Word Learning in Dogs: Object Names and the Gifted Few

1. The Question Behind the Famous Dogs

1.1 Two Different Claims

"My dog knows the names of his toys" and "dogs understand words" are not the same claim, and the research on canine word learning is mostly a story about keeping them apart.

A handful of dogs have demonstrated vocabularies of object names that are, by any reasonable standard, extraordinary. Those cases are real, well documented, and repeatedly replicated in a small group of animals. What they do not establish is what the average dog does when a person says a word. That question has a separate literature, a partly surprising answer, and a set of limits that matter more for training than any of the record-breaking vocabularies do.

1.2 What This Article Covers

This article sets out what the case studies actually demonstrated, how rare the ability is and what the dogs that have it look like, what the exceptional dogs can do beyond simply fetching the right toy, what recent neural work says about ordinary dogs, and the two places where canine word learning reliably breaks down — sound detail and generalization. It closes with what follows from all of it for training.

2. Rico and Chaser: What the Case Studies Showed

2.1 Fast Mapping in a Border Collie

The modern literature starts with a Border Collie named Rico. Kaminski, Call, and Fischer tested him and reported that he not only retrieved objects by name from a vocabulary of roughly two hundred items, but could work out the referent of a new word by exclusion: presented with a set of familiar objects plus one unfamiliar one and asked for a name he had never heard, he retrieved the unfamiliar object — a form of inference by exclusion that has been probed in dogs in other paradigms as well (what dogs appear to know about what they know). He also retained some of those new pairings when retested weeks later.

The authors framed this as evidence for "fast mapping" — the mechanism by which young children attach a new word to a new object on very few exposures. The comparison was deliberate and useful for designing experiments, but it does not imply that dogs acquire language in the human sense.

2.2 Chaser: The Study That Anticipated the Objections

Pilley and Reid's work with a Border Collie named Chaser is the most thorough single-animal study in the field, and its design is worth walking through because each experiment closes a loophole.

Experiment 1 — the vocabulary. Over three years of intensive daily training, four to five hours a day, Chaser learned and retained the names of 1,022 objects. Monthly formal tests with twenty randomly selected items produced accuracy of at least eighteen out of twenty, consistently, and new learning did not displace old.

The three years at four to five hours a day is the number that usually gets dropped when this study is cited. It is not a demonstration of effortless learning. It is a demonstration of what is achievable under conditions almost no dog will ever experience.

Experiment 2 — names are not commands. The obvious objection to any such result is that the dog has not learned a name but a cue for an action. Pilley and Reid tested this by randomly pairing three commands (take, paw, nose) with three object names across fourteen double-blind trials. Chaser was correct on all fourteen. The name specified which object; the command specified what to do with it. The two were independent.

Experiment 3 — categories. Chaser also learned the common nouns "toy," "ball," and "Frisbee," handling one-to-many mappings (one label covering many objects) and many-to-one mappings (several labels applying to one object). A single object could be both a ball and a toy.

Experiment 4 — exclusion, and its limits. Asked for a novel name among familiar objects, Chaser selected the novel object across eight successive trials. But retention of those inferred pairings fell off sharply: strong immediately, reduced after ten minutes, and minimal at twenty-four hours without further rehearsal. Consolidation over that interval is exactly where sleep does its work in other memory paradigms (sleep and memory consolidation).

That last result is the most instructive in the whole paper, and it is almost never quoted. Working out which object must be the new one is not the same as learning its name. The inference is available immediately and mostly gone the next day.

2.3 Why Children Keep Coming Up

Every paper in this field reaches for the same comparison, and it is worth being explicit about why, because the comparison does a lot of unacknowledged work.

Kaminski and colleagues named their finding after fast mapping, the process by which a young child attaches a new label to a new object on very few exposures, without correction and without being told what the word does not mean. Dror and colleagues describe Gifted Word Learner dogs as showing skills functionally similar to those of human infants. The 2024 retention study opens by noting the role long-term memory of words plays in children's developing ability to acquire language. And the shape-bias work is framed explicitly against what human infants do.

The comparison is legitimate as a source of paradigms: the tests used with dogs are adapted from tests designed for pre-verbal children, which is why they rely on retrieval and looking rather than on anything the animal produces. It is a claim about mechanism only where a study tests mechanism.

