Convolutional neural networks are now applied widely to intraoral photographs, and most reports summarise performance as a single accuracy figure. That convention conceals where a model actually fails. We trained an EfficientNet-B0 classifier on 12,320 intraoral photographs covering six conditions (calculus, caries, gingivitis, hypodontia, mouth ulcer, and tooth discoloration) drawn from a public dataset that had not been deduplicated, splitting the images 80/15/5 into training, validation, and test sets. On the 616-image test set the model reached 92.37% accuracy (569/616) with a macro-F1 score of 0.910. Per-class precision, however, ranged from 0.631 for calculus to 1.000 for mouth ulcer and caries, and 40 of the 47 errors (85.1%) were mutual confusions between calculus and gingivitis. These are the two conditions that clinicians separate by probing the gingival margin rather than by inspection, and the model, which saw only photographs, reproduced that same difficulty. For dental artificial intelligence this matters practically: a model that scores above 92% overall may still be unreliable on the class a screening programme most needs to catch. We argue that per-class precision, recall, F1-scores, and confusion matrices belong in the main results of every dental AI paper, not in supplementary material.