Generalization in Dogs: Why a Learned Behavior Is Not the Same Everywhere

1. What Generalization Is

1.1 The Operational Definition

Generalization is what happens when a response learned to one stimulus appears in the presence of a different stimulus that resembles it. A dog trained to sit to a spoken word sits when a different person says the same word, or when the word is said more quietly, or when it is said in a room the dog has never been in. The response has spread beyond the exact conditions under which it was acquired.

That is a description of behavior and not of understanding. It says nothing about whether the dog has grasped a concept, formed a category, or worked out what the handler meant. It says that a response occurred under conditions that differ from the training conditions, and that the size of the difference and the strength of the response can both be measured.

The distinction matters because most disputes about generalization in training are not disputes about the data. They are disputes about what to call the dog when the response fails to appear.

1.2 The Gradient

Responding does not stop abruptly at the edge of the training stimulus. It tapers. The more a new stimulus resembles the trained one, the more likely the response, and the further away it sits along whatever dimension is relevant, the weaker the responding becomes. That taper is the generalization gradient, and it is the oldest quantitative finding in the area.

Two things follow from a gradient that do not follow from a category. The first is that generalization is a matter of degree rather than a switch: a dog does not either transfer a behavior or fail to transfer it, but responds more or less reliably as conditions shift. The second is that the shape of the gradient is itself trainable. Narrow gradients are produced by training that punishes responding to anything but the exact stimulus; broad gradients are produced by training across variation.

A limitation should be stated here rather than buried: the gradient work that established these principles was done in pigeons and rodents, and the canine literature contains no published generalization gradient for any stimulus dimension. What follows in this article about dogs rests on canine studies of discrimination, transfer and context specificity, which are related but not the same thing.

1.3 Generalization and Discrimination Are One System

The complement of generalization is discrimination: responding differently to stimuli that differ. Every training goal sits somewhere between the two. A recall should generalize across locations, distances, distractions and handlers, and should not generalize to the word "come here" spoken to another dog in the park. A muzzle-touch cue should generalize to a different muzzle and should not generalize to every object held near the face.

These are not two skills to be trained separately. They are two readings of the same underlying process. Widening a gradient makes false alarms more likely; narrowing it makes failures under changed conditions more likely. Anyone who has trained a scent-detection dog has met this trade-off in its purest form, and anyone who has trained a household recall has met it without recognizing it as one.

1.4 What Generalization Is Not

It is not obedience, and the failure of a behavior to transfer is not a moral event. A dog that sits reliably in the kitchen and fails to sit at the park has not decided anything. Its behavior is under the control of a stimulus complex that included the kitchen, and the park is a different complex.

This is worth being firm about, because the interpretive error here is expensive. When the dog fails, owners commonly report that the dog knew what to do and chose not to. Attributing knowledge and refusal to the dog is the same class of inference that produces other confident misreadings of canine behavior (the guilty look being the best-documented example), and it leads directly to interventions aimed at the dog's supposed attitude rather than at the training conditions that produced the gap.

It is also not memory failure. A dog that fails at the park may perform perfectly the moment it is back in the kitchen, which rules out forgetting and leaves stimulus control as the explanation.

2. The Canine Evidence That Behavior Stays Where It Was Learned

2.1 Inhibitory Control Does Not Travel Between Tasks

The clearest canine demonstration comes from a study that set out to test whether inhibitory control is a stable trait. Bray, MacLean and Hare (2014) ran dogs through multiple tasks that all required withholding a prepotent response, and examined whether performance on one predicted performance on another. It largely did not. Performance was specific to the task and its conditions rather than reflecting a general capacity the dog carried from one situation to the next.

The finding is inconvenient for the way self-control is usually discussed, because it means a dog described as having good impulse control has demonstrated something narrower than that phrase implies (which the wider self-control literature has to accommodate). It is also directly relevant to generalization: if a capacity as basic as withholding a response is bound to its context, the transfer of a trained behavior across contexts cannot be assumed.

A companion finding from the same group sharpens it further. Bray, MacLean and Hare (2015) reported that raising arousal improved inhibitory control in calm dogs and impaired it in excitable ones, which means the direction of an arousal effect depends on the individual. The same manipulation moves two dogs in opposite directions.

