Systematising Knowledge Representation Challenges for Compliant Autonomous Agents

The paper at hand explores the concept of compliant autonomous agents (CAAs), i.e., AI systems that not only have to meet technical standards, but also adhere to legal requirements akin to norms directed at human agents. The paper proposes a systematisation of knowledge representation challenges involved in embedding legal norms within AI systems, particularly those relying on formal methods such as deontic logic. A novel systematisation (in the form of a taxonomy) is proposed for these challenges, and their impact on the accuracy, robustness, and explainability of CAAs and related systems is highlighted, while briefly outlining applicable mitigation strategies. The conclusion underscores the need for AI literacy and interdisciplinary collaboration between law and technology to ensure the lawful and reliable operation of CAAs and related systems under regulatory frameworks for AI like the EU AI Act.