TL;DR: When an app reads a selfie and returns a personality profile, the available science shows only a small group-level signal: a neural network in Kachur and Osin's 2020 study predicted self-reported Big Five scores somewhat above chance, with modest correlations, the highest around 0.36 for conscientiousness. That signal is not strong enough to predict any one person's personality, let alone fate or the rules of East Asian face reading. No study has been done on Vietnamese people, and none has tested the facial positions of traditional face reading.
What Is an AI Face-Reading App Actually Doing?
When an app reads a selfie and returns a personality profile, technically it is usually running a machine learning model trained on a labelled image set. Those labels may be personality scores the people in the photos reported themselves, scores rated by others, or some label the developer invented. What the app returns is not "reading" the face in the face-reading sense, but an estimated number from the image sample, based on what the model learned from its training data.
The central question of this article is narrower than "is face reading true". Here we only ask: when machine learning works on real photos, how far can it predict? The studies cited below measure on real photographs, mostly abroad, and measure personality using the Big Five model, not fate or the facial positions of traditional face reading. Their results therefore cannot be used to affirm or deny the rules of East Asian face reading; they only speak to the limits of predicting personality from images.
What Does the Largest Study Say About Accuracy?
The largest study in this group is by Alexander Kachur and Evgeny Osin with colleagues, published in 2020 in Scientific Reports. The team recruited 12,447 volunteers, with a total of 31,367 face photos, and each person self-reported their Big Five scores[1]. They then trained a chain of artificial neural networks to predict self-reported scores from the images.
The main result: the model predicted personality traits above chance in a statistically significant way, for both men and women. The highest correlation fell on conscientiousness, around 0.360 for men and 0.335 for women; the average effect size was 0.243. The authors also noted that this result was higher than earlier studies using selfie photos. This was the first time a study achieved significant prediction for the openness to experience trait.
This number needs to be read correctly. A correlation of 0.36 does not mean that looking at a photo tells you whether that person is conscientious. In the study, predictions were tested on a validation set of 3,137 photos from 1,245 individuals. The authors also state an important limitation: their sample was mainly white, Russian-speaking, from one cultural and age group. That means the result has not been tested on Vietnamese people or on other ethnic, cultural, or age groups. In addition, this is a single study at evidence level B: it shows a real phenomenon in that sample, but it has not been widely replicated to become a firm conclusion.
One technical detail worth noting: the authors used only photos with neutral or near-neutral expression, and they explained that weaker effects for extraversion and neuroticism may be because these two traits are tied to positive and negative emotional experience, while the photos used for training had expression filtered out. The team also remarked that real-life photos, especially photos participants chose themselves, are a whole behaviour that may contain many cues beyond static facial features, such as lighting, hairstyle, head angle, and image quality. This is a very important point when judging a commercial app: the model may be learning from the photo context, not only from the face.
What Do Composite-Photo Experiments Show?
Before the 2020 study, Anthony C. Little and David I. Perrett published a composite-photo experiment in 2007 in the British Journal of Psychology. The method was to composite photos of people scoring high and people scoring low on each personality trait, then have participants guess the trait from the composite. The summary result: raters guessed above chance, most clearly for conscientiousness and extraversion; attractiveness, masculinity, and age were mediating cues.
The point to stress: this is guessing on a composite of a whole group, not on individuals. The composite removed or held constant many individual features, so the task is much easier than looking at one real face and predicting that person's personality. Kachur and Osin also noted that studies using composite photos tend to give stronger effect sizes, while studies using real individual photos tend to give more modest effects. So if an app advertises based on "composite photos predicted correctly" results, one must understand that these are group-level results, not individual-level.
Where Does First Impression from a Face Go Wrong, and How Dangerous Is It?
A 2015 review by Alexander Todorov and colleagues in the Annual Review of Psychology shows that people infer personality from faces very quickly and many people agree with one another, but accuracy is low. These impressions still drive real decisions such as elections, hiring, and sentencing[2].
This means the feeling that "you can know a person by looking at their face" is very strong, but strong does not mean correct. When an AI app returns a result that looks plausible, users easily believe it because it matches a pre-existing impression. But according to Todorov, that pre-existing impression is precisely what is often wrong and can lead to wrong decisions about others. This is why researchers stress that predicting personality from faces should not be used as a basis for important decisions about a specific person.
