Two eyes looking out of a socket
The two holes of a socket with a line beneath them, the headlights and grille of a car, the two windows and door of a house: looking at these, most of us see a surprised, smiling or sulking face for a moment. This illusion is called face pareidolia. From the face in the Moon to faces in burnt toast, we meet it constantly in everyday life, and once we have seen it, it often becomes impossible not to.
We drew the 16 objects in this experiment ourselves. In eight of them the parts are arranged like a face; the other eight are objects of the same kind whose parts are not in a face arrangement. That way we can see not only how often you see faces, but how much the face arrangement contributes.
A brain tuned to find faces
Faces matter so much to us that the visual system seems built not to miss them. In 2020 Susan Wardle and colleagues combined brain imaging with measurements of brain waves to see how illusory faces are represented in the brain. At first, illusory faces were represented more similarly to real faces than faceless objects of the same kind were; but within about 250 milliseconds the representation became equivalent to that of ordinary objects. The researchers read this as the trace of a broadly tuned face detection mechanism that favors sensitivity over selectivity.
The tendency is not unique to humans. In 2017 Jessica Taubert and colleagues showed that rhesus monkeys look longer at photos of objects that evoke face pareidolia in humans than at similar objects without faces, and that their eye movements follow the pattern they use when looking at faces. Pareidolia looks less like a habit learned from cartoons and more like a by-product of a face detection system we share with other species.
Expectation summons the face
Whether we see a face depends not only on the object but also on what we expect. In 2014 Liu and colleagues showed participants images made of pure random noise but told them that half contained a hidden face or letter. Participants reported seeing faces on 34% of trials and letters on 38%. When they said they saw a face, the right fusiform face area, one of the brain's face-sensitive regions, responded too.
That is why we keep two separate measures: how often you see a face in face-like objects, and how often you say 'yes' to their matched controls without a face arrangement. The first shows your sensitivity to the face arrangement; the second, your general tendency to say yes.
The expression, age and gender of illusory faces
In their 2022 study in PNAS, Wardle, Paranjape, Taubert and Baker asked a further question: do we merely detect illusory faces, or do we also give them the traits we attribute to real faces? Using 256 photos collected from the internet and their own archives, they ran a series of experiments with 3,815 adults in total. Participants readily gave illusory faces an expression, an age and a gender.
Taking the most common rating for each face, happiness came out on top for 34% of the faces, surprise for 19% and anger for 14%; only 19% of the faces were judged neutral. For 75% of the faces the most common age given was under 30. But the most striking finding concerned gender.
Why are most illusory faces male?
Participants could choose male, female or neutral for gender; about half of the responses were neutral. Of the gendered responses, however, 81.4% were male and 18.6% female. The average rating was on the male side for 90% of the images; 80% of participants showed a male bias, and only 3% a female bias. The bias did not depend on the participant's own gender.
The researchers tested possible explanations one by one. When only the names of the objects were shown (for example, 'potato'), no gender bias appeared. Converting the photos to grayscale left the bias almost unchanged. Matched objects without faces did not explain it either. The authors suggest that a face built from minimal detail is counted as male by default unless another cue suggests otherwise; whether that default comes from language, culture or the faces we see is an open question. In 2023 Wardle and colleagues found the same bias in children from the age of 4.
In Turkish, who is 'o'?
All of Wardle and colleagues' participants lived in the United States and spoke English. In English you usually have to say 'he' or 'she' when referring to a person. In Turkish, the third-person pronoun 'o' does not mark gender. If the language explanation is right, the bias might be weaker among Turkish speakers; but because culture, media and everyday experience may also play a role, such a difference alone would not prove it is the language.
That is why the crowd results show players on the Turkish and English pages separately. This is only the page language; we do not know which language anyone speaks or where they grew up. Still, we think such a comparison is a good place to start testing a widespread assumption.