Hallucination Detection in Large Language Models

Abstract

The concern over hallucination, a situation where a model generates fluent but factually inaccurate or fabricated information-has grown with the steady development of Large Language Models like GPT and Gemini. Such results have the potential to erode safety, dependability, and trust in AI driven fields such as journalism, healthcare, and education. This problem is addressed through the Hallucination Detector project, which offers an online interactive web tool based on Streamlit for analyzing and identifying hallucinated content in model-generated text. This model applies various methods for model uncertainty and reliability metrics calculation, including ensemble scoring Model(1), LLM-as-a-Judge evaluation Model (2), and Black-Box scoring Model(3), powered by the uqlm library (Uncertainty Quantification for Language Models). All these methods combined evaluate models for agreement, confidence, and factual consistency to calculate the probability of hallucinations. It allows users to interactively detect and visualize language model hallucination tendencies in a straightforward, user-friendly manner. This research contributes to a safer and more understandable usage of generative AI systems by advancing transparency, interpretability, and accountability of AI-generated content through integrating several uncertainty-based evaluation paradigms.