The editorial examines several of the most active research fronts where AI is being tested in plastic surgery, and its conclusions are strikingly cautious. The central message is that while AI has already proven useful in non-clinical settings such as communication, accounting, feedback collection, and office management, the technology has not yet accumulated sufficient evidence to justify routine use in clinical practice. That gap between enthusiasm and proof, the author suggests, is where the specialty’s attention should now be focused.

One of the most debated applications is the computerized assessment of aesthetic outcomes. Plastic surgery has long struggled with the subjective nature of judging surgical results, and AI-driven computer vision promises to replace human impression with numbers. A study by Varghaei and colleagues assembled what is likely the largest curated paired pre- and postoperative facial image dataset to date, comprising 7,160 photographs from 1,259 patients. The team introduced a scalable, interpretable computer-vision framework designed to quantify aesthetic outcomes of facial plastic surgery using frontal photographs. On its face, this seems like a decisive step toward objectivity, but the editorial raises a fundamental objection: quantification does not guarantee objectivity.

The problem lies in how these systems learn. Current AI technology is dominated by multimodal large language models trained on human input, and the way humans train these models inevitably shapes the opinions they produce. The editorial coins the term compjective to describe this hybrid state, in which a machine’s judgment is neither fully objective nor entirely subjective. The authors of the facial analysis study themselves acknowledged that facial attractiveness, though inherently subjective, is often associated with geometric regularities such as bilateral symmetry, proportionality, and averageness. That concession raises an uncomfortable question: can computerized assessment of aesthetic outcomes ever truly be objective?

Independent evidence suggests the answer, for now, is no. A comparative evaluation by Haider and colleagues tested leading multimodal large language models, including ChatGPT-4o, GPT-4, Gemini, and Claude, on their ability to assess aesthetic features through both qualitative and quantitative analysis. The verdict was sobering. Without fine-tuning or integration with specialized computer-vision architectures, none of the models could provide a reliable aesthetic assessment. In practical terms, the general-purpose systems that most people call artificial intelligence today do not have the capacity to replace a surgeon’s trained eye and accumulated experience in evaluating an aesthetic result. The editorial’s assessment is that AI-driven computer vision has not yet proven to be a reliable resource for clinical integration and requires substantial advancement, though it may serve as a supplementary tool for quantifying treatment results in research settings.

A second frontier is continuous monitoring, where the case for machine assistance is arguably stronger. Every surgeon knows the classical signs of inflammation, calor, dolor, rubor, tumor, and functio laesa, and every training surgeon has endured the burden of watching flaps overnight with instructions as vague as calling if the color changes. A monitoring system that never gets drowsy and never loses focus could be a welcome alternative to a tired colleague. Kim and colleagues developed and tested an automated free flap monitoring system based on artificial intelligence, and the results were promising: sensitivity for detecting venous insufficiency reached 97.5 percent, while sensitivity for arterial compromise reached 92.8 percent. These figures suggest that supplemental monitoring and early warning systems could genuinely optimize patient care, but the editorial is careful to note that such technologies have yet to be evaluated in a clinical trial setting.

That trial evidence may not be far off. Several clinical studies are currently underway focusing on AI-assisted monitoring, diagnosis, and detection. Among them are WISDOM, a trial using AI within digital wound monitoring of surgical wounds to prioritize non-healing wounds for urgent review, registered as NCT06475703; AnAsToMoSIs, a study of artificial intelligence approaches to the monitoring of microvascular free-flap inpatients, registered as ISRCTN36476735; SeeWound2, a clinical evaluation of a digital wound-assessment device using AI-based image analysis, registered as NCT07211295; and an AI risk assessment model for complication prevention in plastic surgery, registered as NCT06507384. Most of these studies are observational or device precision studies, which means the field is still in the phase of establishing that the tools work as claimed, not yet proving that they improve outcomes.

Beneath these specific applications lies a deeper set of methodological problems that the editorial identifies as endemic to the literature. Many published AI models derive from single centers or demographically narrow datasets, which severely restricts their applicability to larger and more varied patient communities. When such models are ported to different populations, performance will likely degrade, because those populations do not resemble the original training data. There is little published work on cross-ethnicity fairness audits or performance stratification across different phenotypes, and many studies omit the detailed metadata that would make post-hoc analysis, standardization, or reproducibility possible. Patient-reported outcomes are frequently missing from pilot studies, a flaw the editorial calls major, since a patient’s understanding of a successful outcome may differ substantially from that of a computerized assessment program. Ultimately, the argument goes, AI metrics must map to patient-reported outcomes, to the surgeon’s decision-making, or to demonstrable improvements in care, and current models show signs of both selection bias and aesthetic bias.

There are, however, examples of success that point the way forward. The editorial cites work by Ravi Dhawan and colleagues reviewing FDA-approved AI-utilizing tools in the wider surgical field, which found that the devices that have cleared regulatory hurdles share common features, including necessity and efficiency. Future pioneers of AI technology in plastic and reconstructive surgery, the editorial argues, should weigh the necessity, safety, and reliability of their tools in improving patient outcomes, reducing complications, and ensuring minimum disruption to existing clinical standards and practices. A tool that is merely novel, in other words, is not enough.

To close the gap between promise and proof, the editorial lays out a set of steps for responsible integration. Reporting standards should be mandated, covering everything from methods to metadata, with proper study registration, open access data sharing, and detailed transparency as prerequisites for a reliable tool. Validation should be prospective and pre-registered, ideally multi-center, because retrospective validation is subpar for supporting the effectiveness of a new tool, even one designed only as a supplement. A patient-centered approach should govern development, with patient feedback included in assessing effectiveness and sufficient fairness audits, subgroup analyses, and stratification of performance by age, sex, ethnicity, and skin tone to avoid perpetuating aesthetic biases. Surgeon-AI workflow integration trials should investigate how AI metrics influence surgical planning, patient counseling, shared decision-making, and interrater variability, with clinician feedback loops refining the AI’s decision logic during testing. Finally, regulatory and consent frameworks need attention, since existing bodies such as the FDA and CE marking schemes do not define specific approval pathways for AI technology, and without a standardized approach governing patient consent and data protection, these tools cannot be freely deployed. The editorial’s conclusion is blunt: while the specialty is prepared for AI that optimizes care and improves outcomes, existing research does not yet provide sufficient evidence of safety for clinical integration, and future work must incorporate both patient and surgeon feedback while meeting the same methodological and reporting rigor that has become standard in clinical trials.

Subject of Research: Responsible and evidence-based integration of artificial intelligence into plastic and reconstructive surgery

Article Title: Responsible and safe integration of AI technology into clinical practice: accommodating the inevitable technological change

Article References: Sinelnikov, M. Y. (2025). Responsible and safe integration of AI technology into clinical practice: accommodating the inevitable technological change. BMC Plastic and Reconstructive Surgery, 1(1), Article 6. https://doi.org/10.1186/s44452-025-00006-4

Image Credits: AI Generated

DOI: 10.1186/s44452-025-00006-4

Keywords: artificial intelligence, plastic surgery, reconstructive surgery, computer vision, large language models, free flap monitoring, clinical trials, patient outcomes, FDA regulation, aesthetic assessment, fairness audits, surgical innovation