2.2 Aggression Does Not Co-occur Across Contexts

The largest owner-reported survey of human-directed aggression asked whether dogs showing aggression in one context also showed it in others. Mostly they did not. Aggression toward unfamiliar people entering the house, aggression toward family members and aggression toward unfamiliar people outside were not broadly correlated within individuals, which the authors read as evidence that the behavior is learned in relation to specific situations rather than expressing a general characteristic of the dog (Casey et al., 2014).

Read as a statement about generalization, this is one of the strongest results available. A behavior with high emotional loading, considerable reinforcement history and obvious survival relevance still fails to spread across contexts in most individuals (which is why aggression is better read by function than by label). The data are owner-reported and cross-sectional, so they establish a pattern and not a mechanism, but the pattern is the one that matters here.

2.3 Extinction Is Bound to the Context It Happened In

The most thoroughly worked-out case of context specificity is extinction. When a conditioned response is extinguished in one context and the animal is tested in another, the response returns. Bouton (2002, 2004) developed the account behind this: extinction does not erase the original learning but adds new, context-dependent learning that competes with it. Change the context and the original learning is again expressed.

This work is predominantly rodent and human, and the mechanism should be marked as such. Its canine relevance is indirect but substantial, because the applied literature in dogs describes exactly the pattern the account predicts: behavior that has disappeared in the training setting reappears elsewhere or later (a return that is routinely misread as the work having been undone).

What dogs contribute directly is the resistance side. Hall (2017) reviewed persistence and resistance to extinction in dogs, with applications to training, and the practical implication is uncomfortable: the same intermittent history that makes a trained behavior durable makes an unwanted behavior durable.

2.4 What the Three Findings Share

Three literatures with almost nothing else in common — a cognitive task battery, an epidemiological survey and a learning-theory account of relapse — converge on one statement. Behavior is acquired together with the conditions under which it was acquired, and those conditions remain part of what controls it.

The convergence is what makes the claim usable. Any single study would be a finding about a paradigm. Three independent methods producing the same structure is a property of the phenomenon.

3. What Dogs Do Generalize, and How Far

3.1 Perceptual Rules

Dogs do generalize, and the studies that show it are informative about the limits. Müller and colleagues (2014) tested dogs on the support problem, in which food sits either on a board that can be pulled within reach or beside it. Dogs learned to solve the task, and the analysis indicated that the solution rested on perceptual cues — the visible spatial relationship between food and board — rather than on an understanding of physical connection.

That is a real generalization: the dogs applied what they had learned to new configurations. It is also a bounded one. A solution built on a perceptual cue transfers to situations sharing that cue and fails where the cue is absent, which is why the same dog can look insightful on one variant and lost on the next.

3.2 Emotional Expressions and Novel Faces

A cleaner transfer test comes from Müller and colleagues (2015), who trained dogs to discriminate happy from angry human faces using only the upper or only the lower half of the face, then tested them on novel faces and on the untrained half. The dogs transferred the discrimination to faces they had never seen and to the face region they had not been trained on.

This is the strongest evidence in the canine literature that dogs can extract something abstract enough to transfer across substantial changes in the physical stimulus. It should not be inflated: the dogs were trained extensively on a two-choice discrimination under laboratory conditions, and what transferred was a discrimination, not an emotional understanding (a distinction that has to be held throughout this material).

3.3 Words and Fast Mapping

Kaminski, Call and Fischer (2004) documented a Border Collie that retrieved more than two hundred objects by name and inferred the referent of a new word by exclusion, selecting the unfamiliar object when given an unfamiliar label. The dog retained several of these mappings after four weeks.

The result is a genuine transfer of a learning principle to novel material. It is also a single animal with an exceptional training history, and the study has been cited far beyond what one case supports. What it does establish is that the ceiling for canine generalization of arbitrary labels is higher than the median dog's performance suggests — and that reaching it takes years of structured work.

3.4 Social Information

Dogs also transfer information acquired by watching. Fugazza and Miklósi (2014) showed that dogs trained with a demonstrate-then-imitate procedure could reproduce a demonstrated action after a delay, indicating that what was retained was not a motor pattern executed immediately but something recallable later.

Whether this extends to the transfer of trained behaviors across settings is untested. It belongs in this article as evidence that dogs retain and re-apply information across a gap, which is a precondition for generalization rather than a demonstration of it.

