Faces That Aren't There: The Psychology and Neuroscience of Pareidolia

You pass a building you have walked by dozens of times. Today, two dark upstairs windows and the open arch below look like a face caught in a moment of surprise. You can point to each window, explain exactly what it is and still see the expression. A step to one side changes the shadows and the face recedes. From the pavement opposite, it is there again.

This is face pareidolia, the experience of seeing a face in a pattern formed by something else. It raises a question about a skill so familiar that we rarely stop to examine it. How does vision decide that two marks and a gap belong together, and why does the resulting arrangement seem to have a gaze, an age or even a mood?

What exactly is pareidolia?

Pareidolia occurs when an ambiguous sensory pattern takes on a recognisable form. A cloud might look like a dog or a patch of paint like a letter. In face pareidolia, a real stimulus supplies the apparent features. The dark marks on a tree trunk really are there. Their appearance as eyes depends on how the visual system groups them with the bark around them. Ordinary recognition uses many of the same grouping skills and identifies an object that is actually present, such as a person standing beside the tree.

Apophenia is a broader term for finding meaningful connections in unrelated information. It could describe a person reading significance into a series of coincidences. Pareidolia describes a more immediate sensory impression. It also differs from a hallucination, in which a percept-like experience occurs without the corresponding external stimulus. Here, the impression is tied to a particular view of a particular object. Someone can understand its source and still see the face.

The brain’s sensitivity to faces

A face can tell you who has entered a room, whether someone is looking your way and whether it is your turn to speak. Those judgements require different kinds of information. Recognising a friend depends partly on features that remain relatively stable, whereas following their gaze depends on details that can change from one moment to the next. Detecting a face in the first place makes all of these judgements possible.

A network for different facial cues

Using functional magnetic resonance imaging (fMRI), Kanwisher et al. (1997) identified a region in the fusiform gyrus that responded more strongly to faces than to several comparison categories. This became known as the fusiform face area. Haxby et al. (2000) reviewed evidence for a distributed system, with parts of the occipital and temporal cortex contributing to different aspects of face perception. Their account distinguishes relatively stable information, such as identity, from changeable information, such as gaze or expression. Calling a region face-selective describes its preference in a comparison. It leaves room for other regions to contribute and for visually similar objects to elicit a response.

How ordinary objects become faces

Two headlights above a grille, the openings of an electrical socket and two knots above a crack in a tree can all supply a rough facial layout. A cloud offers more freedom because its edges are less distinct. A pair of dark patches at similar heights can become the eyes, and a lower gap can become the mouth. A nose or outline may help, but the experience can survive considerable variation in the texture, colour and shape of the individual features.

Omer et al. (2019) asked one group to rate how face-like photographs of objects appeared and another to identify the features present in those photographs. Eyes and mouths contributed most clearly to face-likeness in that stimulus set. Removing either reduced face-likeness ratings, whereas removing teeth or ears had no comparable effect. The experiment started with photographs already likely to evoke faces, so it identifies influential cues in those examples. The way features are arranged and the object they belong to still matter. A single arch can look like a mouth when two windows sit above it and like any other doorway when seen on its own.

Is face detection automatic?

When asked to look towards the picture containing a real face, participants in a rapid eye-movement task could start a correct movement in roughly a tenth of a second on the fastest trials (Crouzet et al., 2010). For illusory faces, the advantage depends on what people are doing. Caruana and Seymour (2022) used a method in which a changing pattern presented to one eye temporarily suppressed an image shown to the other. Participants detected face-like objects sooner than matched objects in that task. Saurels et al. (2024) found faster eye movements towards food images containing illusory faces when the images were placed in peripheral vision. The advantage was limited to that viewing condition.