The asymmetry that is easy to lose sight of: a child's vocabulary is the foundation of a productive system. The child will go on to combine those words, use them to refer to things that are not present, and ask about them. Nothing in the canine literature demonstrates any of that. Comprehension of object labels is one component of language, examined in isolation, in a species that has no language of its own. Dogs are a comparative model for that one component — which is exactly how the researchers describe them, and rather less than how the coverage usually does.

3. How This Gets Tested

3.1 Forced Choice and Chance

Nearly every result here comes from the same basic structure. A set of objects is placed out of reach or in another room. A name is spoken. The dog retrieves. Correct or incorrect is recorded, and performance is compared against what random selection would produce.

Chance level depends on how many objects are available, which is why the number of distractors matters as much as the number of trials. Two objects means 50% by guessing; eight means 12.5%. A dog described as "usually right" in a household setting may be choosing between two toys that are both within reach, which is a much weaker result than the same description suggests.

The structure also constrains what can be asked. It works for nouns naming portable objects that a dog is willing to pick up and carry. It does not work for words describing actions, places, people, properties, or anything the dog cannot fetch. That limitation is a property of the method, not of dogs, and it explains why the literature is almost entirely about toys.

3.2 Blind Testing Is Not Optional

The confound in this area is the person asking. A handler who knows which object is correct will orient, shift weight, or change breathing, and dogs are unusually good at picking up exactly that kind of information (how dogs read human gestures).

The credible studies handle this structurally. Chaser's formal vocabulary tests and her names-versus-commands experiment were run under double-blind conditions. In the EEG work, the measurement was of brain activity while an object was presented, which removes the choice from the dog altogether. In the toy-exploration comparison, caretakers were instructed to stay passive during the test.

The consequence for anyone evaluating a claim about their own dog, or a video of someone else's, is simple: if the person who knows the answer is visible to the dog, the demonstration does not distinguish word knowledge from cue reading.

3.3 The Dogs Who Do Not Finish

Sample sizes in published papers describe the dogs who completed the protocol, and the gap between recruited and analyzed is informative in its own right.

In the shape-and-texture study, 56 dogs were recruited and 35 completed: 21 were excluded for failure to reach the training criterion, insufficient completed trials, or behavioral biases. In the functional-categorization study, 11 Gifted Word Learner dogs were recruited and 10 analyzed, the exclusion being lack of motivation. In the EEG study, 27 dogs took part and 18 entered the analysis. In the two-year retention study, the protocol called for 12 labeled objects per dog, and the published numbers note one dog tested on 11 and another on 5.

None of that is a flaw in the studies; it is what running cognitive tests on animals looks like, and reporting it is good practice. But it means the published performance figures describe a filtered group. Roughly a third of the dogs recruited into a straightforward object-discrimination task could not be brought to the point of being testable on it. Any general statement about "what dogs do" in this area is a statement about dogs who will reliably fetch a named object in a test room, which is already a selected population (operationalizing behavior for measurement).

4. How Rare Is This?

4.1 The Citizen Science Answer

For nearly twenty years the field consisted of single-animal studies — one or two dogs per paper. Dror, Miklósi, Sommese, and Fugazza changed that with a citizen science model: recruit internationally, test the candidates, and describe the group rather than the individual.

Between February 2018 and March 2023 — five years of open international recruitment — they identified and tested 41 dogs that performed above chance on a toy-name comprehension test. Thirty-five of the owners also completed a questionnaire about how the vocabulary was acquired.

Five years. Forty-one dogs, worldwide. That figure is the single most useful number in this literature, and it belongs next to every viral video.

4.2 What the Group Looked Like

The findings were strikingly consistent across the sample.

Breed. 56.1% were Border Collies (23 dogs) and a further 9.8% Border Collie crosses (4 dogs). The remainder included Labradors, Pomeranians, Corgis, and Australian Shepherds. The concentration is obvious, but a breed that is heavily represented among the gifted is not thereby a breed in which the ability is common (breed and behavior).

Vocabulary size. 56% knew twenty or more toy names. The mean was 29 toys (SD ±20), with a confirmed maximum of 86 and a minimum of 5. At a two-year follow-up, 16 owners reported their dogs now knew more than a hundred.

