To address these gaps, we propose the following framework for the training of AI/ML models in aesthetic evaluation that consists of 6 pillars: data collection and curation, model training methodologies, fairness metrics and evaluation, explainability and transparency, stakeholder engagement, and governance and monitoring.

This framework addresses AI/ML systems used across several distinct aesthetic evaluation tasks, which differ substantially in their fairness stakes and acceptable error thresholds. Attractiveness scoring assigns a rating or ranking to a face and is used primarily for research benchmarking. Preoperative planning uses AI to assess anatomic features and inform surgical approach, where systematic undervaluation of ethnic features could directly influence clinical decision-making. Outcome simulation generates predicted postoperative appearance, where bias has the additional potential to alter rendered ethnicity rather than merely undervalue it, a qualitatively distinct harm addressed further in the framework. Patient counseling involves AI-assisted communication of options, where biased framing may subtly steer patient choices. Fairness requirements and error thresholds should be calibrated to the stakes of the specific use case; the more downstream the application and the more directly it affects patient choice and surgical planning, the more stringent the requirements.

Throughout this framework, recommendations are categorized by their evidence base (Table 2). Practices marked as “established” draw on studies conducted in aesthetic facial evaluation contexts. Practices marked as “adapted” are supported by evidence from adjacent domains, primarily facial recognition, general medical AI, or computer vision, and have been translated to aesthetic evaluation by analogy; these require validation in aesthetic-specific contexts before adoption as standard practice. Practices marked as “proposed” represent conceptual recommendations without empirical validation in any closely related domain and should be treated as research directions.

As a pragmatic baseline informed by existing benchmark datasets, training data should include a balanced representation of at least 7 racial and ethnic categories [31] (White, Black, East Asian, Southeast Asian, Middle Eastern, Latino, and Indian), with additional stratification by gender, age, nationality, and socioeconomic background. This should include multiregional data collection with noted region-specific aesthetic preferences [4,29]. This scheme is explicitly a minimum starting point, not a definitive classification; implementations should adopt more granular schemes as data availability permits. Individuals of mixed ancestry, a rapidly growing population, should be accommodated through multilabel or probabilistic ancestry representation rather than forced assignment to a single category. As the field matures, continuous morphometric representations of facial ancestry, such as principal components of facial geometry derived from diverse reference populations, offer a more biologically-grounded alternative to discrete ethnic labels and should be pursued to reduce the risk of essentialization.

While standardized photographs in a constrained environment are necessary to reduce training inaccuracies from facial expression confounds [30], they introduce a domain-shift risk at deployment: real-world clinical images routinely involve variation in lighting, angle, makeup, and expression that differs systematically from controlled training conditions. To address this, we recommend a hybrid protocol: initial pretraining on standardized images to establish controlled baseline representations, followed by fine-tuning on a curated set of clinically realistic images incorporating documented augmentation strategies—including geometric transformations such as rotation and translation to simulate angle variation, and color space augmentations to simulate lighting variation—to reduce the gap between training and deployment distributions [48]. Explicit evaluation of domain generalization, measuring performance and fairness metric stability across both standardized and nonstandardized image sets, should be required before clinical release.

Even with standardized imaging protocols, acquiring sufficient photographs across all demographic categories remains challenging. Synthetic data generation can bridge these gaps while also protecting privacy, with pretraining gap analysis using established benchmark datasets such as FairFace [31] or Diversity in Faces [32] used to identify demographic imbalances or biases; these have been established in facial attribute classification, though their applicability to aesthetic evaluation requires confirmation. Generative models are themselves trained on real-world data containing existing biases and are susceptible to mode collapse, where the model produces a narrow range of outputs disproportionately representing dominant features, and to hallucination of exaggerated demographic characteristics when conditioned on ethnic labels. These failure modes risk reintroducing the essentialized representations that the framework is designed to prevent. Accordingly, synthetic images must not enter the training set without explicit quality-control steps: adversarial fairness auditing of synthetic outputs to detect feature exaggeration; statistical comparison of synthetic image feature distributions against reference population benchmarks; and human review by the diverse rater panels proposed elsewhere in this framework. Only images passing all 3 criteria should be incorporated.

