Face Value: AI Revives the Discredited “Science” of Physiognomy

In a small London startup office, an experiment is underway that echoes a long-abandoned Victorian belief. A scientist Natalia Segovantseva feeds thousands of human portrait photos into a computer, training a neural network to read personality from a face. The goal sounds like something out of a 19th-century parlour trick – determining if a stranger is kind, clever, or even criminal just by analyzing their features. Once dismissed as quackery, the ancient idea of physiognomy is making a controversial comeback in the age of artificial intelligence. Modern algorithms are doing what was once the province of mystics and charlatans, claiming to judge our character from the contours of a jawline or the arch of an eyebrow. It’s an unsettling revival: can a machine truly discern the soul behind a face, or are we repeating the mistakes of a pseudoscience long thought buried?

From Ancient Faces to Pseudoscience

Physiognomy – the practice of inferring personality from appearance – dates back millennia. In ancient Greece, philosophers like Aristotle speculated that facial traits mirrored inner character. Aristotle wrote that large-headed people were “mean,” while those with small faces were “steadfast,” broad faces signaled stupidity and round faces courage . A colorful legend recounts how a physiognomist examined the famously wise Socrates and pronounced him prone to “intemperance, sensuality, and violent bursts of passion” – shocking his students, who saw Socrates as the model of virtue . Socrates only smiled, admitting that he had indeed been naturally inclined to all those vices, but had trained himself to overcome them . In other words, even antiquity’s greatest mind couldn’t fully escape a snap judgment based on his looks.

After flourishing in Greek and Roman thought (and surfacing independently in Chinese and Indian traditions), physiognomy fell in and out of favor over the centuries. The Renaissance brought a resurgence: in the 1500s, Italian scholar Giambattista della Porta – often dubbed the father of physiognomy – tried to give the practice scholarly legitimacy . Della Porta’s influential 1586 book De Humana Physiognomia even paired illustrations of human and animal heads, implying that a person who resembled a lion might share the lion’s bravery or ferocity .

Comparative physiognomy: A 17th-century illustration by Charles Le Brun draws parallels between a lion’s visage and a bearded man’s profile. Such images reflected the belief that animal-like facial features revealed animal-like temperament . From the shape of one’s brow to the set of one’s jaw, every detail was thought to hold clues to character.

By the 18th century, physiognomy had become a cultural phenomenon in Europe. Swiss theologian Johann Kaspar Lavater published hugely popular essays in the 1770s that claimed to systematize the reading of faces . In Lavater’s view, the face was a living map of the soul’s “motto” – each curve and line a letter in nature’s code . High society thrilled to have their profiles analyzed; silhouette portraits and lavishly illustrated guides were all the rage. Even as Lavater found believers, he also drew skeptics. Enlightenment thinkers bristled at this mystical “science” of appearances. The German scientist Georg Christoph Lichtenberg, Lavater’s fiercest critic, sneered that studying a person’s behavior was far more useful than studying bumps on their head or the cut of their chin .

Indeed, some of history’s greatest minds were unconvinced by physiognomy. Renaissance genius Leonardo da Vinci flatly called it “false” – “a chimera” with “no scientific foundation” . And in 1530, England’s King Henry VIII went so far as to outlaw “subtle, crafty and unlawful games” like physiognomy and palmistry, lumping them in with con-men’s tricks . These early denunciations notwithstanding, belief in reading faces persisted. In the 19th century, the practice took a dark turn: it became entangled with emerging theories of scientific racism and criminology. The Italian criminologist Cesare Lombroso infamously argued that “born criminals” could be identified by physical defects – hawk-like noses, sloping foreheads, or other so-called atavistic traits . He collected skulls and measured facial angles, insisting that biology was destiny. It was an era when measuring skull bumps (phrenology) and scrutinizing profiles passed for cutting-edge science. But those same ideas would soon be used to justify racist and eugenic beliefs, claiming to find biological proof of character and intelligence differences between ethnic groups .

