A stuffed dog may look enough like a dog to trigger investigation, but dogs recognize other animals through combinations of appearance, odor, movement, sound, familiarity, and interaction that a model does not fully reproduce.
A stuffed dog can look enough like a dog to make a dog stop, stare, bark, approach, retreat, or investigate. That does not mean the dog experiences the stuffed animal as equivalent to another dog. The important distinction is between resemblance, classification, recognition, and prediction. A dog may visually classify something as dog-like without treating it exactly as a living dog. A model may also produce behavior that sometimes resembles the dog's response to a live dog. Neither finding, by itself, proves that the model can predict how that dog will behave toward other dogs... or people... in ordinary life. That distinction is well established in the science of animal recognition, and it becomes especially important when a stuffed dog is used to evaluate fear, sociability, or aggression.
In this article, aggression means attack or attempted physical attack. Growling, barking, snarling, staring, postural threats, and other agonistic signals are described separately rather than automatically counted as aggression.
What does it mean for an animal to recognize another animal?
Recognition is not one all-or-nothing ability. Animal-behavior researchers have described recognition as a classification problem. An animal receives information from another animal and sorts that individual into a behaviorally meaningful category. The category might be as broad as "my species" versus "another species," or as specific as one particular mate, offspring, neighbor, or rival.
Wiley described recognition as varying in specificity and multiplicity. At one end, an animal may make a simple distinction between two broad categories. At the other, it may distinguish numerous particular individuals. Beecher's earlier work on recognition systems emphasized the importance of distinctive "signatures," characteristics that allow one animal to identify another when contextual information is no longer enough. Tibbetts and Dale later developed the evolutionary significance of such individual distinctiveness. Wiley Online Library This matters because recognizing "something that has the appearance of a dog" is not necessarily the same cognitive problem as recognizing a living conspecific, distinguishing a familiar dog from a stranger, or identifying a particular individual.
Why do animals that look identical to us recognize each other?
Because the information important to them may exist in a sensory channel that is relatively unimportant to us. Emperor penguins are an unusually clear example. To a human looking across a colony, large numbers of birds can appear almost interchangeable. Emperor penguins also lack fixed nests that could serve as addresses when mates exchange incubation or chick-care duties. Yet they identify one another.
Aubin, Jouventin, and Hildebrand experimentally demonstrated that emperor penguins possess individually distinctive two-voice calls. Interaction between two acoustic frequencies creates a beat pattern containing identity information. Adults responded to the normal call of their mate, and chicks to the normal call of a parent, but experimentally suppressing one voice destroyed the effective identity signal. PubMed
African penguins provide an additional example. They can connect information about a familiar penguin's appearance with that individual's call, evidence of cross-modal recognition, in which identity information from different senses is associated with the same animal. PubMed
The point is not that dogs recognize one another the way penguins do.
The point is that human judgments of visual similarity do not tell us what information another species is using to distinguish individuals.
Can dogs recognize a dog just by looking at it?
Dogs can use visual information alone to classify dog images as dogs.
Autier-Dérian and colleagues trained nine pet dogs to discriminate photographs of dog heads from photographs of humans and other mammals. Despite the enormous differences in appearance among dog breeds, all nine dogs successfully grouped unfamiliar dog photographs into the dog category. PubMed That study is important, but it should not be stretched beyond what it tested. Nine trained dogs categorized static two-dimensional photographs of heads. They were not greeting unfamiliar dogs, deciding whether another animal was socially safe, or identifying particular dogs. The experiment demonstrates that visual information can be sufficient for species-level categorization under those experimental conditions. It does not demonstrate that visual form contains all the information dogs normally use when interacting with another dog.
Is species recognition the same as individual recognition?
No. A dog can apparently classify a wide variety of visual forms as belonging to the category "dog." Recognizing who a particular dog is requires a more specific discrimination. Dogs can perform some of those discriminations through smell alone.
