This paper explores how to fulfill the right to explanation and support the right to challenge in automated decision-making systems by integrating symbolic methods with Large Language Models (LLMs). In cases involving automated decisions based on conflicting arguments, we first model the situation using an abstract argumentation framework. We then apply grounded semantics and discussion games to guide the explanation of the decision and support the right to challenge. Specifically, we prompt OpenAI’s flagship model (o1) to perform these reasoning steps and generate corresponding natural-language explanations. Finally, we ask the model to identify which argument would need to be modified to alter the decision, based on the formal reasoning behind the explanation. To assess the quality of the explanations, we use several state-of-the-art LLMs as evaluators. We compare three types of explanations produced by o1 with a set of criteria: those based on grounded semantics, discussion games, and a baseline explanation (in which o1 generates an explanation without any formal symbolic constraints). The results indicate that explanations based on discussion games are rated higher than those based on grounded semantics, which in turn outperform the baseline explanations. We also discuss the "right to challenge" aspect, showing that explanations based on discussion games effectively identify which arguments can be challenged to alter the decision. Overall, our findings suggest that formally guided LLMs can better fulfill the right to explanation and support the fulfillment of the right to challenge. This supports the view that integrating sub-symbolic, data-driven generative AI with symbolic, knowledge-driven AI is a fruitful way to achieve transparent AI systems that align with our societies’ legal requirements regarding digitalization.

In the devolved legal system of the United Kingdom (UK), legislative provisions may apply differently across regions such as England, Wales, Scotland, and Northern Ireland. Accurately determining this territorial scope is essential for legal interpretation and AI-assisted legal tools. However, metadata capturing jurisdictional applicability is inconsistently format, as only a few Acts include Territorial Application Annexes. This study presents a case study using the Education Act 2005 to evaluate the accuracy of automated methods for identifying territorial scope. We found that only 46.9% of sections matched in jurisdictional coverage. The best-performing approach achieved nearly 80% accuracy, showing that LLMs can effectively support scalable and explainable territorial disambiguation.

Normative stance underlies decisions in law, legal reasoning, policy, and safety-critical settings. A model’s judgment of what is permissible vs. impermissible often determines its downstream behavior. We study how to steer a language model’s normative stances at inference time by adding a tiny, contrastive perturbation to the last-token neural activation in late MLP layers (contrastive last-token steering). For each normative prompt, we construct a contrast direction by comparing its last-token activation to that of a minimally edited variant that implies a more permissive normative stance (e.g., “acceptable” rather than “wrong”). During generation, we add this vector at the last token; a single strength parameter α controls how strongly and in which direction we push the model’s stance (permissive vs. restrictive). Impact is measured as the change in a next-token logit margin between permissive and restrictive continuations. To avoid overclaiming, we calibrate a threshold τ on neutral controls (same layers, tempered strengths with |α|≤1) and count success only when the shift exceeds τ in the expected direction. We also assess specificity by verifying that, on neutral control prompts, steered outputs exactly match unsteered baselines. Beyond component-level tests, we probe neuron-level locality by steering only the top-k contrastive neurons (ranked by last-token contrast) and confirming reversibility on our test set: +α produces the shift and -α reverses it. The method is training-free, uses standard forward hooks, and we report pilot results on Llama-3-8B-Instruct.

This work focuses on the automatic classification of deontic sentences. It presents a novel Machine Learning approach which combines the power of Transfer Learning with the information provided by two famous LegalXML formats. In particular, different BERT-like neural architectures have been fine-tuned on the down-stream task of classifying rules from the European General Data Protection Regulation (GDPR) encoded in Akoma Ntoso and LegalRuleML. This work shows that fine-tuned language models can leverage the information provided in LegalXML documents to achieve automatic classification of deontic sentences and rules.

This paper presents an AI use-case developed in the project “Study on legislation in the era of artificial intelligence and digitization” promoted by the EU Commission Directorate-General for Informatics. We propose a hybrid technical framework where AI techniques, Data Analytics, Semantic Web approaches and LegalXML modelisation produce benefits in legal drafting activity. This paper aims to classify the corrigenda of the EU legislation with the goal to detect some criteria that could prevent errors during the drafting or during the publication process. We use a pipeline of different techniques combining AI, NLP, Data Analytics, Semantic annotation and LegalXML instruments for enriching the non-symbolic AI tools with legal knowledge interpretation to offer to the legal experts.