4. Discrimination: The Other Half of the System

4.1 What Discrimination Studies Measure

Canine discrimination learning has been studied mainly to track cognitive aging rather than to answer training questions, which shapes what is available. Head and colleagues (1998) used visual discrimination tasks in dogs and related performance to beta-amyloid accumulation, establishing the paradigm that most later work uses: two stimuli, one rewarded, learning to criterion, then a reversal in which the contingencies swap.

The paradigm is useful and narrow. It measures how quickly a dog comes to respond to one stimulus and not another under stable laboratory conditions. It does not measure whether the resulting discrimination survives a change of room, handler or day, which is the question practice actually asks.

4.2 Age Changes the Picture

Piotti and colleagues (2018) tested family dogs on discrimination learning, reversal learning and cognitive bias. Young dogs learned both the discrimination and the reversal significantly faster than old dogs. In the reversal, dogs trained on the location of the stimuli learned faster than dogs trained on the characteristics of the stimuli, and most old dogs did not learn the reversal at all within the cut-off of fifty trials.

The second half of that finding is the more useful one. The dimension along which a discrimination is built changes how difficult it is, and the ranking of dimensions is not the same for every dog. A cue that works because the dog is using position will fail when position changes, and the handler who built it may not know which dimension the dog was using (a problem the flexibility literature runs into from the other direction).

A citation correction belongs here, because it recurs in secondary sources. Wallis and colleagues (2016) is frequently cited as a canine reversal-learning study. It is not. It examined discrimination learning, logical reasoning and memory in pet dogs across the lifespan, and its bearing on flexibility runs through perseverative responding in the discrimination tasks rather than through a reversal paradigm.

4.3 Does Recent Discrimination Training Change Later Judgment?

A plausible worry follows from everything above. If a discrimination narrows a gradient, a dog trained on a discrimination shortly before a judgement bias test should treat the ambiguous probes as belonging to the unrewarded class, and the test would then be measuring learning history rather than mood.

The worry has been tested once, and it was not supported. Krahn and colleagues (2024) gave companion dogs a discrimination treatment between two judgement bias tests and found no difference in the change in approach latency to the ambiguous locations relative to a yoked control group that received the same number of rewarded and unrewarded trials. Optimism, as measured in that test, was not altered by the additional discrimination trials.

The final sample was sixteen dogs, split eight and eight, which is small enough that a real effect of moderate size could have gone undetected. The honest summary is that the mechanism is plausible, that one controlled attempt to demonstrate it came back null, and that how recent training history contaminates judgement bias testing remains open (which matters for how judgement bias results are read generally).

4.4 Overgeneralization

The failure mode in the other direction gets less attention in training and more in the clinic. A fear response acquired to one stimulus can appear to a widening set of stimuli that share features with it: one firework becomes all sudden sounds, then all evenings in autumn, then the room in which it happened (which is the developmental course noise sensitivity typically follows).

Nothing about this is pathological in mechanism. It is the same gradient running wide, in a system where the cost of a false alarm is low and the cost of a miss is high. The asymmetry is the point: threat-related learning can generalize broadly, while safety learning often requires more deliberate transfer across contexts, and any account that treats the two as symmetrical will mislead in exactly the cases that matter clinically (a bias built into how fear memories are formed and updated).

5. Where Transfer Predictably Fails

5.1 Location

The most familiar failure. A behavior trained in one place is under the control of a stimulus complex that includes the floor surface, the light, the acoustics, the smells, the distance to walls and the absence of other animals. Moving the dog changes most of these at once.

The practical consequence is that "the dog knows it at home" is not evidence that the dog knows it. It is evidence that the behavior has been trained in one context and not yet tested elsewhere.

5.2 The Handler

People are part of the stimulus. Posture, height, the hand that holds the food, the direction of gaze, the timing of the marker and the sound of the voice all vary between handlers, and a behavior trained by one person frequently fails with another. In a household this shows up as a dog that responds to one family member and not to others, which is usually attributed to authority and is more parsimoniously explained by the training history.

Human-directed cues carry the same problem. Dogs are unusually good at following human pointing (a well-replicated finding with its own interpretive disputes), but responsiveness drops when the form of the gesture changes beyond what the dog has encountered.

5.3 Distance, Intensity, Duration

Three dimensions that are usually varied last and should be varied early. A recall trained at five meters is a different response from a recall at fifty. A settle trained for twenty seconds is not a settle for twenty minutes. A stimulus tolerated at low intensity is not tolerated at high intensity, which is the entire structure of systematic desensitization and the reason it is built as a graded series rather than a single exposure (where the graded structure does specific work).

