Balancing Safety and Information Disclosure in Medical Chatbots: A Signaling Game Approach

Artificial Intelligence (AI) systems and Large Language Models (LLMs) are increasingly used in healthcare applications, including medical chatbots that provide information and guidance to users. However, these systems face an im- portant challenge: they need to give useful answers to normal users while also stopping people who try to misuse them to get harmful or sensitive information. Many current methods mainly rely on filtering content or using fixed rules for safety, but these do not always fully handle the interaction between users and the chatbot. This thesis addresses the central research question: How can game theory help a medical chatbot distinguish between honest and malicious users and respond in a safer and more reliable way? To answer this question, the interaction between the user and the chatbot is modeled as a signaling game with asymmetric information. The user acts as the sender, while the chatbot acts as the receiver. A decision framework based on expected utility and risk thresholds is used to guide the chatbot’s behavior under uncertainty. As part of this work, an existing system, Clinical-ChatBot, is modified and extended by integrating a game-theoretic decision framework. The system is combined with RAG and a vector database (Pinecone) in order to retrieve rele- vant medical knowledge and improve response reliability. To evaluate the proposed approach, datasets containing both normal medical questions and potentially harmful prompts are used. The queries are analyzed and classified according to their level of risk, and the chatbot’s decision strategy is tested through different interaction scenarios. The chatbot can choose between several actions such as Allow, Restrict, or Clarify, depending on the estimated risk level of the user query. In conclusion, this thesis demonstrates the importance of incorporating game- theoretic reasoning into AI-based medical chatbots, improving their ability to manage uncertain user intentions while maintaining useful communication with legitimate users. Furthermore, this work shows how combining Artificial Intelli- gence (AI), Natural Language Processing (NLP), and game theory can contribute to safer and more reliable healthcare chatbot applications.