Facial Expressions in Dogs: Puppy Dog Eyes, DogFACS and What Faces Reveal

1. What a Facial Expression Is, and What It Is Not

1.1 Movement, Expression and Emotion

When people talk about a dog's facial expression, they usually mean three things at once: a movement (a muscle contracts and the ears, lips or brows change position), a pattern (movements that tend to occur together) and an interpretation (the assumption that the pattern reveals what the dog feels or intends).

Movements can be counted and patterns described statistically, but an interpretation is a hypothesis about a state that cannot be observed directly. The general problem of inferring emotion from behavior is treated elsewhere (why behavior does not equal emotion), and it applies to the face with particular force, because the face is where humans feel most confident that they can see feelings.

1.2 Why Faces Are Hard to Read Objectively

Human observers do not see a face as a collection of separate parts. The researchers who built the main coding system for dog faces pointed out that facial expressions are processed as whole units, in an automatic and streamlined way, which makes it difficult for an observer to notice which individual movements are actually present (Waller et al., 2013). A person looking at a dog sees "worried" or "happy" before they see a brow, an ear or a lip corner.

This holistic reading is useful in daily life but a poor basis for science: if the label comes first, the description is shaped by the label. Measuring faces therefore starts with describing what moved before asking what it means, the same logic that underlies any good behavioral definition (how behavior is defined and measured).

2. Measuring Faces: DogFACS

2.1 How the System Was Built

The Facial Action Coding System (FACS) was developed for humans and describes facial movements in terms of the muscles that produce them. Each visible movement is given a numbered code, called an action unit. The dog version, DogFACS, was developed on the same principle. Its authors used about 8 to 10 hours of footage from 28 privately owned dogs of varying breeds as the primary source, together with about 100 clips from online video platforms and footage of 86 dogs at four shelters in the United Kingdom (Waller et al., 2013).

Each facial movement was documented by the change in appearance it produces, by minimal criteria for identifying it, and by comparison with other species in FACS terminology. The muscular basis of each movement was checked against the dissection of one dog specimen and against previously published dissections. The manual is freely available, but its use requires certification (Waller et al., 2013).

2.2 Action Units Instead of Labels

DogFACS codes describe movements, not meanings. AU101, the inner brow raiser, describes the medial part of the brow moving upward. EAD102, the ears adductor, describes the ears being drawn toward each other. AD137 is the nose lick, AD37 the lip wipe, AD19 the tongue show, and AD126 is panting (Caeiro et al., 2017; Pedretti et al., 2024). None of these codes contains the words fear, joy or appeasement.

This separation is the main strength of the system. Studies can be compared at the level of what was actually seen, and meaning becomes something to test: if an action unit occurs more often in one kind of situation, that is a finding; what the dog felt is a further inference.

2.3 What the Method Can and Cannot Do

DogFACS makes observation more objective, not complete. Coding is slow, usually done on clips of a few seconds, and records the face, not the body, the voice or the situation.

That last point matters for almost every study discussed below. To link a facial action to an emotion, researchers need an independent way of knowing which emotion was present. There are two common approaches: designing the situation so that a particular state is likely, such as waiting for a reward or being prevented from getting it, or relying on labels given by human experts (Correia-Caeiro et al., 2025). The first raises ethical limits for negative states; the second brings in human biases. A critical review of the field warned that asking experts or naive observers to validate the emotion in a stimulus, as the only validation step, risks circular reasoning that simply confirms human assumptions (Correia-Caeiro et al., 2023a).

3. The "Puppy Dog Eyes" Story

3.1 The Inner Brow Raiser

The facial movement at the center of the popular story is AU101, the inner brow raiser. It lifts the inner part of the brow and makes the eye region look larger and higher. To human observers, this makes the face look more infant-like. It also resembles a human movement, AU1, that features in human expressions of sadness (Waller et al., 2013). The phrase "puppy dog eyes" refers to the appearance this movement produces.

Three studies from one research group built the story: one linked the movement to faster rehoming of shelter dogs, one to a muscle that dogs have and gray wolves lack, and one showed that dogs move their faces more when a person is looking at them.

3.2 The Rehoming Association

The first study used shelter rehoming as a stand-in for human selection (Waller et al., 2013). To reduce breed variation, it was restricted to the most common breed group, bull breeds: 29 dogs, each filmed for two minutes during a controlled first contact with an experimenter standing in front of the kennel, one hand held out. Faces were coded with DogFACS, and the days until each dog left the shelter were recorded. Two dogs with exceptionally long stays were excluded as outliers, leaving 27.

In this sample, dogs that produced AU101 more often in those two minutes were rehomed faster, and so were dogs that spent more time at the front of the kennel. Tail wagging went in the opposite direction: dogs that wagged more were rehomed more slowly. The relationship between AU101 and rehoming speed was best described by a curve that explained about 39 percent of the variance (Waller et al., 2013).

The authors attached two caveats that are easy to miss. They stated that if Bonferroni corrections for multiple testing were applied, none of the variables would be deemed significant, so the findings were exploratory and "should be taken with caution". And they noted that adopters might have been responding to perceived sadness rather than to an infant-like appearance (Waller et al., 2013). Shelter dogs are the study population here, and the result says nothing about how dogs should be presented or chosen.

