How Analysts Use AI in High-Stakes Crime Linkage: An Industrial Study

Analysts maintain healthy skepticism of AI predictions: Rather than blindly accepting AI recommendations, crime analysts in this study "used the AI predictions selectively and frequently validated them against behavioural (non-AI) evidence," demonstrating partial trust and continued reliance on their established analytical practices—a responsible approach to human-AI collaboration in high-stakes decisions.

Explainability matters, but presentation needs improvement: Analysts "attended to the presented model features and valued their availability," indicating that understanding why the AI made a recommendation is important to them. However, the study identified "opportunities to improve how explanations are presented and integrated into the workflow," suggesting current explanation methods could be more effective.

Real-world evaluation reveals gaps that lab testing might miss: This industrial study using direct observation, eye-tracking, and mouse-tracking in actual work settings uncovered practical insights about how analysts interact with AI that might not emerge from controlled experiments, highlighting "the importance of in-situ evaluation for enginee[ring]" responsible AI systems.

Integration with existing workflows is critical: The findings emphasize "the need for AI-enabled decision-support tools to better integrate explanations and traditional analytical methods," suggesting that AI tools are most trustworthy and effective when they complement rather than replace established professional practices and expertise.