Furthermore, photograph raters should be recruited from diverse cultural backgrounds to avoid annotation bias [29]. Annotation bias occurs when raters systematically apply their own cultural aesthetic standards to evaluate faces from different backgrounds, but the patterns of bias are complex and do not reduce to simple in-group favoritism. Research demonstrates that rater ethnicity influences which facial features are emphasized—for instance, Chinese observers associate lighter skin tones with attractiveness in own-ethnicity faces, whereas European observers prefer warmer tones in Chinese faces [8], and Japanese raters emphasize raised eyebrows in attractive male faces and smaller mouths in attractive female faces more than American raters [9]. However, cross-cultural studies comparing attractiveness ratings across European, East Asian, and African faces found no strong own-race preference in overall attractiveness judgments [49], indicating that annotation bias operates through subtle feature weighting rather than categorical group favoritism. For example, raters trained predominantly in Western aesthetic ideals might systematically underweight features like broader nasal bridges or fuller lips that are attractive within specific cultural contexts, not because of explicit racial preference but because their cultural training emphasizes different facial proportions. Recruiting ethnically diverse rater panels is therefore essential to ensure balanced representation of aesthetic preferences rather than assuming any single demographic composition will eliminate bias.

MTL is central to bias mitigation in AI/ML model training. Models should be simultaneously trained on age, gender, ethnicity, facial expression, and attractiveness ratings to develop a comprehensive and nuanced assessment of faces [36] where feasible. Rater cultural background, region of upbringing, and socioeconomic status should be recorded as covariates to enable analysis of how these dimensions influence annotation.

Several fairness-aware techniques should be implemented during model training, adapted from the facial recognition literature but not yet validated in aesthetic evaluation. Adversarial learning methods [39] should be applied during the training phase to mitigate bias acquired during data collection. Posttraining, centroid fairness loss [40] enables bias measurement and performance alignment across demographic groups without requiring complex model retraining, a significant practical advantage. Skewness-aware reinforcement learning [41] should be used to recognize and adjust imbalances in data distribution or model performance across demographics. Finally, debiasing variational autoencoder [42] can adjust sampling probabilities for underrepresented categories, balancing the effective training data to enable more equitable performance across patient populations.

Beyond these fairness-aware training techniques, AI/ML models can be designed to adapt to regional aesthetic standards for facial evaluation. A meta-learning approach, supported by preliminary evidence from the beauty prediction literature, enables models to “learn how to learn”: models trained on learning tasks from a range of cultures develop the ability to adapt to other cultural preferences more readily [43]. This methodology can account for the subjective nature of beauty perception across cultures and allow customization based on patient population.

While meta-learning enables adaptation to regional standards, the hierarchical structure of aesthetic preferences, with both universal and ethnicity-specific components, suggests opportunities for more sophisticated architectural approaches. Hierarchical Bayesian models could naturally encode this structure through multilevel parameter sharing, where population-level priors capture universal features while group-specific parameters account for cultural variation. Alternatively, the partial invariance framework extends invariant risk minimization by learning features that are invariant within partitions of training environments rather than globally invariant across all environments [50]. Such approaches may encode averageness-related objectives within cultural partitions rather than across the full training distribution, preserving within-group distinctiveness while limiting cross-group homogenization. Universal structural features would be encoded at the population level only when evidence supports genuine cross-cultural validity, while regionally variable features remain governed by culture-specific parameters. Efficient multigroup equivariant techniques that address intersectional fairness across combinations of protected attributes, such as ethnicity and gender, may offer additional methodological directions [51]. However, these techniques have not yet been validated for subjective aesthetic judgments, and their applicability to facial evaluation remains an open empirical question.

Until such validation is established, these advanced architectural approaches should be considered active research directions rather than recommended clinical components. Several debiasing techniques in this section also rest on preprint or single-study evidence whose reproducibility has not been independently established; responsible clinical adoption requires, at a minimum, prospective studies demonstrating improvement in prespecified fairness metrics across multiple independent ethnic groups, independent reproducibility on held-out datasets, and direct comparison against uncontrolled baseline systems under clinically realistic conditions.