Hepper tested kin recognition in domestic dogs and found that olfactory cues were sufficient for puppies to recognize mothers and siblings and for mothers to recognize offspring. Mother-offspring recognition could persist after approximately two years of separation. Interestingly, sibling recognition followed a different pattern, suggesting that even within dogs there may not be one universal mechanism called "recognition." PubMed
That is a useful warning against an overly simple statement such as "dogs recognize each other by smell." Dogs receive different kinds of information through different channels, and the information needed depends on what discrimination the animal is making.
What are identity signatures and recognition templates?
An identity signature is a sufficiently distinctive characteristic that another animal can use it to discriminate individuals. The signature can be visual, acoustic, chemical, or multimodal. Some recognition systems depend on learning a particular individual's signature. Others can involve phenotype matching, in which the receiver compares another animal's characteristics with some learned or self-referent template.
Mateo and Johnston's experiments with golden hamsters, for example, provided evidence of self-referent phenotype matching in kin discrimination. The animal did not have to know every relative individually in advance. Characteristics associated with itself could contribute to the reference against which unfamiliar animals were classified. PMC This is foundational recognition science, but it does not establish that dogs use that particular mechanism in ordinary dog-dog encounters. It shows why recognition cannot simply be reduced to "does it look the same?"
Can biological differences that humans cannot see alter an animal's scent?
Yes. Classic mouse experiments revealed a remarkable relationship between immune-system genes and social odor.
Yamazaki and colleagues showed that differences at the major histocompatibility complex, or MHC, could influence mate preferences in mice. This became part of a much larger literature demonstrating that underlying genotype can contribute to chemically detectable differences among otherwise very similar animals. PMC I would not use this research to claim that dogs identify other dogs by reading their MHC genotype. The evidence cited here does not establish that. Its importance is more fundamental. Animals that appear nearly identical to a human can differ biologically in ways that create sensory information available to another animal. A manufactured stuffed dog does not reproduce those biological processes.
Why does using several senses matter?
Because animal recognition is often multimodal, and the senses do not necessarily contribute equally.
Australian sea lions provide an especially useful demonstration. Wierucka and colleagues experimentally separated acoustic, olfactory, and visual cues involved in mother-pup recognition. Acoustic and olfactory information could each carry identity information, but when cues were combined, their roles changed. In some combinations, acoustic information dominated other information. ScienceDirect Other sea lion experiments using three-dimensional artificial models produced another useful finding. Pups could use visual form to distinguish broad classes, such as an adult female form from a pup form, yet adding that visual model to the mother's call did not improve recognition above the acoustic information alone. Nature
That is close to the issue with a stuffed dog. A model can preserve a meaningful part of the stimulus without preserving the whole animal. That can make a model scientifically useful when the researcher knows exactly what question the model is supposed to answer. It does not make the model equivalent to the living animal.
What information does a living dog provide that an ordinary stuffed dog does not?
An ordinary unscented, unheated, nonrobotic stuffed dog principally reproduces some visual characteristics of a dog. It may reproduce body outline, size, coloration, posture, facial arrangement, ears, legs, and tail. Someone can also move it. A live dog supplies much more. The other dog produces biological odor, changes posture, shifts weight, breathes, looks, turns, approaches, retreats, sniffs, vocalizes, and changes its behavior in response to what the subject dog does. That last difference is particularly important. Movement produced by a living social partner is contingent. The other dog responds. Some experimental models can be scented, heated, mechanically moved, or otherwise modified, so it would be wrong to say that no model can provide any of these cues. But an ordinary stuffed dog used in a temperament test does not reproduce the complete sensory and interactive properties of a conspecific.
Do canine odors really contain information that a stuffed dog lacks?
Yes, although exactly what dogs extract from particular odors remains an active research question.
Hepper's kin-recognition experiments already demonstrate that olfactory information can be sufficient for some canine recognition problems. More recent experiments show that dogs discriminate biologically meaningful conspecific odors.