Debunked, Disgraced, and Discarded

By the early 20th century, physiognomy’s proud “face reading” had been largely exposed as a pseudoscience – and a dangerous one at that. Decades of misuse in justifying racist hierarchies and wrongful prejudice had turned it into a scientific taboo. As one historical review notes, by the latter half of the 1900s physiognomy and its kin (scientific racism and eugenics) were thoroughly debunked as harmful pseudoscience . Academic consensus recognized that no rigorous evidence linked the shape of one’s features to the content of one’s character. In the modern view, judging morals by appearance was no more valid than divining the future from tea leaves.

The fall from grace had been brewing for a long time. Enlightenment and Victorian-era scientists increasingly failed to find any empirical basis for physiognomic claims. Controlled studies (to the extent they were done) showed that observers’ face-based judgments were often just reflecting social biases or random guesses, not genuine insight. By the 1900s, new disciplines like psychology and sociology sought measurable factors in human behavior – personality tests, IQ exams, structured interviews – rather than the quixotic pursuit of reading faces. The very word “physiognomy” came to be used pejoratively, synonymous with superficial prejudice.

It’s telling that even as far back as 1600, a keen observer like da Vinci smelled a fraud, and by the 1800s figures like Charles Darwin (who studied emotional expressions in faces) took care to distinguish expressions from fixed features, wary of grand claims about the latter. In 1886, the British scientist Sir Francis Galton – a cousin of Darwin – tried an experiment: he overlaid multiple photographs of convicted criminals to see if a “criminal face” template would emerge. The composite looked disappointingly ordinary . If anything, Galton’s work inadvertently underscored that faces tell us far less about innate character than physiognomists promised. Little by little, the scientific community relegated physiognomy to the same category as alchemy or astrology: an artifact of our past, not a guide to truth.

By mid-20th century, openly pursuing research on facial features and personality became intellectually disreputable. “Due to its legacy of racism and junk science masquerading as criminology, scientific study or discussion of the relationship between facial features and character has become taboo,” one summary explains . In other words, the very credibility of physiognomy was in shambles. If someone claimed a new way to detect, say, untrustworthiness from a person’s face, most scientists would roll their eyes – or shudder at the echo of old prejudices. The consensus was clear: whatever mild correlations might exist (for example, a lifetime of smiling might produce genuine laugh lines, indicating a cheerful disposition), as a predictive science physiognomy simply did not hold up . That should have been the end of the story.

Yet here we are in 2025, and the story is being rewritten – not by mystics or phrenologists, but by machines.

The AI Revival: Can Algorithms Read Faces?

It turns out you can’t keep a tantalizing idea down for long. In recent years, the rise of artificial intelligence and facial recognition technology has breathed new life into physiognomy’s central question. Researchers and startups around the world are asking, with serious faces (no pun intended): what if advanced algorithms can succeed where old pseudoscience failed? . The promise is seductive: feed a computer millions of human faces and millions of data points about those people’s personalities, and let the machine find patterns too subtle for any human to detect. Modern AI, especially deep learning neural networks, excels at digging out faint signals in vast datasets. Tasks that were impossible before – like recognizing a specific face out of billions, or detecting minute features – are now almost routine. Why not turn that power toward decoding personality?

Indeed, a flurry of studies and products has emerged, essentially rebranding physiognomy as a high-tech endeavor. In 2017, a controversial paper by Stanford researchers claimed an AI could distinguish between gay and straight individuals from facial images with startling accuracy – a claim met with outrage and labeled “junk science” by critics . Around the same time, an Israeli startup called Faception announced it had trained algorithms to identify traits like extroversion, high IQ, even potential terrorists from facial photographs . (One Faception demo infamously included a “Terrorist” classifier defined purely by a facial image , a revival of profiling that many thought consigned to the trash heap of history.) In China, researchers Xiaolin Wu and Xi Zhang reported an AI system that could predict criminality from a mugshot – essentially Lombroso’s 19th-century criminal-physiognomy thesis reborn with silicon chips – claiming over 80% accuracy . The announcement of that study in 2020 sparked such a backlash that the university involved quietly retracted their press release pending “further review” . And the examples keep coming: intelligence agencies exploring face-based “threat” assessment, employers scanning video interviews for personality cues, even dating apps using face analysis to play digital matchmaker.

