Aims: Otoscopic evaluation is critical for diagnosing ear pathologies, yet traditional methods used in primary care settings suffer from diagnostic inaccuracy. Artificial intelligence (AI) offers a promising solution for enhancing diagnostic accuracy. Our convolutional neural network (CNN) classifier was originally trained on a dataset of 10,000 otoscopic images from Aboriginal and Torres Strait Islander children, demonstrating good performance characteristics (AUC of 0.963 to 0.997 on test images). This study aimed to assess the CNN’s cross-population generalisability by evaluating its diagnostic accuracy using adult otoscopic test images not used in its original training. Methodology: This retrospective, cross-sectional study utilised publicly available open-source otoscopic image data from three international databases: Turkey (54 images), Chile (40 images), United States (40 images), each with clinician-derived ground truth classifications. Images were processed sequentially by the CNN, with predictions and confidence levels recorded. Performance was quantitatively evaluated using accuracy, sensitivity and specificity. Results: Across test images, diagnostic accuracy as high as 90% (95% CI 75-100%) was achieved (normal ear in Chile dataset). The model demonstrated overall accuracy of 65.6% (95% CI 57.4-72.7%), sensitivity of 57.1% (95% CI 48.4-65.7%), and specificity of 87.1% (95% CI 83.9-89.8%), indicating good overall discriminative performance. Performance metrics showed greatest accuracy in the identification of normal ear and chronic otitis media across the three geographic cohorts. Conclusion: The convolutional neural network demonstrated good cross-population generalisability for the diagnosis of otoscopic pathology in a heterogeneous adult cohort, despite its original training on a paediatric Indigenous population. Further studies should focus on multi-site, diverse external validation to improve consistent generalisability.