3.3 The Muscle: Dogs and Wolves

The second study looked under the skin (Kaminski et al., 2019). The researchers dissected the heads of four wild gray wolves and six domestic dogs of different breeds: a Labrador retriever, a bloodhound, a Chihuahua, a German shepherd, a Siberian husky and a mixed-breed dog. Facial musculature was largely similar between the two species and differed only around the eye.

The key muscle was the levator anguli oculi medialis (LAOM), which raises the inner brow. It was present as an independent muscle in five of the six dogs; in the Siberian husky it could not be located. In the wolves it was not present as a distinct muscle, but appeared as a small tendon incompletely separated from the ring muscle of the eye. A second muscle, the retractor anguli oculi lateralis, which pulls the outer eye corner toward the ear, was present in all dogs but variable and more slender in wolves (Kaminski et al., 2019). The abstract describes the LAOM as "uniformly present in dogs", which is stronger than the table, where it is marked as variably present.

The same paper compared 27 shelter dogs, mostly Staffordshire bull terriers, with nine park wolves, each filmed for about two minutes by a stranger. Dogs produced more inner brow raises, and the most intense versions were seen only in dogs (Kaminski et al., 2019). The authors hypothesized that humans preferred dogs with expressive brows, so that "puppy dog eyes" result from selection based on human preferences, and raised an alternative: the movement may matter because it exposes more of the white of the eye.

3.4 Faces and Human Attention

The third study asked whether dogs adjust their facial movements to an audience (Kaminski et al., 2017). Twenty-four family dogs were each tested in four conditions: a human either faced the dog or had her back turned, and she either showed food or did not. Each trial lasted two minutes, and the person stood still without reacting to the dog.

Dogs produced more facial movements when the person faced them; the inner brow raiser and the tongue show reached significance on their own, while food, which should raise arousal, had no effect (Kaminski et al., 2017). The authors concluded that dog facial expressions are "potentially active attempts to communicate", and noted that tongue showing can also relate to stress or panting, which shows how quickly one coded movement opens several interpretations.

This fits a broader picture of dogs being highly sensitive to human attention (how dogs read human attention and gestures), but the design could not separate an intention to communicate from a simpler response to being looked at.

4. How Strongly the Story Can Be Told

4.1 The Muscle Is Not Unique to Dogs

The anatomical part of the story was tested by other groups, and the results challenge the idea that the LAOM is exclusive to domestic dogs or first arose through domestication. They do not rule out evolutionary changes in its size or use during domestication. A dissection study of coyotes found a well-developed LAOM, like that of dogs and unlike the modified or absent muscle in gray wolves. Its authors concluded that their findings challenge the hypothesis that the muscle developed through domestication, and suggested instead that it is an ancestral trait that was lost in gray wolves (Cunningham et al., 2024).

A preliminary, qualitative dissection of ten canid specimens, two dogs, three coyotes and five foxes of three species, reached the same direction: both eye muscles linked to dog-human communication were not unique to domestic dogs (Sexton et al., 2024). The sample was small and opportunistic, and no wolf specimens were available. A third study dissected twelve adult specimens of four South American canid species and found both muscles in all of them, which its authors interpret as a broadly conserved trait across the dog family (de Souza-Junior et al., 2026).

A dog-wolf comparison alone therefore cannot show that domestication created the muscle; the better question may be why gray wolves lost or reduced it.

4.2 Brow Raising and Eye Movements

The communicative part of the story was tested in a design built specifically for that question (Bremhorst et al., 2021). Twenty-one family dogs, all Labrador retrievers or a Labrador-like cross to reduce morphological variation, were trained to expect a reward from an apparatus. In one context the reward was delivered automatically. In the other, a person sat inside the apparatus, faced the dog, sought eye contact with a friendly face and handed over the reward. Two certified coders, one of them blind to the hypothesis, coded 276 three-second video samples.

If the inner brow raiser were a communicative signal aimed at people, it should have been more frequent when a person was facing the dog. It was the other way round: the movement was more frequent in the non-social context. A second analysis offered a simpler explanation. Across 211 observations involving the inner brow raiser or eye movements, 94 percent combined eye movements with brow movement or brow tension. Eye movements never occurred without the inner brow raiser, and the brow raise occurred on its own in only 6 percent of cases (Bremhorst et al., 2021). The authors concluded that their result challenges the communicative hypothesis and suggests a lower-level explanation: the brow moves with the eyes.

The study has limits: fixed order of contexts, one breed type, short samples. It does not prove that the brow raise never has a social function, but it shows that a dog looking up at a person will often raise its brows for that reason alone.

4.3 Small Samples, One Group, Correlational Data

Seen as a whole, the evidence base for the original story is narrow: six dogs and four wolves for the anatomy; 27 shelter dogs and nine park wolves for the behavioral comparison, which confounds species with housing and draws on the same four shelters, sample size and breed composition as the rehoming study (whether the footage overlaps is not stated); 27 dogs of one breed group for a correlational rehoming result that did not survive correction for multiple testing; and 24 pet dogs tested under controlled conditions for the audience effect.

The original studies introduced a standardized method and testable hypotheses. But the popular version, that domestication gave dogs a special muscle to win humans over with their eyes, is much stronger than any single finding, and parts of it have been contradicted.