In 2026, Wang and Horowitz tested 43 dogs with body-odor samples from one unfamiliar donor dog collected after situations intended to produce joy, stress, and baseline states. The subjects discriminated joy from baseline and joy from stress under parts of the experimental paradigm, but did not discriminate stress from baseline. The single-donor design is an important limitation. The result therefore should not be inflated into the claim that dogs can universally "smell another dog's emotions." It does demonstrate that changing the donor dog's condition can produce discriminable conspecific odor information. PMC
A stuffed dog does not automatically contain this continuously generated biological information.
What happens when researchers actually compare fake dogs and real dogs?
The results are mixed, which is exactly why a stuffed-dog response should not be treated as self-explanatory.
Barnard and colleagues studied 34 dogs divided among controls, dogs with histories of aggression toward children, and dogs with histories of dog-directed aggression. A fake dog and child-like doll elicited more social behavior than an ambiguous object, and some model responses correlated with owner-reported behavioral history. But the authors concluded that the models did not completely reproduce the corresponding living social stimulus and specifically noted the absence of normal odor, movement, and interaction. ScienceDirect So a model dog plainly can be more than just meaningless furniture. The question is how much the response tells us.
How similar were shelter dogs' responses to stuffed and live dogs?
In one direct comparison, the answer depended strongly on what behavior was being measured.
Shabelansky and colleagues tested 45 shelter dogs with both a single plush dog and a single live stimulus dog. They calculated agreement between the two conditions. For the researchers' combined "friendly" trait, agreement was relatively good: Kappa was 0.60 and prevalence-adjusted bias-adjusted Kappa, or PABAK, was 0.69. For their "fearful" trait, agreement was moderate: Kappa 0.50 and PABAK 0.51. For their "aggressive" trait, agreement was poor: Kappa 0.11 and PABAK 0.38. ScienceDirect Those numbers are more useful than saying simply that the model "worked" or "failed." The same model could correspond fairly well with one broad behavioral dimension and poorly with another.
There is another critical limitation. The investigators were comparing behavior toward a stuffed dog with behavior toward one live test dog. They were not testing whether either response predicted the subject dog's future behavior around dogs in everyday life. ScienceDirect Those are different validity questions.
Haven't other studies found that fake dogs work?
Yes, and those findings belong in any fair treatment of the question.
Barnard and colleagues tested 50 shelter dogs in 2019 using large and small fake dogs and size-matched live stimulus dogs. Responses to fake and live dogs were significantly correlated to the screening process. They nevertheless warned against using the model or temperament test by itself as an assessment or diagnostic tool. MDPI
An even stronger result appeared in 2022.
Reid, Cussen, Collins, and Lockwood studied 292 dogs confiscated in a large dogfighting case. These dogs were an unusual population with more similar selection pressures and histories than a typical shelter population. A life-sized plush model showed meaningful correspondence with same-sex live-dog testing. Seventy-five percent of dogs that showed aggression toward the model also showed aggression toward the same-sex live stimulus dog, while the model detected 81 percent of dogs that showed same-sex aggression toward the live dog. ScienceDirect . That is important counter-evidence…
However… There is an important qualification to these results. Reid and colleagues used the word "aggression" as an operational test category, and that category was broader than violence. A dog could be classified as "Aggressive" after a tense approach or rush followed by behaviors including growling, baring the teeth, snarling, or muzzle punching, as well as snapping or biting. Those behaviors should not automatically be treated as equivalent. Vocalizations, postures, threats, and physical attack can have very different meanings and consequences. Therefore, the reported 75 percent correspondence and 81 percent sensitivity do not mean that a stuffed dog predicted which dogs would violently attack another dog. They mean that the model showed substantial correspondence with a live-dog test using the researchers' broad behavioral category in this unusual population of dogs confiscated from dogfighting. Nor did the study test whether any of these responses predicted future behavior in a home, neighborhood, dog park, or other natural setting. It measured dog-directed responses, not violence toward people. The study therefore provides no basis for using a stuffed dog as a crystal ball for future dog attacks or human-directed aggression. A reaction to the model is evidence about what happened during that test. Any prediction beyond the test requires separate evidence that the test actually predicts that particular outcome.
