Artificial intelligence in orthodontics: a narrative review of clinical and educational applications

Introduction

Background

Artificial intelligence (AI) is rapidly transforming orthodontic clinical practice. From automating cephalometric analysis to enabling remote patient monitoring and more individualized treatment planning, AI technologies are reshaping how orthodontists diagnose, plan, and deliver care (1-5). The convergence of deep learning (DL), computer vision, and natural language processing (NLP) has produced tools that increasingly support expert-level pattern recognition and workflow efficiency, while also opening new possibilities for teleorthodontics, patient communication, and digital education (2-5). At the same time, AI integration into daily practice raises important concerns relating to ethics, data privacy, transparency, and the need for robust clinician training and oversight (5).

In this review, AI is used as the umbrella term for computational systems designed to perform tasks that normally require human-like pattern recognition, inference, or decision support. Within this broad field, machine learning (ML) refers to methods that learn from data to generate predictions or classifications without relying solely on explicitly programmed rules. DL is a more specific subset of ML that uses multilayered neural networks to identify complex hierarchical patterns, and it has been especially influential in image-based orthodontic applications such as cephalometric landmarking, facial analysis, and radiographic classification. Although these terms are often used interchangeably in the literature, maintaining this distinction helps clarify both the technical basis and the clinical scope of the systems discussed in this review.

Within the broader context of digital health, AI has become one of the most transformative technologies in medicine and dentistry, accelerating the shift in orthodontics from analog diagnostics to fully digital workflows (1). In this environment, ML and DL enable automated image interpretation, pattern recognition, and predictive modelling, which enhance diagnostic precision and reduce operator-dependent variability (2,3). AI systems are now embedded across the orthodontic continuum of care, spanning diagnosis and assessment, treatment planning and simulation, clinical workflow and monitoring, and education and training (4). The maturity of AI in orthodontics is uneven across domains. Some applications, particularly automated cephalometric analysis, remote treatment monitoring, and selected digital workflow functions, are already embedded in commercially available platforms, whereas many others, including broader treatment-planning systems, advanced multimodal fusion approaches, and several predictive or generative models, remain primarily research-stage and still require further validation before routine clinical adoption (1-6).

The expanding research activity in this area reflects a rapidly accelerating global interest. To contextualise this growth, a Scopus title search using the search key detailed in the Methods section showed a marked rise in scholarly output between 2020 and 2025, increasing from 5 publications in 2020 to 59 in 2025, with intermediate counts of 26 in 2021, 21 in 2022, 32 in 2023, and 44 in 2024 (Figure 1). This upward trajectory is illustrated in Figure 1, which shows the annual number of Scopus-indexed publications matching this search key over the same period. This sustained upward trajectory suggests a rapidly expanding and increasingly mature research landscape, and further underscores the orthodontic community’s recognition of AI as a major driver of innovation and precision in clinical practice, in line with bibliometric analyses identifying AI as an emerging domain in orthodontics and related maxillofacial disciplines (6,7).

Figure 1 Annual number of Scopus-indexed publications retrieved using the search key TITLE (orthodont* AND (“artificial intelligence” OR “machine learning” OR “deep learning”)) AND PUBYEAR >2019 AND PUBYEAR <2026, showing an increase from 5 publications in 2020 to 59 in 2025 (total n=187).

Rationale and knowledge gap

Despite this momentum, significant barriers still limit safe and effective clinical implementation. Studies repeatedly highlight constraints such as small and homogeneous training datasets, limited external validation, algorithmic bias, and the opaque “black box” nature of many AI systems (8,9). Orthodontists generally report positive attitudes toward AI but remain concerned about reliability, medico-legal responsibility, and the impact of AI on professional judgment and patient trust (10-12). These issues underscore the need for a critical synthesis focused on how AI is actually being deployed in orthodontic settings and what conditions are required for its responsible adoption.

Recent work has included systematic review-level synthesis of clinical applications such as diagnosis, treatment planning, and monitoring, scoping review-level analysis of face-driven orthodontics centred on facial aesthetics, soft-tissue assessment, and individualized planning, and editorial discussion of ethical and medico-legal concerns surrounding orthodontic AI (1,13,14). However, these reviews do not provide the same broad clinician-facing synthesis across diagnosis and assessment, treatment planning and prediction, clinical workflow and monitoring, and education and advanced modalities that the present review aims to offer. This creates a clear need for an integrative review that not only summarises emerging applications but also organises them in a way that is practically interpretable for orthodontists, educators, and researchers.

Objective

The present review was therefore designed to provide an integrative synthesis of AI in orthodontics across the continuum of care and training. Specifically, it maps AI across diagnosis and assessment, treatment planning and prediction, clinical workflow and monitoring, and education and advanced modalities. Within these four higher-order categories, fourteen application domains are used as a clinician-facing framework to organise recurring use cases identified in the literature in a way that is both analytically detailed and practically interpretable. The review evaluates reported performance, benefits, limitations, and key directions for ongoing development, with the aim of providing an evidence-based account of how AI is reshaping orthodontic science, clinical practice, and orthodontic education. We present this article in accordance with the Narrative Review reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0020/rc).

Methods

This review was conducted as an integrative narrative synthesis because the orthodontic AI literature spans heterogeneous study designs, data modalities, and outcome measures across clinical, workflow, educational, and advanced digital applications. A structured literature search was performed in Scopus for English-language publications from 2020 to 2025 using the search key TITLE (orthodont* AND (“artificial intelligence” OR “machine learning” OR “deep learning”)) AND PUBYEAR >2019 AND PUBYEAR <2026 AND (LIMIT-TO (LANGUAGE, “English”)). Scopus was selected as the sole database because the review aimed to provide a broad, interdisciplinary, and reproducible narrative synthesis across the orthodontic continuum of care and training, rather than a multi-database systematic review. The search was restricted to the title field to prioritise specificity and direct topical relevance. Retrieved records were screened for substantive relevance to orthodontic diagnosis, assessment, treatment planning, prediction, workflow, monitoring, education, simulation, or other advanced digital applications. Both peer-reviewed clinical validation studies and proof-of-concept or exploratory AI models were considered eligible when they contributed meaningfully to understanding orthodontic applications, performance, or translational potential. Non-English items, non-peer-reviewed records, duplicates, and papers lacking substantive orthodontic relevance were excluded.

Thematic coding followed a hybrid approach. Four higher-order categories were defined a priori based on the orthodontic workflow and the intended clinician-facing structure of the review: diagnosis and assessment, treatment planning and prediction, clinical workflow and monitoring, and education and advanced modalities. Within these categories, the final application domains were derived inductively during full-text review by grouping studies according to their primary use case, dominant data modality, and main outcome focus. The final fourteen-domain framework was retained because it represented the smallest structure that preserved clinically meaningful distinctions among recurrent application areas without excessive aggregation or fragmentation (Table 1).

Part A: AI in diagnosis and assessment

AI in cephalometric analysis and landmark identification

Cephalometric analysis remains one of the most established orthodontic applications of AI, supporting diagnosis and treatment planning by characterising craniofacial morphology, skeletal relationships, dental inclinations, and growth patterns. Manual tracing is time-consuming and prone to inter- and intra-examiner variability, with outcomes influenced by clinician experience (15). AI-driven automation has therefore been adopted to improve efficiency and reproducibility in cephalometric workflows.

DL, particularly convolutional neural networks (CNNs), has enabled automated landmark detection with performance comparable to experienced orthodontists (1,16). Landmark localisation is typically evaluated using mean radial error (MRE) and success detection rate (SDR). Across studies, MRE values commonly fall within 1.1–1.8 mm (3,15), while SDR is frequently reported at approximately 70–87% within 2 mm and exceeds 90% at 4 mm (1), consistent with thresholds generally considered clinically acceptable for orthodontic diagnosis.

A range of CNN-based architectures has been applied, including models incorporating residual connections, attention mechanisms, and multiscale feature extraction to improve detection of more challenging landmarks (2). Performance tends to be higher for well-defined skeletal landmarks and lower for selected soft-tissue and dental points where anatomical variability and reduced contrast complicate localisation (3,15).

Several commercial platforms, including CephX, Dolphin Imaging, AudaxCeph, and WebCeph, embed AI-based landmark detection and have been both independently evaluated and, in some studies, compared directly with established commercial cephalometric systems. Agreement with manual or semi-automated tracings is commonly high, with intraclass correlation coefficients (ICCs) above 0.90 reported for many cephalometric variables and Bland-Altman analyses typically showing narrow limits of agreement. However, the available comparative evidence does not currently identify one commercial platform as consistently superior across all tasks. Instead, performance appears to vary by landmark and measurement: some studies reported significant differences between WebCeph and Dolphin for selected variables, whereas others found no significant differences between AI-supported systems and Dolphin-based workflows for several standard cephalometric measurements. Similarly, CephX with manual correction has been reported to generate results comparable to CephNinja and Dolphin while remaining time-efficient, and AudaxCeph has shown good agreement with examiner-supported tracing except for selected difficult landmarks such as Porion and incisor apices. These findings suggest that current commercial platforms are clinically promising but not interchangeable in all respects, and that orthodontist review or selective manual correction remains advisable, particularly for challenging landmarks and treatment-critical measurements (1,2,16,17). In addition, direct head-to-head comparisons among commercial cephalometric platforms remain limited, and most currently available systems should still be viewed as clinician-support tools rather than fully autonomous substitutes for orthodontist judgment.

