ETD

• automated decision-making
• black-box AI
• hallucination
• large language models
• legal reasoning
• neuro-symbolic AI

This thesis evaluates the integration of Artificial Intelligence (AI) within civil court systems to address the operational burdens posed by voluminous, complex electronic filings. Standard statistical models, such as Large Language Models, are fundamentally unsuitable for judicial decision support due to their opaque reasoning process, an inability to process legal hierarchies, and hallucinations that fail to meet the strict requirements of the AI Act and the European Commission for the Efficiency of Justice Guidelines. To address these legal and technical impediments, this thesis posits neuro-symbolic AI as a suitable architectural engine for civil court applications. By integrating the capabilities of language models based on neural networks with the inference capabilities of rule-based symbolic models, neuro-symbolic AI provides, in principle, auditable, explainable reasoning chains, while reducing epistemic errors. This thesis makes two central contributions. First, it shows that the technical objections most frequently raised against judicial AI, hallucination and black-box reasoning, are properties of a specific architecture rather than of automated legal reasoning as such, and that neuro-symbolic systems, through mechanisms including rule-constrained training, fact-gating, and fail-closed citation verification, could structurally curtail both. Second, building on this architectural finding and on the observation that a meaningful category of civil court orders, exemplified by default judgments, is already standardised and non-discretionary enough to be a strong candidate for automation, it argues that the categorical prohibition on solely automated decision-making under Article 22 of the General Data Protection Regulation is difficult to justify as a permanent feature of the provision. In its place, this thesis makes a tentative proposal for a differential regulatory framework, calibrating the degree of permissible automation to a case’s procedural risk, factual controversy, and judicial discretion, rather than applying a uniform prohibition regardless of context.