5.4 The Dog's Internal State

Internal conditions belong in the stimulus complex, and they are the part handlers cannot see. Arousal alters what a dog can do, and it does so in a direction that depends on the individual: raising arousal improved inhibitory control in calm dogs and impaired it in excitable ones in the study cited above (Bray, MacLean & Hare, 2015), which is a warning against any general rule about optimal arousal (including the inverted-U account and what the canine data actually supports).

Pain does the same thing less visibly. A dog in discomfort has a different threshold, a different tolerance for handling and a different capacity to sustain a trained response, and the behavior change is frequently the first sign (with pain heavily represented in behavior caseloads). A behavior that generalized last month and fails this month is a medical question before it is a training question.

5.5 Age and Developmental Stage

Transfer is not stable across the lifespan. Discrimination and reversal performance decline with age (Piotti et al., 2018), and at the other end, adolescence brings changes that owners routinely experience as a trained dog coming apart (where the developmental evidence supports less than the popular account claims). Neither is a training failure. Both are conditions under which previously reliable behavior needs retraining rather than enforcement.

6. The Practical Consequence

6.1 Every New Context Is at Minimum a Re-Test

The single operational conclusion from everything above is that a behavior is trained for a set of conditions, and each new set is either a new training situation or, at best, a test whose result cannot be predicted. This is not a counsel of despair. Transfer becomes faster with each context added, because what the dog is learning across contexts is which features are relevant and which are not.

The mistake is to skip the step. A behavior demonstrated in three locations has been tested in three locations, and the fourth is still an open question — a smaller one each time, but open.

6.2 Train Across Variation, Deliberately

If a gradient is shaped by the variation present during acquisition, then the variation should be planned rather than incidental. Vary the location, the handler, the position of the handler's body, the distance, the surface, the time of day, the presence of other animals, and the reinforcer — one dimension at a time where practical, so that a failure identifies which dimension carried the control.

There is a theoretical argument that variability during learning has value beyond the specific variations trained. Špinka, Newberry and Bekoff (2001) proposed that mammalian play functions as training for the unexpected: it generates variable, partly uncontrolled sequences that may prepare an animal for situations it has not encountered. The argument is comparative and evolutionary rather than experimental, and it should be treated as a hypothesis that fits the data rather than a demonstrated training principle.

6.3 What Post-Training Conditions Contribute

Two canine findings suggest that what happens after a session affects what is available later. Affenzeller, Palme and Zulch (2017) reported that a short bout of play after a training session improved performance when Labrador Retrievers were retested, and in family dogs with higher ADHD-like scores, reversal performance improved from before to after a nap in those that spent a substantial part of the recording actually asleep (Kovács et al., 2025).

These are consolidation effects rather than generalization effects, and the distinction should be kept. A behavior can be well consolidated and still context-bound. But a behavior that has not consolidated will not transfer either, which makes sleep and post-session conditions a precondition rather than a shortcut (as the sleep and memory work sets out in detail).

6.4 Reinforcement History Shapes the Gradient

Schedules matter here for a reason that is often stated backwards. Intermittent reinforcement produces responding that persists when reinforcement stops (Humphreys, 1939), and the frustration generated by non-reward during a partially reinforced history is itself part of what builds that persistence (Amsel, 1962). Persistence is not the same as transfer, but a behavior that collapses at the first unreinforced repetition in a new context will never get far enough to transfer (which is the practical case for schedule design).

The counterweight is canine and recent. Cimarelli and colleagues (2021) found that partial reinforcement during clicker training did not improve learning speed in naive dogs and was associated with a more pessimistic-like affective state. Thinning reinforcement early, in an animal that has not yet learned the behavior, buys persistence at a cost and possibly buys nothing at all.

Two further pieces are worth holding together. Extinction bursts are less common than the training literature implies — a review of applied cases found them in a minority of treatments (Lerman & Iwata, 1995) — and the value of a reinforcer is itself context-dependent: food outcompeted social interaction as a reinforcer across the conditions tested in dogs and hand-reared wolves (Feuerbacher & Wynne, 2012). A reinforcer that works in the kitchen may not compete with the environment at the park, and the resulting failure looks like a generalization problem while being a reinforcement problem.

