AI in Aesthetic Consultation: What the Evidence Says
Four different things are sold under the name of an AI consultation, and only one of them has a solid evidence base. Software that measures features on a photograph does what it says, within the limits of the photograph. Software that produces a rating agrees with clinicians moderately on some measurements and much less on aesthetic advice. A simulator that shows a predicted result has no evidence behind it, and a chatbot reply is not a consultation at all. What is established is narrower than the marketing and broader than outright dismissal: image measurement is a real tool, agreement between software and clinicians is measurable and partial, and no algorithm carries responsibility for a decision. This site separates those claims, grades each, and states regulatory status as the separate question it is.
Chongqing West District Hospital, a tertiary general hospital in Jiulongpo District, Chongqing. This is a review of published evidence and regulation, not medical advice, and it is not a recommendation for or against any treatment.
How claims are graded here
Every claim on this site sits in one of four groups, and the group matters more than the wording. Established means the claim follows from evidence that is consistent across studies and is reflected in how the software is regulated and labelled — that an algorithm can measure a named feature on an image, and that the measurement is reproducible for that task, belongs here. Limited means there is real evidence but it is narrow: agreement between software and clinicians has been measured for some tasks, in some populations, against one particular clinician panel. Emerging means the approach is plausible and early work points one way, but the study that would settle it has not been done, which is where triage sits. Not supported means the claim is made confidently and the evidence does not carry it. The reason for grading rather than simply reporting is that these four groups are routinely blended into one voice. Measuring a nose width and advising on an operation are described in the same confident register, and only one of them is something a study can establish. Where a comparison was made against a panel that shares the software’s assumptions, this review says so, because agreement with a panel is not the same as agreement with a person’s own judgement about their face.
Where the line falls, and what it separates
The regulatory line in this area falls between software that describes an image and software that names a condition, and the line is enforced rather than advisory. Software that segments an image, measures a distance or tracks a change over time is regulated as a device for a defined task, and its clearance names that task. Software that outputs a diagnosis — this is rosacea, this is a melanoma — is regulated as a device making a diagnostic claim, and that is a considerably higher bar involving evidence about accuracy in the population where it will be used. A product marketed for the first task while implying the second is the most common mismatch in this field. A rendered prediction of a surgical result is generally neither, because it claims nothing regulated, and an app that returns a generic reply is not a device making a claim about your skin at all. Approval also differs by jurisdiction, so a tool available in one market may have no equivalent clearance in another, and the same product name can sit in different classes in different places. Where a clinic presents one clearance as covering all of these, the clearance is being borrowed for claims it does not make.
What the evidence establishes
These claims survive scrutiny. Each is stated with its own limit rather than as a general reassurance.
An algorithm can measure a named feature
Established as a technical task. Given a suitable photograph, software can locate a landmark, measure a distance or an angle, and track the same measurement across images. That is a real capability and it is what such software is cleared to do. What it establishes is a number, not a judgement about the number.
Software agreement with clinicians is measurable
Limited. Studies have compared software ratings with ratings from clinicians and found moderate agreement on some measurements and much weaker agreement on aesthetic advice. The result is genuine and it is also narrow: it depends on the task, the panel and the population, and it is measured against an opinion rather than against a standard.
Measurement is more consistent than an eye
Established for repeatability. A measurement taken by an algorithm does not drift between two assessments the way an unassisted estimate does, and for tracking a change over time that consistency has real value. Consistency is not accuracy: the same number can be produced reliably and still not mean what it is being used to mean.
Software can shorten the wait for an assessment
Emerging. Triage tools that rank or route patients have shown promise in some settings, and pilot work suggests they can move people toward the right kind of appointment sooner. The studies are early, they come from few centres, and none has shown that a triage decision is as good as an examination rather than merely faster than waiting.
Image data can be handled responsibly
Established as a practice rather than as a property of a product. Facial images can be stored, encrypted and deleted under a defined policy, and consent terms can be written plainly. That any particular app does so is a checkable fact about that app rather than a general property of the category, and it should be asked about rather than assumed.
