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

Abstract: This study examines how LLM-driven non-player characters (LLM-NPCs) affect players' cognitive load and gaming experience, with a particular focus on the underlying psychological mechanisms, differences across task scenarios, and the role of individual traits. Conducting a randomized between-subject experiment (N=130) in a self-developed game prototype "Campus Culture Week", we compared player interactions with LLM-NPCs and traditional pre-scripted NPCs across multiple interactive modules. The results showed that LLM-NPCs significantly increased players' cognitive load (p < .001), an effect mediated by factors such as expressive effort and response uncertainty. However, LLM-NPCs did not yield a statistically significant improvement in overall gaming experience (p = .195); while they positively influenced players' perceived autonomy, they exerted a negative influence on system usability and trust. The effects of LLM-NPCs also significantly varied across task scenarios (p < .001), with stronger increases in cognitive load in more open-ended modules such as content creation and relationship building. The influence of individual differences was generally limited, although the personality traits of extraversion (p = .031) and neuroticism (p = .047) demonstrated some predictive power regarding cognitive load. This study provides empirical evidence for understanding the "double-edged sword" effect of LLM-NPCs on player experience, and highlight the importance of scenario-sensitive and user-sensitive design in intelligent NPC systems.

• The paper establishes that LLM-driven NPCs significantly increase player cognitive load through higher expression cost and response uncertainty.
• Experimental findings reveal that, despite enhanced perceived autonomy, LLM interactions lower system usability and player trust, affecting overall experience.
• Results advocate for a hybrid NPC design that utilizes LLMs in open-ended tasks while minimizing cognitive overload through context-sensitive strategies.

LLM-Driven NPCs in Games: Cognitive Load and Player Experience

Introduction

The integration of LLMs into interactive digital games as non-player characters (LLM-NPCs) is redefining player-NPC interactions, moving beyond static dialogue trees and offering open-ended, context-adaptive engagements. The study titled "The Double-Edged Sword of Open-Ended Interaction: How LLM-Driven NPCs Affect Players' Cognitive Load and Gaming Experience" (2604.10107) delivers a rigorous empirical analysis of how such agents transform cognitive and experiential aspects of gameplay. Through a structured experiment comparing LLM-NPCs to traditional scripted NPCs across diverse in-game interaction modules, the work systematically dissects both benefits and limitations of these approaches.

Methodology and Experimental Design

The authors developed the "Campus Culture Week" game prototype in two parallel versions: one using traditional scripted NPCs and another leveraging GPT-4.1-based LLM-NPCs. They operationalized seven canonical player-NPC interactive scenarios: daily chatting, relationship building, task delegation, collaboration, investigation & reasoning, negotiation & persuasion, and content creation. Employing a between-subjects design, 130 participants were randomly assigned to each prototype, enabling direct comparison across standardized game tasks.

Quantitative evaluation encompassed process-level measures (cognitive load, gaming experience, mechanistic variables) after each module and comprehensive post-test assessments. Tool reliability was validated via Cronbach’s alpha, and statistical analyses employed mixed-effects models, mediation analysis via bootstrapping, and multiple regression for individual difference factors.

Main Findings

Cognitive Load and Gaming Experience

LLM-NPCs significantly increased players' cognitive load: process-level analysis revealed a robust main effect ($p < .001$) with a large effect size (Cohen’s d = 0.884). This elevation is attributable to both higher expression cost (effortful language organization/input) and increased response uncertainty (ambiguity regarding expected NPC feedback).

Contrary to common expectations, the use of LLM-NPCs did not elicit significant improvements in overall gaming experience (post-test, $p = .195$). While perceived autonomy increased (Cohen’s d = 0.620, $p < .001$), this was offset by reductions in system usability (Cohen’s d = -0.866, $p < .001$) and player trust (Cohen’s d = -0.412, $p = .020$). Subjective evaluations cited LLM-NPC dialogues as more time-consuming, labor-intensive, and unpredictably guided compared to menu-driven interactions.

Task Scenario Modulation

Task context strongly moderated LLM-NPC impact on cognitive load ($F = 15.569$, $p < .001$), with load increases most pronounced in modules with high openness and expressive demands (relationship building, content creation). Preset-option interactions outperformed LLM-NPCs in modules emphasizing clarity, efficiency, and low expressive burden (collaboration, task delegation).

Despite these load differentials, real-time gaming experience measures did not vary significantly by interaction type across task scenarios. That is, subjective immersion and satisfaction ratings remained stable, even as cognitive demands diverged.

Mediation Analysis

Mediation results clarify pathways underlying observed effects: expression cost and response uncertainty significantly mediated the increased cognitive load in LLM-NPC conditions, while perceived autonomy positively mediated improved subjective experience. However, these gains were neutralized or outweighed by negative mediations through reduced system usability and trust—explaining the absence of net gaming experience benefit.

Individual Difference Effects

Exploration of individual differences found negligible impact of age, game frequency, AI literacy, and expressive skill on cognitive load or experience. However, higher extraversion and neuroticism independently predicted greater cognitive load in LLM-NPC interactions, indicating personality-driven variance in tolerance to expressive and uncertainty burdens. Notably, game genre preferences did not systematically affect outcome measures.

Subjective and Qualitative Insights

Players preferred natural language interaction in relaxed, open-ended modules (e.g., daily chatting, relationship building), citing emotional resonance and realism. Conversely, modules necessitating precision and logical clarity (e.g., investigation & reasoning, negotiation & persuasion) were deemed better suited to menu-based interfaces. Prominent negative aspects of LLM-NPCs included inefficient operations, lack of clear guidance, and unpredictable behavior; their main advantages were autonomy, immersion, and surprising reactivity.

Implications

Practical Deployment

These findings warrant scenario-sensitive deployment of LLM-NPCs, maximizing their advantages in emotionally expressive, creative, and open-world contexts while retaining traditional option-based interfaces for tasks prioritizing efficiency, clarity, or low cognitive cost. A hybrid system architecture, blending discrete menu options with natural language input as contextually appropriate, is suggested to optimize both user autonomy and cognitive tractability.

Design Recommendations

Design of LLM-NPCs should focus on:

• Reducing expression cost through prompt engineering, contextual scaffolding, and adaptive UI aids.
• Mitigating response uncertainty via enhanced intent-parsing, clarification requests, or fallback strategies.
• Improving usability and trust through consistency checks, fail-safe dialogue flows, and transparent feedback mechanisms.

Scenario-based NPC adaptation and personalization based on personality profiling may further refine interactions, matching interaction style to player disposition.

Future Research Directions

The study highlights critical areas for further exploration:

• Longitudinal adaptation: Whether familiarity with LLM-NPCs reduces cognitive overhang and shifts experience.
• Validation in commercial games: Testing ecological robustness and scalability in real-world, multi-modal environments.
• Extension to multimodal interaction: Incorporating speech, vision, and gesture to potentially reduce expressive burden and augment immersion.
• Personalization frameworks: Automated detection and adaptation of NPC behaviors to player traits and contextual factors.

Conclusion

This research provides robust empirical evidence delineating the double-edged effects of LLM-driven NPCs: while increasing perceived autonomy and potential for immersive, open-ended interaction, they impose consistent and sometimes prohibitive cognitive costs, compounded by reduced usability and trust. Strategic, scenario-driven, and hybrid integration of LLM-NPCs is necessary to harness their strengths while attenuating their weaknesses. Theoretical advances in personalized interaction and adaptive dialogue design, informed by ongoing user studies and cross-modal system integration, will be essential to realizing the full potential of LLM-powered agents in interactive digital games.