“Visible” Intelligence: Designing NPC Decision Explanation Interfaces Based on Web High-Fidelity Prototypes and Empirical Study of Player Trust

Abstract

In today’s video game world, the transformation from predictable state machine to deep reinforcement learning and generalized language model (LLM) has completely changed the mechanism of non-player role (NPC). However, although these generation agents show unprecedented intelligence, they also create a huge transparency vacuum. Players are increasingly talking to “black box” entities, and their reasoning is hidden, which breaks trust when there is AI illusion or unexpected tactical error. Based on the Situational Awareness Agent Transparency (SAT) model, we developed a highly accurate visual interpretation interface system for “translating” machine logic to players to dynamically calibrate trust. We use the Wizard of Oz technology to create a specially designed hybrid Web prototype to simulate the complex information of NPC. The behavioral data of 40 participants recorded with microsecond precision through the performance.now () API shows a very important result: providing complete transparency-especially showing the reasoning of artificial intelligence and its uncertainty-will not cause information overload. On the contrary, when AI fails, it greatly accelerates the intervention speed of players, improves the accuracy of decision-making, and greatly reduces the frustration of users. This study provides a solid framework and practical guide for developers who want to balance AI capabilities and human-centered interpretability in an interactive environment.