7. Unwanted Behavior Generalizes Too

7.1 Predatory Sequences Across Targets

The clearest applied case is chasing. A dog that has run deer will run cyclists, because the element shared between them is movement of a particular speed and direction rather than the species involved. The systematic review of interventions for predatory behavior found, among four consistent limitations across the available studies, that generalization was not demonstrated: the studies that exist do not show that training transfers across targets, environments or stimulus types (McLennan, 2023).

That cuts both ways, and the asymmetry is unpleasant. The unwanted behavior transfers readily across targets, while the training intended to address it has not been shown to transfer at all (which is the honest state of the predatory intervention evidence). A dog that has learned to leave a rabbit in a field has not thereby learned to leave a cat in a garden.

7.2 Guarding Across Items and People

Resource guarding shows a similar structure with a different dimension. What generalizes is not the object but the sequence of events: an approach that predicts loss. Where that contingency has been rehearsed, guarding tends to appear with new items and new people that share the predictive feature, and where it has not, guarding can remain remarkably specific to one item in one location.

The practical reading is that the relevant dimension has to be identified before anything can be trained. Two dogs guarding a bowl may be doing so along entirely different dimensions, and a protocol built for one will fail for the other.

7.3 Punishment Generalizes to Whatever Was Present

An aversive event does not attach itself only to the behavior the handler intended to suppress. It attaches to whatever the dog perceives at that moment, which regularly includes the handler, the location, the equipment and any conspicuous stimulus in view.

The canine evidence on this is unusually consistent. In a controlled study of electronic collars, the group for which the shock bore no clear relationship to the dog's own behavior showed the highest cortisol response, and poor timing produced fear and insecurity rather than learning (Schalke et al., 2007). A broader review of aversive methods found associations with stress and welfare costs across studies (Ziv, 2017). A controlled comparison of training approaches found no efficacy advantage for remote electronic collars over reward-based training (China, Mills & Cooper, 2020), and companion dogs trained with aversive-based methods showed poorer welfare indicators than dogs trained with reward-based methods (Vieira de Castro et al., 2020).

Read through the lens of this article, the fallout is a generalization problem. The suppression is narrow and the association is wide (which is what the neurological work on these methods describes from the mechanism side). That is the worst possible combination of gradients, and it is produced reliably.

7.4 The Asymmetry, Stated Plainly

Across the cases in this section, threat-related learning tends to generalize more readily than appetitive and safety learning. A single frightening event can spread across a category of stimuli; a hundred pleasant repetitions in the kitchen may not survive the trip to the park.

This is not a claim that fear learning is stronger in some general sense, and the canine evidence is not strong enough to put a number on the difference. It is a claim about gradient width under threat, where the cost structure favors false alarms, and its practical consequence is that broad appetitive generalization has to be built deliberately while broad aversive generalization can arrive by accident.

8. Testing for Transfer Instead of Assuming It

8.1 What a Transfer Test Requires

A transfer test is not a repetition in a new place. It requires that the new conditions differ on a specified dimension, that the criterion is set in advance, and that the result is recorded whether or not it is flattering. Without those three, what happens is a demonstration, and a demonstration cannot fail.

The minimum record is the dimension varied, the value at which the behavior was trained, the value at which it was tested, and the outcome. Four entries. Anything less does not permit the comparison that would show whether progress is occurring, which is the same measurement problem the field has everywhere else.

8.2 Habituation as the Model Case

The best-specified account of stimulus generalization in the behavioral literature is not from training but from habituation research. Thompson and Spencer (1966) set out the defining characteristics of habituation, and Rankin and colleagues (2009) revised and updated that description. Stimulus generalization appears in both as a criterial property: habituation to one stimulus transfers to similar stimuli, and the extent of transfer is one of the features that distinguishes habituation from fatigue or sensory adaptation.

Two things follow for practice. The first is that if a dog stops responding to a stimulus and the reduced responding does not transfer at all to similar stimuli, habituation may not be what happened. The second is that this criterion is testable in a session, which makes it one of the few generalization measurements available to a practitioner without laboratory equipment.

8.3 The Failure That Looks Like Success

Worth naming explicitly: a dog that performs perfectly in every training session and fails in the situation the training was for has not learned a behavior with a narrow gradient by accident. The training produced exactly what it selected for. Sessions are highly consistent environments, and consistency is the condition under which narrow gradients form.

