Uncertainty calibration, the process of aligning a model’s predictive confidence with the likelihood of correctness, is crucial for artificial intelligence systems. Despite recent advancements in deep learning, conventional models often struggle to calibrate confidence in a manner similar to human judgment (Guo et al., 2017). Specifically, while increased model capacity facilitates accurate classification, it frequently undermines confidence calibration. For instance, these models tend to exhibit overconfidence when confronted with unknown outlier samples, even in the absence of relevant knowledge, leading to misclassifications or hallucinations. Such miscalibrated confidence poses significant challenges in real-world applications, where incorrect predictions can result in costly or critical outcomes. To address this issue, we introduce a random noise pretraining method (Cheon et al., 2024) inspired by prenatal processes in developing brains, where networks are pretrained with random noise stimuli prior to exposure to training data. Our findings indicate that this approach enables neural networks to calibrate uncertainty, consistently aligning confidence levels with actual accuracy. We observed that randomly initialized, untrained networks display excessively high confidence, even when their knowledge is insufficient. However, training with random noise facilitates effective calibration, standardizing confidence levels to align with chance across input spaces. Consequently, networks pretrained with random noise begin their learning with significantly reduced calibration error, thereby matching confidence to accuracy. During subsequent data training, these networks exhibit increases in confidence levels that closely correspond to improvements in accuracy. The pretrained networks demonstrate a confidence level that accurately reflects the likelihood of correctness, achieving near-ideal calibration. Notably, these pre-calibrated networks also exhibit lower confidence when encountering "unknown data," enhancing their ability to robustly detect out-of-distribution samples based on their confidence levels. Overall, our results suggest that random noise pretraining serves as an effective strategy for addressing uncertainty calibration issues in neural networks, applicable to both in-distribution and out-of-distribution data.