4.4 What Remains

Several things remain reasonably well supported. Dogs can produce an inner brow raise, often and at high intensity. In one laboratory study, dogs moved their faces more when a person was facing them (Kaminski et al., 2017). In a small shelter sample, more brow raising was associated with faster rehoming (Waller et al., 2013). Whether this movement evolved for communication with people, whether dogs use it intentionally, and whether its muscle is a product of domestication are open questions, and the current evidence leans against the strongest versions. Broader ideas about how domestication shaped dog appearance and behavior are discussed separately (what the domestication syndrome does and does not explain), as is the role of mutual gaze in the dog-human bond (what oxytocin and mutual gaze studies show).

5. Do Specific Facial Actions Map onto Specific Emotions?

5.1 Emotional Contexts in Online Videos

The first systematic attempt to link DogFACS codes to emotional states used online videos of 100 family dogs and 50 humans, equally divided across happiness, positive anticipation, fear, frustration and relaxation as a control; videos were chosen for a clear trigger, such as a thunderstorm for fear (Caeiro et al., 2017).

Compared with relaxation, dogs showed more panting in fear contexts; more lip wiping, nose licking and ears adductor during positive anticipation; and more mouth stretch during happiness. Frustrated dogs did not show a higher rate of any facial action. Dogs and humans used different facial movements in comparable situations, and the authors concluded that dogs do not show human-like facial expressions (Caeiro et al., 2017). They also suggested that dogs do not combine several actions into a fixed, stereotypical display as humans do, but produce relatively isolated actions in response to specific situations. In that sample, skull shape, ear shape and breed did not affect the actions that distinguished the contexts.

The emotion was inferred from the situation in the video, not measured independently, and a nose lick seen during anticipation here appears during frustration in other studies. A single action is not a fixed symbol for a single state.

5.2 Anticipation and Frustration

A laboratory program compared Labrador retrievers waiting for a reward with the same dogs when access to the reward was prevented. The original study is described in detail elsewhere (what anticipation looks like in dogs), and the frustration side is part of a broader picture (how frustration works in dogs). A follow-up with 28 new dogs replicated the original pattern: the ears adductor was associated with the positive condition, and ears flattener, blink, lips part, jaw drop and nose lick with the negative condition, with four further actions also more common in the negative condition (Bremhorst et al., 2019, 2022). All but one of these actions occurred regardless of whether the reward was food or a toy.

5.3 Signs, Not Readouts

The follow-up study went one step further and asked how well each facial action would work as a diagnostic test, using the same measures as medical tests: sensitivity (how often the action is present when the state is present) and specificity (how often it is absent when the state is absent) (Bremhorst et al., 2022).

The results are a useful corrective for anyone who reads faces like a dictionary. The ears adductor had excellent specificity for the positive condition (0.90) but low sensitivity (0.50): when it appeared, the dog was usually in the positive condition, but in half of the positive cases it did not appear. Ears flattener and ears downward showed the opposite pattern for the negative condition: high sensitivity (0.89 each) but poor specificity (0.45 and 0.56), so they also appeared often in the positive condition. Nose lick had the lowest sensitivity (0.17) and the highest specificity (0.95). The authors concluded that none of the facial actions, used on its own, would allow a consistent correct classification of the associated emotional state (Bremhorst et al., 2022).

For everyday reading this means that some actions make a state more likely, but none proves it alone. The same logic applies to stress indicators in general (how stress is measured in dogs).

5.4 Audience and Uncertainty

Who is present also changes the face. In a study with 46 dogs, a food reward was first available and then denied in front of a familiar human, another dog, or no visible audience (Pedretti et al., 2024). With another dog present, dogs showed more ears downward, nose licking, lip wiping, panting and whining than without an audience, kept more distance and avoided looking at the stimulus. Cortisol after the test did not differ, and the authors attributed the differences to uncertainty rather than stress level.

The same actions appear across situations that researchers interpret differently, and audience and uncertainty shape the face. Several of these movements are also discussed as possible signals between dogs (what the evidence says about calming signals).

6. Morphology: When the Face Limits the Signal

6.1 Dogs and Wolves Compared

Dogs differ from each other in skull shape, ear carriage, lip length, coat and facial color far more than wolves do. One study asked whether this variation limits what a dog's face can communicate (Hobkirk & Twiss, 2024). The researchers filmed 10 captive wolves and 64 adult dogs housed in small groups at a rehoming center during spontaneous social interactions and reactions to sounds and other stimuli: 559 wolf events and 753 dog events, each usually shorter than ten seconds, coded with DogFACS.

A statistical classifier predicted the observer-assigned state from the facial movements with an overall precision of 71 percent for wolves and 65 percent for dogs, but precision for single states was very low in dogs, for example 6 percent for fear. Half of the dog events labeled as fear were classified as friendly on the basis of facial movements. Brachycephalic and medium skull shapes were associated with nearly 80 percent of the misclassifications, a share that may partly reflect how common these head shapes were in the sample, and floppy ears and pendulous lips were also associated with confusion. Dogs vocalized during 40 percent of events, wolves during 25 percent, and the authors suggested that dogs may compensate vocally for reduced facial signaling (Hobkirk & Twiss, 2024).

An important weakness: each event's state was assigned by one author's subjective appraisal, and the classifier was tested against these labels, so any difficulty the observer had with dog faces is inherited. Species is also confounded with housing and morphology. The safe reading: in this sample, dog facial movements mapped less consistently onto observer-defined states than wolf movements, and certain morphologies were over-represented among the errors.