The counterintuitive implication is that a training program in which the dog succeeds every time is a program that is probably building context-bound behavior. Introducing enough variation to produce occasional failure is what widens the gradient, provided the failures are used as information about dimensions rather than treated as disobedience.

9. Summary at a Glance

Generalization is a gradient, not a switch. Responding tapers as conditions depart from the training conditions, and the width of that taper is a product of the training history.

Canine behavior is context-specific to a degree that surprises most handlers. Inhibitory control did not transfer between tasks (Bray, MacLean & Hare, 2014), and aggression did not co-occur across contexts in a large survey (Casey et al., 2014).

Extinction is contextually bound. Behavior that has disappeared in one setting can return in another, because extinction adds context-dependent learning rather than erasing the original (Bouton, 2002, 2004).

Dogs can transfer abstract discriminations, within limits. Trained on half a face, dogs transferred a happy–angry discrimination to novel faces and the untrained half (Müller et al., 2015).

The dimension the dog is using is frequently not the one the handler intended. In reversal learning, dogs trained on the location of the stimuli learned faster than dogs trained on their characteristics (Piotti et al., 2018), and the dimension actually in use is rarely recorded.

Unwanted behavior generalizes more readily than its treatment. Predatory sequences transfer across targets while interventions have not been shown to transfer at all (McLennan, 2023).

Aversive events generalize widely and suppress narrowly. Unpredictable shock produced the strongest stress response in electronic-collar training (Schalke et al., 2007), with welfare costs documented across the wider literature (Ziv, 2017; Vieira de Castro et al., 2020).

10. Research Gaps and Critical Appraisal

There is no canine generalization gradient in the published literature. The quantitative work that defines the concept was done in other species, and no study has established the shape of a generalization gradient in dogs along any stimulus dimension. Every statement in this article about gradient width in dogs is an inference from discrimination and transfer studies rather than a direct measurement.

Canine discrimination research was built for a different question. The paradigms come largely from cognitive aging work (Head et al., 1998; Wallis et al., 2016; Piotti et al., 2018), where the aim is to detect decline rather than to characterize transfer. Laboratory two-choice tasks under stable conditions are close to the opposite of the situation practice asks about.

The strongest transfer findings rest on small, highly trained samples. The face-discrimination transfer used extensively trained laboratory dogs (Müller et al., 2015), and the word-learning result is a single exceptional animal (Kaminski, Call & Fischer, 2004). Neither supports claims about what an average pet dog will do.

The context-specificity account of extinction is imported. Bouton's work (2002, 2004) is predominantly rodent and human. Its canine application is a reasonable inference supported by convergent applied observations, and it has not been tested directly in dogs at anything like the same resolution.

Few canine studies measure generalization as a primary outcome. The intervention literature reviewed for predatory behavior identified the absence of generalization data as one of four systematic limitations (McLennan, 2023), and that criticism applies well beyond predatory behavior. Training studies typically report acquisition, and transfer is reported far less often.

Recommendations here rest on principle rather than trials. Varying one dimension at a time where practical, planning variation and testing transfer are derived from the structure of the phenomenon and from general learning research. No controlled canine trial has compared a systematically varied training protocol against a single-context protocol on a transfer outcome. That study does not exist, and it would be straightforward to run.

11. Conclusion

The most useful thing to say about generalization is that its absence is the normal case and its presence is the achievement. Behavior is acquired with its conditions attached, and the conditions do not fall away because the handler considers them irrelevant. What looks like a dog that knows a behavior and withholds it is almost always a dog whose behavior is controlled by a stimulus complex that the current situation no longer matches.

That reframing has one practical consequence, and it is not a technique. It is a habit of asking, before any conclusion about a dog's reliability, which conditions the behavior has actually been trained and tested under, and treating everything outside that set as untested. The evidence does not permit a stronger claim, and practice built on the stronger claim fails in precisely the situations that mattered.

Key Insights (Takeaways)

• Generalization is a gradient produced by training history, not a property of the dog. Responding tapers as conditions depart from the training conditions, and the width of that taper is shaped by how much variation was present during acquisition. Narrow training produces narrow gradients reliably.

