Can you read personality from a face? What studies show

“Narrow, sharp eyes mean a cold person; a wide jaw means a stubborn one”: you have probably heard face-reading lines like these. mind.io’s face reading page also measures the proportions and left-right balance of the face in a photo and turns them into light-hearted lines from traditional East Asian face reading, but from the start it says that a face cannot tell your personality or fortune and asks you to enjoy it as a game. So is there any scientific basis for a link between face and personality? This guide answers that question with research.

The short answer is that the claim at the heart of face reading, physiognomy in the Western tradition, that the shape of your features reveals your character and fate, has largely been discredited among researchers (Todorov & Porter, 2014). That is not the end of the story, though. People form an impression from a face within a tenth of a second, and those impressions are fairly similar from one person to the next (Willis & Todorov, 2006; Todorov et al., 2015). Some studies find that a few traits, such as extraversion, can be guessed from photos better than chance (Penton-Voak et al., 2006; Naumann et al., 2009), though the size of that edge has generally been small (Kachur et al., 2020; Todorov, 2018). In a study using full-body photos, cues such as expression, posture and clothing carried personality information (Naumann et al., 2009), and even in a study using face photos only, cues such as makeup, angle and subtle expressions could not be separated from the effect of the face’s built-in structure (Kachur et al., 2020). And when guessing characteristics such as whether someone drinks or which party they support, with other information also available, relying on appearance even made people’s guesses worse (Olivola & Todorov, 2010).

This guide covers where face reading and Western physiognomy came from, how fast and how alike first impressions from faces are, when those impressions match real personality and where that stops, what goes wrong when faces are used to pick people, the controversy over AI face reading, and how to enjoy it without taking it as fact. The research on blood type, on zodiac signs and on palm reading is in the guide to the science of blood type personality, the guide to the science of zodiac personality and the guide to the science of palm reading. mind.io does not offer a test that judges personality from a face.

Attempts to read a person’s character from the face are very old. Alexander Todorov, a psychology professor at Princeton when he wrote in 2018, notes that the first preserved documents referring to it date back to Aristotle’s time (Todorov, 2018). A 1586 book by Giambattista della Porta, published in Naples, set human and animal faces side by side to argue that a pig-like face goes with a pig-like character, and in the 1700s the Swiss theologian Johann Caspar Lavater analysed character from the shape and position of the eyes, brows, mouth and nose (Agüera y Arcas et al., 2017).

Even Charles Darwin nearly fell foul of it. The captain of the Beagle, a fervent physiognomist, doubted that a man with such a nose would have enough determination for the journey, and Darwin almost missed his historic voyage. In his autobiography, Darwin noted dryly: “But I think he was afterwards well-satisfied that my nose had spoken falsely” (Todorov, 2018).

In the 19th century the idea put on scientific clothes. In Italy, Cesare Lombroso measured physical traits such as criminals’ skulls and an “asymmetric face” and claimed that criminals were “born criminals” whose nature showed in their bodies, while in England Francis Galton superimposed photographs of convicts on one plate in search of a “criminal face”. The researchers who reviewed this history describe Lombroso as an early case of using measurement and scientific language to lend authority to his own prejudices, and note that physiognomy is today understood to be pseudoscience (Agüera y Arcas et al., 2017).

East Asian face reading rested on a similar premise, reading even a person’s future in the face. According to a paper by Seo Young Yi of Chonnam National University, classical face reading judged a person’s rank and fortune from the visible features alone, as if they were fixed. The late-Joseon scholar Choi Han-ki embraced it as one way of reading people, but pointed out that it missed what a person actually does (haengsa), which can differ from their inborn looks, and the unpredictable mind, and argued that conduct should be checked alongside it (Seo, 2015). Doubts about judging people from the face alone came even from those who accepted face reading.

Modern psychology asks the face-reading question a little differently. Before asking whether a face can reveal personality, it looks at what impressions people form from a face alone and how fast. In five experiments published in 2006, people who saw an unfamiliar face for a tenth of a second (100 milliseconds) made judgments of attractiveness, likeability, trustworthiness, competence and aggressiveness that correlated highly with the judgments of other people who looked without a time limit. Looking for longer did not significantly raise those correlations, and the researchers suggested that extra time may simply boost confidence in the judgment (Willis & Todorov, 2006).