Detecting a target quickly and being distracted by an irrelevant face are different tests. When Miti et al. (2025) asked people to search for a butterfly, they found no evidence that real or illusory face distractors captured attention in their two versions of the task. When the faces themselves were targets, real faces were found more efficiently than illusory ones or butterflies. Taken together, the studies suggest that a face-like configuration can give an object a head start under some conditions. How conspicuous it is, where it appears and what the viewer is looking for can all influence whether that head start changes behaviour.

What neuroscience reveals

The first quarter of a second

Hadjikhani et al. (2009) used magnetoencephalography (MEG), which detects magnetic fields associated with brain activity, and found that face-like objects evoked an early response at about 165 milliseconds in ventral fusiform cortex. Wardle et al. (2020) compared real faces, objects that elicited pareidolia and matched objects without illusory faces. Their fMRI results located sensitivity to the illusory faces in face-selective parts of the occipitotemporal cortex. In the MEG data, the pattern for illusory faces initially resembled that for real faces more than matched objects did. Within roughly 250 milliseconds, it came to resemble the pattern for ordinary objects. The shift describes patterns in the recordings, while the experience of seeing the face can last much longer.

Recording from inside the brain

Cerrahoğlu et al. (2025) recorded electrical signals directly from the ventral occipitotemporal cortex of 44 participants who already had intracerebral electrodes. Real faces and face-like objects appeared at regular intervals in separate rapid image streams, allowing the researchers to isolate responses at the relevant frequencies. Face-like objects produced weaker responses, yet 89% of the recording contacts selective for face-like objects were also selective for real faces. Both responses began early. This makes it harder to explain the initial response as solely the product of slow, deliberate reinterpretation.

These methods answer different questions. fMRI measures blood-oxygen changes and helps locate responses. MEG has the timing to track how a pattern evolves, though locating its source is less direct. Intracerebral electrodes measure local electrical activity precisely, but their placement samples particular sites in each participant. Wardle’s change in how whole patterns clustered and Cerrahoğlu’s overlapping local responses can coexist. Two kinds of image can recruit much of the same circuitry and still differ in how strongly and in what pattern that circuitry responds.

When objects appear to look at us

A sideways reflection inside a pair of headlights can look like a glance. Takahashi and Watanabe (2013) found that the apparent gaze of a face-like object could cue attention towards a location. The effect occurred when observers saw the object as a face. The same general kind of spatial cue can therefore do different work depending on how the viewer interprets the configuration.

Palmer and Clifford (2020) tested a different prediction. After repeatedly viewing objects that seemed to gaze in one direction, people changed how they judged eye contact in photographs of real faces. Editing out the eye-like and other internal features reduced the effect even though some cues to the object’s orientation remained. The transfer suggests some overlap in the mechanisms used to read apparent gaze in objects and people. The illusion supplies a direction of looking. Intention would require an observer behind those eyes, which the object cannot supply.

The emotions we see in imaginary faces

A car grille that curves upwards can look cheerful. Narrow openings above a low horizontal grille can make another car look irritated. Alais et al. (2021) found that people gave fairly consistent expression ratings to illusory faces, despite large differences in what the objects were made of and how they looked. They also observed serial dependence. A face’s expression tended to be judged closer to the expression of the preceding image. The effect transferred between real and illusory faces when the two appeared in the same sequence.

In an online study by Jakobsen et al. (2025), seven- and eight-year-olds and adults classified happy expressions more quickly and accurately than angry expressions in both human and illusory faces. The results concern particular images and a choice between two labels. Features that make a smile easy to spot may contribute to the advantage. What people agree on is the shape’s expressive appearance. The apparent mood belongs to the reading of the object, just as the surprise on the building belongs to a view of its windows and arch.

The gender bias in face pareidolia

Across a series of experiments involving 3,815 adults, Wardle et al. (2022) found that people assigned male labels to many more illusory faces than they assigned female labels to. Participants also made judgements about apparent age and expression. The researchers tested whether object associations or measured image properties accounted for the gender pattern. Those explanations found little support in the experiments. The finding concerns how people classify a sparse facial pattern, a stimulus with no gender of its own.

