This retrospective study analyzed the errors generated by a convolutional neural network (CNN) when performing automated classification of oral lesions according to their clinical characteristics, seeking to identify patterns in systemic errors in the intermediate layers of the CNN.