Different formalisms for defeasible reasoning have been used to represent knowledge and reason in the legal field. In this work, we provide an overview of the following logic-based approaches to defeasible reasoning: defeasible logic, Answer Set Programming, ABA+, ASPIC+, and DeLP. We compare features of these approaches under three perspectives: the logical model (knowledge representation), the method (computational mechanisms), and the technology (available software resources). On top of that, two real examples in the legal domain are designed and implemented in ASPIC+ to showcase the benefit of an argumentation approach in real-world domains. The CrossJustice and Interlex projects are taken as a testbed, and experiments are conducted with the Arg2P technology.
Pretrial detention is a debated and controversial measure since it is an exception to the principle of the presumption of innocence. To determine whether and to what extent legal systems make excessive use of pretrial detention, an empirical analysis of judicial practice is needed. The paper presents some preliminary results of experimental research aimed at identifying the relevant factors on the basis of which Italian and Brazilian Supreme Courts impose the measure. To analyze and extract the relevant predictive-features, we rely on unsupervised learning approaches, in particular association and clustering methods. As a result, we found common factors between the two legal systems in terms of crime, location, grounds for appeal, and judge’s reasoning.
This article introduces a novel hybrid approach that enhances the accessibility and explainability of symbolic legal systems by integrating them with large language models (LLMs). While symbolic systems provide transparency and justifiability, they suffer from communicability issues and limitations in knowledge representation. Conversely, LLMs offer natural language fluency but struggle with reliability and justifiability in legal reasoning. The article bridges this gap by structuring legal tasks into a modular, interconnected sequence of prompts, supported by contextually relevant memory retrieval, thus creating a chain of prompts and memories. The methodology is validated through two case studies. The first applies it to computable contracts, demonstrating its ability to map code to contractual provisions, thereby making them more accessible. The second case study involves an expert legal system, where the hybrid approach generates accessible explanations of Prolog-based legal reasoning. The findings show that the hybrid approach enhances the clarity and interpretability of legal applications while preserving justifiability, using a modular structure that allows for adaptability across different legal tasks.
Agents, or agentic AI systems, have become capable of managing complex tasks, supporting an augmented form of automation across diverse processes. Yet, the socio-economic impact of their actions and the interconnected nature of their workflows exacerbate existing risks and introduce new technical, legal, and ethical challenges. Their scalability and lawful deployment necessitate comprehensive governing mechanisms to ensure effective risk mitigation and accountability in the event of harm. There is an emerging interest in developing risk management frameworks adapted to the dynamic nature of these systems. However, they often fail to specify at what level of agenticness these frameworks should be considered, creating uncertainty for developers and regulators. This study aims to investigate the level at which AI systems acquire an Augmented Algorithmic Agency, thereby leading to agentic risks and necessitating adapted risk management frameworks. This qualitative threshold acknowledges that agentic risks emerge at the intersection of an agent’s capabilities, identifies the minimum conditions that trigger them, and clarifies how other capabilities serve as risk multipliers. This approach supports developers, deployers, legal officers, and regulators in identifying at what point they need to adapt their AI governance policies to the dynamic, interactive nature of agentic risks.
This work introduces the formal framework DLRisk, based on Defeasible Deontic Logic, for assessing how fundamental rights are balanced under the EU AI Act when new systems are developed. This approach enables qualitative, scenario-aware reasoning by ordering rights, capturing demotions triggered by factual conditions, and detecting collisions between rights and deployment contexts. The result is an explainable assessment of trade-offs that supports compliance analysis and risk mitigation.