This paper presents a framework for evaluating approaches to formalizing legal sources. The framework integrates adoption-related criteria, extending the scope of assessment beyond purely technical dimensions. We build upon three contemporary comparison frameworks for normative specification languages, combining the Technology-Organization-Environment framework with grounded theory and adding criteria from other IT fields. As an illustration, we applied the resulting framework to RegelSpraak (RuleSpeak) from the Dutch Tax Administration.

With the adoption of the AI Act, fundamental rights impact assessment (FRIA) processes become highly relevant for both public and private institutions; yet such processes can be challenging, especially for small- to medium-sized organizations. One recent research that piloted a partial automation of FRIA is Anticipating Harms of AI (AHA!), relying on the use of a large language model and crowd-sourcing; unfortunately, the paper provides limited insights upon its internal working. Therefore, this work presents AFRIA, a processing pipeline that performs specific aspects of FRIA, conceived with AHA! as inspiration. In order to assess to what extent AFRIA is a successful reconstruction of AHA!, we analyzed the percentage of meaningful harms that AFRIA generates and the distribution of harm categories, and compared it to AHA!’s results, finding a satisfactory convergence. Beyond inspiration from AHA!, we also looked into the requirements of the AI Act and scholarly critique to make AFRIA more meaningful for identifying impacts on fundamental rights, targeting categories of human rights impacts, potential harm mitigation measures, and the severity and likelihood of the harms. The results show opportunities, but also limitations in what type of support this technology can bring.

In recent years, several normative systems have been presented in the literature. Relying on formal methods, these systems support the encoding of legal rules into machine-readable formats, enabling, e.g. to check whether a certain workflow satisfies or agents abide by these rules. However, not all rules can be easily expressed (see for instance the unclear boundaries between tax planning and tax avoidance). The paper introduces a framework for norm identification and norm induction that automates the formalization of norms about non-compliant behavior by exploring the behavioral space via simulation, and integrating inputs from humans via active learning. The proposed problem formulation sets also a bridge between AI & law and more general branches of AI concerned by the adaptation of artificial agents to human directives.

Labeling textual corpora in their entirety is infeasible in most practical situations, yet it is a very common need today in public and private organizations. In contexts with large unlabeled datasets, active learning methods may reduce the manual labeling effort by selecting samples deemed more informative for the learning process. The paper elaborates on a method for multi-class classification based on state-of-the-art NLP active learning techniques, performing various experiments in low-resource and imbalanced settings. In particular, we refer to a dataset of Dutch legal documents constructed with two levels of imbalance; we study the performance of task-adapting a pre-trained Dutch language model, BERTje, and of using active learning to fine-tune the model to the task, testing several selection strategies. We find that, on the constructed datasets, an entropy-based strategy slightly improves the F1, precision, and recall convergence rates; and that the improvements are most pronounced in the severely imbalanced dataset. These results show promise for active learning in low-resource imbalanced domains but also leave space for further improvement.

Abstract In this work, we provide an extensive analysis of Hohfeld’s theory of normative relations, focusing in particular on diagrammatic structures. Our contribution is threefold. First, we specify an extensional formal language to represent the main notions in the two families of normative relations identified by Hohfeld (i.e. the deontic and the potestative family). Our primary focus is on the part of the theory concerning potestative relations. In this regard, we assign a key role to the concept of ability, which is treated as a primitive notion and used to formulate three fine-grained definitions of power (outcome-centered, change-centered and force-centered). Second, on the basis of these definitions we build Aristotelian diagrams of opposition for deontic and potestative relations, improving, extending and systematizing previous proposals formulated in the literature. Third, we present a model-theoretic interpretation and a logic programming (ASP) implementation of the proposed framework, elaborating on the procedural dimension of normative reasoning.

With the uptake of digital services in public and private sectors, the formalization of laws is attracting increasing attention. Yet, non-compliant fraudulent behaviours (money laundering, tax evasion, etc.)—practical realizations of violations of law—remain very difficult to formalize, as one does not know the exact formal rules that define such violations. The present work introduces a methodological framework aiming to discover non-compliance through compressed representations of behaviour, considering a fraudulent agent that explores via simulation the space of possible non-compliant behaviours in a given social domain. The framework is founded on a combination of utility maximization and active learning. We illustrate its application on a simple social domain. The results are promising, and seemingly reduce the gap on fundamental questions in AI and Law, although this comes at the cost of developing complex models of the simulation environment, and sophisticated reasoning models of the fraudulent agent.

