ETD

• AI compliance
• EU AI Act
• Explainable AI (XAI)
• High risk AI systems
• Human oversight
• Transparency

The development of Artificial Intelligence (AI) systems based on neural networks has increased the complexity of these models, reducing the ability to understand their outputs. This opacity, referred to as the black box problem, raises concerns about trust, accountability, and the ability of individuals to contest automated decisions. The EU AI Act addresses these concerns by establishing obligations for providers and deployers of high-risk AI systems. Article 11 requires providers to document the logic of the system and the choices made while building it, Article 13 requires that such systems be accompanied by information enabling deployers to understand their outputs, and Article 14 requires human oversight. Taken together, these provisions presuppose a degree of explainability that current AI systems may or may not provide, raising the question of whether Explainable AI (XAI) can serve as a tool for satisfying these obligations. The thesis is structured in three parts. The first part examines the technical foundations of modern AI systems, tracing the evolution from machine learning to deep learning and neural networks, and introduces the field of XAI, analysing how explainability techniques make the reasoning behind AI outputs accessible to developers, regulators, and individuals affected by these systems. The second part examines Articles 11, 13, and 14 of the AI Act, assessing for each provision the type of explanation required by law and whether the techniques discussed in the first part are capable of meeting that requirement, identifying alignment and gaps between legal expectation and technical capability. The third part builds on this analysis to develop a compliance assessment framework for evaluating whether a given XAI implementation satisfies the obligations identified. The framework consists of thirty-seven criteria presented in four tables, each traceable to a provision of the Act or to a numbered point of Annex IV, each written as a question that can be answered from a document the provider or the deployer is already required to produce or to receive, and each scored as evidenced, partially evidenced or not evidenced. The framework is then examined against the conditions it assumes, the positions from which its criteria can be answered, and the limits on its transferability across systems, model types and actors, producing conclusions on its validity, limitations, and potential for real world application.