It tells us that a model can be a useful screening stimulus under some conditions.
It does not tell us that any stuffed dog is equivalent to a live dog, nor did the researchers claim that the result automatically generalized to ordinary pet or shelter populations. They specifically called for research in non-fighting populations. DOI
Does that mean a stuffed-dog test can predict real-life dog aggression?
Not from the evidence above. This is where agreement and prediction need to be separated. If a dog reacts similarly to a stuffed dog and a live test dog, we have learned something about correspondence between those two standardized situations.
That is concurrent agreement.
Predicting whether the same dog will become violent toward an unfamiliar dog weeks later in a home, park, kennel, sidewalk encounter, or greeting is a different question.
The broader shelter-assessment literature gives good reason to be cautious. Patronek, Bradley, and Arps reviewed decades of shelter behavior-evaluation research and concluded that published evidence did not justify broad claims that such evaluations were validated for routine prediction of individual dogs' future aggression. They also emphasized the problem of false positives and the difference between statistical associations in populations and accurate predictions about particular dogs. ScienceDirect
A prospective study of 123 adopted shelter dogs similarly found that some broad qualities observed in the shelter related to post-adoption behavior, while aggression, food guarding, and separation-related problems were not reliably predicted by the standardized assessment. PMC
That does not make behavioral observation useless. It means the question being asked of the observation must match what the evidence can support.
What does my dog's reaction to a stuffed dog actually establish?
It establishes what your dog did in response to that particular stuffed dog under those particular conditions.
Suppose a dog freezes, barks, approaches, sniffs the model, grabs it, shakes it, and then disengages. The observation is the behavioral sequence. One possible inference is that the object's visual properties were sufficient to initiate investigation and that information obtained during closer examination changed the dog's response. What remains unknown is whether the dog initially believed the object was a dog, what emotional state accompanied each behavior, and whether the same behavioral sequence would occur with an unfamiliar living dog. That distinction prevents an observation from quietly becoming a diagnosis. A dog shaking a plush dog has not thereby demonstrated that it will attack dogs. A dog ignoring a plush dog has not demonstrated that it is safe or socially comfortable with dogs. A model response is evidence about the model response. Generalization beyond it requires additional evidence.
Why can a model still be scientifically useful if it is not a real dog?
Because scientific models are often valuable precisely because they remove variables.
The sea lion experiments illustrate this nicely. Researchers used artificial models not because anyone believed polyester and synthetic fur had turned into a sea lion, but because the model allowed visual information to be manipulated separately from sound and odor. Nature
The same logic can apply to a fake dog.
A standardized model may let researchers present approximately the same visual stimulus to many subjects. It can reduce risk to a live stimulus dog. It can help isolate an initial response to canine-like visual form. Those are legitimate uses. The mistake comes when the conclusion outruns the experiment.
A model validated for one screening purpose in one population has not thereby been validated for all dogs, all behavioral questions, all forms of aggression, or prediction of future behavior.
What should I examine instead if my dog has trouble around real dogs?
I would start with the conditions under which the actual behavior succeeds and fails.
Does distance change it? Does movement of the other dog matter? Does familiarity matter? Does the dog's response differ with one handler versus another? Does a barrier matter? Does direction of travel matter? Does the dog respond differently when another dog is approaching, stationary, retreating, or already interacting? Most importantly, where does the response begin to change? Those comparisons tell us more about the live problem than simply assigning a label such as "dog reactive." A label describes a pattern. It does not explain the pattern.