A comprehensive fairness evaluation system integrates group fairness metrics, performance standards, and intersectional assessment. This layered approach prevents the common problem of achieving fairness on average while still having significant disparities in specific subpopulations, which is particularly critical in aesthetic facial evaluation, where cultural considerations vary significantly across intersectional identities.

Fairness metrics are chosen based on the clinical application to facial evaluation. Group fairness metrics measure and reduce bias to ensure that models evaluate different demographic groups equitably [47]. Demographic parity catches outcome bias by ensuring that the models’ aesthetic ratings are independent of ethnicity and equal across different demographic groups. Since the clinical utility of these models requires accuracy as well, equalized odds catches accuracy bias by ensuring that the models perform equally well across all demographic groups, with false positives and false negatives occurring equally. Additionally, the evaluation system should include equal opportunity metrics as an option for clinicians when identifying positive outcomes is the priority and false positives are not an issue. In an aesthetic evaluation scenario, equal opportunity focuses on not missing attractive features while being more flexible about possible overestimation. However, different fairness metrics can oppose each other, which makes it axiomatically impossible to align all of them simultaneously.

We propose the following benchmarks as starting points for community debate and empirical refinement, not as validated thresholds: cultural concordance scores of at least 80%, as reviewed by regional expert review panels [4], feature recognition accuracy of at least 95% [4], and demographic parity in prediction accuracy with no more than 5% variance across groups. These figures represent reasonable aspirational targets informed by analogous fairness benchmarks in other clinical AI applications but require prospective validation in aesthetic evaluation contexts before adoption as standards.

Because fairness metrics will conflict in practice, a decision hierarchy is necessary. Consider the following scenario: a model achieves demographic parity (equal average attractiveness ratings across ethnic groups) but does so by consistently overpredicting attractiveness for underrepresented groups while underperforming on fine-grained feature recognition for those same groups. Demographic parity is satisfied; equalized odds are not. In this scenario, we recommend prioritizing equalized odds because accuracy parity across groups is a prerequisite to the clinical utility of the tool. A model that fails to accurately recognize features in specific ethnic groups cannot serve those patients equitably, regardless of average score distributions. Equal opportunity metrics should then be applied as a secondary check, specifically where the clinical priority is avoiding false negatives, for example, ensuring that attractive features in underrepresented populations are not systematically missed.

This hierarchy assumes deployment in a pluralistic patient population. In settings where the clinical population is demographically homogeneous, the relevant fairness question shifts: the priority becomes within-population accuracy and avoidance of intragroup bias, rather than cross-group parity. In such contexts, fine-tuning on locally representative data may be both technically appropriate and ethically indicated, provided that the resulting model is transparently scoped to its intended deployment population and not generalized beyond it. A model developed and validated for a specific national or regional context, for example, a system trained primarily on Korean patients for deployment in South Korea, should be evaluated against locally derived aesthetic norms and demographic distributions, and its scope of applicability documented accordingly.

Since the above-mentioned techniques evaluate for biases for individual criteria such as race or age, an intersectional assessment should be introduced into model training to evaluate overlapping biases that may compound. In aesthetic facial evaluation, intersectional bias due to the aggregation of multiple social identities, such as race, ethnicity, nationality, socioeconomic background, gender, and age, can incorrectly influence outcomes and may perpetuate stereotypes [14]. An intersectional assessment provides the framework to apply metrics and standards across complex, overlapping demographic categories such as Black women or older Asian men.

Since individuals hold multiple overlapping identities simultaneously, each with associated social norms and expectations, aesthetic preferences cannot be understood by analyzing demographic categories in isolation, a theoretical foundation reinforcing that intersectional assessment addresses the fundamental mechanism through which identity influences perception [52].