Claims the evidence does not carry
These claims are common, and none of them is supported. Naming them is the most useful thing this review does.
That an app can diagnose a skin condition
Not supported. Naming a disease is a diagnostic claim, and a diagnostic claim requires evidence of accuracy in the population where it will be used. Software that describes an image is a different product from software that names a condition, and a product marketed as the first while implying the second has not been assessed for the claim it is making.
That AI predicts your surgical result
Not supported. A predicted image is generated from assumptions about how tissue behaves under a change that has not happened, and no study has shown that such an image forecasts an operative outcome. It can be accurate about a measured feature and simultaneously silent about skin, healing and how a face moves.
That a chatbot reply is a consultation
Not supported, and the mismatch is structural rather than a matter of quality. A consultation involves examining tissue, taking a history, judging what someone actually wants and carrying responsibility for the advice. Software with none of those inputs is producing text, and fluent text is not the same as an assessment.
That a rating reflects a standard of beauty
Not supported. A rating reflects what the model was trained on and what it was asked to optimise. Those choices are made by people, usually without being stated, and the output inherits them. A score presented as an objective measure of a face is a preference expressed as a number.
That uploaded photographs stay private
Not supported as a general assumption. What happens to an image depends on the terms attached to the app and on where the data is processed and stored, and those terms vary widely. A facial image is personal data of a kind that is difficult to withdraw once it has been shared, which is a reason to read the terms before uploading rather than after.
What the risks and limits actually are
The first risk is that a number replaces a question. A rating displayed on a screen has the appearance of a finding, and a finding is difficult to argue with, so a person can be moved toward a treatment by a score rather than by an assessment of what they wanted. That matters more here than in most subjects, because the thing being measured is a face, and a face is not a problem to be solved. The second risk is bias inherited from training data. A model learns from the images it was shown, and if those images came mostly from one population, one age group or one set of faces, its measurements and its ratings will be least reliable for everyone else. This is not a defect that a product update removes; it is a property of learning from examples, and the honest response is to ask what the model was trained on. A third risk is that the output is treated as diagnostic when the product is not cleared to diagnose. A prompt to seek care can be useful; a label applied to your skin is a clinical claim, and a wrong one delays the assessment that would have been correct. There are limits on the evidence itself. Much of the published agreement work compares software against a small panel of clinicians, which measures agreement with an opinion rather than correctness, and the panels are rarely independent of the software being tested. Longitudinal evidence, showing that a tool improves what happens to people over time, is largely absent. And the privacy question is not a technical footnote: a facial image is difficult to withdraw once shared, and how an app stores, reuses or shares it is a matter of its terms rather than of its accuracy.
What else is used, and what software cannot replace
Doing nothing is an option, and where the question is whether to treat at all, it is often the right first one. Where an assessment is genuinely wanted, the alternative to a software-mediated first contact is the older one: a conversation with a clinician who examines the skin, asks about history and listens to what someone is actually asking for. That conversation is slower and it costs more, and it collects things no image contains. Photography remains useful without any algorithm attached, because a standardised photograph taken the same way over time is itself a record of change. Standard dermatological assessment, including dermoscopy and, where indicated, a biopsy, answers diagnostic questions that no photograph-based tool is cleared to answer. Where the appeal of a consultation tool is convenience, a teleconsultation with a clinician is the closer comparison, and it has the advantage of a person on the other end who can be questioned. Where the appeal is measurement, a clinician can measure. The useful way to think about the alternatives is by what each one is good at: software is good at repeatable measurement on an image, a clinician is good at examining tissue and judging a whole situation, and neither substitutes for the other. Whether a tool adds anything to a particular consultation is a clinical question, and this review does not answer it for anyone.
