of AI systems. This confirms the prevalence of hierarchical assurance strategies with limited attention paid to relational assurance strategies. The significance of this disparity and the theoretical and practical implications for assuring the trustworthiness of AI systems are discussed. The article concludes by presenting a researchable framework embedding the combined use of hierarchical and relational assurance strategies within the AI life cycle and beyond.