How they learned. This is the part that upends the intuition. 74% acquired the names spontaneously during play, not through deliberate training. And 54% of owners reported their dog picking up a new toy name in five minutes or less.

So the pattern is not "trained hard and got there." It is closer to the opposite: the owners largely did not set out to teach object names, and noticed after the fact that their dog had them. The authors describe the ability as qualitatively different from what typical family dogs show, not as the top end of a continuum of effort.

5. What the Gifted Dogs Can Actually Do

5.1 Learning Rate and Retention

Dror, Miklósi, Sommese, Temesi, and Fugazza tested six Gifted Word Learner dogs across four experiments on how fast new object names are acquired and whether they survive without further exposure. The dogs learned six names in one week in one condition and twelve in another, and were retested after one and after two months with no intervening practice. They typically learned the new names and typically remembered them.

The same group later pushed the interval much further. Five Gifted Word Learner dogs, all Border Collies, were retested on twelve labeled objects two years after the original learning. Group performance was 44% correct, four of the five performed significantly above chance, and individual dogs retrieved between three and nine of the objects correctly. Performance was broadly stable compared with the one- and two-month tests.

Two years, no rehearsal, still above chance. That is a memory result as much as a language result, and it sits alongside what is known about long-term retention in dogs more generally (episodic memory in dogs) and about how dogs handle long intervals (time perception in dogs).

5.2 Categories Defined by Function, Not Appearance

The most recent finding is the most conceptually interesting. Fugazza, Sommese, and Miklósi worked with ten Gifted Word Learner dogs (eleven were recruited; one was excluded for lack of motivation). During a week of naturalistic play at home, the dogs learned two verbal labels — "pull" and "fetch" — each applied to a category of toys that were deliberately dissimilar in appearance. The label tracked what you did with the toy, not what it looked like.

The dogs were then tested with novel toys from both categories. No labels were supplied; the owners simply played with each new toy in the appropriate way, pulling or fetching. The dogs extended the labels to the new toys according to function.

In other words, these dogs sorted objects into categories defined by use rather than by looks, and attached words to those categories — in a species with no language of its own, during ordinary play. Sorting by use rather than by surface features is a form of flexibility that has been examined in dogs in other contexts (behavioral flexibility). The authors are careful to say that the scope and flexibility of this ability is not yet established, and explicitly raise the question of whether dogs who do not learn labels can classify functionally anyway. That question is open.

5.3 What Distinguishes a Gifted Dog

If 74% of these vocabularies arise spontaneously in play, the obvious question is what these dogs are doing differently. Sommese, Miklósi, Nostri, Temesi, and Fugazza compared ten Gifted Word Learner Border Collies with twenty-one typical Border Collies. Dogs were familiarized with toys over two weeks, then given ninety seconds with labeled, unlabeled, and novel toys while their caretaker stayed passive.

Time spent interacting with toys did not differ between the groups. Both groups strongly preferred novel toys. Only the gifted dogs learned labels during familiarization. The one behavioral difference that stood out was social: the gifted dogs were markedly more likely to interact with their caretaker while holding a toy in their mouth — particularly when the toy was new.

That is a small finding with a large implication. The distinguishing feature was not more interest in objects. It was a stronger tendency to bring the object into a joint activity with a person — to make the toy the subject of an exchange. Object names live in that exchange, which connects word learning to the broader literature on how dogs use human attention and social information (how dogs read human gestures, social learning). With ten and twenty-one dogs in a single breed, it is a lead rather than a conclusion.

6. The Ordinary Dog: What the Brain Says

6.1 A Result That Cuts Against the Whole Framing

Everything above concerns a rare group. The question most owners actually have is about their own dog, and until recently the behavioral answer was discouraging: typical dogs tested on object names in the laboratory perform at chance.

Boros, Magyari, Morvai, Hernández-Pérez, Dror, and Andics approached it differently. Instead of asking the dog to choose an object, they recorded brain activity while the dog was shown one. Twenty-seven dogs took part and eighteen were included in the analysis. The owner said a word the dog was reported to know — "Zara, look, the ball" — and then presented either the matching object or a different one, while EEG recorded the response.