6.2 Facial Markings and Plain Faces

Coat pattern on the face may matter as well. A study of 103 dogs living in households, aged six months to twelve years and from eight breed groups, coded facial movements with DogFACS and rated the complexity of facial markings such as eyebrow spots, patches and widow's peaks (Sexton et al., 2023). Dogs with plainer faces tended to produce more facial movements, but the correlation was weak (r = -0.326). For adult dogs, owners estimated their dog's expressiveness more accurately when the dog had a plainer face. The authors themselves named the pandemic-era data collection, sample size and unequal breed and age groups as limitations.

6.3 Short Muzzles, Ears and Lips

Evidence on specific features is thin and partly contradictory. In the online-video study, skull and ear shape did not affect the actions that distinguished contexts (Caeiro et al., 2017); in the rehoming-center study, short and medium skulls, floppy ears and pendulous lips were associated with errors (Hobkirk & Twiss, 2024). Methods and populations differ, so the disagreement cannot be resolved here. The health side of short-muzzled breeding is covered elsewhere (what brachycephaly means for dogs).

No verified study was found on how cropped ears affect facial communication. Given how often ear actions appear in the studies above, a loss of information is plausible, but it is an inference, not a finding.

7. What Humans Read in Dog Faces

7.1 Experience and the Fear Problem

People recognize happy dogs fairly well and fearful dogs much worse. In an online study with 2,163 participants, videos pre-categorized by dog behavior professionals as happy or fearful were labeled correctly as happy with a probability of .90 to .93 regardless of experience, but as fearful with a probability rising from about .30 among people who had never lived with a dog to more than .70 among professionals. More experienced respondents used more physical features, especially the ears (Wan et al., 2012). The videos showed whole dogs, but the pattern fits the facial studies.

Experience does not help uniformly. When photographs of a dog's face, taken under behaviorally defined conditions and rated by experts, were shown to people with and without dog experience, both groups could read the emotions, but experienced people were less accurate at reading aggressiveness and better at identifying the behaviorally defined situations (Bloom & Friedman, 2013). Sample sizes are not given in the abstract, and the stimuli appear to come from one dog.

7.2 Culture

One study compared adults and five- to six-year-olds: non-Muslim Europeans with and without dogs, Muslims without dogs living in Europe, and Muslims without dogs in Morocco, who judged frontal photographs of dog, chimpanzee and human faces as happy, sad, angry, fearful or neutral (Amici et al., 2019).

Children's recognition of dog emotions did not depend on their experience with dogs, and they recognized angry and, to some extent, happy dog faces better than sad or fearful ones. Among adults, people who grew up in a culture with a dog-positive attitude recognized dog emotions better than those who did not, whether or not they owned a dog; the difference did not appear for chimpanzee faces. Adults were generally good at happy and angry dog faces and poor at fearful ones, and all dog emotions were recognized less well than human emotions (Amici et al., 2019). The authors concluded that the ability is mainly acquired through experience and cultural exposure. The stimuli were selected by the authors as typical for each context, so the "correct" answer was itself a human judgment.

7.3 Empathy and Personality

The observer matters too. Among 34 observers, ratings of dog faces followed the same pattern as ratings of human faces; more empathetic observers rated threatening faces of both species as more negative and aggressive, and experience with dogs went with more positive ratings of pleasant and neutral dog faces (Kujala et al., 2017).

7.4 Children

Children deserve particular attention because they are often at face height with dogs. In a study with 89 children aged four to seven and 30 adults, participants interpreted photographs of "angry," "happy" and "neutral" dog and human faces. Adults made hardly any errors with aggressive faces. Of the four-year-olds, 67 percent misunderstood aggressive dog faces, with close to 40 percent errors; 70 percent of five-year-olds showed 35 percent errors, 58 percent of six-year-olds 25 percent, and just under half of seven-year-olds 17 percent. The most frequent error was reading an aggressive dog face as happy, and aggressive human faces were not misread in this way (Meints & Racca, 2026). An eye-tracking study with children aged four to ten found that they looked at dog mouths as much as or more than at dog eyes, were better at judging valence than naming emotions, and performed worse with dog than with human faces; the authors suggested that casual familiarity with dogs is not enough and that explicit training may be needed (Correia-Caeiro et al., 2023b).

7.5 Can Reading Be Taught?

The encouraging finding is that reading dogs can be learned. In a longitudinal study, children aged three to five and their parents were tested on how they interpreted dogs' distress signals, then taught to link what they saw with the correct interpretation, and tested again. Understanding improved significantly in children and adults, older children learned better, and the effects lasted over time (Meints et al., 2018). The study included 124 children and 40 parents, of whom 88 children completed the full protocol.

In practice: In our experience, owners learn to read their dog's face fastest when they first learn to describe it neutrally: where the ears are, whether the mouth is open or closed and tight or loose, whether the brow or the skin above the eyes is tense, where the eyes are directed. Only then do we ask what the situation was. This is a practice method, not a tested intervention, but it follows the same order as the research: description before interpretation.

8. How Dogs Look at Faces

The face is not only something dogs show but also something they read. Briefly: dogs trained to discriminate happy from angry human faces, using only the upper or only the lower half of the face, transferred the discrimination to new faces and to the other half of the face, and dogs rewarded for choosing happy faces learned faster than those rewarded for angry ones (Müller et al., 2015). In a looking-time study, dogs looked longer at a face whose expression matched the emotional valence of a simultaneous vocalization, for both dog and human faces (Albuquerque et al., 2016).

