Countering AI Hallucination by Utilizing a Concept-Aware Model
Yousif Ahmed Mohamed, Yee‐Loo Foo · 2025
This paper showcase how to reduce artificial intelligence hallucination by making information retrieval and task execution more efficient, AI-based virtual assistants have revolutionized organizational communication. However, they are still susceptible to hallucinations, which can result in outputs that are confidently expressed but inaccurate or deceptive. Current defenses, like context enrichment and rule-based filters, produce inconsistent outcomes and uneven accuracy, which erodes efficiency and confidence. To significantly lower hallucinations, we present a hybrid framework that combines uncertainty-aware screening with a multi-stage self-familiarity model. While Bayesian deep learning and Monte Carlo dropout quantify response confidence and allow the system to verify or withhold uncertain answers, the self-familiarity component strengthens conceptual understanding across processing stages. Our method improves overall communication effectiveness in organizational settings, increases user trust, and produces more accurate and dependable responses.