The mismatching object produced a different neural response from the matching one, in a pattern resembling the human N400 effect, which in people is widely taken as a marker of semantic processing. The effect was stronger for words the dogs knew better. And — the finding that matters most here — semantic expectations emerged irrespective of vocabulary size. It did not matter how many object words a dog was reported to understand.

6.2 An Earlier Imaging Study That Points Somewhere Else

The EEG result did not arrive into an empty field, and the study that preceded it gives a more restrained picture.

Prichard, Cook, Spivak, Chhibber, and Berns trained twelve dogs to retrieve two objects by name, then scanned them awake in an fMRI and compared the response to those trained words against "oddball" pseudowords the dogs had never heard.

The main effect ran in the direction of novelty rather than of meaning: activation was greater for the pseudowords than for the trained words, bilaterally in parietotemporal cortex. The authors interpret this as novelty detection — the unfamiliar sound is what the brain flags.

They then ran a searchlight multivoxel pattern analysis to look for representations of the trained words themselves. This found clusters of informative voxels distinguishing the two trained words in a subset of the dogs, in left temporal cortex and amygdala, left caudate nucleus, and thalamus. Their stated conclusion is that dogs' processing of human words uses basic processes such as novelty detection, and for some dogs may also include auditory and hedonic representations.

Three things in that are worth holding onto. Novelty detection is a much weaker capacity than semantic representation: registering that a sound is new is not understanding what a familiar sound refers to. Word-specific patterns appeared in only some of the twelve dogs. And the involvement of amygdala and caudate points toward the reward and emotional value attached to a word rather than toward its referent — which is the mechanism the whole field has been trying to rule out (reward learning and canine neurochemistry).

Different methods, samples and comparisons mean the two studies are not in direct contradiction. But read together they place the ordinary dog somewhere less flattering than the EEG headline alone suggests: clear evidence of detecting that a word is unfamiliar, partial evidence of representing which object a familiar word names, and a real possibility that much of what looks like word knowledge is the value a word has acquired.

6.3 How to Read That Honestly

The conclusion the authors draw is that hearing a known object word activates a mental representation of that object, rather than merely triggering an association with an action. That is referential understanding in the relevant sense, and on this evidence it is not confined to the gifted few.

But the tension in the result has to be stated plainly, because the researchers state it themselves: most dogs do not demonstrate this behaviorally. They perform at chance when asked to choose. Why the capacity does not translate into performance is unresolved — and it is the central open question in this area.

The plausible candidates are all about the gap between having a representation and acting on it: attention, motivation, the demands of the test format, the difficulty of inhibiting a competing response (inhibitory control in dogs). None of them has been isolated. What is clear is that "my dog doesn't fetch the right toy" is now poor evidence that the dog does not know the word (why behavior is not a direct readout of an internal state).

6.4 What the EEG Result Does and Does Not License

Because that study is the one most likely to be over-read, it is worth separating what it supports from what it does not.

It supports the claim that a familiar object word activates something about the object it names, not merely a disposition to act. That is what the mismatch response is evidence for, and it is the distinction the study was built to make.

It supports the claim that this is not restricted to exceptional dogs. Vocabulary size did not modulate whether the effect appeared — only how pronounced it was for better-known words.

It does not establish how detailed the representation is. A response to a mismatch shows that the presented object violated an expectation. It does not show whether the expectation was specific ("the blue rope toy"), categorical ("something to tug"), or coarser still.

It does not establish that the dog will act on the word, and the authors say so. The behavioral gap is the finding's companion, not a footnote to it.

And it does not establish that the effect is the same phenomenon as the human one. The similarity is in the pattern of the response, measured with different equipment on a different skull, in a species whose brain organization for this is not well characterized (the neurology of dog behavior). Analogy at the level of a measured signal is a reason to look further, not a demonstration of a shared mechanism.

7. Where Word Learning Breaks Down

7.1 Sound Detail

Magyari, Huszár, Turzó, and Andics developed a procedure for measuring event-related potentials non-invasively in awake dogs and applied it to seventeen animals. The dogs heard three kinds of stimulus: familiar instruction words, nonsense words that were phonetically similar to them, and nonsense words that were phonetically dissimilar.

