Physiognomy

Greek and Roman treatises

The historian Kevin van Bladel defines physiognomy as the science of assessing personal character through the observation of external physical features. In a review of Simon Swain's 2007 edited volume on Polemon, he writes that two influential ancient Greek treatises are extant: one attributed falsely to Aristotle and one by the scholar and politician Polemon, who lived about AD 88 to 144. Polemon's work is lost in its original form and survives in abridgments in different languages. The points below are the review's account of that scholarship, not a first-hand reading of the treatises.

According to the review, Swain places Polemon among educated men seeking the patronage of Roman emperors, who faced intense personal scrutiny while delivering speeches, and reads Polemon's system as an instrument of that scrutiny. The reviewer's own impression is of a work by someone keen to use claimed expertise to make allies and shame enemies. The review adds that Jas Elsner found it hard to establish links between the treatises and Roman portrait sculpture.

Arabic reception

The review reports that, on the testimony of Arabic authors, Polemon's text had been translated into Arabic by the middle of the ninth century, and that the Arabic word for physiognomy is firasa. Most Arabic authors on the subject were physicians. Avicenna (died 1037) ranked physiognomy with medicine in the second rank of the natural sciences, since both involve reading signs in the body, and the pseudo-Aristotelian treatise, in its Arabic translation by Hunayn ibn Ishaq (died 873), appears to have been more influential in this medical setting than Polemon. The reviewer notes that how far Arabic firasa continued Greek physiognomy remains unresolved.

Printed treatises and Lavater

The Wellcome Collection catalogs a 1586 folio edition of Giambattista della Porta's De humana physiognomonia libri IIII, printed at Vico Equense with engraved illustrations and portraits. The record does not describe the book's argument, and the sources reviewed for this page do not set out della Porta's methods.

Historian Melissa Percival, in a book on physiognomy in eighteenth-century France, writes that a study of the period can begin in 1775, when Johann Caspar Lavater published the first volume of his Physiognomische Fragmente. She cites Georg Christoph Lichtenberg, who called the public response a 'Raserei für Physiognomik', a physiognomical frenzy, that lasted well into the following century. Thomas Holcroft translated the work into English as Essays on Physiognomy; the Wellcome Collection holds a third edition printed in London in 1840. The sources reviewed do not set out the content of Lichtenberg's criticism.

Physiognomy and neighboring terms

In the historical sources reviewed, physiognomy means the inference of character from external features, above all of the face. The Arabic word firasa is used for it, although the reviewer of Swain's volume notes that the term refers fundamentally to discrimination through scrutiny and that the relation between firasa and Greek physiognomy is an open question. Face Reading is the family page for this area; this page covers the Western lineage and the research literature that has grown up around it.

Phrenology and criminal anthropology are related but distinct. Marc Renneville's abstract notes that Gall, the founder of phrenology, rejected Lavater's ideas, even though Gall's followers often combined study of the skull with face reading. Lombroso's born-criminal theory belongs to the criminal anthropology tradition that Renneville describes as drawing on physiognomic signs.

Modern usage

Psychologists describe work on how people judge traits from faces as research on first impressions or social attributions. Bowyer and colleagues point out that this research concerns what the average observer would say about a face, a consensus about traits such as trustworthiness and dominance, and that agreement among observers does not show the judgments are true. They cite a preface by Alexander Todorov stating that modern claims of a 'new physiognomy' are nearly as exaggerated as those of the eighteenth and nineteenth centuries. Critics also apply the word physiognomy to machine-learning systems that infer traits from face images.

Historical method

The sources reviewed describe physiognomy as the observation of external features as signs of character. Swain's reading of Polemon treats it as a tool for judging public speakers, and Elsner's chapter, as summarized in the review, notes a discussion of Polemon's terminology for eye colors. In the Arabic medical setting, Ghersetti is reported as describing a shared semiotic procedure: the observation of signs in the body and inference from them. The sources reviewed do not list the specific feature-to-trait rules of the treatises, so this page does not reproduce them.

Computational form

In a 2018 study, Wang and Kosinski used deep neural networks to extract features from 35,326 facial images, entered the features into a logistic regression, and classified sexual orientation; the features included fixed ones such as nose shape and transient ones such as grooming style. Bowyer and colleagues describe the 'criminality from face' concept in similar terms: images labeled Criminal or Non-Criminal train an algorithm, whose accuracy is then estimated on a held-out subset of the same data.

The sources reviewed describe no current practitioner-led service, register or training standard for physiognomy, and no statutory regulation or training standard was found in them.

Historical claims

Swain's reading of Polemon casts physiognomy as a way for elite speakers to be assessed and for rivals to be discredited. In the Arabic literature it was classed beside medicine as sign-reading, and in the nineteenth century, per Renneville, it served as a basis for criminal typologies. These are descriptions of claimed uses in the sources, not findings that the methods worked.

