Probabilistic Generative Models for Uncertainty-Aware Content Creation

Abstract

Generative AI systems are increasingly deployed in high-stakes applications where the uncertainty of generated outputs is as important as their quality: a medical imaging system that generates plausible-looking but uncertain diagnoses, or a legal document assistant that produces confident-sounding but factually unreliable text, can cause harm precisely because the system fails to communicate its epistemic limitations to users. Uncertainty-aware content generation -- the principled quantification and communication of model uncertainty alongside generated content -- has received limited systematic attention despite its practical importance. This paper proposes the Uncertainty-Aware Generative (UAG) framework, a methodology for integrating epistemic and aleatoric uncertainty quantification into large generative models across text, image, and code generation domains. UAG comprises three integrated components: a Bayesian uncertainty estimator (BUE) that provides calibrated uncertainty estimates for LLM token predictions through Monte Carlo dropout ensembles; a semantic uncertainty propagator (SUP) that aggregates token-level uncertainties into claim-level and document-level uncertainty representations; and an uncertainty-adaptive generation controller (UAGC) that modulates generation behaviour -- triggering abstention, caveat insertion, or alternative generation -- based on uncertainty thresholds. UAG is evaluated on calibration quality, uncertainty communication effectiveness, and generation quality across four domains. UAG achieves a 34.8% improvement in uncertainty calibration (Expected Calibration Error reduction from 0.148 to 0.097) and a 28.6% improvement in user trust calibration accuracy (users correctly identifying high-uncertainty outputs) relative to standard LLM baselines. The study contributes the UAG specification, a Generative Uncertainty Quality (GUQ) evaluation framework, and empirical evidence for the practical value of explicit uncertainty communication in high-stakes generative AI deployment.