The responses separated the familiar instructions from the phonetically dissimilar nonsense words, in a late window 650 to 800 milliseconds after word onset and — in one of two artefact-cleaning procedures the authors compared — in an early window of 200 to 300 milliseconds as well. There were no differences at all between the familiar instructions and the phonetically similar nonsense words. Dogs that heard the instructions more often showed larger differences between instructions and dissimilar nonsense words.

Two caveats belong with that. The early effect survived only one of the two cleaning procedures, which the authors report openly, and they close by saying the ERP correlates they describe should be confirmed by further work. The finding of interest is the null one: at the level of processing measured, a phonetically similar nonsense word was not distinguished from a known instruction.

The authors' interpretation is that this reflects limited automatic access to phonetic detail rather than an inability to hear the difference — dogs can discriminate such sounds when the task requires it, which makes this a matter of what they attend to by default (hearing in dogs). They suggest it may be part of why canine vocabularies stay small.

The practical reading is direct. A dog responding correctly to a cue may be responding to its rough sound shape, its intonation, its position in a routine, and whatever else is happening at the same time. Intonation in particular carries information dogs are demonstrably sensitive to (vocal communication in dogs). Cues that are phonetically close to one another are, from the dog's side, at risk of being the same cue.

7.2 Learning a Label From Referential Cues Alone

If dogs pick up object names in play without being taught, the obvious hypothesis is that they are reading the human's referential behavior — working out what the person means by what the person does. Tempelmann, Kaminski, and Tomasello tested that hypothesis directly, and it did not hold up.

They ran three studies, each modelled on an established study with human infants, using four word-experienced dogs. The design removed the easy route: there was no spatial or temporal contiguity between hearing the word and seeing the object, so simple association could not do the work. The only information available was referential action by the human, displaced in time from the object it referred to.

No dog managed to link an object to a label reliably from social-pragmatic cues alone across all the tests. One dog did show some ability in some of the tests, which the authors note as a possible indication that the capacity exists.

Four dogs is a small sample and a negative result in four animals does not establish a general incapacity. But it is the study that tested the appealing explanation most directly, and the appealing explanation did not survive it. Dogs read human referential behavior exceptionally well in other contexts (how dogs read human gestures) — and here, for attaching a word to an object across a gap in time, that skill was not enough.

Which leaves the acquisition mechanism open. The gifted dogs plainly learn names in play; how they do it is not explained by the referential account as tested here.

7.3 Generalization

Children learning object words generalize predominantly by shape: shown a novel object called a "dax," they extend the word to other objects of the same shape rather than the same color or material. This shape bias is one of the best-established regularities in early word learning.

Van der Zee, Zulch, and Mills tested whether a dog does the same. Their subject, Gable, learned a novel label for a novel object and was then offered alternatives matching on shape, size, or texture. He did not generalize by shape. The dimensions he used were not the ones a child would use.

Fugazza, Jacques, Nostri, Kranzelic, Sommese, and Miklósi extended this with a larger sample: 35 dogs completed the study out of 56 recruited, the exclusions being for training failure, insufficient trials, or behavioral biases. Dogs were trained to retrieve a target object, then given trials pitting shape-similar against texture-similar alternatives. The result was a sequence rather than a single bias: dogs first approached objects of similar shape, then predominantly chose objects of similar texture.

Shape got them to the object; texture decided it. That fits an animal that concludes an inspection with its mouth rather than its eyes, and it means a dog's category boundary around a named object is not where a human would draw it (generalization in dogs). Note also the attrition: 21 of 56 dogs could not complete a fairly simple object-discrimination task, which is itself a data point about typical dogs and object work.

7.4 What the Dog Learns Instead of the Word

If phonetic detail is not attended to by default and object categories are drawn on unexpected dimensions, the question becomes what a dog responding correctly to a spoken cue has actually learned.

The plausible answer is a compound. Part of it is the rough sound shape of the word — its length, rhythm, and stress rather than its individual phonemes. Part of it is intonation, which dogs are demonstrably sensitive to. Part of it is the posture, gaze, and movement of the person speaking. Part of it is where in a sequence of events the word arrives: after the lead comes off, before food, at the door. And part of it is the physical setting, which is why cues so often fail on relocation (generalization in dogs).

None of those components is an error on the dog's part. Attending to the whole situation rather than to one acoustic feature of it is a sensible strategy for an animal for whom the situation is usually more informative than the sound. It becomes a problem only when a human assumes the word is doing the work alone, and then removes everything else.