Practically, however, current dataset sizes preclude stable estimates across all intersectional combinations, necessitating a prioritization hierarchy. High-risk intersections with documented performance disparities—such as older women with higher Fitzpatrick skin types—should be evaluated as the primary tier. Compositional approaches offer a promising avenue for addressing certain intersectionality challenges more tractably: multigroup equivariant network designs that use product groups can provide fairness guarantees across intersectional demographic combinations with computational complexity proportional to the sum rather than the product of group sizes, as demonstrated in natural language generation debiasing tasks [51]. Second, Bayesian hierarchical modeling of rare subgroups enables partial pooling of statistical strength from related intersections, providing more stable estimates for low-frequency cells; this approach ties naturally to the hierarchical architectures recommended in the Model Training Methodologies section. Third, multitask regularization can share statistical strength across related intersectional categories during training. Intersectional combinations not covered by the primary tier should be explicitly designated for future work rather than omitted without acknowledgment.

Explainability and transparency are prerequisites for trustworthy AI/ML systems in clinical use. Three different techniques adapted from general medical AI would be beneficial for training models used in aesthetic facial evaluation: gradient-weighted class activation mapping (Grad-CAM), local interpretable model-agnostic explanations (LIME), and Shapley additive explanations (SHAP) [53]. Grad-CAM provides visual and spatial explanations through heatmaps that highlight which facial regions contribute most to a model’s aesthetic predictions. Clinicians can identify which anatomic feature the model prioritizes and can also verify that the model emphasizes culturally appropriate features rather than defaulting to Eurocentric beauty standards. LIME provides case-specific explanations of a model’s individual predictions by showing which features influenced that specific assessment. This technique is model-agnostic and therefore flexible and versatile, meaning it can be applied to any AI/ML architecture—whether convolutional neural networks, transformer models, or future technologies—making it adaptable as the field evolves. SHAP provides information about how each evaluated feature contributes to the specific output of a model. It can explain individual predictions and provide a global overview of which features are most important in a dataset. The practical value of SHAP-based interpretation has also been demonstrated in surgical predictive AI, where feature-level explanation was used to identify key perioperative risk factors and improve transparency of model behavior for clinical decision-making [54]. While these methods continue to evolve with ongoing algorithmic refinements, these tools are useful but not sufficient for verifying cultural appropriateness in aesthetic evaluation. They identify which facial regions or features influence a model’s output, but do not explain why those features are aesthetically valued within a specific cultural context, which is the central interpretive question this framework is designed to address. Human expert review by culturally knowledgeable clinicians is, therefore, a required complement to explainable AI (XAI) output, not an optional one.

Intrinsically interpretable approaches, which incorporate domain knowledge directly into model architecture rather than applying post hoc explanation methods, represent a more direct path toward clinical-grade cultural interpretability. For facial aesthetic evaluation, this includes explicitly encoding geometric relationships (such as facial proportions, angles, and distances) as structured features within the model, and physics-based models that incorporate established anatomical principles and morphometric relationships.

Transparency standards should ensure that stakeholders have access to essential information about model development and validation. Transparency requirements for AI/ML training models should include the documentation of training data demographics and model architecture, interpretable explanations for aesthetic assessments, and disclosure of identified biases and performance disparities across demographics.

For generative outcome simulation systems specifically, XAI tools must address an additional interpretive requirement: clinicians should be able to verify that simulated postoperative appearances reflect only the intended surgical modifications and do not introduce ethnically incongruent features as artifacts of model bias. This requires comparison of presimulation and postsimulation facial geometry at the feature level, which current Grad-CAM and LIME implementations are not designed to provide; bespoke evaluation protocols for generative systems are needed.

The development of AI models should involve AI developers, patients, clinicians, cultural consultants, and ethicists as equal partners from the conceptualization stage through implementation and postdeployment monitoring [46]. This cocreation process ensures that diverse perspectives shape decision-making at every stage, from selecting training datasets to interpreting model outputs to refining algorithms based on real-world performance. Workshops and iterative feedback sessions should be conducted throughout the development lifecycle to gather input on critical questions such as which facial features should be prioritized for analysis, how to define culturally appropriate aesthetic outcomes, and whether model recommendations align with patient values and clinical judgment. Participant recruitment should deliberately include patients who identify across multiple marginalized dimensions simultaneously, such as older immigrant women from non-Western countries, as their aesthetic priorities and experiences of algorithmic bias are likely to differ from those captured by single-axis demographic sampling. It is vital to recognize and address potential power dynamics to ensure that underrepresented stakeholders’ voices are valued equally alongside technical experts and established institutions.

