Analyzing the Believability of Game Character Behavior Using the Game Agent Matrix

This paper examines how convincingly non-player characters (NPCs) in modern single-player games behave as human-like agents, a property the authors call believable behavior. Its central argument is that NPC believability depends on whether characters act consistently with their social context, and that existing analysis tools lack precision for diagnosing this. To address that gap, the authors refine the Game Agent Matrix (GAM), an analytical instrument adapted from the Model Social Agent matrix, which classifies NPC behavior by agent type (acting, reacting, interacting) and by social environment (single agent through cultural-historical).

The study observed NPC behavior across 20 AAA role-playing and action titles released within roughly the past decade, recording gameplay sessions and grading each observed behavior as either strengthening or weakening immersion. Behaviors involving language, tool use, and navigation were consistently convincing, while problems clustered around adaptation, awareness, interruptibility, and responsiveness to other agents. Common failures included archers who fail to reposition effectively, NPCs who ignore nearby dangers or fights, and characters who continue fighting until incapacitated. More complex social behaviors, such as norm maintenance, sanctions, etiquette, and cooperation, were often present but stilted, as if NPCs shared a collective cognition rather than acting as individuals.

The authors conclude that the middle columns of the matrix, covering social structure and social goals, are largely unrepresented in current games, indicating a substantial gap in NPC social capability. Based on their observations, they propose refinements to the GAM, including relocating the Face to Face value and adding a Memory of Previous Interactions value, since NPCs frequently repeat themselves and forget past encounters. They also suggest that NPCs should more visibly acquire information rather than appearing to share it invisibly. The authors acknowledge limitations, including a black-box methodology, a narrow game selection, and the absence of player-perception testing, and they outline future work involving broader game samples, developer interviews, and player studies.