Beyond accuracy, AI provides substantial time savings and standardisation. Manual tracings often require 10–20 minutes per radiograph, whereas AI systems can complete analyses within seconds (1,3). Reduced examiner-dependent variance, including effects related to fatigue or differing tracing conventions, may support more reproducible diagnoses, particularly for novice clinicians (8,16,18).

A recurring limitation across cephalometric AI is dependence on training data quality and representativeness. Models perform best when trained on large, heterogeneous datasets with consistently annotated landmarks spanning diverse ages, malocclusion types, and imaging devices (3,11,12). Many systems remain based on retrospective single-centre datasets, external validation is still inconsistent, and comparatively few studies have examined downstream clinical impact on treatment planning or outcomes rather than landmark accuracy alone.

Recent work has extended AI-assisted cephalometry from two-dimensional (2D) lateral cephalograms to three-dimensional (3D) applications. Automated segmentation of craniofacial structures on cone-beam computed tomography (CBCT) using AI has shown high accuracy and reduced processing time, enabling more comprehensive evaluation of volumetric relationships and asymmetries (1,19). Although less mature than 2D approaches, these developments support a gradual shift toward more three-dimensional AI-assisted diagnostic frameworks in orthodontics.

Beyond cephalometry and skeletal assessment, recent meta-analytic evidence also suggests that CBCT-based AI can support detection and quantification of orthodontically induced root resorption, with a pooled sensitivity of 0.93, a specificity of 0.89, and an AUC of 0.95, indicating that volumetric orthodontic diagnostic AI is expanding into additional treatment-relevant radiographic outcomes (20).

AI in skeletal maturation and growth prediction

Optimising the timing of orthodontic intervention remains a longstanding clinical challenge. Conventional approaches, including cervical vertebral maturation (CVM) assessment and hand-wrist radiographs, are vulnerable to interobserver variability and interpretative bias. AI offers a more standardised approach, and CVM staging from lateral cephalograms also avoids additional radiation associated with separate hand-wrist imaging (21).

Most AI-based systems in this domain focus on automated CVM assessment. DL models trained on labelled CVM stages have demonstrated high classification performance, with reported accuracies often exceeding 90% in comparison with expert-labelled staging, supporting faster and more consistent distinction of prepubertal, pubertal, and postpubertal phases (21).

ML has also been applied to hand-wrist radiographs, with studies reporting accurate detection of maturity indicators such as ossification patterns and epiphyseal fusion. Although routine use of hand-wrist radiographs has declined due to dose considerations, AI-supported interpretation may improve reliability when such imaging is available (3,19).

A clinically important extension is craniofacial growth prediction. Growth reflects complex and nonlinear interactions between genetic, hormonal, and environmental factors, and conventional models based on population averages or simple regression often fail to capture individualised trajectories. AI models can detect subtle morphological patterns and complex relationships in large datasets, supporting more personalised forecasting (3,21). Promising applications include prediction of mandibular growth direction, rotational patterns, and evolving skeletal relationships, using either sequential cephalograms to learn longitudinal trends or models integrating demographic variables, dental parameters, and early growth markers (2).

AI-based prediction is also increasingly discussed in relation to growth-modification decisions, for example, estimating whether a Class II patient is more likely to benefit from growth modification versus camouflage or orthognathic pathways (3,19,21). Key limitations mirror those observed across diagnostic AI: performance depends on imaging quality and consistent acquisition protocols, and many models remain trained on retrospective single-centre datasets with limited external validation. In addition, because CVM staging itself is based on categories that may be interpreted inconsistently across clinicians, AI systems trained on labelled CVM datasets may reproduce annotator subjectivity, and uneven representation of maturation stages may further increase the risk of class imbalance and overfitting (15).

AI in dental and skeletal malocclusion classification

Beyond landmark detection, AI models demonstrate strong performance in malocclusion classification. DL systems can categorise sagittal skeletal patterns (Class I, II, III) and vertical facial types from cephalograms with reported accuracies commonly in the 90–97% range, which in several studies was comparable to expert orthodontist classification under the evaluated conditions (1,3). Validation work supports reliable classification of sagittal discrepancies and vertical facial patterns such as hyperdivergent, normal, and hypodivergent types (2).

In three-dimensional imaging, ML applied to CBCT and 3D intraoral scans can assess transverse discrepancies, skeletal asymmetry, and occlusal relationships volumetrically, and can classify skeletal bases and dental arch morphology with high precision, increasingly aligning with computer-aided design/computer-aided manufacturing (CAD/CAM) planning and virtual orthodontic setups (19).

A key development is multimodal modelling. By combining radiographs, photographs, and digital scans, these systems integrate skeletal and dental information in a way that reflects clinical reasoning. Improved skeletal discrepancy classification was reported when cephalometric and photographic inputs were combined compared with single-modality models (22). Despite promising results, performance remains sensitive to protocol variability, population differences, and the broader clinical context (soft tissue, function, airway, and patient-centred factors) that current AI systems only partially capture (8).

AI in facial and intraoral image analysis

Facial and intraoral photographs are central to orthodontic diagnosis, treatment planning, and monitoring, yet interpretation can vary with clinician experience and subjective judgement (23). AI has strengthened this domain by automating image classification and extracting clinically relevant patterns with high reproducibility.

DL managed to classify orthodontic photographs into nine clinically relevant categories (including frontal and profile facial views, buccal intraoral views, and occlusal photographs), achieving accuracy up to 98% (24). Such performance can streamline digital workflows by supporting rapid triage and organisation of large image databases, even when resolution, angulation, or lighting vary.

AI has also been applied to extract diagnostic features from intraoral photographs, supporting screening and more standardised documentation. Models have demonstrated accurate identification of occlusal features such as midline discrepancies, overbite, overjet, spacing, and crowding from intraoral photographs, and decision-support applications have been described for tasks including extraction planning using intraoral images alongside radiographic data (17,25). A 2026 DL study further demonstrated that Angle’s classification can be derived directly from intraoral occlusal images with strong internal and external performance, reporting 97.41% internal accuracy and 94.3% external accuracy, together with Grad-CAM-based interpretability, supporting the growing diagnostic value of standardized intraoral photography in orthodontics (26).

Facial image analysis further extends AI utility, with CNN-based models used to evaluate facial symmetry, vertical proportions, lip competence, chin projection, and related soft tissue parameters relevant to orthodontic planning (27). In addition, experimental work has suggested that emerging bimodal mechanoluminescent and biosensing visual sensors may enhance future data acquisition and potentially improve AI-driven diagnostic performance (28).

AI in airway and skeletal morphology assessment

Assessment of airway dimensions and craniofacial skeletal morphology informs orthodontic diagnosis and planning, including evaluation of obstructive sleep apnea risk, mandibular deficiency, vertical dysplasia, and asymmetry. AI has been applied to automate and standardise these assessments, particularly within CBCT-based workflows (1,3).

In airway evaluation, DL models are commonly used for automated segmentation of the upper airway on CBCT scans. Given that manual segmentation is labour-intensive and prone to error due to irregular airway anatomy and boundary ambiguity, AI-driven segmentation can rapidly delineate airway regions and compute volume, minimum cross-sectional area, and linear dimensions. AI-based airway segmentation and related upper-airway assessment tasks have shown strong reported performance, including a Dice similarity coefficient of 97% for pharyngeal airway segmentation on CBCT and F1 scores of 0.88–0.90 for upper-airway obstruction and adenoid hypertrophy detection, while also offering faster and more standardised analysis than manual workflows (19).

AI has also been applied to automated segmentation and measurement of skeletal structures, including the maxilla, mandible, cranial base, and temporomandibular joint (TMJ) components on CBCT (19). These applications support evaluation of asymmetry, cortical thickness, condylar morphology, and skeletal discrepancies relevant to orthognathic planning and complex cases, and early work has also suggested utility for identifying radiographic features associated with condylar degeneration and potential screening of temporomandibular disorders (3). An additional recent development is AI-assisted multimodal registration of CBCT and magnetic resonance imaging (MRI) for advanced TMJ and craniofacial assessment. This approach is clinically relevant because CBCT and MRI provide complementary information, with CBCT offering superior visualization of osseous structures and MRI providing better depiction of soft tissues such as the articular disc. Recent work has shown that AI-supported registration pipelines can improve the anatomical coherence of fused CBCT-MRI images compared with manual alignment, thereby supporting more confident interpretation of condylar morphology, disc position, and surrounding joint relationships in complex TMJ presentations (29). Such workflows remain technically demanding because of differences in voxel size, slice thickness, field of view, head positioning, and image geometry between modalities, but AI-assisted orientation, segmentation, resampling, and registration are beginning to reduce these barriers (30). A closely related emerging direction is the generation of MRI-derived synthetic computed tomography (CT), which may eventually strengthen radiation-sparing multimodal workflows and CT-like hard-tissue representation from MRI data, although its orthodontic clinical use remains less established at present than CBCT-MRI registration.

