Are Hallucinations Bad Estimations?

2025-09-25

AI Summary

This paper addresses hallucination in generative models—outputs that lack any plausible real-world causal explanation. It argues that hallucination is not due to model incapacity, but rather stems from a structural misalignment between the loss-minimization objective and human-acceptable outputs, manifesting as systematic miscalibration-induced estimation error. Method: The authors formally define hallucination as “an estimate unattributable to any plausible causal source” and derive a high-probability lower bound on the hallucination rate under general data distributions. They empirically validate their theory across three diverse tasks: coin-aggregation, open-domain question answering, and text-to-image generation. Contribution/Results: Their analysis demonstrates that hallucination persists even under optimal loss minimization, establishing its intrinsic inevitability. This work provides a novel theoretical framework for understanding the root causes of hallucination and rigorously characterizes its fundamental limits in generative modeling.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Natural Language Processing: Generation

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

We formalize hallucinations in generative models as failures to link an estimate to any plausible cause. Under this interpretation, we show that even loss-minimizing optimal estimators still hallucinate. We confirm this with a general high probability lower bound on hallucinate rate for generic data distributions. This reframes hallucination as structural misalignment between loss minimization and human-acceptable outputs, and hence estimation errors induced by miscalibration. Experiments on coin aggregation, open-ended QA, and text-to-image support our theory.

Problem

• Hallucinations are defined as ungrounded generative model outputs
• Optimal estimators still produce hallucinations despite loss minimization
• Hallucinations stem from misalignment between loss functions and human expectations

Innovation

• Formalizes hallucinations as estimation-causation failures
• Proves optimal estimators still hallucinate under loss minimization
• Reframes hallucinations as structural misalignment with human expectations