Artificial Intelligence in Orthodontics: From Laboratory Benchmarks to Clinical Care

Artificial intelligence (AI) is now used in routine orthodontic practice, not only in research, but how well it works depends heavily on the application. This narrative review synthesises peer-reviewed studies published between 2019 and early 2026 across seven task domains: cephalometric landmark detection, cervical vertebral maturation (CVM) staging, treatment-decision support for extraction and orthognathic surgery, cone-beam computed tomography (CBCT) segmentation, aligner monitoring, treatment-outcome prediction, and large language model (LLM) patient communication, which differ markedly in their level of maturity. In cephalometric landmark detection, mean radial errors ran from roughly 1.0 mm in three dimensions to about 1.37 mm in two dimensions, and pooled detection reached about 81% at the 2 mm threshold. CBCT segmentation reached pooled Dice similarity coefficients of 0.93 for teeth and 0.91 for the maxilla on multicentre data. Extraction-decision models performed with a sensitivity of 0.70 and a specificity of 0.90, yet lost as much as 20% of their accuracy once tested across institutions. Remote aligner monitoring reduced in-office attendances by roughly 1.68 to 3.5 visits over a treatment course, although the same platforms detected periodontal status poorly, with sensitivities of only 0.53, 0.35, and 0.22 for plaque and calculus, gingivitis, and recession, respectively. Soft-tissue prediction after orthognathic surgery remained inaccurate at the lip and chin. Patients tended to prefer LLM-generated information about their treatment, whereas orthodontic experts rated the same material less favourably. Three problems recurred across every domain examined: training data confined to single centres, narrow demographic representation, and an absence of independent external validation. None of these applications removed interpretive responsibility from the clinician, who retained the analytical decision, even where AI reduced the computational burden.