Quality of Experience Based Trustworthiness for LLMs : New Approach and Use Case

Assuring the trustworthiness of AI-based systems, especially in video surveillance, is critical in smart cities today. Large Language Models have emerged as the strongest tool for analyzing videos by offering human-readable descriptions of complex scenes. However, their effectiveness in sensitive tasks like abnormal behavior detection depends heavily on trust and system reliability. This paper proposes a new approach for enhancing trustworthiness in LLMs using a QoE-based framework. We propose a system that couples these Vid-LLMs with a real-time feedback loop, which incorporates both user corrections and validations to improve accuracy and user trust. We show how such a system can be trained and evaluated using the UCA Crime database and demonstrate that it is able to describe abnormal events in a highly reliable way. Our results indicate that the QoE-based trust model significantly improves user satisfaction, serving as a valued approach in real-world surveillance applications or other similar use cases.