Modern claims

Bowyer and colleagues quote recent papers reporting accuracy as high as 97% for labeling faces Criminal or Non-Criminal, and a press release stating that software with 80 percent accuracy could help law enforcement prevent crime; the release was later removed at the request of the faculty involved. Wang and Kosinski framed their sexual-orientation results as exposing a threat to privacy and safety rather than as a service. Olivola and Todorov review research indicating that voters rely heavily on candidates' appearance, a finding about voting behavior and not a claim that appearance reveals competence.

First impressions of faces

Todorov and colleagues reported in 2005 that inferences of competence made from candidates' faces alone predicted the outcomes of U.S. congressional elections better than chance, for example 68.8% of 2004 Senate races, and related linearly to the margin of victory after exposure of about one second. Olivola and Todorov's 2010 review reanalyzed earlier data and described facial competence as a robust and specific predictor of political preferences. These studies measure what observers infer and how that affects votes; they do not test whether the inferences are accurate.

The 2015 Annual Review of Psychology article by Todorov, Olivola, Dotsch and Mende-Siedlecki argues that the diagnostic validity of face-based attributions has been greatly overstated. Bowyer and colleagues make the same distinction, stating that consensus about a face does not imply the judgment is true.

Facial width-to-height ratio

A 2015 meta-analysis by Haselhuhn and colleagues covered 14 articles, 19 studies and 4,327 men, and found a small positive correlation between facial width-to-height ratio and aggression (r = .11, 95% CI .08 to .14). The effect was larger in laboratory studies (r = .21, N = 603) than in field or archival studies (r = .09, N = 3,538), a funnel plot suggested small null studies may be missing, and the authors caution against predicting behavior from faces. Geniole and colleagues' 2015 meta-analysis of 56 manuscripts found threat behavior in men related at r = .16, perceived threat at r = .46, and noted that the data are correlational and that body mass index may explain part of the link.

Kosinski's 2017 study of 137,163 participants and 55 psychometric scales found the ratio was not substantially related to any self-reported behavioral tendency. Those measures are self-reports, not observed aggression, so the study and the meta-analyses address different outcomes.

Machine-learning claims and critiques

Wang and Kosinski reported that, from one image, their classifier separated gay and heterosexual men in 81% of cases and women in 71%, against 61% and 54% for human judges, rising to 91% and 83% with five images; grooming-style features were among those used. Leuner's replication on 20,910 dating-site photos reported 68% for men and 77% for women and said it could not separate facial morphology from grooming and presentation. Gelman, Mattson and Simpson criticize the underlying measurement as decontextualized and falsely precise.

Bowyer and colleagues argue that high accuracy in criminality studies comes from extraneous differences between image sets, such as mugshots and ID photos from different sources. In their illustration, a simple classifier trained to separate mugshots from a different public face dataset reached 95.86% accuracy, and applied to a third dataset of presumably non-convicted people it labeled 63% as criminal.

Offshoots and popular culture

Renneville describes physiognomy as central to nineteenth-century efforts to build a science of the criminal, although disputed, with French physicians such as Lacassagne, Topinard and Manouvrier proposing a criminal anthropology built on professional type and degeneration. He reports that the search for visual signs of criminality spread through popular books, novels by Balzac and Zola, crime journalism and silent film, and that writers such as Victor Hugo and famous trials were also sources of rejection.

Decline as science

In Renneville's account, Lombroso's born-criminal theory marked the peak of visual criminology and also signaled its decline in scientific circles, a decline beginning around the eve of World War I that he attributes more to social and disciplinary change than to decisive scientific proof. Some criminologists moved the search for physical difference to the brain and genes, which he calls an internal physiognomy. Bowyer and colleagues state that the idea of a born criminal measurable from the face is regarded today as having no scientific foundation.

Documented harms

Buolamwini and Gebru tested three commercial gender classification systems on a dataset balanced by gender and skin type. Darker-skinned women had error rates up to 34.7%, compared with a maximum of 0.8% for lighter-skinned men, and two common benchmarks were 79.6% and 86.2% lighter-skinned. This concerns gender classification, not trait inference, and should not be transferred to other tasks, but it shows how facial analysis performance can differ by group.

Bowyer and colleagues warn of a large social cost if people come to believe in criminality from face, and note that a mugshot records an arrest, not a conviction. Wang and Kosinski and Leuner both state that tools to infer sexual orientation from faces threaten the privacy and safety of gay men and women. The reviewer of Swain's volume describes ancient physiognomy as a tool for shaming enemies.

Limits of the sources and care boundary

The sources reviewed do not quantify harms from historical racial or criminal pseudoscience, and they contain no trial data for physiognomy as a health or counseling practice. Nothing reviewed supports using facial appearance as a basis for judgments about an individual's character, health or conduct, and facial appearance should not delay or replace assessment by an appropriate professional.