AI-based systems can further assist with construction of three-dimensional reference planes and quantification of asymmetry, supporting planning for orthognathic surgery and skeletal expansion (7,19). Predictive approaches have also been explored to estimate how orthodontic or orthognathic interventions may influence airway space, supporting interdisciplinary planning in selected patients (1,3).

Key limitations are airway-specific as well as generalisable across imaging AI. Airway morphology is dynamic and affected by head posture, respiratory phase, and tongue position; static CBCT-based models may not fully capture this variability (8). Differences in CBCT protocols across clinics (voxel size, field of view, exposure parameters) can also influence model performance (19). Nonetheless, recent studies support the value of DL-based segmentation for efficient evaluation of craniofacial structures and airway morphology (31,32).

AI in aesthetic and treatment need evaluation

Aesthetic assessment and determination of treatment need influence clinical decision making, patient motivation, and eligibility for publicly funded orthodontic care. Conventional approaches, including the aesthetic component (AC) of the index of orthodontic treatment need (IOTN), smile analysis, and profile evaluation, rely on clinician judgement and are therefore subject to variability (23). AI offers an avenue for more standardised, image-based scoring and improved consistency (3).

AI systems trained on labelled datasets can analyse frontal and smiling photographs to evaluate smile harmony, gingival display, incisal show, lip curvature, buccal corridors, profile convexity, and facial proportions. A study reported that AI can predict the IOTN AC with a sensitivity of 77%, a specificity of 88%, and an overall accuracy of 82%, comparable to expert evaluation (27). Such objective scoring may also have public health relevance in systems where care allocation depends on IOTN thresholds, potentially reducing bias and supporting more equitable resource distribution (8,27).

AI-based aesthetic simulation represents a further development. Generative models and soft tissue prediction algorithms can simulate post-treatment facial and smile outcomes, allowing realistic previews and supporting informed consent and treatment acceptance (1,19). Multimodal systems that integrate cephalometric data, photographs, and three-dimensional scans may further strengthen this domain by relating aesthetic outputs to underlying skeletal and dental morphology (19). This trend is reinforced by recent face-driven orthodontic literature, which emphasizes AI-assisted facial analysis, soft-tissue evaluation, 3D reconstruction, and treatment simulation as central components of individualized, aesthetics-oriented treatment planning, while also noting persistent limitations in external validation and clinical applicability (13).

Limitations remain important, particularly because aesthetic preferences vary across cultures and individuals, and models trained on nonrepresentative datasets may generalise poorly. Photographic variability (lighting, posture, camera angle, resolution) can also affect performance (8). However, as datasets expand and become more diverse, AI-based aesthetic assessment is increasingly positioned as a useful adjunct for both individual clinical care and population-level orthodontic service planning.

Part B: AI in treatment planning and prediction

Within Part B, domains were separated according to the primary role of AI in decision-making. Discrete high-stakes case-selection tasks, such as extraction versus non-extraction decisions and orthognathic versus camouflage pathways, are discussed separately from broader comprehensive planning systems that integrate multiple data streams to support overall treatment design, sequencing, and digital setup optimisation.

AI in extraction and orthognathic surgery decision-making

Decisions regarding tooth extraction and orthognathic surgery are among the most complex and consequential steps in orthodontic treatment planning, with long-term implications for facial aesthetics, occlusal stability, periodontal health, and airway function. Traditionally, these decisions have relied heavily on clinician experience and subjective interpretation of diagnostic records (3). AI has therefore gained traction as an adjunct that can add objective, data-driven decision support and improve consistency in borderline cases (1,19).

AI-based extraction prediction models typically apply ML and DL to cephalometric radiographs, dental models, intraoral photographs, and relevant demographic or clinical variables to identify patterns linked to crowding, incisor proclination, dental protrusion, and soft tissue imbalance, all central to conventional extraction reasoning (17). Across studies, AI systems have achieved extraction predictions comparable to expert orthodontists, reducing inter-examiner variability and improving reproducibility when the indication is less clear (3,19). Beyond extraction selection, some tools extend into mechanotherapy planning by estimating anchorage demands, anticipated tooth movement sequences, and space-closure challenges, and by highlighting potential risks such as bite deepening or anchorage loss (2,19). This supports a more individualised approach aligned with precision orthodontics.

In the orthognathic domain, AI has shown value in distinguishing cases appropriate for surgery from those likely to be managed with orthodontic camouflage. Because orthognathic decision-making integrates skeletal discrepancy severity, soft tissue harmony, function, and airway considerations, multimodal AI systems are particularly relevant (2,3). Models trained on combinations of cephalometric data, CBCT images, and facial photographs have demonstrated high accuracy in classifying surgical versus non-surgical cases, adding diagnostic objectivity and supporting clinician confidence (19). DL-based simulation tools also enable prediction of postoperative skeletal changes and soft tissue adaptations, providing realistic visualisations that support patient understanding and interdisciplinary communication (1). In parallel, airway-focused prediction has become increasingly important in orthognathic planning, and AI models have been explored to estimate how proposed jaw movements may alter airway dimensions, supporting safer surgical design (3,19). Overall, evidence suggests that AI can contribute meaningful support for extraction and orthognathic decision-making by improving consistency, accelerating interpretation of complex records, and enhancing patient-centred planning and communication (1,5,17,19).

Important limitations remain. Many extraction and orthognathic decision-support studies are retrospective and based on selected datasets, often centred on non-complex malocclusions, fully dentate patients, or narrowly defined surgical groups, which may restrict generalisability to more diverse real-world cases. In addition, these models are typically trained against expert-defined treatment decisions, yet expert opinion itself may vary across clinicians and treatment philosophies, raising questions about the stability of the reference standard. As a result, current systems are best understood as decision-support tools that may improve consistency and efficiency, but still require orthodontist oversight and patient-specific clinical judgement.

AI in comprehensive treatment planning

AI-driven treatment planning is evolving from single-task tools toward systems that synthesise multiple data streams, including cephalometric analysis, facial and intraoral photographs, and digital scans, to propose more holistic, evidence-informed strategies (3). By integrating cephalometric measurements, dental arch form, occlusal relationships, and soft tissue characteristics, ML and DL models can identify diagnostic patterns linked to particular malocclusion presentations and recommend planning options consistent with those patterns (17,19). This has particular relevance for integrated planning tasks that require simultaneous consideration of cephalometric findings, arch form, occlusal relationships, soft tissue characteristics, biomechanical feasibility, and anticipated treatment efficiency. Rather than replacing discrete decision-support tools discussed separately elsewhere in this review, these systems aim to synthesise multiple inputs into a more holistic treatment strategy.

This direction is supported by a 2026 treatment-planning study that evaluated multiple machine-learning and deep-learning models across extraction, non-extraction, functional appliance, and orthopedic treatment categories, with the best-performing artificial neural network achieving balanced accuracy of 0.83, F1-score of 0.84, and area under the receiver operating characteristic curve (ROC-AUC) of 0.90; Shapley Additive Explanations (SHAP) analysis further identified clinically meaningful predictors such as vertical face proportions and mandibular plane angle (33). However, the study also highlights persistent translational barriers, including single-institution data and a lack of external validation.

AI is also increasingly embedded in digital orthodontic workflows, particularly clear aligner therapy, where treatment success depends on the feasibility and staging of tooth movements. DL systems can automate tooth segmentation, evaluate planned movements for predictability, and optimise staging by flagging rotations, extrusions, or torque movements that may require auxiliaries or refinements (2). These capabilities can improve the accuracy of digital setups and reduce refinement rates, supporting more efficient and predictable aligner care (19). Similar principles apply to fixed appliance planning, where AI-driven tools have been explored to support bracket positioning, archwire sequencing, and anchorage strategies by anticipating biomechanical challenges and helping minimise undesirable side effects such as anchorage loss or occlusal cant development, while maintaining clinician control over final decisions (3).

Predictive analytics constitutes a further planning layer. ML models have been used to estimate treatment duration, the probability of mid-course corrections, and risks related to compliance, particularly relevant in aligner therapy, where adherence strongly influences outcomes (34,35). Such forecasting can support realistic expectation setting, targeted patient communication, and personalised monitoring plans. AI can strengthen interdisciplinary planning in orthodontic-restorative and orthodontic-surgical care by integrating skeletal morphology, soft tissue analysis, smile aesthetics, and occlusal considerations into a more unified overview of patient needs, while simulation outputs can facilitate communication between specialties and support shared decision-making with patients (1,19).

AI in virtual simulation and predictive modelling

Virtual simulation and predictive modelling represent rapidly advancing applications of AI in orthodontics, enabling visualisation of anticipated outcomes, optimisation of staging sequences, and prediction of skeletal and soft tissue responses. Traditional approaches based on manual cephalometric tracing, regression models, or clinician intuition often lack patient-specific precision and can be variable. By contrast, AI models leverage large datasets and pattern recognition to generate more consistent patient-specific predictions (1,3).

A key area is the prediction of tooth movement with greater biological realism. DL models can analyse dental morphology, periodontal support, contact relationships, and occlusal patterns to anticipate how teeth may respond to orthodontic forces (2). This is especially important for clear aligner therapy, where small incremental movements require careful staging. AI tools can identify movements that are less predictable, such as rotations of cylindrical teeth, extrusions, or torque corrections, and indicate when auxiliaries or modified designs may be needed (19). By flagging likely problem areas early, AI may reduce refinement rates and improve the accuracy of virtual setups (19). In fixed appliance therapy, predictive models have similarly been explored to forecast anchorage requirements, tooth movement trajectories, and risks such as bite deepening or midline deviation, supporting more tailored biomechanical strategies (3).

