In short. What these models are trained on, and why a score reflects its training data rather than a standard of beauty anyone agreed to.
Key points
- Studies that compare software ratings with clinician ratings measure how closely two sets of judgements line up.
- That is a useful thing to know and it is not the same as correctness, because the clinicians are also expressing preferences.
- A model can agree closely with a panel and still be describing the panel.
A rating looks like a measurement, and it is produced by a different process. Understanding how a model arrives at a score makes it easier to see why its output is a statement about the training data rather than about a face.
What the model is trained on
A model learns from examples. A large set of images is assembled, each is given a label by somebody — attractive, symmetrical, youthful, or a number on a scale — and the model learns the statistical pattern that connects the image to the label. Both steps involve choices. Which faces were included, how they were photographed, who did the labelling and what they were asked to judge all shape the result before any learning happens.
Why the score is a preference
The labels encode judgements. If the raters favoured a particular proportion, that preference is what the model learns and what the score reports. Nothing in the process converts a preference into a standard, and the output is presented as a number precisely because a number looks like it was measured rather than chosen.
Where the measurement stops and the inference begins
Some outputs are measurements: a distance, an angle, a ratio between two landmarks. Those are reproducible and they mean what they say. Where the software moves from a ratio to a rating — this face scores well, this one does not — it has made an inference from the measurement, and the inference carries the assumptions of the training step. The two outputs can appear side by side on the same screen, which is how a preference acquires the authority of the number next to it.
Agreement with clinicians is agreement with an opinion
Studies that compare software ratings with clinician ratings measure how closely two sets of judgements line up. That is a useful thing to know and it is not the same as correctness, because the clinicians are also expressing preferences. A model can agree closely with a panel and still be describing the panel.
Why the population question is unavoidable
A model performs best on faces that resemble the ones it learned from. Where a training set drew heavily on one population, one age group or one set of photographic conditions, the tool is least reliable for everyone else, and it usually fails quietly rather than announcing its uncertainty. The practical question is not whether a model is biased, which all models trained on examples are, but which faces it was trained on and whether anyone measured how it performs on faces it has not seen.
Three questions worth asking about any rating
- What were the labels, and who decided what they meant?
- Which faces were in the training set, and how does the tool perform on faces that were not?
- Is this output a measurement, or an inference drawn from one?
A rating is a preference expressed arithmetically. Knowing that does not make it useless, and it does make it something other than an answer.
Frequently asked questions
How does facial analysis software produce a score?
It is trained on a set of images that have been labelled in some way, and it learns the pattern that links image features to those labels. The labels come from whoever assembled the set, and the score is a compressed version of their preferences rather than a measurement against any agreed standard.
Why can two tools give different ratings for the same face?
Because they were trained on different images, asked to optimise different things, and evaluated differently. A rating is an output of a particular model rather than a property of a face, and two models can disagree without either being broken.
Does a large training set remove bias?
No. A large set that is unrepresentative is still unrepresentative, and it makes the bias harder to see because the results look stable. What matters is which faces were included, how they were labelled, and whether anyone checked how the tool performs on people who were not in the training data.
Risks & Limitations
- Software measures an image; it does not examine tissue, and an examination is what a diagnosis usually turns on.
- A rating reflects what a model was trained on rather than a standard of beauty, and it is least reliable for faces least represented in its training data.
- Agreement between software and clinicians is measured against an opinion, not against correctness.
- A predicted image is a drawing of a plan expressed in a body the model does not describe.
- No algorithm takes a history or carries responsibility for the advice it gives.
When to Seek Care
- A new or changing lesion, a mole changing in size, shape or colour, or a sore that is not healing — have it examined by a clinician rather than assessed from a photograph.
- A skin change you are worried about, whether or not a tool has told you it is nothing.
- A reaction, infection or worsening after treatment — contact the service that holds your records.
- If you have uploaded a facial image and want to know how it is held or want it deleted, ask the service that holds it under the terms you accepted.
References & Further Reading
- U.S. Food and Drug Administration (FDA) — guidance and authorisations for software as a medical device and for image analysis
- European Commission and notified bodies — Medical Device Regulation and artificial intelligence guidance
- National Medical Products Administration (NMPA, China) — approved software and its named claims
- World Health Organization (WHO) — guidance on the ethics and governance of artificial intelligence for health
- Cochrane Library — systematic reviews of diagnostic and decision-support software
This page is patient education, not medical advice, and it does not cite individual studies. It reviews evidence and regulation; it is not a booking page and does not recommend a treatment. Discuss your own case with a qualified clinician. To ask a question about this review, see Contact & Consultation.