Driving this revival is not only technological capability but also a treasure trove of data. Billions of images of human faces – from social media, CCTV cameras, driver’s licenses, you name it – are now available to train AI models. Neural networks can comb through these images and, if given some kind of labeled trait for each face, attempt to learn correlations. For example, one recent study used a database of college students who had taken personality tests. Their ID photos were fed into a deep neural network which then attempted to predict the students’ Big Five personality traits (openness, conscientiousness, extroversion, agreeableness, neuroticism) from their facial features . The researchers reported accuracy better than random guessing – enough to suggest a real signal, though far from perfectly reliable . The study concluded that “machine learning can recognize five-dimensional personality features based on static facial features” , but also acknowledged limitations (their sample was relatively homogeneous, and using more varied profile images might improve accuracy) .

Those nuances, however, often get lost in translation when the technology reaches the marketplace. Companies eager to capitalize on the allure of AI face-reading have not been shy with bold claims. Dating apps, in particular, have jumped in with both feet – after all, matchmaking is one arena where reading someone’s true personality is the holy grail. Why rely on fickle swipes and sketchy bios if an AI could find you a soulmate by literally looking at your face?

Love at First Sight? AI Matchmaking and “Face Diagnostics”

One of the most headline-grabbing developments in this field is the emergence of AI-powered dating platforms that promise to find your perfect partner using facial analysis. Forget lengthy questionnaires or endless swiping – these apps ask for nothing more than a selfie. Snap a photo, and let the algorithm do the rest, as one marketing pitch goes . Among the pioneers is SciMatch, a U.S.-based dating app launched in 2023. SciMatch’s premise is straight out of a sci-fi romance: its AI (cheekily named “A.I. Ruby”) scans your facial features to deduce your personality traits, then compares them against other users to suggest highly compatible matches . The app’s underpinning is explicitly rooted in modern physiognomy research – the founders cite a “collective body of research” showing deep learning algorithms can extract the Big Five traits from facial images . In practice, SciMatch claims, “our unique face match app accurately reads each user’s face, analyzing their personality traits, and connecting them with their perfect match” . It’s a bold assurance that sounds almost magical: love unveiled by a glance into the camera.

Another rising player is SoulMatcher, an international dating platform that has gained traction in Europe (including a presence in Britain and EU). On the surface, SoulMatcher’s philosophy is a bit different – it emphasizes depth psychology and clinical personality tests in tandem with photographs . The app requires users to complete voluntary psychological assessments measuring traits like narcissism, empathy, and borderline personality tendencies. The results of these tests are then “overlaid” on the user’s profile photos, giving potential matches a snapshot of one’s psychological makeup alongside their looks . “We don’t just want people to choose by appearance; it’s better to take personal qualities into account,” explains Natalia Sergovantseva, SoulMatcher’s co-founder . In an interview, Sergovantseva stressed that traditional dating apps overly reward the most attractive 10% of users – leading 80% of the “likes” to go to those lucky few . SoulMatcher’s solution is to counterbalance good looks with real character data: “What if that handsome guy is a narcissist?” she notes pointedly . By displaying a user’s psychological profile right on their pictures, the app nudges people to consider compatibility beyond just a pretty face.