Beyond dental movement, AI has strengthened soft tissue and facial prediction. Whereas traditional models often rely on simplified linear assumptions, AI can learn from large sets of pre- and post-treatment images to generate more nuanced predictions of facial change following orthodontic or orthognathic interventions (19). DL systems can also segment skeletal structures from CBCT scans, simulate osteotomies, and predict postoperative skeletal and soft tissue changes, facilitating comparison of alternative surgical scenarios (19). Airway-focused predictive tools further contribute by estimating how planned jaw movements may affect upper airway dimensions, which is increasingly important for safe orthognathic planning (3). Finally, predictive modelling has been applied to estimate treatment duration and identify patients at higher risk of delays due to poor compliance or oral hygiene issues, enabling more personalised monitoring and communication strategies (34,35).

Beyond prediction alone, recent AI-driven open-source workflows are also improving quantitative evaluation of treatment outcomes. A 2026 multicenter study on automated 3-dimensional craniofacial superimposition in growing patients found high agreement between AI-driven and conventional registration approaches, with most absolute average differences under 1.5 mm for linear and 1.5° for angular measurements, supporting reliable and faster reference-explicit outcome assessment (36).

Part C: AI in clinical workflow and monitoring

AI in remote treatment monitoring

Remote monitoring is one of the most visible AI-enabled shifts in orthodontics, particularly alongside digital workflows, clear aligner therapy, and telehealth. Conventional monitoring depends on periodic in-office visits, which can delay detection of tracking problems or appliance failures and increase time demands on patients and clinics. AI-driven systems now support higher-frequency assessment through patient-generated images and videos, enabling timely identification of deviations. DL models can analyse smartphone intraoral photographs to detect aligner tracking issues, broken appliances, plaque accumulation, and inflammatory signs, supporting continuity of care in aligner-based and hybrid remote treatment pathways (8,34,35).

Platforms such as Dental Monitoring^® use CNNs to analyse smartphone intraoral photographs and track progress, detect aligner tracking issues, identify oral hygiene concerns, and monitor appliance integrity in near real time (34). Algorithms can flag insufficient aligner seating, lagging rotations, broken brackets, archwire disengagement, and early signs of gingival inflammation or plaque accumulation, often before problems would be identified at routine appointments (35). This is particularly valuable in aligner therapy, where predictability depends on fit and adherence. AI-based monitoring can detect deviations from planned movement, estimate the likelihood of refinements, and alert clinicians when adherence appears suboptimal, supporting early intervention (3). Recent multicenter evidence also supports this application in clear aligner therapy: a 2026 study of DentalMonitoring’s AI reported sensitivity and specificity of 93.2% and 86.2% for seat-versus-unseat detection, and 91.1% and 90.5% for noticeable-versus-slight unseat classification, indicating strong detection of clinically meaningful tracking discrepancies in real-world monitoring workflows (37). AI-assisted tracking may reduce unnecessary in-person visits and help prevent prolonged tracking failure (19). Comparable systems have also been used in fixed appliance care to flag broken or loose components and wire-related issues, reducing the risk of unwanted movement and associated complications (34). Automated tools have additionally been explored for monitoring overall treatment progression and post-treatment retention with high consistency (38). New 2026 data also extend this evidence to fixed appliance care: DentalMonitoring’s AI showed sensitivity/specificity of 98.4%/99.6% for bracket debonding, 93.3%/96.5% for tie loss, and 91.1%/88.3% for open self-ligating clips, supporting the value of AI-assisted remote surveillance for common brace-related emergencies (39).

Remote monitoring may also influence behaviour through automated reminders, progress feedback, and visualisation of movement, supporting engagement and adherence (35). Some platforms incorporate gamification elements, which have been associated with improved compliance and performance in digitally monitored patients (34). From a workflow perspective, AI can triage large volumes of patient submissions by highlighting cases requiring clinician review, allowing prioritisation of chair time for patients who need mechanical adjustments while maintaining virtual oversight of stable cases (1). Limitations remain important: performance depends on image quality (lighting, angulation, camera resolution, and cooperation), and effective telemonitoring presupposes digital literacy and reliable internet access, which can amplify inequalities in care (8,11,12). Despite these challenges, current evidence indicates that AI-assisted monitoring can support better engagement and compliance while enabling more flexible hybrid care models (34,35).

AI in clinical documentation and workflow automation

Clinical documentation and administrative workflow are essential but time-consuming components of orthodontic practice, encompassing record organisation, reporting, progress notes, scheduling, and financial processes. These activities often require substantial manual effort and can be vulnerable to inconsistency and clerical error, reducing time available for direct patient care. AI has therefore been increasingly used to automate and standardise routine documentation and operational tasks (3,19).

One established application is automated image classification and record organisation. DL models can categorise orthodontic photographs into clinically relevant groups with high accuracy, supporting consistent labelling, reducing administrative work, and improving dataset organisation for diagnostics and digital workflows (24). In parallel, AI-enhanced radiographic analysis embedded in commercial software can generate cephalometric measurements, skeletal classifications, and soft tissue analyses rapidly, enabling streamlined report production and integration into electronic records (1,2,17). NLP has also been described as a route to more efficient documentation, enabling transcription and structuring of clinical notes and summarisation of key findings into navigable formats that support clinical decision-making (11,12).

AI-driven predictive analytics has been explored for scheduling and appointment management, using patterns of attendance and treatment progress to improve chair utilisation and reduce conflicts, while also identifying patients at higher risk of missed appointments and triggering targeted communication (34,35). Administrative operations such as fee estimation, insurance processing, and resource management may also benefit from AI-supported trend analysis and error reduction (19). Finally, quality assurance applications can identify missing or inconsistent records, flag overdue assessments, and support protocol adherence, potentially strengthening regulatory compliance and risk management (8). Overall, AI-enabled workflow automation is positioned to reduce administrative burden and support more consistent documentation, provided that interoperability, governance, and clinician oversight remain robust (3,19).

Important limitations also apply in this domain. The performance of documentation and workflow automation depends heavily on interoperability between software platforms, the quality and structure of input data, and appropriate integration into existing clinical systems. Errors in automated transcription, summarisation, classification, or record linkage may propagate through the workflow if outputs are not reviewed carefully. In addition, administrative AI raises important concerns regarding data privacy, transparency, accountability, and governance, particularly when patient records or decision-support outputs are handled across multiple digital platforms. These tools may reduce administrative burden, but they should be implemented with secure data handling, clear oversight, and continued clinician review.

AI in integration with 3D printing and CAD/CAM systems

The integration of AI with CAD/CAM and 3D printing has expanded the capabilities of digital orthodontics by supporting customised appliance fabrication and more scalable digital workflows. Traditional digital workflows often depend on manual segmentation, modelling, and design steps that are labour-intensive and prone to human variability. AI has increasingly automated or augmented these processes, improving efficiency and reproducibility (3,19).

A major contribution is automated segmentation and digital model generation. Segmentation of dental and craniofacial structures from CBCT or intraoral scans has historically required substantial manual input, whereas DL models have demonstrated high accuracy in segmenting teeth, roots, alveolar bone, and craniofacial anatomy, reducing time and improving reliability of downstream digital models used for appliance design (1). AI also supports virtual setups and tooth movement planning by evaluating root morphology, crown shape, periodontal support, and occlusal relationships to suggest feasible movements and biologically appropriate staging, which is particularly relevant in clear aligner therapy, where poor staging can increase refinements (2,19). In appliance design, AI has been described as contributing to the customisation of aligners, retainers, indirect bonding trays, customised archwires, and lingual systems by optimising design parameters linked to biomechanical efficiency, including attachment configuration and force delivery patterns (3). Within 3D printing workflows, AI can optimise print orientation, support placement, and material usage, while automated checks can identify segmentation errors or print anomalies earlier in the pipeline, reducing waste and remakes (19).

AI-enabled CAD/CAM integration is also relevant in orthognathic and complex interdisciplinary cases. AI-assisted CBCT segmentation combined with CAD-based planning software can support semi-automated creation of surgical splints, osteotomy guides, and virtual surgical simulations, and predictive modelling can provide foresight into postoperative skeletal repositioning and soft tissue adaptation (1,3). Limitations remain, including restricted interoperability between proprietary systems, the need for clinician oversight in complex cases (for example, anomalies, impacted teeth, or asymmetries), and barriers to adoption related to costs, licensing, and training (8,11,12).

Part D: AI in education and advanced modalities

AI in orthodontic education and simulation tools

AI is increasingly influencing orthodontic education by supporting how learners acquire diagnostic skills, practice clinical procedures, and engage with complex concepts. Traditional training (i.e., lectures, manual cephalometric tracing, supervised clinics, and case-based discussion) remains effective but can be time-intensive and constrained by instructor availability and case exposure. AI-based educational tools address some of these constraints by enabling automated feedback, personalised learning pathways, simulation-based practice, and longitudinal performance tracking (40,41).