The practical test for whether a cue is a cue: does it still work when the handler is seated, silent otherwise, facing away, in a different room of the house, and at a point in the day when nothing else predicts it? Most cues that are reported as reliable fail at least one of those, and the failure locates what the dog was actually using.

8. What This Means for Training

8.1 Names Are Cheap, Discrimination Is Not

Teaching a dog to fetch one named object is usually straightforward. Teaching it to distinguish that object from a second named object is a different task, and it is where most attempts stall. The discrimination — not the association — is the work, and it depends on the dog reliably being told which of two choices was right (prediction error and learning).

8.2 Choose Cues That Sound Different

Given the phonetic finding, cue words should differ from one another by more than a single sound, and ideally in rhythm, stress and length as well. Pairs that look clearly distinct written down can be phonetically close, and a human ear reading them on a page is a poor guide to whether a dog will treat them as two cues.

8.3 Expect the Category to Be Wrong

If a dog learns the name of one ball, it may extend that name to things that feel similar rather than things that look similar. Anyone teaching object names should plan to test the boundary deliberately with objects that vary on one dimension at a time, rather than assume the dog has divided the world the way the trainer has.

8.4 Use Play, Not Drills

Three quarters of the documented gifted vocabularies were acquired spontaneously during play. Whatever mechanism is at work, it operated in joint activity with a person and without an explicit teaching protocol. And the one behavioral marker that distinguished gifted dogs from typical ones was bringing the object to the human. Someone wanting to try object names has a better prospect making the toy the subject of a shared game than running formal trials. Play is worth arranging for its own sake in any case (the neurobiology of play), and joint activity of this kind sits squarely in the behavioral-interactions side of welfare assessment (welfare science and quality of life).

8.5 Do Not Diagnose Intelligence From It

A dog that does not learn object names is not thereby a less capable dog. Object-label learning is one narrow skill, concentrated in one breed group, and the neural evidence suggests the underlying representational capacity is widespread even where the behavior is absent. Individual dogs differ along several partly independent dimensions (temperament and coping styles). Canine cognition is not a single scale (cognitive abilities in dogs), and the comparisons that matter are within an individual over time, not between dogs (operationalizing behavior for measurement).

8.6 Beware the Clever Hans Problem

Section 3.2 applies to any home demonstration: if the person asking knows the answer and the dog can see them, the result does not distinguish word knowledge from cue reading — the same trap that made one famous canine behavior look like guilt (the guilty look).

8.7 If You Want to Try It

Nothing in this literature amounts to a validated teaching protocol, and the honest framing is that the documented cases largely happened rather than being produced. But the conditions under which they happened are described well enough to be imitated.

Use toys the dog already wants, and let it choose which ones. Use the name while the dog is engaged with that object, not as an instruction to go and find it. Keep the set small at first and physically and texturally distinct — not two ropes of different colors. Say the name as part of ordinary talk during the game rather than in a training voice. Let the dog bring the toy to you and make the arrival of the toy the interesting part, since that exchange is the one behavior that separated gifted dogs from typical ones.

Then test it under the conditions in section 3.2, and record correct and incorrect trials rather than an impression. If the dog is at chance, that is information about the test, not a verdict on the dog.

9. Summary at a Glance

The case studies are real and were carefully controlled. Rico worked out new referents by exclusion; Chaser learned 1,022 object names, kept names and commands independent under double-blind testing, handled category labels, and inferred novel names by exclusion.

Chaser's vocabulary took three years at four to five hours a day. The training load is part of the result and is usually omitted.

Exclusion learning is not durable learning. Chaser's inferred name-object pairings were strong immediately, reduced after ten minutes, and minimal at twenty-four hours without rehearsal.

The ability is rare. Five years of open international recruitment produced 41 dogs that tested above chance — 56.1% Border Collies plus 9.8% Border Collie crosses.

Most of these vocabularies were not taught. 74% were acquired spontaneously in play; 54% of owners reported new names learned in five minutes or less; the mean vocabulary was 29 toys.

Gifted dogs remember for years. Five dogs retested on twelve objects after two years without rehearsal averaged 44% correct, with four of five above chance.