Eye-tracking adds detail. In 31 dogs viewing photographs, the eyes were the most likely first target; threatening dog faces drew more attention, threatening human faces avoidance (Somppi et al., 2016). With whole bodies in naturalistic videos, the 92 analyzed dogs attended more to bodies than heads, whereas the 129 humans focused on heads (Correia-Caeiro et al., 2021). A critical review noted that most studies use faces only and reported mixed results on whether dogs discriminate smiling faces (as reviewed by Correia-Caeiro et al., 2023a). How dogs see in general is a separate topic (what dogs can and cannot see).

9. Pain and the Face

In several species, pain changes the face in characteristic ways, and grimace scales based on these changes have been developed for horses, cattle, pigs, sheep, donkeys, rabbits, rats, mice and cats. For dogs, a 2025 review states that there is no developed and validated grimace scale (Mota-Rojas et al., 2025). Behavior-based pain scales for dogs include some facial items; the Colorado State University Canine Acute Pain Scale, for example, considers droopy ears, arched brows and darting eyes, described as a "worried facial expression" (Mota-Rojas et al., 2021, 2025).

The review notes that coat length, color and facial traits might influence such evaluations in dogs (Mota-Rojas et al., 2025). One review also calls animal facial expressions involuntary (Mota-Rojas et al., 2021), which the audience studies above suggest is too simple for dogs.

For owners, the practical consequence is that the absence of a recognizable pain face does not rule out pain. Pain in dogs often shows in changes in activity, posture, tolerance of touch and behavior, not only in the face (why pain in dogs is often missed).

In practice: We regularly see dogs whose face changed with pain only in hindsight, once owners compared current photos with older ones: a tighter look around the eyes, ears held differently, less mobile expression. Comparing a dog's face with its own earlier face is often more informative than comparing it with a general picture of what pain should look like. If you notice such a change together with other behavior changes, a veterinary examination is the next step, not a facial diagnosis.

10. Automated Analysis

Manual DogFACS coding is slow, and automated systems are being developed. Using the Labrador data set from the anticipation and frustration experiments (29 dogs, 164 balanced three-second videos), a pipeline based on DogFACS variables distinguished the positive from the negative condition with an accuracy above 71 percent, and a deep-learning model working directly on images above 89 percent, validated on dogs not used for training (Boneh-Shitrit et al., 2022). With manual codes, the ears flattener alone gave similar performance to the full set.

A second project built a data set of 3,732 dog images annotated with 46 facial landmarks grounded in canine anatomy and DogFACS, as a basis for detecting facial movements and classifying expressions automatically (Martvel et al., 2025). Its authors identify the variety of textures, shapes and morphology across breeds as the main technical challenge and point out that data sets built from internet images with keyword labels have questionable ground truth.

These limits matter more than the accuracy figures. The best-documented model used one breed in one controlled set-up, and its authors note that generalizing to other breeds and natural settings is a known difficulty (Boneh-Shitrit et al., 2022). A model is only as good as its labels, so an app claiming to read emotion from a photo inherits every problem described here.

11. Expressions That Are Easily Overinterpreted

11.1 The "Smile"

A relaxed open mouth with drawn-back lip corners is often read as a smile and in many contexts coincides with a relaxed or playful dog. But panting from heat or exertion produces a similar mouth, and a tense, drawn-back mouth occurs in stress. That children's most frequent error with aggressive dog faces was reading them as happy (Meints & Racca, 2026) shows how error-prone the "smile" category is. In play, an open mouth usually comes with loose movement and other play signals (how dogs signal play).

11.2 The "Guilty Look"

The "guilty look", with lowered head, ears back and averted gaze, is probably the most famous facial misreading. In a study with 14 dogs, the look did not depend on whether the dog had actually eaten a forbidden treat; it appeared more when owners scolded, and the effect of scolding was stronger when the dog had in fact been obedient (Horowitz, 2009). The full evidence is discussed elsewhere (what the guilty look really shows).

Scolding can stop a behavior in the moment, and the lowered posture can look like a sign that the message has arrived. The more useful question is what the dog has learned: in this study, the look was a response to the owner's behavior, not a sign of understanding a misdeed. A dog that is regularly scolded on the basis of its facial expression may learn that the owner's return or the owner's tone predicts trouble, at the cost of trust in exactly the moments when the owner wants the dog to approach.

12. What Follows for Practice

12.1 Principles Supported by Research

The studies support a few clear principles: single facial actions are signs with limited sensitivity and specificity (Bremhorst et al., 2022); the same actions occur across contexts and audiences (Pedretti et al., 2024); morphology can limit what a face shows (Hobkirk & Twiss, 2024); people read fear poorly (Wan et al., 2012); dogs themselves attend strongly to bodies (Correia-Caeiro et al., 2021); and reading can be taught (Meints et al., 2018).

12.2 The Face as Part of the Whole Dog

The most robust practical conclusion is that the face should be read as one part of the whole dog, in its situation, and over time. Ears, eyes, mouth and brow add information to body posture, tension, movement, tail, vocalization and context; they do not replace them (how to read canine body language as a whole). Questions about emotion in general are covered in a separate overview (what is known about emotions in dogs).