One possibility is that people learn to treat male as a default category when an image lacks cues they associate with female faces. Familiar conventions in illustration, where a few extra features are used to indicate gender, offer a way to imagine how such an expectation might develop. The experiments did not trace where that expectation came from. The photographs were selected for their face-like qualities, and the label choices set boundaries on what people could report. The strength and even the form of the bias could vary with the images, language and conventions familiar to different groups of observers.

Is face pareidolia an evolutionary adaptation?

In a crowded scene, detecting a face early can help an animal orient towards a possible social partner, rival or threat. A system sensitive to faint or incomplete cues will sometimes flag a pattern with no face in it. The trade-off is plausible. Missing a partly hidden face and briefly checking a face-like tree trunk can have very different costs. An ordinary false alarm could arise as a by-product of a detection system that works well across varied conditions.

Taubert et al. (2017) found that rhesus macaques looked preferentially at objects that people also judged face-like, compared with matched objects. Their fixation patterns showed some of the organisation seen when looking at real faces. The result makes a story based solely on human language or cartoon conventions less plausible. Looking patterns are indirect evidence about the macaques’ experience. They also leave open how much the response depends on inherited sensitivities, years of seeing faces or broader visual abilities shared by primates.

Why some people notice more faces than others

Two people can differ in how strongly they pick up a pattern, how willing they are to call it a face or how long they spend examining it. Wang and Yang (2025) found that a gender difference among university students depended on the task used. The authors suggested that the groups differed in how readily they judged an ambiguous object to be face-like. Their results say more about the judgements elicited by those tasks than about how often people see faces during an ordinary day. The number of faces someone reports combines what they notice with the threshold they use to report it.

From infancy to childhood

Kato and Mugitani (2015) used a sound and measured where infants looked on a face-like pattern. Ten-month-olds increased their looking towards the region that could serve as a mouth when the pattern was upright. Looking towards a potential source of the sound offers indirect evidence about how infants organised the image. Macchi Cassia et al. (2004) showed that newborn preferences for some face-like displays can also be explained by a broader preference for patterns with more elements in the upper half. The tasks probe different abilities at different ages. Together with the expression findings in school-age children, they show how much depends on what researchers ask a child to look at or do.

Pareidolia and predictive perception

Bottom-up processing begins with the available sensory cues, such as contrast, position and shape. Top-down influences can include what you expect to find or which feature you have been asked to look for. Predictive accounts propose that prior expectations help the visual system interpret uncertain input. The strength and timing of those influences vary across tasks (de Lange et al., 2018). Expecting a face could affect what you see, where you look or how much evidence you require before saying you see one. Those possibilities can yield similar answers in an experiment (Firestone & Scholl, 2016).

Liu et al. (2014) asked people to search visual noise for faces or letters under the impression that some images contained them. Reports followed the category people were looking for. Face reports were associated with a response in the right fusiform face area, and the noise patterns behind those reports, averaged together, resembled a face. Expectations mattered in a task built around them, and features in the noise also mattered. The rapid responses to everyday face-like objects in other experiments suggest that plausible visual cues can do considerable work even without such instructions.

Context can also make a face harder to see. Bourgaux et al. (2026) found weaker face-like neural responses when illusory faces appeared in rapid streams containing human faces, compared with streams containing houses. In a separate task, people were slower and less accurate at detecting the illusory faces in the human-face context. Balas et al. (2025) found that reversing light and dark in photographs of natural textures increased face reports in both adults and children. That result fits the possibility that disrupting the ordinary reading of a scene gives an alternative shape a chance to stand out. Both studies show that the visual cues surrounding a candidate face can affect whether it is picked out. For a related question about what prompts a second look, see Inside Behaviour’s article on the psychology of curiosity.

