Approaching the AI Act... with AI: LLMs and knowledge graphs to extract and analyse obligations

The EU Artificial Intelligence Act (AIA) exemplifies the growing complexity of digital regulation in the domain of computer technologies. Characterised by abstract terminology, multi-layered provisions, and intersecting regulatory requirements, the AIA poses significant challenges for the identification and interpretation of legal obligations, making compliance a demanding and potentially error-prone endeavour for legal professionals and organisations alike. Recent advances in Artificial Intelligence (AI), particularly in the fields of Natural Language Processing (NLP) and Large Language Models (LLMs), offer promising support for addressing these challenges. By automating the extraction and structuring of legal rules, AI-based tools have the potential to assist regulatory compliance activities and provide more systematic insights into complex legislative frameworks. This paper presents an experiment combining NLP techniques and LLMs to automate the extraction and structuring of legal obligations from the AIA. The approach is based on a modular workflow comprising four main stages: identification of obligations, filtering of deontic statements, analysis of deontic content, and the construction of searchable knowledge graphs. The experiment employed the LLaMA 3.3 70B model, supported by more traditional NLP tools. Five experts (4 Ph.D. students and 1 post-doc in legal informatics and philosophy) evaluated the system’s performance on a subset of cases. The results indicate a precision of 93% in the obligation filtering phase and over 99% accuracy in classifying obligation types, addressees, and predicates. A quantitative analysis of the extracted and analysed obligations revealed a predominance of prescriptive obligations (603 out of 729 total), among which 136 are imposed on the European Commission, while 88 consist of informative duties. The results are in line with current discussions around the AI Act regulatory approach. These findings underscore the potential of LLM-based tools to enhance regulatory compliance and analysis. Future research will focus on extending the system to additional EU regulations and integrating formal ontologies to enable more advanced representations of legal obligations.