
In 2016 an Israeli start-up called Faception said its software could analyze a photo of a face and tell whether the person was likely to be a terrorist, a pedophile, a genius or even a professional poker player. The company said it was already working with a homeland security agency. Scientists were blunt in response: there is no good evidence that personality or criminal intent can be read from facial features, and the idea is a modern version of physiognomy, a discredited 19th-century pseudoscience.
The story is worth revisiting because the same promise keeps coming back with each new wave of artificial intelligence. Here is what Faception claimed, what evidence it offered, why even an "80 percent accurate" system would fail badly at catching terrorists, and how to judge similar claims.
What Faception said its software could do
Faception, then about two years old, described itself as a facial personality profiling company. According to reporting by The Washington Post in May 2016, the firm said its computer-vision and machine-learning system could take an image of a face and produce a profile of the person's character without any other information.
The company said it used 15 "classifiers", each representing a type of person with a particular set of traits. Its examples included extrovert, genius, professional poker player, pedophile and terrorist. The software scored each face for how well it matched each type, using cues the company said were too subtle for the human eye.
The pitch was aimed squarely at security services: identify dangerous people before they act. Faception said it had signed a contract with a homeland security agency to help identify terrorists. Some articles at the time described that customer as the US Department of Homeland Security, but the agency was not named in the original reporting, and that detail was never confirmed.
Chief executive Shai Gilboa argued that personality is shaped by our DNA and that this shows in the face, which, he said, is why a machine can read it.
The evidence offered for the claims
Faception pointed to a few demonstrations.
- A poker tournament. The company said it compared photos of 50 amateur players with its database of professional players and picked four as the strongest. Two of those four were among the tournament's three finalists.
- An accuracy figure. Gilboa said the classifiers judged certain traits with about 80 percent accuracy.
- The Paris attackers. Britain's Mirror reported that the software had flagged 9 of the 11 men involved in the November 2015 Paris attacks without being told who they were. That claim came from the company and has never been independently verified.
None of this amounts to scientific proof. A single poker tournament with 50 players is a tiny sample; luck alone can produce a result like that. Identifying known attackers after the fact is also very different from spotting unknown ones in advance, especially when photos of the attackers had already circulated widely in the news. No peer-reviewed study showing that the system worked was released.

Why "80 percent accurate" is nowhere near good enough
Eighty percent sounds respectable. For finding something rare, like a terrorist, it is close to useless, because of a statistical trap known as the base rate fallacy.
Imagine screening one million people, of whom 100 are genuine terrorists. Suppose the software catches 80 of those 100 and wrongly flags 20 percent of everyone else. That second number is the problem: 20 percent of 999,900 innocent people is almost 200,000 false alarms. So out of roughly 200,000 flagged faces, only 80 belong to real terrorists. Fewer than one alarm in 2,000 would be correct.
Even a far better system runs into the same wall. A classic textbook example imagines a face-recognition alarm that catches 99 percent of terrorists and wrongly flags only 1 percent of others. In a city of one million with 100 terrorists, it still produces about 10,000 false alarms for 99 real hits. When the thing you are looking for is extremely rare, false positives swamp everything else, and each one is a real person who could be stopped, questioned or worse.
Put another way, a system that is wrong 20 percent of the time could mislabel one person in five. The labels in question, terrorist and pedophile, are among the most damaging a person can carry.
Physiognomy: an old idea with a dark history
The belief that character is written on the face is ancient. A treatise called Physiognomonica came out of Aristotle's school, and the practice was popular for centuries before falling out of favor. In 1772 the Swiss pastor Johann Kaspar Lavater revived it with widely read essays, and his ideas fed into early criminology.
The most notorious example is the Italian doctor Cesare Lombroso, who in the 19th century claimed that "born criminals" could be identified by physical features such as the shape of the skull or jaw, describing them as evolutionary throwbacks. His theories are now regarded as flawed and steeped in racist and sexist assumptions. Physiognomy and related ideas were also used to justify scientific racism, which is one reason serious researchers treat any new claim of this kind with great caution.
Today, academics classify physiognomy as a pseudoscience. That does not mean faces carry no information at all; they reveal age, some health conditions and emotion in the moment. It means there is no reliable evidence that stable traits like honesty, intelligence or a tendency to violence can be read from bone structure.
What experts said at the time
Researchers who study faces were unimpressed. Alexander Todorov, a Princeton psychology professor and one of the leading experts on how people form first impressions from faces, told The Washington Post: "The evidence that there is accuracy in these judgments is extremely weak." He compared Faception's approach to a revival of physiognomy, an idea he said science had abandoned around a century earlier.
Pedro Domingos, a professor of computer science at the University of Washington, raised the ethical problem in a single question: "Can I predict that you're an ax murderer by looking at your face and therefore should I arrest you?"
Todorov's own research helps explain why such systems are seductive. People make snap judgments about trustworthiness or competence from a face in a fraction of a second, and they tend to agree with each other. But agreement is not the same as accuracy. Those impressions mostly reflect stereotypes, expressions and lighting, and they often turn out to be wrong.
The bias problem in face-reading AI
Machine learning finds whatever patterns exist in its training data, including the wrong ones. If a system is trained on photos of convicted criminals and photos of everyone else, it may learn differences in camera quality, lighting, expressions, age or ethnicity rather than anything about character. Since criminal-justice data already reflect who is policed and arrested most, a model can simply automate those biases and present them as objective.
Later AI face studies show how easy this is to get wrong. A 2017 study claimed AI could detect sexual orientation from faces more accurately than humans; follow-up work found much of that apparent accuracy came from non-facial cues such as grooming, glasses and photo angle. Critics of such research warn that tools like these can be misused by governments or companies to profile people who have done nothing wrong.
How to judge claims like this
The questions raised by Faception's 2016 claims still apply whenever a product says it can read minds from faces. A few checks help:
- Is there peer-reviewed evidence? Marketing demonstrations and company statements are not the same as independent studies.
- What is the false-positive rate? For rare outcomes, this matters far more than headline accuracy.
- What was the system trained on? Biased or unrepresentative data produce biased results.
- What happens to people who are wrongly flagged? A mistake in a film recommendation is trivial; a mistake in a terrorism screening can ruin a life.
The old saying still holds: you cannot judge a book by its cover, and a faster computer does not change that.
Sources: AI Incident Database: report on The Washington Post's May 2016 coverage of Faception; Wikipedia: Physiognomy; Wikipedia: Base rate fallacy. Earlier coverage by Michaela Whitton (The Anti-Media), via Anon HQ.

COMMENTS