FROM RECOGNITION TO RELIABILITY: A FRAMEWORK FOR TRUSTWORTHY MULTIMODAL AI IN CULTURAL HERITAGE (THE MUSEEPAL CASE)

The integration of Artificial Intelligence (AI) into cultural heritage offers transformative potential for visitor engagement but faces a critical challenge: the "hallucination" of facts by Generative models in high-stakes domains.This paper presents MuseePal, an AI-powered tourist assistant designed to mitigate these risks through a formalized, constraint-based architecture.Unlike generic chatbots, MuseePal operates on a rigorous Grounding-Aware framework that synthesizes real-time vision with a strict retrieval-augmented generation (RAG) pipeline.We introduce a theoretical formalism to define a generalized Cultural Knowledge Mapping function, ensuring that generated outputs are structurally bound to a curated ontology and verifiably attributed to source documents.While our proof-of-concept implementation utilizes the YOLOv12 architecture and Gemini models, the proposed framework is model-agnostic, adaptable to various perception and reasoning engines.Evaluation within the Grand Egyptian Museum context demonstrates how these formal constraints bridge the gap between generative AI capability and reliability.