Your face as a medical test: AI revives physiognomy, this time with some science behind it

THE STUDY THAT WENT VIRAL. On 19 October, The Lancet Digital Health published work by researchers at Mass General Brigham in Boston. The Guardian summed it up in a widely shared headline: a selfie can predict whether you will survive cancer. Behind it is a 'foundation model' called FAHR-Face. A foundation model is an AI that first learns from huge amounts of data with no specific task and is then adapted to particular uses. FAHR-Face learned from more than 40 million face images. Two tools come from it. FAHR-FaceAge estimates biological age, meaning how much the body has actually aged. FAHR-FaceSurvival estimates the probability of survival directly.

The numbers are striking. According to Scientific American, in a group of 3,820 cancer patients, 64 of the 122 the model placed at highest risk died within 90 days of the photo: 52.5%. In the lowest-risk group of 975 patients, about 2% died. In another cohort of more than 38,000 patients, those who looked older than their actual age were more likely to die. There was also an independent check on nearly 5,000 patients in the Netherlands. That is the step that makes any model credible: testing it on data different from what it learned from.

This did not come out of nowhere. In May 2025, the same group presented FaceAge in the same journal. It was trained on 58,851 portraits of healthy adults and tested on 6,196 cancer patients in the US and the Netherlands. On average, the patients looked about five years older than they were. In one test, eight doctors tried to guess which terminally ill patients would die within six months. They did barely better than chance. With the FaceAge score, they improved clearly.

Our reading: this is serious science, and the line 'a selfie predicts whether you will survive' overstates it. The model does not see cancer in the face. It sees the body's overall decline: muscle loss, fatigue, poor nutrition or smoking. Doctors already judge this by eye with scales such as ECOG, which scores from 0 to 5 how much the disease limits a patient's daily life. Study author Raymond Mak says so himself: 'We're trained in medical school to walk into a room, look at the patient and make an assessment.' The AI turns that subjective impression into a number. That is valuable, but it is not magic.

FROM LAVATER TO LOMBROSO: A HISTORY THAT CALLS FOR CAUTION. The idea that the face reveals who we are has 250 years of bad precedent. Between 1775 and 1778, the Swiss pastor Johann Kaspar Lavater published his 'Physiognomic Fragments'. They were hugely popular and inferred moral character from the nose or the forehead. In 1876, the Italian physician Cesare Lombroso published 'Criminal Man'. He argued that there were 'born criminals', recognisable by 'atavistic' physical traits, meaning traits of primitive humans. In the same period, Francis Galton overlaid photos of prisoners to produce a 'typical' criminal face. None of these ideas survived statistics. They did survive as justification for scientific racism and discriminatory policies.

The algorithm has revived the temptation. In 2016, Xiaolin Wu and Xi Zhang of Shanghai Jiao Tong University posted on arXiv, a repository of papers without peer review, a classifier they said could spot criminals from photos of 1,856 Chinese men. In 2017, Blaise Agüera y Arcas, Margaret Mitchell and Alexander Todorov replied with an essay called 'Physiognomy's New Clothes'. Their argument: models like this may be picking up differences in where the photos came from and how they were taken, not differences between people. And the label 'criminal' does not measure a trait of the person. It measures who was convicted, with all the biases of the justice system.

THE 'GAYDAR' CASE AND THE LABEL TRAP. In 2017, Yilun Wang and Michal Kosinski of Stanford claimed that a neural network could guess sexual orientation from dating-site photos. Groups such as GLAAD called it 'junk science' and warned it could be used to out people without their consent. Critics noted that the sample included only white faces. They also noted that the label came from what users said they were looking for on the site, which is not the same as orientation. The authors replied that the study was peer-reviewed and that they meant to warn about a risk.

In 2019, a replication by John Leuner (University of Pretoria) using 20,910 photos reached 68% accuracy for men and 77% for women. But the telling detail is elsewhere. With blurred, almost unrecognisable images, the model still scored 63% and 72%. That suggests much of the signal lies in brightness, colour or how people present themselves to the camera, not in facial bone structure. Leuner himself admits he cannot separate how much is morphology and how much is grooming or lifestyle.

