Bridging Algorithmic Transparency Obligations and Practice: Can eXplainable Artificial Intelligence achieve the policy objectives behind multi-dimensi

Complex AI systems may harbor proxy discrimination and spurious correlations in multi-layered process structures. These biases, reflecting implicit values, undesired weights, intersectional discrimination, or intellectual property infringement necessitate regulatory intervention through robust identification. Notably, a meaningful explanation of models' logic and outputs, both accessible and understandable to all stakeholders in its AI value chain and lifecycle is not only desirable but mandatory. In response, my research critically investigates how Explainable AI (XAI) advances policy objectives behind multi-dimensional transparency obligations across high-risk AI applications. It examines AI governance approaches to bridge policy aspirations and technological solutions to operationalize algorithmic transparency obligations in existing and emerging legislation. The project will scrutinize these solutions through participatory, human-centered approaches to consider how different actors-ranging from domain experts like clinicians and directly or indirectly affected groups such as patients-interact with Algorithmic Decision-Making and Decision Support Systems. It seeks to examine XAI and governance frameworks' limitations in addressing individual and collective needs of meaningful explanations (available, accessible, and appropriately communicated) by acknowledging varying levels of technical literacy (i.e. AI, XAI, Digital, and Data). Particularly, XAIs own lack of interpretability, interoperability and associated automation bias hindering effective Human-AI partnerships through an HCI lens. The project will discuss 'algorithmic transparency' pragmatically by facilitating interdisciplinary practices at the Responsible NLP CDT. The dual output will assist technical communities (e.g. IEEE's INNS) to navigate the complex regulatory landscape and incorporate ethical reflection throughout design and iterative course correction, and inform policy to mitigate mismatch and...