What Does the Meta-Analysis on Trustworthiness Say?
Y. Z. Foo and colleagues in 2021 published in Personality and Social Psychology Bulletin two meta-analyses on accuracy in judging trustworthiness from faces. The summary result: there is accuracy but it is modest, with correlations around 0.14 at the face level and 0.27 at the rater level, and it depends on the behavioural domain[3]. The authors asked whether this is a "kernel of truth" or "modern phrenology", and concluded against reading personality confidently from faces.
This is evidence level A, meaning a synthesis of many studies, so it carries more weight than a single study. The numbers 0.14 and 0.27 show a real but very small signal: enough to say that faces carry a little information about how others perceive someone, but not enough to conclude about an individual's true personality.
Comparison Table of the Studies
| Study | Sample size / data scope | Traits predicted | Correlation level | What it does not measure |
|---|---|---|---|---|
| Kachur and Osin et al. (2020), Scientific Reports | 12,447 people, 31,367 photos; white Russian speakers; validation set of 3,137 photos from 1,245 people | All five Big Five traits, above chance; strongest is conscientiousness | Highest around 0.360 (men) and 0.335 (women) for conscientiousness; average effect size 0.243 | Does not measure fate, does not test face-reading rules, does not measure on Vietnamese people |
| Little and Perrett (2007), British Journal of Psychology | Composite photos of high-scoring and low-scoring groups | Conscientiousness and extraversion most clearly; mediating cues: attractiveness, masculinity, age | Above chance | Guessing on group composites, not individuals |
| Todorov et al. (2015), Annual Review of Psychology | Systematic review | Social attributions from faces: fast, high consensus, low accuracy | Does not measure fate; stresses impressions drive real decisions | |
| Foo et al. (2021), PSPB | Two meta-analyses | Trustworthiness from faces | Around 0.14 at face level, 0.27 at rater level | Depends on behavioural domain; does not support reading personality confidently from faces |
Why Is There Still a Small Signal?
Kachur and Osin and colleagues give four groups of possible reasons for the link between face photos and personality. First, some personality traits are tied to frequent emotional expressive behaviour, and frequent expression may shape static facial features, creating wrinkles or developing facial muscles. Second, appearance may influence how others treat a person, which in turn influences personality development, in a self-fulfilling or self-defeating prophecy pattern. Third, facial morphological features may be related to some personality traits. Fourth, real-life photos also contain cues beyond the face, such as lighting, hairstyle, and camera angle.
Because there are so many pathways, the small signal in the research should not be read as "the face determines personality". It only says that in a specific dataset, there is a weak link between photos and self-reported scores. That weak link does not allow predicting a specific person correctly, and even less does it allow inferring fate.
What This Article Does Not Say
This article does not evaluate any specific face-reading app, because there is no independent validation data on those apps. This article also does not test the facial positions of East Asian face reading such as forehead, eyes, nose, chin, or rules like three courts and five mountains. The studies cited here measure personality using the Big Five model, a Western framework, and do not measure face-reading concepts. Their results therefore cannot be used to affirm whether face reading is true or false; they only speak to the limits of predicting personality from images by machine learning.
There is currently no research on Vietnamese people about predicting personality from face photos, and no research testing East Asian face-reading rules by machine learning methods. This is an evidence gap that needs to be stated clearly, rather than broadly extrapolating from foreign results to Vietnamese people.
The Effect Is Real but Born from Belief
In folk belief, many people believe that looking at a face can predict personality, and some practitioners of face reading rely on that belief. This belief can produce a real effect through another path: when people believe a description about themselves, they may change their behaviour, and changed behaviour can lead to different outcomes. This is a real effect, but it is born from belief and behaviour, not from the system "predicting correctly" from the face. In other words, two things must be distinguished: the system predicting correctly, and the believer changing behaviour. Only the second has a clear mechanism, and it does not turn the system into an accurate measurement tool.
The right way for app users to read it: treat the result as entertainment, not as a basis for decisions about a specific person's work, love, or money. For developers, the worthwhile thing is to publish the sample size, data scope, and true correlation level, instead of only saying "AI predicts accurately".
Link to an Existing Article
The article What Science Says About Face Reading presented the foundation on Todorov and Foo, that is, the reviews on the accuracy of impressions from faces. This article adds the machine learning angle on photos, focusing on the question of AI apps and the size of the signal in studies using real photos.
