Effectiveness of Large Language Model-Based Assistance in Escape Room Gameplay: Puzzle-Solving and Enjoyment

Artificial intelligence (AI), particularly large language models (LLM), has developed rapidly in recent years. An increasing amount of applied research on AI in the gaming field has been observed. In this study, we constructed a simple escape room game prototype and used an LLM chatbot to create the escape room virtual assistant (EVA). The EVA can provide hints based on the player's questions and the current decryption status, affecting the player's next action. EVA also automatically gives players prompts every 3 min based on the player's current decryption status to prevent them from becoming stuck. We describe the production method of the EVA and the experiences and findings gained during various trial runs. Our hypotheses were as follows: 1) EVA can improve players' puzzle-solving efficiency, and 2) EVA can increase the enjoyment of players. Based on the semantic differential questionnaire responses from the participants, our results suggest that EVA is perceived as helpful and enjoyable, and that user reliance on EVA was high, indicating the strong potential of LLMs as puzzle assistants and highlighting future development opportunities.