Regional expert review panels convened to evaluate cultural concordance must be constituted with explicit accountability requirements. Panel composition should include representatives from diverse geographic regions within each cultural community, not only urban or elite centers, and should reflect variation in socioeconomic background, age, and gender. Selection criteria should be documented and publicly disclosed. Formal mechanisms for recording minority and dissenting views are required; panel reports should distinguish consensus from majority positions and preserve dissenting opinions for review. Periodic panel rotation and external audit of panel composition guard against the entrenchment of a single institutionalized aesthetic perspective. These panel requirements must be operationalized through rigorous rater training and calibration protocols.

Rater training and calibration protocols are essential to annotation quality. Before scoring, raters should complete structured training that includes: an orientation to the study’s cultural equity goals; exposure to diverse face exemplars across all demographic categories to be rated; and explicit instruction to evaluate attractiveness according to within-group cultural standards rather than a universal ideal. Calibration should be conducted using a standardized set of anchor images—rated in advance by a culturally matched expert panel—against which individual rater scores are benchmarked. Raters whose scores diverge systematically from calibration anchors by more than a prespecified threshold (for example, mean absolute deviation greater than 1.0 on a 10-point scale) should receive additional training before contributing to the primary dataset. For images where rater scores span more than 3 points on a 10-point scale, the image should be flagged for adjudication by a culturally matched expert reviewer rather than resolved by averaging. Averaged scores obscure genuine aesthetic disagreement that may itself be informative about cultural variation. Ongoing audit of rater score distributions by demographic subgroup should be conducted throughout data collection to detect systematic drift in individual rater calibration.

A human-centered design framework positions clinicians and patients as essential collaborators while respecting their cultural contexts and prior experiences. Patients who have undergone aesthetic procedures provide experiential knowledge about how cultural identity influences aesthetic goals, what features they sought to preserve vs modify, and how algorithmic recommendations might have impacted their decision-making. This bidirectional learning process builds cultural competency across all stakeholders and creates the foundation for AI systems that serve diverse populations equitably.

Additionally, informed consent architecture is a prerequisite for ethical dataset development. Individuals depicted in training images must provide explicit consent for secondary use of their facial photographs in AI training, with the right to withdraw consent and have their images removed from future training cycles. This requirement applies regardless of whether images are sourced from clinical records, publicly available datasets, or social media platforms, and must account for jurisdiction-specific biometric privacy regulations and state-level statutes. However, consent withdrawal raises a technically significant challenge: once a model is trained on data, removing a data point’s influence from a deployed model without full retraining is computationally costly. The emerging field of machine unlearning addresses this problem through methods such as approximate unlearning and influence function-based data removal, though these techniques have not yet been validated or operationalized in medical AI governance contexts. Until practical machine unlearning protocols are established for clinical AI, consent frameworks should, at a minimum, guarantee removal from future retraining cycles and document this limitation transparently as a residual risk. Raters whose aesthetic judgments become training labels should similarly provide informed consent, be compensated equitably, and retain the right to withdraw their ratings from the dataset. Evolving Food and Drug Administration (FDA) guidance on training data provenance under the Software as a Medical Device (SaMD) framework should be monitored for additional requirements as it develops. Documentation of consent procedures, including the current technical limitations of consent withdrawal from deployed models, should be included in the transparency disclosures required elsewhere in this framework.

A governance structure for AI models should be based on multidisciplinary committees including clinicians, ethicists, data scientists, and patient representatives. This allows for ongoing ethical review with diverse stakeholder input and mandates human oversight for all AI-driven recommendations, requiring clinician review of model outputs before they inform patient consultations or treatment planning. To ensure regulatory compliance, developers should align with the FDA’s SaMD framework [55] and remain aware of evolving federal regulations for AI in health care.

