AI cosmetic-surgery simulations: accuracy, privacy, and consent

An AI or 3D cosmetic-surgery simulation is an illustrative, versioned communication tool—not a guaranteed result, treatment plan, validated informed-consent substitute, or proof of surgical skill. Its value depends on the exact software, procedure-specific validation, inputs and manual edits, uncertainty disclosure, and whether the clinic has a clear contract for facial-image access, vendor processing, model training, retention, deletion and reuse.⁵⁶

Save the preview as a data artifact, not a promise

A simulation may help a patient and surgeon discuss size, proportion, profile, asymmetry or priorities. It cannot model every biological variable: tissue quality, swelling, scarring, healing, implant behavior, weight change, aging, surgical compromise and complications. The 2026 literature describes communication benefits while emphasizing limited objective accuracy assessment and unresolved ethical and legal questions.⁵

Preserve the date, software name and version, device/camera, source images, selected procedure, user inputs and every manual edit. If the surgeon changes the preview by hand, that is part of the artifact. A screenshot detached from those coordinates is not reproducible.

Accuracy must be defined before it can be measured

One tool may estimate breast volume, another modify nasal contours, and another generate a whole-face “after.” Accuracy could mean millimeters at landmarks, three-dimensional surface distance, volume, blinded similarity ratings or satisfaction. These are not interchangeable.

In a 2025 image-generation study, plastic-surgery clinicians rated three general AI models differently on realism and clinical-value measures; all three performed poorly for healing and scarring prediction. The study did not compare simulations with actual postoperative outcomes.⁶ Ask whether a clinic’s tool has been validated with the same procedure, imaging protocol, patient population, follow-up interval and software version it uses. Internal testing on selected examples is weaker than external testing on consecutive cases.

The postoperative comparison also needs a stable endpoint. Early swelling, camera angle, expression, focal length and head position can make a real result look closer to or farther from the preview.

A face image is not anonymous because a name was removed

HHS lists full-face photographs and comparable images among identifiers in the HIPAA Safe Harbor method.¹ Cropping a background or deleting a filename does not necessarily de-identify a recognizable face. Three-dimensional scans, landmarks, device IDs, timestamps and associated procedure interests can add re-identification value.

HIPAA applies to covered entities and business associates in defined circumstances, not to every consumer app. When a covered practice uses a cloud vendor to maintain electronic protected health information, HHS describes continuing duties and generally a business-associate arrangement.² If the tool is consumer-facing or outside that relationship, the FTC Act and Health Breach Notification Rule may still matter.⁴ “HIPAA certified” is not a universal federal approval seal.

Treatment consent is not blanket AI-training consent

Permission to take photographs for care does not automatically authorize public marketing, model training, vendor product improvement or indefinite storage. Those purposes should be separated in understandable language. Refusing optional marketing or training should not be hidden as refusal of necessary clinical documentation.

FTC’s Everalbum matter illustrates why facial-recognition and retention promises must match actual practice, including handling of derived models and data after a requested deletion.³ It is not a finding about every cosmetic-simulation vendor; it demonstrates that downstream model use can be part of the privacy question.

The surgeon must remain the source of the plan

A simulation should follow examination and a feasible surgical range, not lead with an unconstrained ideal. Ask the surgeon to mark which features are likely, possible, uncertain or not advisable and which limitations come from anatomy, tissue, prior procedures or safety. Save those annotations.

The informed-consent discussion still needs the exact operation, alternatives, material or implant when relevant, scars, anesthesia, recovery, common and serious risks, uncertainty, revision possibilities and the option not to proceed. A generic variability disclaimer is not enough if staff repeatedly describe the preview as what the patient will look like.

Version drift can invalidate old claims

AI software may change its model, training data, image pipeline or default settings. A validation study of version 2 may not establish version 5 performance. Ask how the clinic records updates, revalidates important uses, handles generated artifacts after an update and identifies changed outputs.

The same applies to consumer tools used before consultation. Bring the image as a preference prompt, but disclose that it may be generated from an unvalidated app. A surgeon should not infer that a cosmetically plausible output is anatomically achievable.

1. Define the preview's job Use it to discuss a bounded feature or tradeoff, not as a predicted final photograph.
2. Record provenance Save software/version, inputs, camera setup, settings, edits, operator and date.
3. Demand relevant validation Match procedure, outcome metric, population, follow-up and current version; ask for error, not only average similarity.
4. Separate privacy choices Distinguish care documentation, vendor processing, model training, research and marketing permissions.
5. Return to clinical consent Have the surgeon state feasible range, uncertainties, risks, alternatives and revision limits independently of the image.

The best simulation does not make the future certain. It makes the present conversation more precise while preserving the patient’s control over both expectations and facial data.