They can form categories by function. Ten gifted dogs extended labels learned for "pull" and "fetch" toys to novel toys according to how the toys were used, not how they looked.

What sets them apart may be social, not object-directed. Gifted and typical Border Collies did not differ in time spent with toys; the gifted dogs were more likely to bring a toy to their person, especially a new one.

Ordinary dogs show referential processing. An N400-like response to mismatching objects appeared in dogs regardless of reported vocabulary size — while most dogs still perform at chance behaviorally.

Phonetic detail is not attended to by default. In seventeen dogs, neural responses separated known instructions from phonetically dissimilar nonsense words but showed no difference at all from phonetically similar ones.

Imaging gives a more restrained picture than EEG. In twelve dogs scanned awake, activation was greater for unheard pseudowords than for trained words — novelty detection — and word-specific patterns appeared in only a subset of the dogs.

Reading the human's reference was not enough. Four word-experienced dogs tested with referential cues displaced in time failed to link label to object reliably; one showed partial ability.

Dogs do not generalize like children. No shape bias in the single-subject test; in the larger study, dogs approached by shape and then chose by texture.

Published samples are filtered. In the shape-and-texture study 56 dogs were recruited and 35 completed; the EEG study analyzed 18 of 27. Performance figures describe dogs who reached testability.

Blind testing separates word knowledge from cue reading. The credible studies are double-blind or measure the brain rather than the choice; a demonstration in which the person asking knows the answer does not distinguish the two.

10. Research Gaps

The performance gap is unexplained. Typical dogs show neural signatures of referential understanding and behavioral performance at chance. Whether that is attention, motivation, response competition, or test format has not been isolated, and it is the most consequential open question here.

Nobody knows the prevalence. Forty-one dogs from five years of self-selected international recruitment gives no denominator. The ability is clearly rare; how rare is not a number anyone can currently state.

The samples are small and one breed dominates. Five, six, ten dogs per experiment, mostly Border Collies. Whether the findings describe gifted dogs or gifted Border Collies is not settled.

Functional categorization has not been tested in typical dogs. The authors of the 2025 study raise this themselves: it is unknown whether dogs that do not learn labels classify objects by function anyway.

Cause and direction are unclear in the social finding. Gifted dogs bring toys to their people more. Whether that disposition enables label learning, results from a history of it, or reflects something both share is not established by a cross-sectional comparison.

The two imaging literatures have not been reconciled. An fMRI study found novelty detection as the dominant effect with word-specific patterns in only some dogs; an EEG study found a mismatch response independent of vocabulary size. Different methods and samples, and no study has yet tested them against each other.

How the gifted dogs acquire names is unexplained. The referential account was tested and largely failed; no alternative mechanism has been established.

The grain of the representation is unknown. A mismatch response shows that an expectation was violated. Whether the expectation was of a specific object, a functional category, or something coarser has not been resolved.

Almost nothing is known about words for anything but objects. Actions, places, people, properties — the literature is overwhelmingly about nouns for toys, because that is what the paradigm affords.

11. Conclusion

The honest summary is a split one. A small number of dogs — overwhelmingly Border Collies, identified in the dozens rather than the thousands after years of looking — acquire object names at a speed and in quantities that invite comparison with toddlers, retain them for years without practice, and can sort objects into functional categories and attach words to those. That is a genuine finding and it has survived double-blind testing since 2004.

The average dog does not do this. What the average dog appears to do is something less visible and in some ways more interesting: hearing a familiar object word activates a representation of that object, whether or not the dog will then go and get it. The behavior and the representation have come apart, and the research has only recently been able to see the difference.

For anyone training a dog, the useful conclusions are not about vocabulary size. Cue words need to sound genuinely different; a dog's idea of what a name covers will not match a human's; discrimination between names is the real work; and the context surrounding a word is doing more than the word is. A dog who never learns the name of a single toy has told you nothing about how much it understands.

Key Insights

Extraordinary dogs show what is possible, not what is typical. Rico and Chaser establish an upper bound. Five years of international recruitment turning up 41 dogs establishes the rarity.

Chaser's 1,022 names cost three years of daily hours. Cite the training load whenever citing the number.

Working out which object is new is not learning its name. Inferred pairings were largely gone within a day without rehearsal.