How this review is kept current
This subject moves faster than any other on this cluster, and the fastest clock is not the evidence. Products are revised, retrained and relaunched, and a model that was tested in a published study may have been replaced by the time the study is read. Regulation moves more slowly and more consequentially, because a change in what a tool is cleared to claim is the fact that determines whether a marketing sentence is permitted. The evidence moves slowest of all, because agreement studies take years and the longitudinal work that would show a benefit to patients has barely begun. This review is therefore dated and revisited on a short cycle, and it covers three things. The first is whether any claim has moved between the four grading groups, in either direction, because a claim can become supported as well as cease to be. The second is regulatory status, since a clearance or a warning is a checkable fact that a reader is entitled to have current. The third is whether the evidence base has changed in kind: another agreement study against another small panel is a different event from a trial that follows people over time, and only the second can show that a tool makes a difference. Where this review is wrong, the correction is made to the page rather than noted elsewhere, and the date of the most recent review is shown on every page. Where a question is genuinely unresolved, this review says that it is unresolved rather than resolving it by tone.
Claims about AI in aesthetic consultation, set against the evidence
Every row is a claim you are likely to meet, placed against what the evidence supports and what regulators have actually done about it. Evidence and regulatory status are two separate columns because they answer two separate questions.
The evidence guides
Each guide answers one question about the claim rather than about a treatment. Read the two or three that match what you are trying to establish.
What an AI Aesthetic Consultation Actually Is
Image measurement, outcome simulation, chatbots and triage forms are four different products sold under one name, and only the first has a solid evidence base.
Read the guide →Why AI Is Being Offered in Clinics Now
The commercial logic of a fast, cheap, always-available first contact, and the parts of an assessment a first contact cannot include.
Read the guide →How Facial Analysis Algorithms Reach a Rating
What these models are trained on, and why a score reflects its training data rather than a standard of beauty anyone agreed to.
Read the guide →Regulatory Status of AI in Aesthetic Medicine
Where the line falls between describing an image and naming a disease, and how that line is enforced across different markets.
Read the guide →What the Evidence Supports for AI Facial Analysis
Moderate agreement with clinicians on some measurements, and much weaker evidence for aesthetic advice or for any benefit to a patient.
Read the guide →What Is Not Established About AI Consultation
It cannot examine tissue, take a history, judge expectations or carry responsibility for a decision, and no study shows it improves care.
Read the guide →A question about the consultation evidence?
This site is an evidence review rather than a booking page. It does not recommend for or against a treatment, and it does not offer an assessment, a skin analysis or a consultation in relation to anything it describes. If you have a question about this review, or about your own care, contact the team and ask.
Frequently asked questions
What is actually established about AI in aesthetic consultation?
That software can measure named features on an image reproducibly, and that its agreement with clinicians has been measured and is moderate for some measurements and much weaker for aesthetic advice. Beyond that the evidence thins quickly.
Can an app diagnose a skin condition?
Naming a disease is a diagnostic claim, and a diagnostic claim needs evidence of accuracy in the population where it will be used. Software that describes an image is a different product from software that names a condition, and a product marketed as the first while implying the second has not been assessed for what it is implying.
Is a chatbot the same as a consultation?
No, and the difference is structural rather than a matter of quality. A consultation examines tissue, takes a history, judges what someone actually wants and carries responsibility for the advice. Software with none of those inputs is producing text, and fluent text is not an assessment.
What is a facial analysis score actually measuring?
What the model was trained on and what it was asked to optimise. Those are choices made by people, and the output inherits them. A score presented as an objective measure of a face is a preference expressed as a number rather than a standard anyone agreed to.
What happens to photographs I upload to a consultation app?
It depends on the terms attached to that app and on where the image is processed and stored, and those terms vary widely between products. A facial image is difficult to withdraw once it has been shared, so the terms are worth reading before an upload rather than afterwards.
Does AI triage shorten the wait for an assessment?
That is emerging rather than established. Pilot work in some settings suggests a triage tool can route people toward the right kind of appointment sooner. No study has shown that a triage decision is as good as an examination, only that it can be faster than waiting for one.
Part of Regenerative and Emerging Treatments — 9 sites on this topic