Uncertainty, Disagreement, and Information Fusion in Modern AI Systems: A Comprehensive Survey

Abstract

The reliability of modern artificial intelligence systems critically depends on their ability to quantify uncertainty, interpret disagreement, and fuse information from multiple models or data sources. While classical machine learning emphasized probability calibration for discriminative classifiers [10], contemporary AI systems increasingly operate in open-ended, generative, and multimodal settingswhere uncertainty is semantic, subjective, and decision dependent [17]. This survey provides a comprehensive review of uncertainty estimation, disagreement modeling, and information fusion in modern AI systems. We formalize aleatoric and epistemic uncertainty [4] and review foundational evaluation metrics including negative log-likelihood, proper scoring rules, and calibration error [8]. We then survey uncertainty estimation methods spanning Bayesian neural networks [5], [22], deep ensembles [18], calibration techniques [10], consistency-based approaches [27], and semantic uncertainty measures for generative models [17]. Disagreement is analyzed as an informative signal rather than noise, covering ensemble disagreement [12], human annotation variability, and multi-agent systems. Finally, we review uncertainty-aware information fusion techniques for multimodal and multi-model systems [6]. Throughout the survey, we highlight empirical trade-offs between computational cost and reliability, discuss evaluation practices, and identify open challenges related to scalability, distribution shift, hallucination detection, and decision-theoretic integration. This work aims to serve as a unified reference for researchers and practitioners designing AI systems that are not only accurate, but reliably aware of their own limitations.