Wanneer het gaat over Responsible AI gaat het niet alleen over het identificeren wat een ethisch risico is en wat er allemaal niet kan. Het vinden van oplossingen is des te belangrijker. Binnen de Responsible AI module van de Open Universiteit heb ik me hiervoor verdiept in het gebruik van AI voor onderzoek binnen het juridisch domein.
Hieronder vind je een korte samenvatting van het artikel.
Artificial intelligence has the potential to significantly improve legal research, but its application in high-stakes domains raises important ethical concerns regarding transparency, fairness, accountability, and human oversight. This paper proposes a responsible Retrieval-Augmented Generation (RAG) architecture to support legal research while preserving human responsibility in judicial decision-making. Building on the principles of Value-Sensitive Design, the proposed architecture incorporates human-in-the-loop verification, a transparent non-trainable retriever, and a generation component restricted to document summarization. To improve retrieval quality and legal validity, the system combines semantically enhanced queries with user verification, retrieves complete legal documents and their citation networks, and relies on Dutch legal data sources and a Dutch-language language model. The design further aims to mitigate automation bias by introducing deliberate friction during user validation and prioritizing accountability over efficiency. Although the proposed architecture has not yet been implemented or empirically evaluated, it provides a conceptual framework for developing trustworthy AI systems in the legal domain. Future work should focus on validating the effectiveness of the proposed safeguards, assessing remaining sources of bias, and extending the architecture to additional legal systems and multimodal legal information.
Enkele bestanden die je kan downloaden.