Contestable AI for criminal intelligence analysis: improving decision-making through semantic modeling and human oversight

Criminal investigation analysis involves processing large amounts of data, making manual analysis impractical. Artificial intelligence (AI)-driven information extraction systems can assist investigators in handling this data, leading to significant improvements in effectiveness and efficiency. However, the use of AI in criminal investigations also poses significant risks to individuals, requiring the integration of contestability into systems and processes. To meet this challenge, contestability requirements must be tailored to specific contexts. In this work, we analyzed and adapted existing requirements for criminal investigation analysis, focusing on the retrospective analysis of police reports. For this purpose, we introduced a novel information extraction pipeline based on three language modeling tasks, which we refer to as semantic modeling. Building on this concept, we evaluated contestability requirements and integrated them into our system. As a proof of concept, we developed an AI-driven information extraction system that incorporates contestability features and provides multiple functionalities for data analysis. Our findings highlight three key perspectives essential for contestability in AI-driven investigations: information provision, interactive controls, and quality assurance. This work contributes to the development of more transparent, accountable, and adaptable AI systems for law enforcement applications.