Artificial intelligence and face detection

Gupta and Dobs (2025) trained deep convolutional neural networks on different combinations of tasks, including face detection, face identification, object detection and object categorisation. They compared how the models represented real faces, illusory faces and matched objects with patterns in human MEG recordings. Models trained on object categorisation generally aligned better with the human responses. In their feature analyses, the model trained jointly on face identification and object categorisation was the one that relied on features such as eye-like regions to classify pareidolia images as faces.

Training a model on different tasks lets researchers ask which demands might make face-like errors emerge. A system that has learnt to distinguish individual faces while also sorting many kinds of object may acquire representations in which the categories partly overlap. The comparison concerns patterns of activity and the information used for a classification. A model can reproduce aspects of that pattern without supplying evidence of a conscious impression of a face, and the fit leaves open other ways the human visual system could arrive at a similar result.

What pareidolia teaches us about the mind

Perception often makes a useful guess from partial information. If the visual system required a sharp outline and a complete set of features before recognising anything face-like, a person glimpsed through a bus window would be hard to pick out. The same tolerance for incomplete evidence lets the windows of a building act as eyes. Seeing the face in the building is different from deliberately imagining someone’s face. In the illusion, the shape and position of features in the scene constrain what you see.

It also differs from a false memory. The question there is how the mind reconstructs an event after it has happened. Here, the question is how information is organised while the object is in view. Both give us an opportunity to separate the available evidence from the experience it supports. Inside Behaviour explores the other side of that distinction in its article on reconstructive memory.

Finding faces in the ordinary

Walk past the building in a different light and the surprised face might disappear. The windows and doorway remain. From the right angle, their arrangement gives vision enough to suggest a face, and familiar ways of reading faces supply some of its character. That impression can sit alongside a perfectly clear understanding of the building. The next time it catches your eye, the interesting part is how quickly two windows and a gap become something you feel you could make eye contact with.

Written by Arkadiusz Dawidowicz for Inside Behaviour.

References

Alais, D., Xu, Y., Wardle, S. G., & Taubert, J. (2021). A shared mechanism for facial expression in human faces and face pareidolia. Proceedings of the Royal Society B: Biological Sciences, 288(1954), Article 20210966. https://doi.org/10.1098/rspb.2021.0966

Balas, B., Morton, M., Setchfield, M., Roshau, L., & Westrick, E. (2025). Contrast negation increases face pareidolia rates in natural scenes. Journal of Vision, 25(13), Article 5. https://doi.org/10.1167/jov.25.13.5

Bourgaux, L., Rekow, D., Leleu, A., & de Heering, A. (2026). Context matters: Human faces hinder face pareidolia. Journal of Cognitive Neuroscience, 38(2), 213–225. https://doi.org/10.1162/JOCN.a.94

Caruana, N., & Seymour, K. (2022). Objects that induce face pareidolia are prioritized by the visual system. British Journal of Psychology, 113(2), 496–507. https://doi.org/10.1111/bjop.12546

Cerrahoğlu, B., Jacques, C., Rekow, D., Jonas, J., Colnat-Coulbois, S., Caharel, S., Leleu, A., & Rossion, B. (2025). The neural basis of face pareidolia with human intracerebral recordings. Imaging Neuroscience, 3, Article imag_a_00518. https://doi.org/10.1162/imag_a_00518

Crouzet, S. M., Kirchner, H., & Thorpe, S. J. (2010). Fast saccades toward faces: Face detection in just 100 ms. Journal of Vision, 10(4), Article 16. https://doi.org/10.1167/10.4.16

de Lange, F. P., Heilbron, M., & Kok, P. (2018). How do expectations shape perception? Trends in Cognitive Sciences, 22(9), 764–779. https://doi.org/10.1016/j.tics.2018.06.002

Firestone, C., & Scholl, B. J. (2016). Cognition does not affect perception: Evaluating the evidence for “top-down” effects. Behavioral and Brain Sciences, 39, Article e229. https://doi.org/10.1017/S0140525X15000965