Most gifted vocabularies were never taught. Three quarters emerged in play, often within minutes — which means drilling is probably the wrong model of how it happens.

Your dog may know more words than it will show you. Neural evidence of referential understanding appeared independently of vocabulary size, in dogs that perform at chance when asked to choose.

A word that sounds close is not a different word. Dogs' default processing did not separate familiar instructions from phonetically similar nonsense words at all.

Dogs finish deciding with their mouth. They approached by shape and chose by texture — so a named object's category will not have a human's boundaries.

Bring the object into a shared game. The behavior that distinguished gifted dogs was offering the toy to their person, not playing with it longer.

Test blind or do not claim a result. If the person asking knows which toy is right, the demonstration shows less than it seems to.

Some of what looks like word knowledge may be value, not reference. The imaging work found word-related patterns in the amygdala and caudate, which points at what a word is worth to the dog rather than at what it names.

Object names are not an intelligence test. One narrow skill, concentrated in one breed group, with the underlying capacity apparently widespread regardless.

References

Boros, M., Magyari, L., Morvai, B., Hernández-Pérez, R., Dror, S., & Andics, A. (2024). Neural evidence for referential understanding of object words in dogs. Current Biology, 34(8), 1750–1754.e4. https://doi.org/10.1016/j.cub.2024.02.029

Dror, S., Miklósi, Á., & Fugazza, C. (2024). Dogs with a vocabulary of object labels retain them for at least 2 years. Biology Letters, 20(9), 20240208. https://doi.org/10.1098/rsbl.2024.0208

Dror, S., Miklósi, Á., Sommese, A., & Fugazza, C. (2023). A citizen science model turns anecdotes into evidence by revealing similar characteristics among Gifted Word Learner dogs. Scientific Reports, 13, 21747. https://doi.org/10.1038/s41598-023-47864-5

Dror, S., Miklósi, Á., Sommese, A., Temesi, A., & Fugazza, C. (2021). Acquisition and long-term memory of object names in a sample of Gifted Word Learner dogs. Royal Society Open Science, 8(10), 210976. https://doi.org/10.1098/rsos.210976

Fugazza, C., Jacques, E., Nostri, S., Kranzelic, U., Sommese, A., & Miklósi, Á. (2024). Shape and texture biases in dogs’ generalization of trained objects. Scientific Reports, 14, 28077. https://doi.org/10.1038/s41598-024-72244-y

Fugazza, C., Sommese, A., & Miklósi, Á. (2025). Dogs extend verbal labels for functional classification of objects. Current Biology, 35(19), 4820–4826.e3. https://doi.org/10.1016/j.cub.2025.08.013

Kaminski, J., Call, J., & Fischer, J. (2004). Word learning in a domestic dog: Evidence for ‘fast mapping’. Science, 304, 1682–1683. https://doi.org/10.1126/science.1097859

Magyari, L., Huszár, Z., Turzó, A., & Andics, A. (2020). Event-related potentials reveal limited readiness to access phonetic details during word processing in dogs. Royal Society Open Science, 7(12), 200851. https://doi.org/10.1098/rsos.200851

Pilley, J. W., & Reid, A. K. (2011). Border collie comprehends object names as verbal referents. Behavioural Processes, 86, 184–195. https://doi.org/10.1016/j.beproc.2010.11.007

Prichard, A., Cook, P. F., Spivak, M., Chhibber, R., & Berns, G. S. (2018). Awake fMRI reveals brain regions for novel word detection in dogs. Frontiers in Neuroscience, 12, 737. https://doi.org/10.3389/fnins.2018.00737

Sommese, A., Miklósi, Á., Nostri, S., Temesi, A., & Fugazza, C. (2026). Toy exploration in gifted word learner dogs and typical dogs. Animal Cognition, 29(1), 24. https://doi.org/10.1007/s10071-026-02047-3

Tempelmann, S., Kaminski, J., & Tomasello, M. (2014). Do domestic dogs learn words based on humans’ referential behaviour? PLoS ONE, 9(3), e91014. https://doi.org/10.1371/journal.pone.0091014

van der Zee, E., Zulch, H., & Mills, D. (2012). Word generalization by a dog (Canis familiaris): Is shape important? PLoS ONE, 7, e49382. https://doi.org/10.1371/journal.pone.0049382