That impressions from faces come out similar from one person to the next has been known since the early 20th century (Todorov et al., 2015). The tendency shows up very early: in one study, when 3- to 4-year-olds looked at two computer-generated faces and picked which one was nice and which was mean, their choices converged with adults’ choices (Cogsdill et al., 2014). What these studies measured, though, is how much people agree with each other, not whether the impressions match real personality.

Why are impressions so fast and so alike? Researchers explain it as overgeneralisation, reading a passing state as a lasting trait. Just as we take a smiling face to be a friendly person, people stretch a momentary state, such as the expression they can see right now, into a stable personality (Todorov & Porter, 2014). Even for so-called neutral faces, impressions such as trustworthiness depend on how much the face resembles an emotional expression, and researchers see this as our ability to read someone’s intentions in the moment spilling over into a judgment of what that person is “like” in general (Agüera y Arcas et al., 2017).

That is one reason face reading can feel accurate. An impression many people share easily feels like an objective fact, and expressions do carry something real about a person’s mood or intention at that moment. Researchers stress, however, that agreement in judgments of faces does not by itself mean those judgments are accurate (Todorov & Porter, 2014). The belief is not rare in Korea either: in a 2026 Gallup Korea survey, 40% of adults said they believe in fortune-telling, saju, face reading, name divination and the like (3% “mostly”, 37% “somewhat”) (Gallup Korea, 2026). The question asked about face reading together with fortune-telling, saju and name divination, not on its own.

So do impressions ever match real personality? Some studies report results better than chance. A 2006 study photographed the faces of 146 men and 148 women, each of whom also completed a Big Five personality questionnaire. Perceived extraversion in the faces was related to the extraversion people reported themselves, and for male faces there was some accuracy for emotional stability and openness too. When the faces of high and low scorers were blended into composite faces, the composites could be told apart on agreeableness, extraversion and, for male faces only, emotional stability. The authors’ title calls this a “kernel of truth” in impressions from faces (Penton-Voak et al., 2006).

A 2009 study used full-body photographs rather than faces alone: observers judged 10 personality traits from the photos, and their judgments were compared with ratings from the people themselves and from people who knew them. When posture and expression were held constant, the judgments were accurate only for extraversion, self-esteem and religiosity, but when people were photographed with a spontaneous pose and expression, they were accurate for almost all of the traits. Cues such as clothing style and cues such as facial expression and posture both carried information about personality (Naumann et al., 2009).

Artificial intelligence has produced similar results. A 2020 study trained and validated neural networks on 31,367 photos from 12,447 Russian-speaking adult volunteers, together with their Big Five test results. Participants were asked to take front-facing photos with a neutral expression and no makeup, and the photos were screened by angle, expression and makeup. The correlations between the scores predicted from photos and the scores people reported ranged from 0.14 to 0.36, with a mean of 0.243. Put the way the researchers did, that means correctly guessing which of two randomly chosen people scores higher 58% of the time, where chance would be 50%. The researchers took this as evidence that all five Big Five traits are associated with facial cues and claimed that the network does better than an average person judging a stranger met face to face. They also called the correlations modest and said that personality assessment from facial images can hardly replace traditional personality tests (Kachur et al., 2020).

Todorov points out that in studies claiming to read various characteristics from face photos, the guesses are barely better than chance and often less accurate than guesses based on more general knowledge (Todorov, 2018). It also matters where the cues came from. In the full-body photo study, allowing a natural expression and pose greatly increased the number of traits judged accurately (Naumann et al., 2009). The AI researchers tried to limit such cues through their instructions and screening, but stated that the photos participants chose themselves could still carry cues such as makeup, angle, subtle expressions and the choices people make in taking and picking photos, which cannot be separated from the effect of the face’s built-in structure. The participants were also adult volunteers from one Russian-speaking cultural group, and three of the authors were employed by the company that provided the data (Kachur et al., 2020). Even if a photo of a face carries faint traces of personality, that does not support the face-reading rule that fixed features reveal character and fate.