In cephalometric and diagnostic training, AI systems can generate automated tracings and provide immediate feedback, allowing objective comparison with expert-defined standards. Novices who trained with AI-assisted cephalometric tools achieved higher tracing accuracy and faster learning curves than peers using traditional methods (18). AI-enabled platforms can also expose learners to larger and more diverse banks of radiographs, photographs, and digital models, broadening experience across malocclusions and clinical presentations beyond routine teaching clinic case-mix (23). DL-based classification and automated scoring further support skill development by providing rapid evaluation of diagnostic interpretations, including malocclusion classification and treatment need assessment (3,24).

Simulation-based learning has also benefited from AI integration, particularly through digital practice environments that support repeated rehearsal, objective feedback, and more standardised skills assessment. In orthodontic education, the main value of these systems lies in strengthening skills acquisition, reducing grading variability, and expanding training opportunities beyond the constraints of instructor time and clinical case availability. The immersive and interface-specific contributions of virtual reality (VR)/augmented reality (AR)-supported simulation are discussed separately in the following section.

AI-driven adaptive learning platforms further personalise education by analysing student performance, identifying patterns of misunderstanding, and tailoring content difficulty and sequencing to individual needs, approximating aspects of one-to-one tutoring (40). Large language models such as ChatGPT have also been introduced as supplementary learning tools for explaining biomechanical concepts, generating case scenarios, and supporting revision, with reported gains in perceived efficiency and confidence when use is guided and outputs are verified against authoritative sources (40,42). AI may also strengthen competency-based assessment through automated evaluation of cephalometric tracings, bracket positioning, treatment planning decisions, and written responses using NLP (11,12). Such approaches can provide educators with longitudinal data on competency attainment and potential curriculum gaps (41). These applications suggest that AI in orthodontic education is evolving along three main lines: standardization of assessment, expansion of simulation-based training, and integration of adaptive support into the broader curriculum.

Important challenges remain. Over-reliance on AI may weaken critical thinking or hands-on clinical reasoning if not carefully integrated with supervised patient care (9). Dataset representativeness is also essential to avoid embedding diagnostic or aesthetic bias into educational tools (8). Academic integrity concerns persist, as generative tools may be misused in assignments or assessments, reinforcing the need for explicit institutional policies, assessment redesign, and guidance on appropriate use (11,12).

AI combined with VR and AR

The integration of AI with VR and AR is discussed here as a cross-cutting enabling modality rather than a standalone decision domain, because immersive environments can support visualisation, simulation, guidance, and communication across orthodontic diagnosis, planning, and education. While VR/AR can enhance visualisation on their own, AI adds automation (notably segmentation and landmarking), responsiveness, and predictive modelling, with implications for training, patient communication, and complex case planning (19,41).

AI-enhanced VR/AR platforms enable high-fidelity 3D visualisation of craniofacial structures by using segmentation algorithms to extract skeletal, dental, and soft tissue anatomy from CBCT scans and digital impressions, reducing manual processing time (1). Within VR environments, these models can be explored to assess asymmetry, occlusal relationships, and skeletal discrepancy severity. Embedded AI tools can also assist with landmark identification and quantitative analysis (for example, asymmetry quantification), supporting more consistent diagnostic workflows in complex cases such as craniofacial anomalies, impacted teeth, and orthognathic planning (3).

In education and skills training, AI-supported VR/AR simulations allow repeated practice of selected orthodontic procedures while algorithms track performance metrics and deliver personalised feedback in real time (40,41). This supports scalable training with standardised assessment and may reduce the supervisory burden for high-volume skill acquisition. In orthognathic planning, AI-driven segmentation and predictive models can support simulation of osteotomies and prediction of skeletal and soft tissue responses within immersive environments, facilitating interdisciplinary communication and comparison of alternative scenarios (19). Airway implications of jaw movements can also be explored within these planning workflows, reflecting the growing relevance of airway considerations in orthognathic decision-making (3). AR additionally enables overlay of predicted outcomes onto a patient’s facial image to support understanding and treatment acceptance.

Beyond planning and education, AI-supported AR can provide real-time clinical guidance by projecting digital information onto the clinical field. Examples include AR-based bracket placement guides displayed via tablet interfaces or smart glasses to support accurate bonding, and analogous guidance for procedures such as temporary anchorage device placement by aligning overlays with relevant anatomical landmarks (1). AR can also support consultations by visualising planned tooth movements or predicted outcomes in a way that facilitates shared decision-making (19). From a patient perspective, personalised 3D visualisations may enhance communication, expectation management, and acceptance of complex interventions, with broader dental evidence suggesting that immersive visualisation can improve satisfaction and compliance (3,41).

Limitations remain substantial. High hardware and software costs may hinder adoption, particularly in smaller clinics and training centres (11,12). Some users experience discomfort or motion sickness with prolonged VR exposure. Technical performance is also dependent on segmentation accuracy, meaning that errors in AI-generated anatomy can propagate into planning and simulation outputs, underscoring the need for validation, user training, and continued integration with conventional diagnostic and planning methods (8).

Discussion

AI has moved from experimental proof-of-concept to a mature, clinically relevant technology in orthodontics. Across the fourteen domains synthesized in this review, AI systems based on ML, DL and CNNs now support cephalometric analysis, malocclusion classification, aesthetic assessment, growth prediction, treatment planning, teleorthodontics, CAD/CAM workflows and education. These findings are consistent with recent scoping and narrative reviews that describe AI as an increasingly central component of orthodontic diagnostics, treatment planning, predictive modeling, and workflow optimization, rather than an isolated research niche (5,43-47).

From a clinical perspective, the strongest and most consistent evidence relates to gains in diagnostic accuracy and efficiency. Multi-stage CNNs and ensemble models now achieve MREs commonly around 1.1–1.8 mm for lateral cephalometric landmark detection, with SDRs often reported at 70–87% within 2 mm and above 90% at 4 mm, while automated malocclusion classifiers have reached accuracies in the 90–97% range in selected datasets. These results are often comparable to expert performance for specific tasks, although generalisability across datasets, devices, and populations remains more limited than these headline figures may suggest (21,44,48,49). These improvements are reflected in commercially available platforms such as WebCeph, CephX, Dolphin and AudaxCeph, which provide automated tracings, skeletal classifications, growth staging and aesthetic indices. The breadth of these applications is summarized in Table 2, which collates key AI application areas, representative commercial systems, typical performance ranges and supporting literature, illustrating how single-task models are increasingly embedded within integrated clinical software environments.

Evidence from teleorthodontics and remote monitoring adds an important behavioural and outcome-oriented dimension. AI-driven monitoring platforms improve plaque control, gingival health, and reduce the incidence of white spot lesions when compared with conventional in-office follow-up, while simultaneously enhancing compliance and shortening treatment time in selected cohorts (55-58). Studies in aligner therapy report high agreement between AI-assisted remote assessments and chairside evaluations, as well as high levels of patient satisfaction with digitally supervised care (7,59,60). These findings align closely with the observations in Part C of this review, where AI-supported teleorthodontic systems were shown to triage patients effectively, reduce unnecessary visits and support hybrid models of care.

Educational and training applications form another rapidly expanding strand of work. AI-assisted cephalometric training and diagnostic simulators have been shown to improve landmarking accuracy and accelerate learning in preclinical orthodontic education (18,40). Large language models and AI tutors are increasingly used to supplement teaching, create formative questions and support reflective learning, with early studies reporting generally positive attitudes among students and postgraduate trainees (61-64). At the same time, surveys from multiple regions indicate that acceptance remains contingent on clear guidance about appropriate use, transparency regarding limitations, and explicit integration of AI literacy into dental curricula (65-69). These educational findings support the position, reinforced throughout this review, that clinician-in-the-loop models are central to safe and effective AI adoption.

Viewed in relation to adjacent dental specialties, orthodontic educational AI appears to be developing along the same broader trajectory, but with a literature that is still less explicitly structured. Recent reviews in endodontic education and dentomaxillofacial radiology education describe AI not only as a diagnostic aid but as a framework for assessment standardization, simulation-based training, adaptive learning, performance tracking, and curriculum transformation. In endodontics, proposed applications extend from radiographic interpretation and treatment planning to case-difficulty assessment, advanced simulation, real-time clinical guidance, personalized remediation, and calibration/standardization. In dentomaxillofacial radiology, emphasis has similarly been placed on AI-guided simulation, automated assessment, OSCE and mini-CEX support, learning analytics, and scalable feedback systems. Orthodontics already shows clear movement in the same direction through AI-assisted cephalometric training, simulation-supported skill acquisition, adaptive learning platforms, and competency-oriented assessment. However, compared with these adjacent fields, the orthodontic literature has so far placed more emphasis on specific use cases than on fully articulated curriculum models. Future work should therefore move beyond proof-of-concept adoption toward multicenter validation of educational AI tools, clearer explainability standards for systems used in teaching and assessment, stronger faculty development and governance frameworks, and thoughtful integration with digital learning environments and, where appropriate, de-identified clinical data infrastructures that can support authentic case-based training and longitudinal competency evaluation (70,71).