Under the hood, SoulMatcher still leverages AI to make the experience seamless. “We use machine learning for training models,” Sergovantseva says, describing how AI improves the accuracy of personality assessments and match suggestions . As more users join, their interactions (likes, passes, successful conversations) feed back into the algorithm, allowing it to “fine-tune the AI so users see people who they find attractive when they open their accounts” . It’s an intriguing blend: on one hand SoulMatcher wants to break users of the habit of judging solely on looks; on the other, its AI explicitly learns who you tend to find attractive to better serve up appealing faces. The company argues this hybrid approach – mixing validated psychological diagnostics with pattern-hunting AI personalization – leads to more meaningful relationships. Essentially, SoulMatcher is betting that technology can reveal the soul behind the selfie, without falling for the superficial. And it’s not alone. From major sites tapping AI to vet profile pictures to experimental apps that animate your face and gauge micro-expressions, the dating industry is riding the AI wave to try to solve an age-old mystery: who among these countless faces might be “the one”?

Of course, matchmaking is a relatively benign application of AI face analysis (the worst outcome, perhaps, is an awkward date or a mismatch). Other uses are far more consequential – and worrisome. When algorithms claim to identify criminals, or diagnose mental illness, or evaluate job applicants based on facial “fit,” the specter of the old pseudoscience looms large. Is this truly a new scientific frontier, or just new bottles for very old snake oil? As AI-driven physiognomy moves from labs to real-world deployment, plenty of experts are urging caution.

The New Face of an Old Question

The resurrection of physiognomy in digital form forces us to confront tough questions: What if the idea wasn’t entirely wrong, just ahead of its time? Could there be kernels of truth in face-personality correlations that only a complex AI can detect? Or is this a dangerous delusion, a high-tech mirror of our own biases that risks automating prejudice under the guise of objective analysis?

For now, the verdict is very much in flux. What’s clear is that AI has made it technically possible to analyze faces at a scale and depth never before imaginable. Whether this should be done, and how, is another matter. Some companies, like SoulMatcher, tread carefully – blending AI with human psychology and explicitly warning against shallow appearance-based judgments. Others, like Faception or more extreme applications, have barreled ahead, sometimes pulled back only after public outcry. “The most accurate way to judge character is by real-life observation of behavior,” SoulMatcher’s founder herself admits , acknowledging that even her advanced app cannot escape the truth that knowing a person requires time and interaction, not just an algorithmic guess.

As an investigative journey, the trail from ancient physiognomy to modern AI is a cautionary tale of scientific hubris and human bias. It teaches us that our desire for quick reads and simple answers about people can easily lead us astray. The London Times spoke with Dr. Eleanor Watson, a UK-based AI ethicist, who encapsulated the dilemma: “We can program a computer to find patterns in faces, but we must be very careful about the stories we then tell about those patterns. The danger is seeing what we want to see – reviving old myths with new tools.” In other words, if we ask an AI to perform physiognomy, we shouldn’t be surprised if it delivers… physiognomy. The risk of a self-fulfilling prophecy – teaching our machines our own biases and then believing their outputs as “scientific truth” – is real.

Still, the pursuit continues, with investment pouring in and consumers intrigued by the concept. SoulMatcher, SciMatch, and their ilk will undoubtedly refine their algorithms. Maybe they will produce success stories – couples happily matched by AI insight, or friendships formed through deeper compatibility screening. And in domains like security or hiring, it’s possible (though many would say unlikely) that carefully validated AI tools could add a layer of useful information – perhaps flagging non-verbal cues of deception in an interrogation video, or noticing signs of stress in a driver’s face to prevent an accident. These more modest uses of face analytics are a far cry from the grand claims of reading an entire character from a still image.

What is certain is that society will have to decide where to draw the line. How much should we allow algorithms to judge about us from how we look? At what point does it trespass on privacy, or revive social prejudice, or simply become bad science? The ghosts of physiognomists past remind us that the line between science and pseudoscience can be perilously thin when it comes to human beings. As artificial intelligence gazes deeply into our faces, seeking the secrets within, we would do well to gaze back with a healthy dose of skepticism – and perhaps recall that old saying: “Never judge a book by its cover.” In the end, we might program our smartest machines to do just that, but the moral judgment remains ours to bear.