The idea that one photo represents a person has a trap in it too. A 2014 study took five frontal photos each of 20 people from a face-recognition image database and had 800 participants each judge one of eight attributes, including attractiveness, competence, extraversion and trustworthiness. Although the photos were chosen to have similar lighting and face orientation, for most judgments other than attractiveness, the differences between photos of the same person were as large as or larger than the differences between people. Within the same photo set, raters’ extraversion judgments were consistent (reliability of 0.88 or more), yet the correlation between judgments of the same people made from different photos averaged only 0.20 and was as low as 0.03 (Todorov & Porter, 2014).

Relying on looks can even backfire. A 2010 study analysed data from a website where users posted photos and strangers guessed facts about them. The 1,005,406 guesses about 901 people did better than chance on all 11 characteristics, such as whether the person drinks, owns a gun or has a college degree. These are not personality traits, but they are still guesses about a person from a photo. Yet on 10 of the 11, the guesses were less accurate than simply always picking the more common answer, and survey respondents who never saw the photos would have done better on 7 by answering only from their own sense of which answer is more common (Olivola & Todorov, 2010).

In a second experiment in the same paper, 1,005 people guessed whether candidates in US House of Representatives elections were Democrats or Republicans from their facial photos alone. When half the candidates were from each party, they were right 55% of the time, slightly better than chance (50%). But when told that 70% of the candidates were Democrats, they were still right only 55% of the time on average. Answering “Democrat” every time would have got 70% right; they gave the photos too much weight and used that information too little. The researchers concluded that in real situations, where several cues are available, relying on appearance may make us worse at predicting other people’s characteristics (Olivola & Todorov, 2010).

The same held when money was at stake. In a 2022 experiment with students at the University of Zurich in Switzerland, 131 people playing a trust game with real money decided whether to entrust money to 31 people seen only in neutral-expression photos. Those who saw the original photos entrusted money more often to faces that looked trustworthy, but their choices were unrelated to how much those people actually returned. They earned no more than they would have by deciding at random, and in this sample trusting no one would have earned more. A second experiment, in which 266 people bet money on predicting the others’ behaviour, found the same with photos cropped to the face and with uncropped photos. The researchers noted the limits of a young student sample and added that results may differ when more cues are available, such as after a brief interaction (Jaeger et al., 2022). How well people spot lies when they can watch expressions and body language and hear what is said, rather than look at a photo, is covered separately in the guide on spotting lies.

Impressions from faces change outcomes even when they are not accurate. Todorov summarises that people are more likely to vote for politicians who look competent, to invest in people who look trustworthy, and to give harsher prison sentences to people who look the opposite (Todorov, 2018). Reviewing studies showing that impressions from faces predict important outcomes such as investment decisions and leader selection, his team argued that claims about the accuracy of those impressions have been greatly overstated in the research literature (Todorov et al., 2015).

Recently there have also been attempts to hand such judgments to artificial intelligence. In 2016, two computer scientists posted a paper online that had not been peer reviewed, claiming that an algorithm trained on ID photos of 1,856 Chinese men could pick out people with criminal convictions (Wu & Zhang, 2016; Agüera y Arcas et al., 2017). The authors said they had controlled for facial expressions, among other things (Wu & Zhang, 2016), but according to the researchers who critiqued it, in the sample photos shown by this paper, which claimed nearly 90% accuracy, the “criminals” were frowning while the “non-criminals” were not and appeared to wear white collared shirts. With only three examples of each, it is unclear whether these are typical, but a powerful algorithm easily picks up differences like these (Agüera y Arcas et al., 2017).

Todorov explains that seemingly trivial differences, such as the distance between camera and head, camera settings, a slight tilt of the head and subtle expressions, all disturb the measurement of what are meant to be stable facial features. When these differences are not controlled for, AI studies simply amplify human biases (Todorov, 2018). The critics pointed out that the paper’s language echoes Lombroso’s and warned that prejudice passed through an algorithm can look like an objective result (Agüera y Arcas et al., 2017).