The Scopus-based synthesis also sharpens understanding of structural limitations that run across the fourteen domains. A recurrent problem is the dependence on relatively small, single-centre and often demographically narrow datasets, which constrain external validity and heighten the risk of biased predictions in under-represented groups (23,45,72). Recent efforts to develop larger, multi-centre and multimodal benchmarks such as MMDental, WebCeph2k and the FDDI dual-modality dataset are important steps toward more generalisable models, but these resources remain the exception rather than the norm (73-75). This review’s findings on cephalometry, growth prediction and malocclusion classification highlight precisely this tension: high accuracy within constrained datasets, but limited evidence that performance is maintained across devices, ethnicities and clinical settings. Although multicenter validation remains uncommon, recent work on AI-driven 3D craniofacial superimposition shows that external validation across institutions is feasible and clinically informative (36).

A further concern is that bias in orthodontic AI may arise not only from dataset imbalance, but also from the reference standards used to train and validate models. In objective tasks such as landmark localisation, this issue is partly bounded by measurable spatial error. In more subjective or partially standardised domains, however, including CVM staging, aesthetic assessment, treatment-need scoring, extraction decisions, and treatment planning, the labels themselves may reflect clinician preference, institutional culture, or historically uneven treatment norms rather than a universally stable ground truth. In CVM assessment, for example, disagreement in stage assignment and uneven distribution of cases across maturation classes may cause models to learn annotator preference patterns or overfit majority stages rather than capture a robust developmental signal. Under such conditions, AI may not simply inherit bias, but also amplify it by reproducing the same patterns with greater consistency and scale. This means that strong internal performance does not necessarily indicate fairness or clinical neutrality. Mitigating this risk requires more than larger datasets. It also requires clearer reporting of demographic and device composition, attention to class imbalance, subgroup-level performance evaluation, multi-annotator or consensus-based labelling where feasible, external validation across centres, and continued clinician-in-the-loop review during deployment (14,19). In practical terms, the goal should not be to train AI to imitate existing judgment uncritically, but to develop systems whose outputs remain transparent, contestable, and accountable within a human-supervised clinical workflow.

A further cross-cutting limitation is variability in imaging acquisition and preprocessing. Although not all fourteen domains reviewed here are equally image-dependent, many of the core applications, including cephalometric landmarking, skeletal maturation assessment, malocclusion classification, airway analysis, facial assessment, remote monitoring, and several digital workflow tasks, are directly influenced by the technical characteristics of the source data. Differences in imaging modality, device manufacturer, detector resolution, field of view, voxel size, patient or head positioning, soft-tissue posture, motion artefacts, exposure parameters, and preprocessing or annotation pipelines can alter landmark visibility, contrast, segmentation boundaries, and geometric consistency. In practical terms, this means that strong performance reported in a single-centre or tightly standardised dataset may not translate unchanged to other clinics, scanners, or imaging workflows. Future orthodontic AI studies should therefore prioritise external validation across multiple devices and centres, together with clearer reporting and greater standardisation of acquisition protocols, to improve robustness and clinical transferability.

Ethical and governance questions are equally prominent. Data protection, image privacy, secondary use of radiographs and photographs, and consent for AI training are recurrent concerns in recent dental and medical ethics reviews (76-78). At the same time, work from craniofacial and orthodontic researchers has begun to articulate the specific problem of algorithmic bias, including the potential for AI models to amplify existing racial or socioeconomic inequities in access to treatment or eligibility scoring (79,80). Broader health-informatics literature points to individual-level fairness frameworks and domain-generalized training strategies as promising, though still incomplete, approaches to mitigating these risks (81,82). The converging message is that orthodontic AI research needs not only larger datasets, but also better characterised and more transparently reported ones, with explicit attention to demographic composition and device diversity.

A further challenge lies in interpretability and medico-legal responsibility. Many of the most accurate models, especially deep neural networks for imaging and integrated decision support, operate as opaque systems that output classifications or recommendations without human-readable reasoning. This “black-box” nature complicates error analysis, undermines trust and raises questions about liability when AI-assisted decisions contribute to adverse outcomes (83,84).

A related ethical risk is overreliance on AI outputs once systems become embedded in routine workflows and their errors become less frequent, less obvious, or more difficult to detect (85,86). Under these conditions, clinicians may gradually shift from critically evaluating AI recommendations to passively accepting them, especially when outputs are presented in polished, quantitative, or seemingly authoritative formats. This creates a form of automation bias in which the efficiency gains of AI are accompanied by reduced vigilance and weakened independent judgment. In orthodontics, the danger is not only technical but professional: if treatment recommendations, aesthetic assessments, airway analyses, or surgical predictions are accepted with insufficient scrutiny, errors may become harder to contest and responsibility more difficult to assign. Mitigating this risk requires more than clinician availability alone. It requires training that explicitly addresses AI limitations, workflow designs that encourage verification rather than deference, and a clear understanding that AI outputs should remain advisory and contestable within a human-led decision process.

Although the term AI is retained throughout this review because it remains the dominant term in the orthodontic literature, the clinical reality is often better described as augmented intelligence. In this framing, algorithmic systems are used to support perception, standardisation, pattern recognition, and decision support, while responsibility for interpretation, contextual judgement, patient communication, and final treatment decisions remains with the orthodontist. This distinction is especially important in orthodontics, where apparently high-performing models still operate within bounded tasks and may not fully account for patient-specific biological, aesthetic, behavioural, or ethical considerations (87). Evidence from clinician-in-the-loop systems in orthodontics and related fields suggests that close human oversight not only improves diagnostic accuracy but also reduces ethical and legal risk by ensuring that AI outputs are treated as recommendations rather than prescriptions (4,5,88,89). The domains examined in this review illustrate how indispensable such oversight remains when decisions carry significant functional, aesthetic or surgical consequences. This cautious, augmented-intelligence framing is also consistent with recent professional-attitude data. In a 2026 survey of orthodontists and oral radiologists, most participants welcomed fully automated cephalometric landmark software, yet orthodontists were more likely to support using AI diagnosis together with clinician diagnosis rather than as an autonomous replacement (90).

The included literature remains promising but methodologically uneven. Much of the evidence base is dominated by retrospective, single-centre studies, often with relatively small or demographically narrow datasets, limited external validation, heterogeneous outcome metrics, and inconsistent reporting of calibration, subgroup performance, or real-world clinical impact. In several domains, especially treatment planning, aesthetic assessment, and skeletal maturation, the reference standards themselves may depend on clinician judgement, which introduces further uncertainty regarding reproducibility and bias propagation. These features mean that strong reported performance should be interpreted cautiously, particularly when claims of clinician-level equivalence are based on internally validated or tightly standardised datasets. The present review also has strengths and limitations. Its main strength lies in providing a broad clinician-facing synthesis across fourteen application domains organized within four higher-order categories, including educational and advanced-modality applications that are often not integrated within a single orthodontic review. However, the review is limited by its integrative narrative design, Scopus-only search strategy, English-language restriction, and the heterogeneity of the included studies, which precluded formal quality appraisal and meta-analysis. Accordingly, the review should be interpreted as a structured synthesis of the field’s current scope and direction rather than as a quantitative estimate of effect or certainty.

Conclusions

This review maps AI across fourteen domains of orthodontic practice, showing that AI is now capable of reliably supporting diagnosis, treatment planning, monitoring, education, and digital workflows. What emerges across these domains is not just a catalogue of tools, but a system-level pattern: AI performs best when it operates within clearly defined, supervised tasks, on well-curated data, and inside workflows that keep the clinician firmly “in the loop”. The key take-home message is that the question is no longer whether AI works in orthodontics, but under what conditions it should be trusted. By foregrounding dataset quality, interpretability, and integration with training and governance, this review shifts the focus from technical performance alone to practical readiness while identifying the concrete requirements that must be in place for AI to deliver safe, equitable, and genuinely useful orthodontic care.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0020/rc