The initial training goal is therefore not necessarily to make the dog get past another dog. The first goal is to identify conditions under which the dog can perceive the other dog while remaining capable of participating, then determine what changes as the situation becomes more difficult.
Can a stuffed dog still be useful in training?
Yes, if its role is narrow and explicit. I might use a model while teaching a person leash handling, positioning, timing, or some other mechanical part of an exercise where exposing another dog would add unnecessary difficulty or risk. A stuffed dog can also be useful for observing what the subject dog does when presented with that particular visual stimulus. What I would not do is silently change the question from "What does this dog do around this model?" to "What will this dog do around dogs?" Those are not interchangeable questions.
When does professional assessment make sense?
Professional assessment makes sense when reactions to actual dogs are strong, worsening, difficult to interrupt, involve lunging or escape, create a risk of injury, or occur at distances where the owner cannot find a manageable starting point.
It also makes sense when medical discomfort has not been ruled out or when the available history does not explain the differences among situations.
The value of an in-home or on-site assessment is not simply having somebody watch the dog. It is being able to compare conditions systematically and determine where behavior changes while controlling distance, movement, handling, barriers, familiarity, and other relevant variables.
If your dog is having a similar problem and you are not sure what changes between the situations where the behavior succeeds and fails, an in-home assessment can help identify the relevant variables and build the training plan from there.
You can learn more about working with me at SamTheDogTrainer.com.
The important distinction is that a stuffed dog can be a useful model without being perceptually equivalent to a dog. Visual resemblance can initiate categorization or behavior, and under some validated conditions a model may even be useful as a screening instrument. But animal-recognition research shows that living animals provide identity and social information through multiple sensory channels, and the model-dog literature shows that correspondence varies with the population, the test, and the behavior being measured.
Start with the question the test can actually answer. Do not turn a response to a model into a conclusion about a living dog without evidence that the inference is valid.
Glossary
Aggression: Attack or fighting behavior in which an animal uses, or attempts to use, physical force against another individual to injure, overpower, seize, or forcibly displace it. Threat displays, warning signals, defensive postures, growls, snarls, and other communicative behaviors may precede, prevent, or accompany aggression, but are not by themselves physical aggression.
Agonistic behavior: The broader class of behavior associated with social conflict, including threat displays, challenge, attack, fighting, avoidance, flight, submission, and appeasement. Aggression is one possible component of an agonistic interaction, not a synonym for the entire sequence.
Conspecific: A member of the same species.
Concurrent validity or agreement: The extent to which two measurements taken under different test conditions correspond. Agreement between a fake-dog test and a live-dog test is not automatically evidence that either predicts future behavior.
Cross-modal recognition: Linking information arriving through different senses as belonging to the same animal.
Identity signature: A sufficiently distinctive characteristic, or combination of characteristics, that can be used to discriminate one individual from others.
Individual recognition: Distinguishing a particular individual and responding on the basis of that individual's identity.
MHC: Major histocompatibility complex, a group of genes with central immune functions. In some animals, MHC variation also contributes to chemically detectable differences.
Multimodal recognition: Recognition involving information from more than one sensory system.
Phenotype matching: Comparing another individual's characteristics with a learned or self-referent recognition template.
Predictive validity: The degree to which a test accurately predicts a later outcome. A test agreeing with another test is not the same as predicting behavior outside the test.
Species discrimination: Distinguishing members of one's own species from members of other species.
Bibliography
Aubin, T., Jouventin, P., & Hildebrand, C. (2000). Penguins use the two-voice system to recognize each other. Proceedings of the Royal Society B: Biological Sciences, 267(1448), 1081-1087. DOI: 10.1098/rspb.2000.1112.
Autier-Dérian, D., Deputte, B. L., Chalvet-Monfray, K., Coulon, M., & Mounier, L. (2013). Visual discrimination of species in dogs (Canis familiaris). Animal Cognition, 16, 637-651. DOI: 10.1007/s10071-013-0600-8.