So whether it is a human eye or an app, using a face alone to pick or rule out a date, a job candidate or a colleague can turn a weakly grounded impression into an unfair label. In the trust game above, too, choices based only on faces did no better than choosing at random (Jaeger et al., 2022).

Face reading and AI face reading make handy conversation topics, for passing photos around and laughing with friends or for breaking the ice with someone new. What the research shows is that a face alone is a poor guide to personality, not that you should never talk about it. A few habits make it much easier to enjoy.

First, try it with several photos of yourself. mind.io’s face reading page itself explains that angle, light and expression change the measured proportions, so two photos of the same person can read differently and no single photo is the “right” one. It is the easiest way to see for yourself the finding that impressions of the same person change a lot from photo to photo (Todorov & Porter, 2014).

Second, turn impressions into questions. Instead of “You looked so cold I didn’t dare talk to you”, ask “What are you like around people you’ve just met?”, and the same topic helps you actually get to know the person. Third, laugh off the readings that do not fit, and never use them to pick or rule out a date or a colleague. Treat a first impression as a guess, and revise it as you get to know someone.

Fourth, on days when you want to know yourself better, try a test that measures your current tendencies from your answers. The Big Five personality test measures the five traits two of the studies above used as their yardstick, and the MBTI test shows a type built from four preferences. Unlike a face, both give results from the items you answer. How far you can trust the MBTI is covered in the guide to MBTI accuracy.

Q. Is there any scientific evidence for face reading? A. The face-reading rule that the shape of your features determines your character and fate has largely been discredited among researchers (Todorov & Porter, 2014). There are, however, studies in which impressions from faces were better than chance for some traits, such as extraversion (Penton-Voak et al., 2006; Naumann et al., 2009). The edge has generally been small (Kachur et al., 2020; Todorov, 2018) and mixed in with cues in the photos such as expression, posture, clothing, makeup and angle (Naumann et al., 2009; Kachur et al., 2020). When guessing characteristics such as whether someone drinks or which party they support, with other information available, relying on appearance even made guesses worse (Olivola & Todorov, 2010).

Q. Can first impressions be trusted? A. First impressions form within a tenth of a second and are similar from one person to the next (Willis & Todorov, 2006), but being similar does not make them accurate. In an experiment where people guessed party affiliation from facial photos alone, they were right 55% of the time, only slightly better than chance (50%) (Olivola & Todorov, 2010), and in a trust game with money at stake, people could not pick out trustworthy partners from faces alone (Jaeger et al., 2022). Treat a first impression as a guess and revise it with experience.

Q. Is AI face reading more accurate? A. In a 2020 study where artificial intelligence predicted Big Five scores from face photos, the correlation between predicted and self-reported scores averaged 0.243. The researchers claimed their network beats an average person judging a stranger met face to face, but also said it could hardly replace personality tests. Despite instructions and screening, cues in the photos such as makeup, angle and subtle expressions could still remain, so they could not be separated from the effect of facial structure (Kachur et al., 2020). Critics of the claim that faces can reveal criminals pointed out that the algorithm may have picked up differences in expression and clothing in the photos (Agüera y Arcas et al., 2017). mind.io’s face reading page measures facial proportions to give a reading for fun; it does not judge personality or fortune.

Q. Can your impression change? A. Yes. The impression the same person gives can change a lot from photo to photo: in one study, for most judgments other than attractiveness, the differences between photos of the same person were as large as or larger than the differences between people (Todorov & Porter, 2014). Impressions such as trustworthiness depend on how much a face resembles an emotional expression, and computer-generated “trustworthy” faces had more positive expressions than “untrustworthy” ones (Agüera y Arcas et al., 2017). A changed impression, though, does not mean a changed personality.