• Context specificity is the canine default, and it has been measured three separate ways. Inhibitory control did not transfer between tasks (Bray, MacLean & Hare, 2014), aggression did not co-occur across contexts in roughly four thousand owner reports (Casey et al., 2014), and extinguished responding returns when the context changes (Bouton, 2002, 2004).

• Dogs can transfer something abstract, but the ceiling has been demonstrated in unusual animals. Transfer of a happy–angry discrimination to novel faces and an untrained face region is real (Müller et al., 2015), as is inference of a new label by exclusion — in one Border Collie with a two-hundred-object vocabulary (Kaminski, Call & Fischer, 2004).

• The failure is diagnostic, not moral. A behavior that works at home and fails at the park identifies which dimension was carrying stimulus control. Reading it as knowing-and-refusing points the intervention at the dog's attitude instead of at the training conditions, and nothing improves.

• Unwanted behavior transfers more easily than its treatment. Chase sequences generalize across targets while the intervention literature cannot show that training generalizes at all (McLennan, 2023), and aversive events attach to whatever is present while suppressing narrowly (Schalke et al., 2007; Ziv, 2017).

• A training program with no failures is probably building context-bound behavior. Consistent sessions are the condition under which narrow gradients form. Planned variation, one dimension at a time where practical, with the result recorded, is what turns a demonstration into a test.

References

Affenzeller, N., Palme, R., & Zulch, H. (2017). Playful activity post-learning improves training performance in Labrador Retriever dogs (Canis lupus familiaris). Physiology & Behavior, 168, 62–73. https://doi.org/10.1016/j.physbeh.2016.10.014

Amsel, A. (1962). Frustrative nonreward in partial reinforcement and discrimination learning: Some recent history and a theoretical extension. Psychological Review, 69(4), 306–328. https://doi.org/10.1037/h0046200

Bouton, M. E. (2002). Context, ambiguity, and unlearning: Sources of relapse after behavioral extinction. Biological Psychiatry, 52(10), 976–986. https://doi.org/10.1016/S0006-3223(02)01546-9

Bouton, M. E. (2004). Context and behavioral processes in extinction. Learning & Memory, 11(5), 485–494.

Bray, E. E., MacLean, E. L., & Hare, B. A. (2014). Context specificity of inhibitory control in dogs. Animal Cognition, 17(1), 15–31. https://doi.org/10.1007/s10071-013-0633-z

Bray, E. E., MacLean, E. L., & Hare, B. A. (2015). Increasing arousal enhances inhibitory control in calm but not excitable dogs. Animal Cognition, 18(6), 1317–1329. https://doi.org/10.1007/s10071-015-0901-1

Casey, R. A., Loftus, B., Bolster, C., Richards, G. J., & Blackwell, E. J. (2014). Human directed aggression in domestic dogs (Canis familiaris): Occurrence in different contexts and risk factors. Applied Animal Behaviour Science, 152, 52–63. https://doi.org/10.1016/j.applanim.2013.12.003

China, L., Mills, D. S., & Cooper, J. J. (2020). Efficacy of dog training with and without remote electronic collars vs. a focus on positive reinforcement. Frontiers in Veterinary Science, 7, 508. https://doi.org/10.3389/fvets.2020.00508

Cimarelli, G., Schoesswender, J., Vitiello, R., Huber, L., & Virányi, Z. (2021). Partial rewarding during clicker training does not improve naïve dogs' learning speed and induces a pessimistic-like affective state. Animal Cognition, 24(1), 107–119. https://doi.org/10.1007/s10071-020-01425-9

Feuerbacher, E. N., & Wynne, C. D. L. (2012). Relative efficacy of human social interaction and food as reinforcers for domestic dogs and hand-reared wolves. Journal of the Experimental Analysis of Behavior, 98(1), 105–129. https://doi.org/10.1901/jeab.2012.98-105

Fugazza, C., & Miklósi, Á. (2014). Deferred imitation and declarative memory in domestic dogs. Animal Cognition, 17(2), 237–247. https://doi.org/10.1007/s10071-013-0656-5

Hall, N. J. (2017). Persistence and resistance to extinction in the domestic dog: Basic research and applications to canine training. Behavioural Processes, 141(Pt 1), 67–74. https://doi.org/10.1016/j.beproc.2017.04.001

