Background/Objectives: Accurate prediction of craniofacial growth is essential for establishing orthodontic diagnoses and planning treatment. However, the ultimate extent of jaw growth remains largely judged subjectively based on clinical experience. In this study, we developed a convolutional neural network (CNN) model to predict chronological age from lateral cephalograms and to investigate whether artificial intelligence (AI) can autonomously learn growth-related morphological features. We also reevaluated which anatomical regions are most informative for predicting growth. Methods: We retrospectively analyzed 2116 cephalograms from patients 5–30 years old with malocclusion. After excluding patients with craniofacial syndromes or systemic diseases, 2014 images were used for training and 102 for testing. The mean age of the training dataset was 19.51 years (standard deviation [SD]: 4.71), whereas that of the test dataset was 18.20 years (SD: 7.33). In addition to the entire cephalograms, five regional datasets were generated (mandible, maxilla, cervical vertebrae, frontal region, and cranial base). All images were resized to 256 × 256 pixels and trained with a ResNet50-based CNN. Performance was evaluated using mean absolute error (MAE), Pearson’s correlation coefficient (r), and coefficient of determination (R²). To assess growth-related learning, test data were divided into growth (5–19 years) and post-growth (20–30 years) groups, and age trends were analyzed using a sliding window approach with an 8-year window. Results: The model trained on entire cephalograms achieved high accuracy in the younger group (MAE 1.16 years, r 0.952, R² 0.884), but performance declined markedly in the older group. Among the regional models, accuracy was highest for the mandible, followed by the cervical vertebrae and maxilla. Conclusions: The CNN model predicted chronological age with high accuracy, particularly in patients <20 years old, and may have captured age-associated craniofacial features related to growth. Prediction was most accurate in the mandible and cervical vertebrae—regions clinically used to assess growth. As a proof of concept, this approach may provide a foundation for future studies of craniofacial growth prediction through transfer learning, although further validation is required.