Q. How is face reading different from blood type or zodiac signs? A. Unlike your blood type or birthday, how you appear in a photo carries information you show in that moment, such as expression, posture and clothing, and some studies found judgments from photos to be better than chance (Naumann et al., 2009). But the idea that fixed features, blood type or birthday decide your personality and fate is not supported by research for any of the three. The research on blood type and zodiac signs is in the guide to the science of blood type personality and the guide to the science of zodiac personality.

Q. Is there a test that reads personality from your face? A. mind.io does not offer a test that judges personality from a face. The face reading page is a game that measures the proportions and left-right balance of the face in a photo and describes them in the language of traditional face reading, and as the page explains, photos are analysed only in your browser and are never uploaded or stored on a server. If you want to explore your tendencies from your own answers, try the Big Five personality test.

References: Todorov, A., & Porter, J. M. (2014), Misleading first impressions: different for different facial images of the same person, Psychological Science 25(7), 1404-1417, https://doi.org/10.1177/0956797614532474; Agüera y Arcas, B., Mitchell, M., & Todorov, A. (2017), Physiognomy’s new clothes, Medium, 7 May 2017, archived at https://web.archive.org/web/20260930013957/https://medium.com/@blaisea/physiognomys-new-clothes-f2d4b59fdd6a; Todorov, A. (2018), Can we read a person’s character from facial images?, Scientific American, 14 May 2018, https://www.scientificamerican.com/article/can-we-read-a-persons-character-from-facial-images/; Seo, Y. Y. (2015), The experiential evolution of physiognomy: focusing on Choi Hanki’s human assessment [in Korean], Korean Studies Quarterly 38(4), 39-62, https://doi.org/10.25024/ksq.38.4.201512.39; Willis, J., & Todorov, A. (2006), First impressions: making up your mind after a 100-ms exposure to a face, Psychological Science 17(7), 592-598, https://doi.org/10.1111/j.1467-9280.2006.01750.x; Todorov, A., Olivola, C. Y., Dotsch, R., & Mende-Siedlecki, P. (2015), Social attributions from faces: determinants, consequences, accuracy, and functional significance, Annual Review of Psychology 66, 519-545, https://doi.org/10.1146/annurev-psych-113011-143831; Cogsdill, E. J., Todorov, A. T., Spelke, E. S., & Banaji, M. R. (2014), Inferring character from faces: a developmental study, Psychological Science 25(5), 1132-1139, https://doi.org/10.1177/0956797614523297; Gallup Korea (2026), Do you believe in fortune-telling and saju? 1991-2026 [in Korean], Gallup Report, 30 April 2026, https://www.gallup.co.kr/gallupdb/reportContent.asp?seqNo=1638; Penton-Voak, I. S., Pound, N., Little, A. C., & Perrett, D. I. (2006), Personality judgments from natural and composite facial images: more evidence for a “kernel of truth” in social perception, Social Cognition 24(5), 607-640, https://doi.org/10.1521/soco.2006.24.5.607; Naumann, L. P., Vazire, S., Rentfrow, P. J., & Gosling, S. D. (2009), Personality judgments based on physical appearance, Personality and Social Psychology Bulletin 35(12), 1661-1671, https://doi.org/10.1177/0146167209346309; Kachur, A., Osin, E., Davydov, D., Shutilov, K., & Novokshonov, A. (2020), Assessing the Big Five personality traits using real-life static facial images, Scientific Reports 10, 8487, https://doi.org/10.1038/s41598-020-65358-6; Olivola, C. Y., & Todorov, A. (2010), Fooled by first impressions? Reexamining the diagnostic value of appearance-based inferences, Journal of Experimental Social Psychology 46(2), 315-324, https://doi.org/10.1016/j.jesp.2009.12.002; Jaeger, B., Oud, B., Williams, T., Krumhuber, E. G., Fehr, E., & Engelmann, J. B. (2022), Can people detect the trustworthiness of strangers based on their facial appearance?, Evolution and Human Behavior 43(4), 296-303, https://doi.org/10.1016/j.evolhumbehav.2022.04.004; Wu, X., & Zhang, X. (2016), Automated inference on criminality using face images, arXiv preprint 1611.04135v1, not peer reviewed, https://arxiv.org/abs/1611.04135v1