Head, E., Callahan, H., Muggenburg, B. A., Cotman, C. W., & Milgram, N. W. (1998). Visual-discrimination learning ability and beta-amyloid accumulation in the dog. Neurobiology of Aging, 19(5), 415–425. https://doi.org/10.1016/S0197-4580(98)00084-0

Humphreys, L. G. (1939). The effect of random alternation of reinforcement on the acquisition and extinction of conditioned eyelid reactions. Journal of Experimental Psychology, 25(2), 141–158. https://doi.org/10.1037/h0058138

Kaminski, J., Call, J., & Fischer, J. (2004). Word learning in a domestic dog: Evidence for "fast mapping." Science, 304(5677), 1682–1683. https://doi.org/10.1126/science.1097859

Kovács, T., Reicher, V., Csibra, B., Csepregi, M., Kristóf, K., & Gácsi, M. (2025). Repeated task exposure and sufficient sleep may mitigate ADHD-related cognitive flexibility impairments in family dogs. Animals, 15(21), 3074. https://doi.org/10.3390/ani15213074

Krahn, J., Azadian, A., Cavalli, C., Miller, J., & Protopopova, A. (2024). Effect of pre-session discrimination training on performance in a judgement bias test in dogs. Animal Cognition, 27(1), 66. https://doi.org/10.1007/s10071-024-01905-2

Lerman, D. C., & Iwata, B. A. (1995). Prevalence of the extinction burst and its attenuation during treatment. Journal of Applied Behavior Analysis, 28(1), 93–94. https://doi.org/10.1901/jaba.1995.28-93

McLennan, T. (2023). Review of literature on interventions aimed at resolving problems caused by predatory behaviour in dogs (Canis familiaris). Applied Animal Behaviour Science, 266, 106037. https://doi.org/10.1016/j.applanim.2023.106037

Müller, C. A., Riemer, S., Virányi, Z., Huber, L., & Range, F. (2014). Dogs learn to solve the support problem based on perceptual cues. Animal Cognition, 17(5), 1071–1080. https://doi.org/10.1007/s10071-014-0739-y

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

Piotti, P., Szabó, D., Bognár, Z., Egerer, A., Hulsbosch, P., Carson, R. S., & Kubinyi, E. (2018). Effect of age on discrimination learning, reversal learning, and cognitive bias in family dogs. Learning & Behavior, 46(4), 537–553. https://doi.org/10.3758/s13420-018-0357-7

Rankin, C. H., Abrams, T., Barry, R. J., Bhatnagar, S., Clayton, D. F., Colombo, J., Coppola, G., Geyer, M. A., Glanzman, D. L., Marsland, S., McSweeney, F. K., Wilson, D. A., Wu, C.-F., & Thompson, R. F. (2009). Habituation revisited: An updated and revised description of the behavioral characteristics of habituation. Neurobiology of Learning and Memory, 92(2), 135–138. https://doi.org/10.1016/j.nlm.2008.09.012

Schalke, E., Stichnoth, J., Ott, S., & Jones-Baade, R. (2007). Clinical signs caused by the use of electric training collars on dogs in everyday life situations. Applied Animal Behaviour Science, 105(4), 369–380. https://doi.org/10.1016/j.applanim.2006.11.002

Špinka, M., Newberry, R. C., & Bekoff, M. (2001). Mammalian play: Training for the unexpected. The Quarterly Review of Biology, 76(2), 141–168. https://doi.org/10.1086/393866

Thompson, R. F., & Spencer, W. A. (1966). Habituation: A model phenomenon for the study of neuronal substrates of behavior. Psychological Review, 73(1), 16–43. https://doi.org/10.1037/h0022681

Vieira de Castro, A. C., Fuchs, D., Morello, G. M., Pastur, S., de Sousa, L., & Olsson, I. A. S. (2020). Does training method matter? Evidence for the negative impact of aversive-based methods on companion dog welfare. PLoS ONE, 15(12), e0225023. https://doi.org/10.1371/journal.pone.0225023

Wallis, L. J., Virányi, Z., Müller, C. A., Serisier, S., Huber, L., & Range, F. (2016). Aging effects on discrimination learning, logical reasoning and memory in pet dogs. Age, 38(1), 6. https://doi.org/10.1007/s11357-015-9866-x

Ziv, G. (2017). The effects of using aversive training methods in dogs—A review. Journal of Veterinary Behavior, 19, 50–60. https://doi.org/10.1016/j.jveb.2017.02.004