Human-Centric AI Architecture Improves Collaborative Decision-Making

The research addresses the challenges in human-AI collaborative decision-making, where humans often mis-rely on AI recommendations and AI systems are poorly calibrated to human expectations. It proposes a novel framework called the Human-Centric Reflective Architecture (HCRA) to improve decision-making effectiveness and align AI agents with human preferences and needs, particularly in safety-critical applications. HCRA conceptualizes the collaborative decision-making task as a stochastic game between an AI agent and a human. The architecture integrates models calibrated to human behavior with reinforcement learning agents that process linguistic feedback through an iterative, reflective process. Evaluation results indicate that HCRA significantly enhances the effectiveness of decision-making and generates higher-quality recommendations, mitigating risks associated with AI non-determinism.