The Double-Edged Sword of Open-Ended Interaction: How LLM-Driven NPCs Affect Players' Cognitive Load and Gaming Experience

Why Read This from an HCI Perspective

This article shows that LLM-driven NPCs are not just “smarter characters,” but systems that can shape players’ cognitive burden, trust, and sense of autonomy. In particular, when evaluating in-game conversational AI, focusing only on model performance can cause teams to overlook interaction costs that are easy to miss—making this highly relevant for HCI/UX practitioners and researchers. By examining differences across scenarios, it also offers design clues about when AI helps and when it instead becomes a burden.

CIT's Commentary

The core of this study is that LLM-NPCs can be both a “feature that adds fun” and a “mechanism that increases what users have to do in their own heads.” As speech becomes more open-ended, players must continuously guess what they should expect next, and the cost of those guesses can translate into cognitive load. As a result, the effects may swing more dramatically in situations with no clear answers—such as open-ended content generation or relationship-building. These findings raise important questions when translating the work into real products: rather than asking whether the AI can carry on a conversation well, it may be more important to ask how easily users can understand the system’s state and intervene. In particular, trust and usability depend not on the AI being “better,” but on how clearly failure modes and guidance are handled.

Another interesting point is that the influence of task context was larger than individual differences. This suggests that designing LLM-based interactions by looking only at an “average user” can be risky. In real services, it may be necessary to adjust autonomy differently by module and, when uncertainty is high, provide shorter, clearer responses or a path for human intervention. Ultimately, this study is a good starting point for asking not just how naturally to make an AI character, but when and how much to delegate—and when users should be able to take back control.

Questions to Consider While Reading

• Q.When increasing the response freedom of LLM-NPCs, what minimal interface cues can reduce cognitive load?
• Q.How should the role of LLM-NPCs be designed differently for open-ended versus closed-ended tasks so that user trust doesn’t collapse?
• Q.If we apply these results to services outside games—for example, community platforms or learning apps—what new research questions would emerge?