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0020/prf

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-1-0020/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Dipalma G, Inchingolo AD, Inchingolo AM, et al. Artificial Intelligence and Its Clinical Applications in Orthodontics: A Systematic Review. Diagnostics (Basel) 2023;13:3677. [Crossref] [PubMed]
2. Kunz F, Stellzig-Eisenhauer A, Boldt J. Applications of artificial intelligence in orthodontics - an overview and perspective based on the current state of the art. Appl Sci 2023;13:3850.
3. Liu J, Zhang C, Shan Z. Application of Artificial Intelligence in Orthodontics: Current State and Future Perspectives. Healthcare (Basel) 2023;11:2760. [Crossref] [PubMed]
4. Miranda F, Barone S, Gillot M, et al. Artificial intelligence applications in orthodontics. J Calif Dent Assoc 2023;51:2195585.
5. Mupparapu M, Tussie C, Nath S, et al. Artificial Intelligence and its Applications in Orthodontics and Dentofacial Orthopedics. Dent Clin North Am 2026;70:471-86. [Crossref] [PubMed]
6. Wong KF, Lam XY, Jiang Y, et al. Artificial intelligence in orthodontics and orthognathic surgery: a bibliometric analysis of the 100 most-cited articles. Head Face Med 2023;19:38. [Crossref] [PubMed]
7. Polizzi A, Boato M, Serra S, et al. Applications of artificial intelligence in orthodontics: a bibliometric and visual analysis. Clin Oral Investig 2025;29:65. [Crossref] [PubMed]
8. Nordblom NF, Büttner M, Schwendicke F. Artificial Intelligence in Orthodontics: Critical Review. J Dent Res 2024;103:577-84. [Crossref] [PubMed]
9. Haliloğlu Özkan T, Acar A, Özkan E, et al. Are artificial intelligence-based chatbots reliable sources for patients regarding orthodontics? APOS Trends Orthod 2025;15:141-6.
10. Al-Taie NN, Al-Talabani SZ. Perception and application of dental artificial intelligence in orthodontic clinical practice: a cross-sectional survey of orthodontists in Iraq. Al-Rafidain J Med Sci 2025;9:39-47.
11. Hellyer P. Artificial intelligence in orthodontics. Br Dent J 2024;237:48. [Crossref] [PubMed]
12. Sharp C. Orthodontics in the city: artificial intelligence in orthodontics. BDJ Team 2025;12:168-71.
13. Heribanová B, Janáková K, Tomášik J, et al. Current Applications and Future Perspectives of Artificial Intelligence in Face-Driven Orthodontics: A Scoping Review. Biomimetics (Basel) 2026;11:146. [Crossref] [PubMed]
14. Leonardi R, Vaiid N. Artificial Intelligence in Orthodontics: Concerns, Conjectures, and Ethical Dilemmas. Int Dent J 2025;75:20-2. [Crossref] [PubMed]
15. Subramanian AK, Chen Y, Almalki A, et al. Cephalometric analysis in orthodontics using artificial intelligence—a comprehensive review. Biomed Res Int 2022;2022:1880113. [Crossref] [PubMed]
16. Kunz F, Stellzig-Eisenhauer A, Widmaier LM, et al. Assessment of the quality of different commercial providers using artificial intelligence for automated cephalometric analysis compared to human orthodontic experts. J Orofac Orthop 2025;86:145-60. [Crossref] [PubMed]
17. Evangelista K, de Freitas Silva BS, Yamamoto-Silva FP, et al. Accuracy of artificial intelligence for tooth extraction decision-making in orthodontics: a systematic review and meta-analysis. Clin Oral Investig 2022;26:6893-905. [Crossref] [PubMed]
18. Lin J, Liao Z, Dai J, et al. Digital and artificial intelligence-assisted cephalometric training effectively enhanced students’ landmarking accuracy in preclinical orthodontic education. BMC Oral Health 2025;25:623. [Crossref] [PubMed]
19. Gracea RS, Winderickx N, Vanheers M, et al. Artificial intelligence for orthodontic diagnosis and treatment planning: A scoping review. J Dent 2025;152:105442. [Crossref] [PubMed]
20. Ardila CM, Pineda-Vélez E, Vivares-Builes AM. Artificial Intelligence Models for the Detection and Quantification of Orthodontically Induced Root Resorption Using Cone-Beam Computed Tomography: A Systematic Review and Meta-Analysis. Dent J (Basel) 2026;14:79. [Crossref] [PubMed]
21. Lyros I, Vastardis H, Tsolakis IA, et al. Growth Prediction in Orthodontics: ASystematic Review of Past Methods up to Artificial Intelligence. Children (Basel) 2025;12:1023. [Crossref] [PubMed]
22. Yu HJ, Cho SR, Kim MJ, et al. Automated Skeletal Classification with Lateral Cephalometry Based on Artificial Intelligence. J Dent Res 2020;99:249-56. [Crossref] [PubMed]
23. Bichu YM, Hansa I, Bichu AY, et al. Applications of artificial intelligence and machine learning in orthodontics: a scoping review. Prog Orthod 2021;22:18. [Crossref] [PubMed]
24. Ryu J, Lee YS, Mo SP, et al. Application of deep learning artificial intelligence technique to the classification of clinical orthodontic photos. BMC Oral Health 2022;22:454. [Crossref] [PubMed]
25. Li S, Guo Z, Lin J, et al. Artificial Intelligence for Classifying and Archiving Orthodontic Images. Biomed Res Int 2022;2022:1473977. [Crossref] [PubMed]
26. Kavousinejad S, Ghazanfari S, Manshadi SAH, et al. Artificial Intelligence–Based Multi-Stage System for Automated Angle’s Classification of Malocclusion from Intraoral Images in Orthodontics. J Imaging Inform Med 2026; Epub ahead of print. [Crossref]
27. Stetzel L, Foucher F, Jang SJ, et al. Artificial Intelligence for Predicting the Aesthetic Component of the Index of Orthodontic Treatment Need. Bioengineering (Basel) 2024;11:861. [Crossref] [PubMed]
28. Feng H, Lv Y, Yang X, et al. Bimodal visual sensors based on mechanoluminescence and biosensing for artificial intelligence-assisted orthodontics. Adv Funct Mater 2025;35:2416437.
29. Caleme ED, Cevidanes L, Mattos C, et al. Aligning MRI and CBCT for Advanced TMJ Diagnostics: Case Series Using AI-Powered Registration in Dentistry and Orthodontics. Semin Orthod 2025; Epub ahead of print. [Crossref]
30. Emami H, Dong M, Nejad-Davarani SP, et al. Generating synthetic CTs from magnetic resonance images using generative adversarial networks. Med Phys 2018; Epub ahead of print. [Crossref]
31. Ueda A, Tussie C, Kim S, et al. Classification of Maxillofacial Morphology by Artificial Intelligence Using Cephalometric Analysis Measurements. Diagnostics (Basel) 2023;13:2134. [Crossref] [PubMed]
32. Liu J, Chen Y, Li S, et al. Machine learning in orthodontics: Challenges and perspectives. Adv Clin Exp Med 2021;30:1065-74. [Crossref] [PubMed]
33. Subramanya S, Mohan A, Premalatha A, et al. Artificial intelligence in orthodontics: modeling decision support systems for treatment planning. IAES Int J Artif Intell 2026;15:97-105.
34. Thurzo A, Kurilová V, Varga I. Artificial Intelligence in Orthodontic Smart Application for Treatment Coaching and Its Impact on Clinical Performance of Patients Monitored with AI-TeleHealth System. Healthcare (Basel) 2021;9:1695. [Crossref] [PubMed]
35. Surovková J, Haluzová S, Strunga M, et al. The new role of the dental assistant and nurse in the age of advanced artificial intelligence in telehealth orthodontic care with dental monitoring: preliminary report. Appl Sci 2023;13:5212.
36. Huang Y, Anchling L, Bates WR, et al. Automated artificial intelligence-driven 3-dimensional craniofacial superimposition for clinically useful treatment outcomes in growing patients: A multicenter study. Am J Orthod Dentofacial Orthop 2026;169:753-65. [Crossref] [PubMed]
37. McCray JF, Dabney W, Handlin D, et al. Accuracy of dentalmonitoring’s artificial intelligence in detecting aligner tracking issues: a retrospective multi-centric study. BMC Oral Health 2026;26:239. [Crossref] [PubMed]
38. Strunga M, Urban R, Surovková J, et al. Artificial intelligence systems assisting in the assessment of the course and retention of orthodontic treatment. Healthcare 2023;11:683.
39. McCray JF, Smith L, Handlin D, et al. Accuracy of Dental Monitoring’s artificial intelligence in detecting common orthodontic braces treatment related emergencies. Sci Rep 2026;16:6627. [Crossref] [PubMed]
40. Ardila CM, Pineda-Vélez E, Vivares Builes AM. Integrating artificial intelligence into orthodontic education: a systematic review and meta-analysis of clinical teaching application. J Clin Med 2025;14:5487. [Crossref] [PubMed]
41. Baxmann M, Kárpáti K, Baráth Z. The potentials and challenges of integrating generative artificial intelligence in dental and orthodontic education: a systematic review. BMC Oral Health 2025;25:905. [Crossref] [PubMed]
42. Yousuf AM, Ikram F, Gulzar M, et al. Performance assessment of artificial intelligence chatbots (ChatGPT-4 and Copilot) for sharing insights on 3D-printed orthodontic appliances: a cross-sectional study. Int Orthod 2025;23:100992. [Crossref] [PubMed]
43. Akdeniz S, Tosun ME. A review of the use of artificial intelligence in orthodontics. J Exp Clin Med 2021;38:157-62.
44. Kazimierczak N, Kazimierczak W, Serafin Z, et al. Artificial intelligence in orthodontics: revolutionizing diagnostics and treatment planning—a comprehensive review. J Clin Med 2024;13:344.
45. Fawaz P, Sayegh PE, Vannet BV. What is the current state of artificial intelligence applications in dentistry and orthodontics? J Stomatol Oral Maxillofac Surg 2023;124:101524. [Crossref] [PubMed]
