Hybrid AI transforms European Union legal drafting processes
A study by the University of Bologna reveals how integrating hybrid AI into the LEOS editor enhances the precision and transparency of EU legislation. By combining rule-based logic with Large Language Models, the project successfully automates complex tasks like citation retrieval and the mapping of legal recitals. This strategic shift toward machine-readable legal documents ensures greater interoperability and consistency across the European Union’s regulatory landscape.
Points clés
- The University of Bologna conducted the study for the European Commission (DG DIGIT), concluding in December 2025.
- The project utilizes “hybrid AI,” blending symbolic rule-based systems with statistical Large Language Models (LLMs) to ensure explainability.
- All documents use the Akoma Ntoso (AKN) XML standard to maintain machine-readability and legal interoperability.
- The REFERA tool suggests valid normative references using temporal metadata and EUROVOC descriptors to eliminate citation errors.
- The DEFINA assistant achieved 72% accuracy in identifying and reusing established legal definitions from a 10-year EU corpus (2010–2021).
- RECONTA employs a fine-tuned ModernBERT transformer model to map the interpretative relationships between non-binding recitals and binding articles.
- The study analyzed 891 “Reporting Requirements” to create standardized templates for institutional actions like Commission reports.
- The TRENDS module extracts semantic entities such as Agents, Actions, and Deadlines into RDF formats for automated monitoring.
- The architecture is built on RESTful microservices, allowing AI features to be integrated directly into the LEOS drafting environment.
- The methodology prioritizes “white-box” transparency to prevent the risks associated with “black-box” AI in the sensitive legal domain.
À retenir
So, it turns out that making European law readable for humans was too difficult, so we’ve decided to make it readable for machines instead. If you’re a non-expert, just know that the “black-box” robots aren’t taking over yet; we’ve given them a “white-box” leash and some XML homework. My advice? Don’t worry about the definitions anymore—DEFINA is 72% sure it knows what it’s talking about, which is honestly a higher percentage than most of us on a Monday morning. Just keep clicking the templates and let the metadata worry about the deadlines.
Sources
Quiz sur le document: 10 questions