Gupta, P., & Dobs, K. (2025). Human-like face pareidolia emerges in deep neural networks optimized for face and object recognition. PLOS Computational Biology, 21(1), Article e1012751. https://doi.org/10.1371/journal.pcbi.1012751

Hadjikhani, N., Kveraga, K., Naik, P., & Ahlfors, S. P. (2009). Early (M170) activation of face-specific cortex by face-like objects. NeuroReport, 20(4), 403–407. https://doi.org/10.1097/WNR.0b013e328325a8e1

Haxby, J. V., Hoffman, E. A., & Gobbini, M. I. (2000). The distributed human neural system for face perception. Trends in Cognitive Sciences, 4(6), 223–233. https://doi.org/10.1016/S1364-6613(00)01482-0

Jakobsen, K. V., Hickman, C. M., & Simpson, E. A. (2025). A happy face advantage for pareidolic faces in children and adults. Journal of Experimental Child Psychology, 251, Article 106127. https://doi.org/10.1016/j.jecp.2024.106127

Kanwisher, N., McDermott, J., & Chun, M. M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. The Journal of Neuroscience, 17(11), 4302–4311. https://doi.org/10.1523/JNEUROSCI.17-11-04302.1997

Kato, M., & Mugitani, R. (2015). Pareidolia in infants. PLOS ONE, 10(2), Article e0118539. https://doi.org/10.1371/journal.pone.0118539

Liu, J., Li, J., Feng, L., Li, L., Tian, J., & Lee, K. (2014). Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. Cortex, 53, 60–77. https://doi.org/10.1016/j.cortex.2014.01.013

Macchi Cassia, V., Turati, C., & Simion, F. (2004). Can a nonspecific bias toward top-heavy patterns explain newborns’ face preference? Psychological Science, 15(6), 379–383. https://doi.org/10.1111/j.0956-7976.2004.00688.x

Miti, F., Ciaramidaro, A., Rubichi, S., & Iani, C. (2025). Attentional capture by real and illusory faces: A failure to replicate. Psychological Research, 89(6), Article 181. https://doi.org/10.1007/s00426-025-02211-3

Omer, Y., Sapir, R., Hatuka, Y., & Yovel, G. (2019). What is a face? Critical features for face detection. Perception, 48(5), 437–446. https://doi.org/10.1177/0301006619838734

Palmer, C. J., & Clifford, C. W. G. (2020). Face pareidolia recruits mechanisms for detecting human social attention. Psychological Science, 31(8), 1001–1012. https://doi.org/10.1177/0956797620924814

Saurels, B. W., Peluso, N., & Taubert, J. (2024). A behavioral advantage for the face pareidolia illusion in peripheral vision. Scientific Reports, 14(1), Article 10040. https://doi.org/10.1038/s41598-024-60892-z

Takahashi, K., & Watanabe, K. (2013). Gaze cueing by pareidolia faces. i-Perception, 4(8), 490–492. https://doi.org/10.1068/i0617sas

Taubert, J., Wardle, S. G., Flessert, M., Leopold, D. A., & Ungerleider, L. G. (2017). Face pareidolia in the rhesus monkey. Current Biology, 27(16), 2505–2509.e2. https://doi.org/10.1016/j.cub.2017.06.075

Wang, H., & Yang, Z. (2025). Gender differences in face pareidolia: The effect of cognitive style and judgment criteria. i-Perception, 16(4), Article 20416695251364206. https://doi.org/10.1177/20416695251364206

Wardle, S. G., Paranjape, S., Taubert, J., & Baker, C. I. (2022). Illusory faces are more likely to be perceived as male than female. Proceedings of the National Academy of Sciences of the United States of America, 119(5), Article e2117413119. https://doi.org/10.1073/pnas.2117413119

Wardle, S. G., Taubert, J., Teichmann, L., & Baker, C. I. (2020). Rapid and dynamic processing of face pareidolia in the human brain. Nature Communications, 11(1), Article 4518. https://doi.org/10.1038/s41467-020-18325-8