We propose here a conceptual framework by which to analyze legal-regulatory problematics of algorithmic decision-making systems, focusing on mechanisms of value production in their design and deployment. An aim of our intervention is to develop an investigative model for application to algorithmic decision systems with regulatory effects, including predictive artificial intelligence applications and recommender systems that filter data and suggest courses of action. Technical systems that integrate complex algorithmic techniques perform critical and sensitive functions that are both object and instrument of regulatory governance, functions such as predicting behavior, steering information flows, assessing risk, etc. These functions, however, are not simple or static phenomena, but rather contextual, partial performances of complex socio-technical dynamics. One of our interests is to discern what is valorized in this new regulatory ecology. Accordingly, we are sketching a framework to target terms and tokens of value as they are produced, reproduced, incorporated, and translated among design processes, legal practices and background conditions structuring their use. Rather than asking which values AI should satisfy in contested governance contexts, we address conceptually prior questions concerning how values manifest and ‘map’ among context-sensitive computational and social processes in the first place. Furthermore, current research often takes for granted that an AI application is produced against the backdrop of a stable and pre-defined set of values and legal practices. Existing research does not yet adequately account for the ways in which laws and values as produced in and through the ecology of the AI application differ from idealized presuppositions assumed to preexist development of the latter. For the purpose, our contribution engages three broad lines of inquiry: one, we take forward calls for a materialized study of law, such as put forward broadly by Alain Pottage, and as put forward more recently and specifically with respect to computational technologies by Mireille Hildebrandt, among others; two, we contribute to the elaboration of a critical practice for AI, in the tradition of Philip Agre; and three, our attention to assemblages potentially contributes to debates over techno-regulation or regulation by design.

Normative positions are sometimes illustrated in diagrams, in particular in didactic contexts. Traditional examples are the Aristotelian polygons of opposition for deontic modalities (squares, triangles, hexagons, etc.), and the Hohfeldian squares for obligative and potestative concepts. Relying on previous work, we show that Hohfeld’s framework can be used as a basis for developing several Aristotelian polygons and more complex diagrams. Then, we illustrate how logical theories of increasing strength can be built based on these diagrams, and how those theories enable us to determine in a computably efficient way whether a set of normative positions can be derived from another set of normative positions.

The combination of smart contracts with blockchain technology enables the authentication of the contract and limits the risks of non-compliance. In principle, smart contracts can be processed more efficiently compared to traditional paper-based contracts. However, current smart contracts have very limited capabilities with respect to normative representations, making them too distant from actual contracts. In order to reduce this gap, the paper presents an architectural analysis to see the role of computational artifacts in terms of various ex-ante and ex-post enforcement mechanisms. The proposed framework is assessed using scenarios concerning data-sharing operations bound by legal requirements from the General Data Protection Regulation (GDPR) and data-sharing agreements.

This paper aims to set up a conceptual framework for studying the second-order guidance problem—that is, designing coordination mechanisms for autonomous actors by means of adequate monitoring and enforcement measures—in a way which is sensible for designers and users of data-sharing infrastructures such as digital market-places. The paper outlines a minimal, but reusable and extensible computational model to test the sustainability of diverse norm implementations, evaluating it against relevant higher-level models presented in the literature.

In order to effectively implement guidance structures in a computational social system, directives which are specified in general terms of duties and rights need to be transformed in terms of powers and liabilities attributed to social parties. The present paper is a work in progress report on an axiomatization of power structures in a logic programming setting, covering the intentional level in specifying actions, the connection between productive characterization of actions and causation, the default nature of action specifications, failures and omissions, the relations of causation and power, and the concept of interfering actions.

Responsibility, as referred to in everyday life, as explored in moral philosophy and debated in jurisprudence, is a multiform, ill-defined but inescapable notion for reasoning about actions. Its presence in all social constructs suggests the existence of an underlying cognitive base. Following this hypothesis, and building upon simplicity theory, the paper proposes a novel computational approach.

To align representations of law, of implementations of law and of concrete behaviours, we designed a common ground representational model for the three domains, based on the notion of position, building upon Petri nets. This paper reports on work to define subsumption between positional models.

The paper introduces elements of a methodology for the acquisition of descriptions of social scenarios (e.g. cases) and for their synthesis to agent-based models. It proceeds along three steps. First, the case is analyzed at signal layer, i.e. the messages exchanged between actors. Second, the signal layer is enriched with implicit actions, intentions, and conditions necessary for the story to occur. This elicitation is based on elements provided with the story, common-sense, expert knowledge and direct interaction with the narrator. Third, the resulting scenario representation is synthesized as agent programs. These scripts correspond to descriptions of agent-roles observed in that social setting.

Rather than as abstract entities, jural relations are analyzed in terms of the bindings they create on the individual behaviour of concurrent social agents. Investigating a simple sale transaction modeled with Petri Nets, we argue that the concepts on the two Hohfeldian squares rely on the implicit reference to a “transcendental” collective entity, to which the two parties believe or are believed to belong. From this perspective, we observe that both liabilities and duties are associated to obligations, respectively of an epistemic or practical nature. The fundamental legal concepts defined by Hohfeld are revisited accordingly, leading to the construction of two Hohfeldian prisms.

This article presents a conceptual framework intended to describe and to abstract cases or scenarios of compliance and non-compliance. These scenarios are collected in order to be animated in an agent-based platform for purposes of design and validation of both new regulations and new implementations, or to be used as reference base for a diagnosis tool. In our approach, legal narratives become a source of agent-roles descriptions, i.e. abstractions of individual characters/agents from singular stories, feeding the target applicative framework.