In practice: When a dog's face is ambiguous, we look for change rather than for a single picture. What did the face look like a few seconds before, and what happens next? A dog whose ears flatten briefly and then return to a loose position while it moves toward a person is in a different situation from a dog whose ears stay flat, whose mouth closes and whose body stiffens. Sequences tell more than snapshots, and this is the main thing we teach owners.

In practice: Every dog has its own neutral face, and we encourage owners to learn it: how the ears sit when the dog is relaxed at home, how the mouth looks when it rests, whether the breed's skin folds or lip shape create a permanent "worried" or "smiling" impression. Deviations from that baseline are more informative than comparisons with photographs of other dogs, especially for breeds with extreme facial features.

13. Summary at a Glance

Facial expressions in dogs can be measured objectively with DogFACS, which codes movements, not meanings (Waller et al., 2013); linking them to emotions requires an independent criterion for the emotion, the field's main methodological problem (Correia-Caeiro et al., 2023a, 2025).

The "puppy dog eyes" story rests on small samples. More inner brow raising went with faster rehoming in 27 shelter dogs, in exploratory statistics (Waller et al., 2013). The brow muscle was found in five of six dogs and not as a distinct muscle in four wolves (Kaminski et al., 2019), but also exists in coyotes, foxes and South American canids (Cunningham et al., 2024; Sexton et al., 2024; de Souza-Junior et al., 2026). Dogs moved their faces more when watched (Kaminski et al., 2017), yet in a social versus non-social comparison the brow raise was more frequent without a person and was tied to eye movements (Bremhorst et al., 2021).

Facial actions are associated with situations but none is diagnostic alone (Caeiro et al., 2017; Bremhorst et al., 2022); audience and morphology shape the face (Pedretti et al., 2024; Hobkirk & Twiss, 2024; Sexton et al., 2023). People read happy dogs well and fearful dogs poorly, shaped by experience, culture and empathy (Wan et al., 2012; Amici et al., 2019; Kujala et al., 2017); children often misread aggressive faces as happy but can be taught (Meints & Racca, 2026; Meints et al., 2018). There is no validated grimace scale for dogs (Mota-Rojas et al., 2025), and automated analysis is largely limited to one breed and controlled settings (Boneh-Shitrit et al., 2022; Martvel et al., 2025).

14. Research Gaps and Critical Appraisal

14.1 Sample Size

Most facial studies in dogs are small: six dogs and four wolves dissected (Kaminski et al., 2019), two dogs, three coyotes and five foxes in the challenge (Sexton et al., 2024), 27 dogs for the rehoming association (Waller et al., 2013), 24 for the audience effect (Kaminski et al., 2017), 21 for the eye-movement study (Bremhorst et al., 2021) and 28 for diagnostic accuracy (Bremhorst et al., 2022). The largest samples come from perception studies, not from experiments on emotional states.

14.2 Populations

Key studies used shelter or rehoming-center dogs (Waller et al., 2013; Kaminski et al., 2019; Hobkirk & Twiss, 2024), and the experimental program on emotional states used almost only Labrador retrievers (Bremhorst et al., 2021, 2022; Boneh-Shitrit et al., 2022). Kenneled dogs differ from family dogs in stress and experience, and a single breed with moderate morphology is the best case for facial coding, not the typical one.

14.3 Ground Truth

Every study linking faces to emotions needs a criterion for the emotion: induced situations, video context or observer judgments (Caeiro et al., 2017; Hobkirk & Twiss, 2024; Amici et al., 2019). When the criterion is a human judgment, findings about how well humans read dogs, or how well dogs express states, are partly circular (Correia-Caeiro et al., 2023a).

14.4 Design

The rehoming result is correlational and exploratory (Waller et al., 2013), the dog-wolf comparison confounds species with housing (Kaminski et al., 2019), and the eye-movement study used a fixed order of contexts (Bremhorst et al., 2021). Experiments that vary a single facial movement and measure human responses were not found in the material reviewed.

14.5 Morphology

Two studies disagree on whether skull and ear shape matter (Caeiro et al., 2017; Hobkirk & Twiss, 2024); facial markings were examined in one correlational study with a weak effect (Sexton et al., 2023); studies on cropped ears were not found.

14.6 Faces Without Bodies

Most work isolates the face, while dogs themselves rely heavily on bodies (Correia-Caeiro et al., 2021, 2023a). Coding face, body and voice together in naturalistic situations would come closer to how communication works.

14.7 Pain and Automation

A validated pain face scale for dogs does not exist (Mota-Rojas et al., 2025), and automated systems have been tested mainly on one breed under controlled conditions (Boneh-Shitrit et al., 2022; Martvel et al., 2025). Both fields need breed-diverse, independently labeled data.

15. Conclusion

Dog faces move in measurable, structured ways, and DogFACS has made it possible to study them without immediately assigning meaning. That work has shown that dogs do not simply show human expressions, that facial actions are associated with situations, and that human readers are less reliable than they believe, particularly when a dog is afraid or a child is looking.

The "puppy dog eyes" story shows how a few interesting studies can become a confident narrative. The inner brow raise is real, dogs do it often, and in one small shelter sample it went with faster rehoming. But the muscle is not unique to dogs, the movement is tied to eye movements, and whether it was shaped by human preference or serves communication remains open.

For everyday life, the conclusion is modest and practical: read the face as part of the whole dog, in its situation and over time; do not treat any single expression as proof of a feeling; learn your own dog's baseline; and be especially careful with smiles, guilty looks and children's interpretations.

