Transparency through tensions: an integrated multi-method framework for advanced interpretations and robust auditing of artificial intelligence systems

<p dir="ltr">The present period in Artificial Intelligence adoption is characterised by AI systems playing a crucial role in aiding decision-making processes, with little human involvement. This extensive adoption of AI gives rise to numerous challenges, especially regarding transparency, accountability, and potential biases. These issues highlight the urgent need for auditing mechanisms at multiple levels to ensure that AI systems comply with established regulatory and ethical standards. This thesis addresses critical obstacles in the auditing of AI systems by proposing innovative methodologies and techniques designed to improve the assessment process. The research is categorised into four primary domains: the examination of interdisciplinary tensions in AI system development, the organisation of AI system components for more effective audits, the identification of factors crucial for successful AI audits, and the development of a comprehensive auditing framework for AI systems. To address these challenges, this work offers several contributions. First, it introduces an Interdisciplinary Tension-Aware Value Chain for AI systems, which addresses the sociotechnical challenges and aids in designing robust audit programs for AI systems. Second, it applies an Ontology-Based Approach to structure both AI systems and their audit components for more comprehensive audits. Third, the research explores the application of Critical Success Factors to increase the success and effectiveness of AI system audits. Lastly, it proposes an overarching Auditing Framework aimed at enhancing the robustness of AI system audits. Together, these contributions help improve auditing practices and approaches for AI systems, facilitating the adoption of AI systems that are more transparent, responsible, and trustworthy.</p>