Bias in the human–AI decision system extends beyond model development to the point of clinical use. Even a well-calibrated, fairness-aware model can produce inequitable outcomes if clinicians or institutions apply its outputs selectively or inconsistently across patient groups. Research on physician implicit bias [56] suggests that providers may differentially override algorithmic recommendations based on patient demographics, accepting recommendations for patients who resemble the provider’s implicit reference population while discounting them for others. Governance structures should include mandatory documentation of AI recommendation acceptance or rejection by patient demographic category, regular review of override rates stratified by patient race, ethnicity, gender, age, and language, and structured clinician training on the mechanisms of implicit bias in AI-assisted decision-making. Where systematic override disparities are detected, the governance committee should determine whether the source is model error for specific subgroups or clinician bias in interpretation. Fairness in aesthetic AI, therefore, requires monitoring both algorithmic outputs and human responses to those outputs.

Let drift be a statistically significant change, exceeding prespecified control limits, in 1 or more of the following: prediction accuracy by demographic subgroup, fairness metric values (demographic parity, equalized odds, cultural concordance), or the distribution of model inputs relative to the training distribution. Demographic-subgroup-specific degradation that does not affect overall accuracy is a particularly important drift signal, as aggregate metrics can mask emerging disparities. Institutional responsibility for drift response should be assigned explicitly at deployment: a designated AI clinical lead bears primary responsibility for reviewing automated alerts, convening the multidisciplinary governance committee, and authorizing remediation. Triggered remediation actions should follow a tiered protocol keyed to severity: minor drift triggers increased monitoring frequency and a targeted data audit; moderate drift triggers model recalibration or posttraining correction without full retraining; severe drift, including any demographic subgroup falling below institutionally defined minimum performance thresholds, triggers suspension of AI-assisted outputs for affected use cases pending full model retraining and revalidation. The specific thresholds delineating these severity tiers should be defined prospectively by each deploying institution based on clinical context, use case stakes, and available monitoring infrastructure, rather than adopted from universal benchmarks for which no empirical basis currently exists in aesthetic AI. All drift events and remediation actions should be documented in an institutional AI governance log and reported in periodic transparency disclosures.

Postdeployment monitoring of the AI/ML models should include continuous tracking of prediction accuracy, fairness metrics across demographic groups, and concordance with clinical judgment. A tiered monitoring approach is recommended. Continuous monitoring using statistical process control on prespecified fairness metrics, with automated alerts when metrics exceed defined control limits described above, provides the first line of detection. Quarterly structured reviews should assess fairness metric trends and flag emerging disparities for clinical review. An annual deep audit evaluates model architecture, training data composition, rater panel diversity, and alignment with updated regulatory guidance. Where continuous monitoring infrastructure is not feasible, drift-triggered audits, initiated automatically when prediction distributions shift beyond a prespecified threshold, represent a minimum acceptable alternative. The appropriate monitoring intensity scales with deployment volume: high-volume systems serving diverse patient populations require continuous monitoring; lower-volume or single-institution implementations may operate on a quarterly plus annual cycle with drift-triggered escalation. When audits identify performance disparities exceeding 5% variance across demographic groups, retraining (or appropriate posttraining) should be initiated. Feedback loops integrating clinician and patient input, as well as regular retraining cycles incorporating new, diverse data, allow for ongoing model improvement.

Ultimately, this framework is directed primarily at academic and institutional developers and is intended as aspirational guidance and input to regulatory deliberation. Most currently deployed aesthetic AI tools are commercial, including consumer-facing filters, practice-management platforms, and direct-to-consumer assessment applications, and fall outside the scope of institutional governance structures. We recommend that regulatory bodies consider disclosure-based accountability as a lighter-touch regulatory instrument: commercial developers would publicly document which framework components their systems satisfy, analogous to transparency requirements in other regulated industries, enabling clinicians and patients to assess compliance. We acknowledge that even this approach requires formal regulatory action and cannot be implemented through voluntary adoption alone. Long-term, enforcement mechanisms aligned with the FDA’s evolving SaMD framework and equivalent international regulations will be necessary to extend these standards to commercial systems.