Key Insights (Takeaways)

• DogFACS codes facial movements by their muscular basis, not by meaning; linking movements to emotions requires an independent criterion for the emotion, which is the field's central problem (Waller et al., 2013; Correia-Caeiro et al., 2023a).

• The "puppy dog eyes" story rests on small samples: 27 shelter dogs for the rehoming association, six dogs and four wolves for the anatomy, and 24 dogs for the audience effect (Waller et al., 2013; Kaminski et al., 2019, 2017).

• The inner brow muscle also occurs in coyotes, foxes and South American canids, so it is unlikely to be a product of domestication alone (Cunningham et al., 2024; Sexton et al., 2024; de Souza-Junior et al., 2026).

• In a controlled comparison, the inner brow raise was more frequent without a person present and occurred independently of eye movements in only 6 percent of cases (Bremhorst et al., 2021).

• No single facial action reliably identifies an emotional state; ears flattener, for example, had high sensitivity but poor specificity for frustration (Bremhorst et al., 2022).

• Skull shape, ears, lips and facial markings can make faces harder to read, though studies disagree on how much (Hobkirk & Twiss, 2024; Sexton et al., 2023; Caeiro et al., 2017).

• People recognize happy dogs well and fearful dogs poorly; experience, cultural background and empathy shape the reading, children often misread aggressive faces as happy, and reading can be taught (Wan et al., 2012; Amici et al., 2019; Meints & Racca, 2026; Meints et al., 2018).

• There is no validated grimace scale for dogs, and automated facial analysis has so far been tested mainly on one breed in controlled settings (Mota-Rojas et al., 2025; Boneh-Shitrit et al., 2022).

References

Albuquerque, N., Guo, K., Wilkinson, A., Savalli, C., Otta, E., & Mills, D. (2016). Dogs recognize dog and human emotions. Biology Letters, 12(1), 20150883. https://doi.org/10.1098/rsbl.2015.0883

Amici, F., Waterman, J., Kellermann, C. M., Karimullah, K., & Bräuer, J. (2019). The ability to recognize dog emotions depends on the cultural milieu in which we grow up. Scientific Reports, 9, 16414. https://doi.org/10.1038/s41598-019-52938-4

Bloom, T., & Friedman, H. (2013). Classifying dogs' (Canis familiaris) facial expressions from photographs. Behavioural Processes, 96, 1–10. https://doi.org/10.1016/j.beproc.2013.02.010

Boneh-Shitrit, T., Feighelstein, M., Bremhorst, A., Amir, S., Distelfeld, T., Dassa, Y., Yaroshetsky, S., Riemer, S., Shimshoni, I., Mills, D. S., & Zamansky, A. (2022). Explainable automated recognition of emotional states from canine facial expressions: The case of positive anticipation and frustration. Scientific Reports, 12, 22611. https://doi.org/10.1038/s41598-022-27079-w

Bremhorst, A., Mills, D. S., Stolzlechner, L., Würbel, H., & Riemer, S. (2021). 'Puppy dog eyes' are associated with eye movements, not communication. Frontiers in Psychology, 12, 568935. https://doi.org/10.3389/fpsyg.2021.568935

Bremhorst, A., Mills, D. S., Würbel, H., & Riemer, S. (2022). Evaluating the accuracy of facial expressions as emotion indicators across contexts in dogs. Animal Cognition, 25(1), 121–136. https://doi.org/10.1007/s10071-021-01532-1

Bremhorst, A., Sutter, N. A., Würbel, H., Mills, D. S., & Riemer, S. (2019). Differences in facial expressions during positive anticipation and frustration in dogs awaiting a reward. Scientific Reports, 9, 19312. https://doi.org/10.1038/s41598-019-55714-6

Caeiro, C. C., Guo, K., & Mills, D. S. (2017). Dogs and humans respond to emotionally competent stimuli by producing different facial actions. Scientific Reports, 7, 15525. https://doi.org/10.1038/s41598-017-15091-4

Correia-Caeiro, C., Guo, K., & Mills, D. (2021). Bodily emotional expressions are a primary source of information for dogs, but not for humans. Animal Cognition, 24(2), 267–279. https://doi.org/10.1007/s10071-021-01471-x

Correia-Caeiro, C., Guo, K., & Mills, D. S. (2023a). Visual perception of emotion cues in dogs: A critical review of methodologies. Animal Cognition, 26(3), 727–754. https://doi.org/10.1007/s10071-023-01762-5

Correia-Caeiro, C., Lawrence, A., Abdelrahman, A., Guo, K., & Mills, D. (2023b). How do children view and categorise human and dog facial expressions? Developmental Science, 26(3), e13332. https://doi.org/10.1111/desc.13332

Correia-Caeiro, C., Zamansky, A., Karl, S., & Bremhorst, A. (2025). Research methods for the analysis of visual emotion cues in animals: A workshop report. Animals, 15(21), 3142. https://doi.org/10.3390/ani15213142

