Calibration Collapse: On the Disconnect Between Predictive Probability and Ground Truth in Deep Neural Networks

Modern deep neural networks often produce high-confidence predictions that are incorrect — aphenomenon known as miscalibration. This paper examines calibration collapse in deeplearning systems. We formalize calibration, identify its causes including overparameterizationand limited training coverage, and empirically demonstrate the confidence-accuracy gap onMNIST using Expected Calibration Error (ECE) and reliability diagrams. We then connectmiscalibration to AI safety, arguing that overconfidence threatens high-stakes domainsincluding healthcare, autonomous vehicles, and criminal justice. We survey mitigationstrategies and their limitations. Calibration is not optional — it is essential for trustworthy AI.