Artificial Intelligence in Facial Harmonization: Uses and Limits

Artificial intelligence is already in the apps patients use, in social media filters and, more and more, in office tools. It detects facial landmarks, calculates proportions, generates "after" images and organizes schedules. For harmonization, this is both useful and risky: useful because it gives visual language to a hard conversation, risky because it is easy to mistake a convincing image for a prediction.

This page separates what AI does well, where it fails and what the professional must watch in ethics, LGPD (Brazil's data protection law) and communication. For the specific layer of simulation, see digital simulation in HOF. For the background on patients who arrive influenced by filters, see patient brings a filter photo.

Where does AI show up in the routine?

Each use has its own risk and benefit. What they share is that AI is a support tool, and the professional remains responsible for the act.

What does AI do well?

• Standardization. By detecting landmarks and aligning photos, it helps compare before and after under similar conditions. See standardized clinical photography.
• Visualization. A simulated image helps the patient understand concepts such as volume, proportion and intensity, and opens the conversation about what they actually want.
• Organization. It fills in fields, summarizes records, reminds about follow-ups and cuts administrative tasks.
• Education. It generates materials to explain procedures, as long as they are reviewed.
• Triage of repetitive questions, in service channels, under supervision.

Where does AI fail?

The limits deserve attention, because they are hard to see:

1. Hallucination. Generative models can invent plausible details, such as pores, shadows, teeth or contours that do not exist. See hallucination in image AI.
2. False precision. Measurements and scores with decimal places give an impression of exactness, but the photo, the angle and the expression change the result.
3. Biases. Training data reflect cultural and ethnic patterns. The tool may treat as "ideal" a profile that is not the patient's.
4. No clinical context. AI does not see layers, vessels, procedure history, medications or the real quality of the skin.
5. Inconsistency. The same photo can generate different results in different attempts.
6. Changes beyond the request. Some models alter skin, expression or background without warning, which weakens the honesty of the simulation.
7. Opacity. It is not always clear how the model decides or what it does with the image.

What are the ethical and legal precautions?

A photo of the face is personal data and, in many uses, biometric data and sensitive data. This brings obligations:

• Legal basis and consent. The patient must know that the image will be processed by AI, for what purpose and by whom. See consent and LGPD for photos and LGPD.
• Where the photo goes. If the tool sends the image to a third-party server, that must be clear in the notice and in the consent. Prefer tools that disclose how data is handled and how long it is stored.
• On-device processing reduces exposure when possible; see on-device processing.
• Transparency of use. Never present the image as an actual result. Advertising a simulation as if it were a result is inappropriate, and the advertising rules of the professional councils in Brazil apply. See what you may advertise: CFO and CONAR.
• Provenance. Label AI-generated images as such, so as not to mislead others. See AI image provenance.
• Do not use AI to deceive. Altering real before-and-after photos in marketing materials is a serious ethical breach.
• Patient rights. The patient may ask for their images to be deleted, under the terms of the law and of the patient record rules.

How do you use the simulation honestly?

The simulation is for conversation, not for selling a promise. A few practices:

• State that it is illustrative before showing it.
• Show the change at the volume you actually plan, not an exaggerated image.
• Ask the patient to say what they see before commenting.
• Do not use it to convince an undecided patient. See simulation for an undecided patient.
• Avoid showing it to patients with signs of dysmorphia. See simulation for a patient with suspected dysmorphia and when not to show the simulation.
• Never promise that the result will be the same. See patient wants a result identical to the simulation.

How do you choose a tool?

Some practical evaluation criteria, in order of importance:

And AI in marketing?

Using AI to create promotional content has its own rules. Generated images cannot be presented as real patients or as results. The professional must follow the rules of their council and those of CONAR (Brazil's advertising self-regulation body), avoid promises and protect privacy. Caution is especially important with photos of people who exist: creating or altering someone's image without permission is an ethical and legal problem, and the risk of deepfakes is growing. See what you may advertise.

And in Harmy?

Summary

AI supports harmonization with measurements, visualization and organization, but it does not assess the patient or predict the result. Use it with consent, attention to LGPD, transparency that it is a simulation, and a critical eye on biases and hallucinations. The clinical act, and the responsibility for it, remain with the professional.