Cunningham, P., Shankar, M., vonHoldt, B., Brzeski, K. E., & Kienle, S. S. (2024). Coyotes can do 'puppy dog eyes' too: Comparing interspecific variation in Canis facial expression muscles. Royal Society Open Science, 11(10), 241046. https://doi.org/10.1098/rsos.241046

de Souza-Junior, P., Souza, E. C., Viotto-Souza, W., Costa, M. E. R., Crivelaro, A. Z., Kasper, C. B., & Abidu-Figueiredo, M. (2026). Beyond domestication: Occurrence of levator anguli oculi medialis and retractor anguli oculi lateralis in four Neotropical canid species. Anatomia, Histologia, Embryologia, 55(3), e70121. https://doi.org/10.1111/ahe.70121

Hobkirk, E. R., & Twiss, S. D. (2024). Domestication constrains the ability of dogs to convey emotions via facial expressions in comparison to their wolf ancestors. Scientific Reports, 14, 10491. https://doi.org/10.1038/s41598-024-61110-6

Horowitz, A. (2009). Disambiguating the "guilty look": Salient prompts to a familiar dog behaviour. Behavioural Processes, 81(3), 447–452. https://doi.org/10.1016/j.beproc.2009.03.014

Kaminski, J., Hynds, J., Morris, P., & Waller, B. M. (2017). Human attention affects facial expressions in domestic dogs. Scientific Reports, 7, 12914. https://doi.org/10.1038/s41598-017-12781-x

Kaminski, J., Waller, B. M., Diogo, R., Hartstone-Rose, A., & Burrows, A. M. (2019). Evolution of facial muscle anatomy in dogs. Proceedings of the National Academy of Sciences, 116(29), 14677–14681. https://doi.org/10.1073/pnas.1820653116

Kujala, M. V., Somppi, S., Jokela, M., Vainio, O., & Parkkonen, L. (2017). Human empathy, personality and experience affect the emotion ratings of dog and human facial expressions. PLOS ONE, 12(1), e0170730. https://doi.org/10.1371/journal.pone.0170730

Martvel, G., Zamansky, A., Pedretti, G., Canori, C., Shimshoni, I., & Bremhorst, A. (2025). Dog facial landmarks detection and its applications for facial analysis. Scientific Reports, 15, 21886. https://doi.org/10.1038/s41598-025-07040-3

Meints, K., Brelsford, V., & De Keuster, T. (2018). Teaching children and parents to understand dog signaling. Frontiers in Veterinary Science, 5, 257. https://doi.org/10.3389/fvets.2018.00257

Meints, K., & Racca, A. (2026). Look, he's smiling! Children's misinterpretation of dogs' facial expressions. Anthrozoös, 39(2), 315–342. https://doi.org/10.1080/08927936.2025.2551434

Mota-Rojas, D., Marcet-Rius, M., Ogi, A., Hernández-Ávalos, I., Mariti, C., Martínez-Burnes, J., Mora-Medina, P., Casas, A., Domínguez, A., Reyes, B., & Gazzano, A. (2021). Current advances in assessment of dog's emotions, facial expressions, and their use for clinical recognition of pain. Animals, 11(11), 3334. https://doi.org/10.3390/ani11113334

Mota-Rojas, D., Whittaker, A. L., Coria-Avila, G. A., Martínez-Burnes, J., Mora-Medina, P., Domínguez-Oliva, A., Hernández-Avalos, I., Olmos-Hernández, A., Verduzco-Mendoza, A., Casas-Alvarado, A., & Grandin, T. (2025). How facial expressions reveal acute pain in domestic animals with facial pain scales as a diagnostic tool. Frontiers in Veterinary Science, 12, 1546719. https://doi.org/10.3389/fvets.2025.1546719

Müller, C. A., Schmitt, K., Barber, A. L. A., & Huber, L. (2015). Dogs can discriminate emotional expressions of human faces. Current Biology, 25(5), 601–605. https://doi.org/10.1016/j.cub.2014.12.055

Pedretti, G., Canori, C., Costantini, E., Palme, R., Valsecchi, P., & Marshall-Pescini, S. (2024). Intra and interspecific audience effect on domestic dogs' behavioural displays and facial expressions. Scientific Reports, 14, 9546. https://doi.org/10.1038/s41598-024-58757-6

Sexton, C. L., Buckley, C., Lieberfarb, J., Subiaul, F., Hecht, E. E., & Bradley, B. J. (2023). What is written on a dog's face? Evaluating the impact of facial phenotypes on communication between humans and canines. Animals, 13(14), 2385. https://doi.org/10.3390/ani13142385

Sexton, C. L., Diogo, R., Subiaul, F., & Bradley, B. J. (2024). Raising an eye at facial muscle morphology in canids. Biology, 13(5), 290. https://doi.org/10.3390/biology13050290

Somppi, S., Törnqvist, H., Kujala, M. V., Hänninen, L., Krause, C. M., & Vainio, O. (2016). Dogs evaluate threatening facial expressions by their biological validity: Evidence from gazing patterns. PLOS ONE, 11(1), e0143047. https://doi.org/10.1371/journal.pone.0143047

Waller, B. M., Peirce, K., Caeiro, C. C., Scheider, L., Burrows, A. M., McCune, S., & Kaminski, J. (2013). Paedomorphic facial expressions give dogs a selective advantage. PLOS ONE, 8(12), e82686. https://doi.org/10.1371/journal.pone.0082686

Wan, M., Bolger, N., & Champagne, F. A. (2012). Human perception of fear in dogs varies according to experience with dogs. PLOS ONE, 7(12), e51775. https://doi.org/10.1371/journal.pone.0051775