Baciadonna, L., Solvi, C., La Cava, S., Pilenga, C., Gamba, M., & Favaro, L. (2021). Cross-modal individual recognition in the African penguin and the effect of partnership. Proceedings of the Royal Society B: Biological Sciences, 288, 20211463. DOI: 10.1098/rspb.2021.1463.
Barnard, S., Kennedy, D., Watson, R., Valsecchi, P., & Arnott, G. (2019). Revisiting a previously validated temperament test in shelter dogs, including an examination of the use of fake model dogs to assess conspecific sociability. Animals, 9(10), 835. DOI: 10.3390/ani9100835.
Barnard, S., Siracusa, C., Reisner, I., Valsecchi, P., & Serpell, J. A. (2012). Validity of model devices used to assess canine temperament in behavioral tests. Applied Animal Behaviour Science, 138(1-2), 79-87. DOI: 10.1016/j.applanim.2012.02.017.
Beecher, M. D. (1982). Signature systems and kin recognition. American Zoologist, 22(3), 477-490. DOI: 10.1093/icb/22.3.477.
Hepper, P. G. (1994). Long-term retention of kinship recognition established during infancy in the domestic dog. Behavioural Processes, 33(1-2), 3-14. DOI: 10.1016/0376-6357(94)90056-6.
Mateo, J. M., & Johnston, R. E. (2000). Kin recognition and the 'armpit effect': Evidence of self-referent phenotype matching. Proceedings of the Royal Society B: Biological Sciences, 267(1444), 695-700. DOI: 10.1098/rspb.2000.1058.
Patronek, G. J., Bradley, J., & Arps, E. (2019). What is the evidence for reliability and validity of behavior evaluations for shelter dogs? A prequel to "No better than flipping a coin." Journal of Veterinary Behavior, 31, 43-58. DOI: 10.1016/j.jveb.2019.03.001.
Reid, P. J., Cussen, V. A., Collins, K. A., & Lockwood, R. (2022). The utility of model dogs for assessing conspecific aggression in fighting dogs. Applied Animal Behaviour Science, 254, 105710. DOI: 10.1016/j.applanim.2022.105710.
Shabelansky, A., Dowling-Guyer, S., Quist, H., D'Arpino, S. S., & McCobb, E. (2015). Consistency of shelter dogs' behavior toward a fake versus real stimulus dog during a behavior evaluation. Applied Animal Behaviour Science, 163, 158-166. DOI: 10.1016/j.applanim.2014.12.001.
Tibbetts, E. A., & Dale, J. (2007). Individual recognition: It is good to be different. Trends in Ecology & Evolution, 22(10), 529-537. DOI: 10.1016/j.tree.2007.09.001.
Wang, A., & Horowitz, A. (2026). Dogs (Canis familiaris) distinguish conspecific emotional chemosignals. Scientific Reports, 16, 11176. DOI: 10.1038/s41598-026-41426-1.
Wierucka, K., Pitcher, B. J., Harcourt, R., & Charrier, I. (2018). Multimodal mother-offspring recognition: The relative importance of sensory cues in a colonial mammal. Animal Behaviour, 146, 135-142. DOI: 10.1016/j.anbehav.2018.10.019.
Wiley, R. H. (2013). Specificity and multiplicity in the recognition of individuals: Implications for the evolution of social behaviour. Biological Reviews, 88(1), 179-195. DOI: 10.1111/j.1469-185X.2012.00246.x.
Yamazaki, K., Boyse, E. A., Miké, V., Thaler, H. T., Mathieson, B. J., Abbott, J., Boyse, J., Zayas, Z. A., & Thomas, L. (1976). Control of mating preferences in mice by genes in the major histocompatibility complex. Journal of Experimental Medicine, 144(5), 1324-1335. DOI: 10.1084/jem.144.5.1324.
No comments:
Post a Comment