Our reading: this is the key to the whole debate. A model can be statistically 'accurate' and still not measure what it claims to measure. What separates medicine from physiognomy is not the power of the algorithm. It is the label. 'Died within 90 days' or 'has a mutation confirmed by a genetic test' are facts measured independently. 'Criminal', 'gay' or 'good hire' are social categories. They are also collected through data shaped by how the photo was taken, who posted it and why.

WHAT THE MACHINE CAN READ. Some signals have a physiological basis, and a few are already in use. Face2Gene, from the company FDNA, helps geneticists suspect rare syndromes from facial features. In 2019 it published its DeepGestalt system in Nature Medicine, trained on more than 17,000 images covering 216 syndromes. In a test on 502 images, the correct syndrome appeared among its top ten suggestions in 91% of cases. FDNA itself stresses that it is an aid, not a diagnosis. That makes sense: children with a rare disease can go years without a diagnosis, and a good lead shortens that search.

The voice leaves traces too. Parkinson's disease impairs control of the speech muscles, and depression flattens intonation. In 2025, a study in the Annals of Family Medicine evaluated the voice tool from the company Kintsugi on 14,898 adults. It detected moderate to severe depression with a sensitivity of 71.3%: it found about 7 in 10 cases. Its specificity was 73.5%: it correctly cleared about 7 in 10 healthy people. But the reference was a questionnaire filled in by participants themselves, not a doctor's diagnosis. Several authors also work for the company, and they call the results preliminary.

Our reading: these signals are real but modest, and nearly all of them serve one purpose. They help with triage, meaning deciding who to look at first. They are not verdicts. This fits what we have already seen in clinical AI: its value lies in allocating the specialist's time, not replacing the specialist.

THE STATISTICAL FINE PRINT. There are five questions worth asking of any headline like this. The first is sample diversity. According to Scientific American, the survival model was trained on mostly white patients from Boston. Face2Gene shows why this matters. In a study of African children, the tool recognised Down syndrome in 80% of children of European descent but only 36.8% of African children. After retraining on African photos, it reached 94.7%. And the GestaltMatcher reference database is still 67% European.

The second question is external validation, meaning testing in other hospitals and populations. FAHR-Face partly has it, with the Dutch cohort. The third is correlation versus causation. The model does not explain what biology it detects. Miroslaw Bober of the University of Surrey has stressed this point. The authors admit that smoking or heavy drinking may explain part of the predictions.

The fourth is confounding: hidden variables that drive both the photo and the outcome. Lighting, camera, hospital and socioeconomic status all play a part, and status affects both skin and survival. The fifth, and most overlooked, is real clinical value. The eye-catching figures compare the extremes: 52.5% against 2%. Most patients sit in the middle, where the signal is weaker. That is why outside experts are asking for more. Wilbur Lam of Emory University wants prospective trials, which follow patients from now on instead of analysing old data. Bober wants a head-to-head comparison with doctors' judgement. The decisive question is not whether the model is accurate. It is whether it changes decisions and improves outcomes.

IMPLICATIONS: THE FACE AS HEALTH DATA. The short-term risk is concrete. Matthew Liao of New York University's Center for Bioethics warns of two dangers. One is that an error leaves someone who needs care without it. The other is that insurers or employers use these tools. A face is data you cannot change, and we leave it on social media, cameras and video calls. If a photo reveals biological age or risk of death, it becomes health data and should be protected as such.

Europe has already drawn a useful line. Since February 2025, Article 5(1)(f) of the EU AI Act bans inferring emotions from biometric data in the workplace and in education, including hiring. Medical and safety uses are exempt. HireVue, the video interview platform, dropped facial analysis in 2021 after criticism, although it still analyses candidates' speech. Our reading: this is the right approach. Ban what lacks a scientific basis and demand evidence for what has one. That is regulating evidence, not fear.

In the long run, the outlook is hopeful. A camera and a microphone cost almost nothing and sit in every pocket. If these biomarkers are validated rigorously and across diverse populations, they could cut years off the search for a rare diagnosis. They could also help catch Parkinson's earlier, match the intensity of a treatment to what each body can take, and bring triage to places without specialists. That is part of the promise of cheaper, earlier and more abundant medicine. The condition is not to forget Lombroso: your face can tell how your body is doing, but never who you are.