Dog owners’ facial expressions shift with body shape, study finds

People whose body shape falls into overweight or obesity categories show measurably different facial expressions when they talk to their dogs, according to a new study of Hungarian speakers, and the pattern is strikingly specific: the differences appear when the conversation partner is a dog, but largely vanish when the same people speak to infants or other adults. The research, published in Frontiers in Veterinary Science, is the first to examine whether body condition is linked to visual prosody, the facial movements that accompany speech and carry emotional information alongside the voice.

The study was led by researchers at the HUN-REN Research Center for Natural Sciences in Budapest, working with colleagues at Eötvös Loránd University, Pázmány Péter Catholic University and the University of Ljubljana. It analysed video recordings of 146 adult speakers, aged 18 to 59, interacting with their own dogs, their own 6-to-18-month-old infants, or an adult human partner such as a spouse or an experimenter. Two speakers had to be excluded because the facial analysis software could not detect their facial movements, leaving 144 people in the final dataset. The recordings had originally been collected for earlier studies and were made in semi-naturalistic settings: participants were given instructions, such as playing a game, solving an easy task, teaching an object label or reciting a nursery rhyme, but the tasks were designed to mimic real-life interaction. Speakers faced their interlocutor rather than the camera, spoke Hungarian, and the clips ranged from 20 seconds to five minutes, recorded either at home or in the laboratory.

To quantify facial behaviour, the team used Noldus FaceReader, an automated facial coding program built on Ekman’s Facial Action Coding System, which registers individual muscle movements and assigns intensity values. Four measures were analysed: the intensity of the Happy facial expression, an overall Arousal score reflecting the general activation level of facial movements, a Valence score indicating positive or negative deviation from a neutral state, and the intensity of Action Unit 6, the cheek raiser muscle involved in forming the Duchenne smile. Each variable was scored on a scale from 0 to 1, with Valence ranging from minus 1 to 1.

Body condition was assessed visually rather than measured directly. An experimenter who was blind to the FaceReader outputs rated each speaker’s body shape contour from the video recordings on an eight-point scale running from severe thinness to class III obesity, in line with established clinical guidelines for classifying overweight and obesity. For statistical analysis, speakers were grouped into a normal-weight category with scores of 1 to 4, comprising 83 participants, and an overweight category with scores of 5 to 8, comprising 63 participants. Because the sample was randomly collected, the distribution across the eight individual scores was uneven, with no participants in the two thinnest categories and only two in the most severe obesity category. Interrater agreement on the scoring, checked on a subset of 13 recordings, was high, with an intraclass correlation of 0.861, although the authors note that agreement between raters does not by itself validate the scoring system.

The statistical approach reflected the sheer size of the data. The full dataset contained 964,610 rows, one for every 0.25-second video frame, which exceeded computational limits, so the team analysed a balanced random 10 percent subsample of 96,461 rows. They verified the subsample’s representativeness by running all models again on an independently generated control subsample and finding no differences in the results, and by performing additional participant-level control analyses. Linear mixed models were fitted with communication condition, obesity category, gender and their two-way interactions as fixed effects, participant identity as a random intercept and condition as a participant-specific random slope, with post-hoc comparisons carried out using the Tukey method.

The clearest and most consistent finding concerned facial arousal. Speakers with higher body condition scores showed lower facial arousal across all three communication contexts, whether they were talking to a dog, an infant or another adult. At the reference levels of the interacting factors, the model estimated a significant reduction, and the participant-level analysis found the largest standardized difference of the whole study for arousal in the dog-directed condition, at 0.798 with a 95 percent confidence interval of 0.282 to 1.315. The authors suggest two non-exclusive explanations: excess facial adipose tissue may physically mask the salience of muscle movements, or the difference may reflect a lower tendency to express emotions outwardly, greater expressive suppression, or different emotion regulation strategies, patterns previously reported in laboratory studies of overweight and obese individuals.

The positive emotional measures told a different story. Happy facial expression and facial Valence differed between body condition groups only during dog-directed communication. In the dog condition, participants in the normal-weight category showed higher Happy scores than those in the overweight category, with estimated marginal means of 0.230 versus 0.176, a significant difference, and higher Valence scores of 0.061 versus minus 0.016, also significant. In the infant-directed and adult-directed conditions, by contrast, the estimated differences were negligible and statistically indistinguishable from zero. This specificity surprised the researchers, because previous work has generally emphasised similarities between dog-directed and infant-directed communication, including more exaggerated facial movements than in adult-adult talk. One interpretation offered is that dog-directed communication already operates within a narrower expressive range, since humans may attenuate certain facial signals toward dogs to avoid transmitting cues that can carry agonistic or threatening meanings in canids, such as visible teeth, so additional factors like body-condition-related differences become more apparent. Another possibility is that dogs rely heavily on facial and emotional signals, making subtle differences more relevant in that context.

The cheek raiser results ran in the opposite direction to expectations. Speakers with higher body condition scores displayed stronger AU6 intensity during infant-directed and adult-directed communication, where no difference had been predicted, but no difference at all during dog-directed communication. The authors argue that if body condition merely reduced the visibility of facial movements, lower rather than higher AU6 intensity would be expected, so the pattern suggests that obesity influences different components of facial expression in distinct ways. They list several candidate explanations: facial morphology may have altered the mapping between muscle activation and the software’s output, with muscles activating without producing visible skin displacement through intervening soft tissue; the program may model faces of different shapes and sizes with differing accuracy; the groups may simply differ in habitual communicative style or expressivity rather than emotional processing; the communication contexts may amplify or mask body-condition differences; or a compensatory mechanism may exist. Importantly, no direct muscle activation was measured, and the authors call for studies using electromyography to clarify the pattern.

The study also replicated and extended earlier findings on gender. Female speakers displayed more intense positive facial expressions and greater facial arousal than male speakers across several contexts, a pattern earlier studies had often lacked the statistical power to detect robustly. By combining data from multiple datasets, the present analysis revealed more consistent gender-related variation, although these gender differences were reduced in dog-directed interactions. The authors frame the broader message as evidence that dog-directed communication should not be viewed as a fixed behavioural pattern but as a flexible communicative system shaped by both situational and individual factors, including the speaker’s body condition and gender.

The authors are careful about the limits of what the study can show. It is inherently correlational and cannot establish causal mechanisms; body shape was scored visually from video rather than measured anthropometrically; the body shape groups were unevenly represented; and computational constraints required analysis of a subsample, albeit one validated against a control subsample. A further caveat concerns missing data: frames in which FaceReader could not detect facial movements were excluded, and body composition itself could be one cause of non-detectable movements, so the ratio of missing values might differ between groups, a question the authors leave to future work on facial coding software development. They also stress that the significance of these body-condition-related alterations for the effectiveness of face-to-face communication, and for social relationships between humans and their dogs, remains to be examined, ideally through studies of how dogs behaviourally respond to variation in their owners’ facial expressions.

Subject of Research: The association between body condition and visual prosody in dog-directed, infant-directed and adult-directed human communication

Article Title: Body condition-linked alterations in dog owners' visual prosody

Article References: Gergely, A., Filep, L. A., Kukovič, T., & Kis, A. (2026). Body condition-linked alterations in dog owners' visual prosody. Frontiers in Veterinary Science, 13, Article 1962197. https://doi.org/10.3389/fvets.2026.1962197

Image Credits: AI Generated

DOI: 10.3389/fvets.2026.1962197

Keywords: visual prosody, dog-directed communication, obesity, facial expression, FaceReader, human-animal interaction, facial arousal, emotional expression, cheek raiser, body condition, Frontiers in Veterinary Science, correlational study