46. Azizi S, Hatampoor S, Tahamtan S. Applications of artificial intelligence in diagnosis and treatment planning of orthodontics: a narrative review. Saudi Dent J 2025;37:70. [Crossref] [PubMed]
47. Olawade DB, Leena N, Egbon E, et al. Artificial intelligence-driven advancements in orthodontics for precision and patient outcomes. Dent J 2025;13:198.
48. Kim J, Kim I, Kim YJ, et al. Accuracy of automated identification of lateral cephalometric landmarks using cascade convolutional neural networks on lateral cephalograms from nationwide multi-centres. Orthod Craniofac Res 2021;24:59-67. [Crossref] [PubMed]
49. Sahlsten J, Järnstedt J, Jaskari J, et al. Deep learning for 3D cephalometric landmarking with heterogeneous multi-center CBCT dataset. PLoS One 2024;19:e0305947. [Crossref] [PubMed]
50. Foroozandeh M, Salemi F, Shokri A, et al. Comparative accuracy of artificial intelligence-based AudaxCeph software, Dolphin software, and the manual technique for orthodontic landmark identification and tracing of lateral cephalograms. Imaging Sci Dent 2025;55:11-21. [Crossref] [PubMed]
51. Navone C, Doldo T. Accuracy of artificial intelligence in the prediction of cervical vertebrae maturation stages in orthodontics: a systematic review. J Osseointegration 2025;17:20-5.
52. Khanagar SB, Al-Ehaideb A, Vishwanathaiah S, et al. Scope and performance of artificial intelligence technology in orthodontic diagnosis, treatment planning, and clinical decision-making—a systematic review. J Dent Sci 2021;16:482-92. [Crossref] [PubMed]
53. Tahir K, Abul Barakaat A, Fida M, et al. In the contemporary era of artificial intelligence, the trajectory of orthodontics: past and future perspectives—a narrative review. J Calif Dent Assoc 2024;52:2400420.
54. Patil S, Eisenhuth G, El-Bialy T, et al. Reliability and performance of four large language models in orthodontic knowledge assessment. J Dent Educ 2026;90:462-5. [Crossref] [PubMed]
55. Snider V, Homsi K, Kusnoto B, et al. Effectiveness of artificial intelligence-driven remote monitoring technology in improving oral hygiene during orthodontic treatment. Orthod Craniofac Res 2023;26:102-10.
56. Sangalli L, Savoldi F, Dalessandri D, et al. Effects of remote digital monitoring on oral hygiene of orthodontic patients: a prospective study. BMC Oral Health 2021;21:435. [Crossref] [PubMed]
57. Sharma A, Sharma T, Tomer G, et al. Assessment of artificial intelligence-based monitoring method in assessing oral hygiene during orthodontic treatment. J Orthod Sci 2025;14:17. [Crossref] [PubMed]
58. Asarsa SK. Effects of digital health interventions on oral hygiene and treatment compliance in orthodontic patients: a systematic review. Int J Orthod Rehabil 2025;16:48-62.
59. Martínez Gil-Ortega A, Cintora-López P, Pérez Rodríguez LM, et al. Pilot case control evaluation of artificial intelligence-assisted orthodontic monitoring and pediatric patient perception. Sci Rep 2025;15:40515. [Crossref] [PubMed]
60. Varela JP, Vila ML, Hirschhaut M, et al. Remote aligner supervision and STL generation with dental monitoring®: clinical integration and case report. Semin Orthod 2025; Epub ahead of print. [Crossref]
61. Patil SR, Karobari MI. Exploring artificial intelligence for enhanced endodontic practice: applications, challenges, and future directions. Adv Public Health 2024;2024:8075515.
62. Martina S, Cannatà D, Paduano T, et al. Reliability of Large Language Model-Based Chatbots Versus Clinicians as Sources of Information on Orthodontics: A Comparative Analysis. Dent J (Basel) 2025;13:343. [Crossref] [PubMed]
63. Rui YU, Minmin SI, Zhiyuan FE. Prospects of ChatGPT in orthodontic education and clinical practice. J Prev Treat Stomatol Dis 2024;32:478-84.
64. Hanenkrath J, Park JH, Bay C. Training, use, and modifications related to artificial intelligence in postgraduate orthodontic programs in North America. Am J Orthod Dentofacial Orthop 2025;167:89-94.
65. Yüzbaşıoğlu E. Attitudes and perceptions of dental students towards artificial intelligence. J Dent Educ 2021;85:60-8. [Crossref] [PubMed]
66. Jeong H, Han SS, Jung HI, et al. Perceptions and attitudes of dental students and dentists in South Korea toward artificial intelligence: a subgroup analysis based on professional seniority. BMC Med Educ 2024;24:430. [Crossref] [PubMed]
67. Ghasemian A, Salehi M, Ghavami V, et al. Exploring dental students’ attitudes and perceptions toward artificial intelligence in dentistry in Iran. BMC Med Educ 2025;25:725. [Crossref] [PubMed]
68. Alam ST, Bhanotia D, Trehan M, et al. Assessing the artificial intelligence acumen: a study on the knowledge, attitude and behaviour of dental students towards artificial intelligence. J Contemp Orthod 2024;8:290-3.
69. Elemam RF, Al-Bayati SA. Awareness of GMU dental students regarding the use of artificial intelligence in dentistry. Ain Shams Dent J 2024;36:27-35.
70. Negrete D, Lopes SL, Barretto MD, et al. Artificial intelligence and dentomaxillofacial radiology education: innovations and perspectives. Dent J 2025;13:245.
71. Aminoshariae A, Nosrat A, Nagendrababu V, et al. Artificial intelligence in endodontic education. J Endod 2024;50:562-78. [Crossref] [PubMed]
72. Majstorović NV, Dimitrijević S. Artificial intelligence in orthodontics diagnosis and treatment. In: Karabegović I. editor. New Technologies, Development and Applications VII. Cham: Springer Nature Switzerland; 2025:57-66.
73. Wang C, Zhang Y, Wu C, et al. MMDental—a multimodal dataset of tooth CBCT images with expert medical records. Sci Data 2025;12:1172. [Crossref] [PubMed]
74. Wang X, Rigall E, An X, et al. A new benchmark and low computational cost localization method for cephalometric analysis. IEEE Trans Circuits Syst Video Technol 2024;34:8939-52.
75. Yu Z, Li M, Yang J, et al. A benchmark dual-modality dental imaging dataset and a novel cognitively inspired pipeline for high-resolution dental point cloud synthesis. Cognit Comput 2023;15:1922-33.
76. Mörch CM, Atsu S, Cai W, et al. Artificial intelligence and ethics in dentistry: a scoping review. J Dent Res 2021;100:1452-60. [Crossref] [PubMed]
77. Tuygunov N, Samaranayake L, Khurshid Z, et al. The transformative role of artificial intelligence in dentistry: a comprehensive overview part 2: the promise and perils, and the international dental federation communique. Int Dent J 2025;75:397-404. [Crossref] [PubMed]
78. Rokhshad R, Ducret M, Seifi S, et al. Artificial intelligence: limitations, safety, and regulatory considerations in dentistry. In: Rakhshan V, Tebyani M. editors. Artificial Intelligence for Oral Health Care. Cham: Springer Nature Switzerland; 2025:165-71.
79. Allareddy V, Oubaidin M, Rampa S, et al. Call for algorithmic fairness to mitigate amplification of racial biases in artificial intelligence models used in orthodontics and craniofacial health. Orthod Craniofac Res 2023;26:124-30. [Crossref] [PubMed]
80. Jones RK. Algorithmic bias and fairness in biomedical and health research. In: Gupta S. editor. Ensuring Secure and Ethical STM Research in the AI Era. Hershey: IGI Global; 2025:287-324.
81. Anderson JW, Visweswaran S. Algorithmic individual fairness and healthcare: a scoping review. JAMIA Open 2025;8:ooae149. [Crossref] [PubMed]
82. Gong H, Lu Y, Wan X, et al. Domain generalized medical landmark detection via robust boundary-aware pre-training. Proc AAAI Conf Artif Intell 2025;39:3140-8.
83. Trakulmututa S, Win KT. Machine learning in orthodontics: Transforming Invisalign treatment planning through precision, interpretability, and ethical practices. Int J Med Inform 2026;206:106139. [Crossref] [PubMed]
84. Kumar SS, Michelogiannakis D, Xu X, et al. Patient trust in artificial intelligence for orthodontic advice: a systematic review. J Am Dent Assoc 2025;156:931-944.e6.
85. Khan FA. AI in clinical diagnostics: Is overreliance eroding clinical expertise? PLOS Digit Health 2025;4:e0000959. [Crossref] [PubMed]
86. Patil SV, Myers CG, Lu-Myers Y. Calibrating AI reliance—a physician’s superhuman dilemma. JAMA Health Forum 2025;6:e250106. [Crossref] [PubMed]
87. Jagtap R. Dentistry - Overview of Artificial and Augmented Intelligence Uses in Dentistry. American Dental Association, SCDI White Paper No. 1106. 2023. Available online: https://www.ada.org/-/media/project/ada-organization/ada/ada-org/files/resources/practice/dental-standards/ada_1106_2022.pdf
88. Kapoor S, Shyagali TR, Kuraria A, et al. An artificial neural network approach for rational decision-making in borderline orthodontic cases: A preliminary analytical observational in silico study. J Orthod 2023;50:439-48. [Crossref] [PubMed]
89. Mohamed M, Ferguson DJ, Venugopal A, et al. An artificial intelligence-based referral application to optimize orthodontic referrals in a public oral healthcare system. Semin Orthod 2021;27:157-63.
90. Dal’Ongaro Savegnago G, Boessio Vizzotto M, Delamare EL, et al. Oral radiologists and orthodontists’ attitude regarding the use of artificial intelligence for cephalometric analysis. J Orofac Orthop 2026;11:1-8.