The Clever Hans Effect in Dogs: How Human Cues Shape Behavior, Training and Research

1. Clever Hans: The Original Case

1.1 A Horse That Seemed to Count

In the early twentieth century, a horse called Hans attracted wide public attention. His owner, Mr. von Osten, presented him as an animal that could answer arithmetic problems by tapping his hoof the right number of times, and many observers were convinced. The case was investigated systematically by the psychologist Pfungst, whose report has been available in English translation since 1911 (Pfungst, 1911).

Pfungst treated the performance as a phenomenon to be explained and designed experiments to find out which information Hans was actually using. That is why the case still matters: it showed how to separate an animal's apparent understanding of a task from its sensitivity to the people around it.

1.2 What Pfungst Varied

The central manipulation was simple: in some trials the questioner knew the answer, in others not. Pfungst called this procedure "with knowledge" and "without knowledge". Pfungst also used blinders to vary whether Hans could see the questioner, and worked with several questioners, including von Osten (Pfungst, 1911).

The result was striking. In the series where the questioner knew the answer, 90 to 100% of Hans's responses were correct. In the series where the questioner did not know it, at most 10% were correct (Pfungst, 1911). Hans did not need von Osten in particular. He needed someone who knew the answer and whom he could see.

1.3 The Cue Nobody Intended

Pfungst then looked for the signal itself. When the horse reached the correct count, the questioner made a slight upward jerk of the head, and Hans stopped tapping. The questioners were not aware of this movement. According to Pfungst, the cue was given without any knowledge on their part, and they did not even know that they were being observed (Pfungst, 1911).

Once he had identified the cue, Pfungst produced it deliberately and found that, by voluntarily giving the necessary signs, he could make the horse respond as he wished (Pfungst, 1911). That completed the argument: the information that guided Hans's answers came from the people around him, not from an understanding of arithmetic.

1.4 What the Case Shows and What It Does Not

The Clever Hans case concerns a single animal, studied with early twentieth-century methods and without inferential statistics. It is not evidence about dogs. What it provides is a mechanism and a method. The mechanism is that an animal can solve a task by attending to small, unintended movements of a person who knows the answer. The method is to compare performance when the people present know the answer with performance when they do not.

Hans was not "stupid": he detected tiny changes in human posture and used them to obtain a reward. The error lay with the observers who read that ability as arithmetic. The Clever Hans effect is a problem of interpretation and design, not a judgment about the animal.

2. What the Effect Is and Is Not

2.1 A Working Definition

In this article, a Clever Hans effect means that an animal's behavior in a task is guided by cues from a person who knows, expects or wants a particular outcome, and that these cues are given without the person intending or noticing them. Two elements matter: the information comes from a person rather than from the task, and the person is not aware of providing it.

This separates the effect from deliberate cueing. An owner who points at the correct cup or nudges the dog toward it is giving information intentionally. That is social influence, and it can be strong, but it is not a Clever Hans effect in the narrow sense. The distinction has a practical side: deliberate cues can be prevented by instructions, unintended cues cannot, because the person producing them does not know they exist.

2.2 Belief Versus Visible Behavior

A second distinction is between what a person believes and what a person visibly does. A belief can only influence a dog if it changes something the dog can perceive: posture, gaze, movement, timing, voice or possibly odor. Experiments that manipulate owners' beliefs therefore test an indirect route. Experiments that manipulate a person's visible behavior, such as demonstrating a preference or using attention-getting signals, test a direct one.

As sections 4 and 5 show, owners' beliefs alone did not measurably change dogs' choices in two pointing studies, while visible human choices and communication did shift dogs.

2.3 The Reverse Direction: Observer Bias

A related problem runs in the opposite direction. In the classic case, the animal reads the human. In observer bias, the human misreads the animal: expectations shape how a behavior is perceived, scored or reported, regardless of what the animal did. Owners asked whether their dog had broken a rule could not tell from the dog's greeting behavior (Ostojić, Tkalčić & Clayton, 2015). Here, human interpretation is the source of error.

The two directions can occur together. A handler who expects an alert at a certain location may move differently there and may also be quicker to read a hesitation as an alert. Separating the two requires designs in which neither the dog nor the person scoring its behavior has access to the expected answer (how behavior is defined and measured objectively).

2.4 Channels Beyond Vision

Pfungst's cue was visual, and most research on dogs has focused on pointing, gaze and body orientation. But people transmit information through other channels too. In one experiment, sweat collected from people watching fear-inducing video raised dogs' heart rate and stress behaviors and changed their behavior toward a stranger, compared with odor from happy or neutral conditions (d'Aniello, Semin, Alterisio, Aria & Scandurra, 2018). In another, the odor of an unfamiliar stressed person shifted dogs toward more "pessimistic" choices in a cognitive bias test (Parr-Cortes et al., 2024).

These studies do not show that odor guides answers in cognitive tasks, but a human state can reach the dog through a channel the person cannot suppress (how human stress affects dogs).

3. Dogs' Sensitivity to Human Cues

3.1 Following Human Gestures

A Clever Hans effect requires an animal that attends to people closely enough to pick up small signals. Dogs are a strong candidate. In a widely cited comparison, dogs outperformed chimpanzees in using human communicative cues to find hidden food, and puppies with little human contact already showed the skill, while human-reared wolves in that study did not (Hare, Brown, Williamson & Tomasello, 2002). Later work showed that socialized wolves could follow momentary distal points under other testing conditions (Udell, Dorey & Wynne, 2008), which suggests that cue sensitivity depends on rearing, experience and test details rather than on species alone (what the research on gesture comprehension shows).

An animal that follows a deliberate point is likely to notice smaller, unintended versions of it too: a glance, a shift of weight, a turn of the shoulders.

3.2 Reading Human Attention

Dogs also track whether a person is paying attention. In a forbidden-food paradigm with 12 dogs, a human either watched or did not, because the person left the room, turned away, was distracted or closed their eyes. When watched, the dogs retrieved less food, approached it more indirectly and sat rather than lay down more often, and they behaved differently in most of the no-watching conditions (Call, Bräuer, Kaminski & Tomasello, 2003).

A study in the dogs' homes gave 16 dogs a "down" command, after which the owner looked at the dog, read, watched television, turned their back or left the room during 60-second trials. With food present, the condition had a significant effect, and dogs stood up faster when the owner's back was turned or the owner was absent than with eye contact (Schwab & Huber, 2006). The dogs used eye contact and the orientation of eyes, head and body; the authors noted that results may be context dependent.

A further study found that dogs distinguished attentive from inattentive people by body orientation, head direction and eye visibility, depending on context, and hesitated to approach blindfolded owners (Gácsi, Miklósi, Varga, Topál & Csányi, 2004).

3.3 Attention Changes the Dog's Own Signals

In a study of 24 dogs, the dogs produced more facial movements when a person faced them than when the person turned away, and food made no difference (Kaminski, Hynds, Morris & Waller, 2017). Any measure based on scoring a dog's expressive behavior can therefore be affected by whether a person in the room is attending to the dog (how dogs use body language and facial expressions).

3.4 Ostensive Cues and Emotional Information

How a person frames a demonstration can matter too. In a manipulative social learning task, dogs learned more when the demonstration came with eye contact and verbal address, so-called ostensive cues (Range et al., 2009). Such signals tell the dog that what follows is directed at it, and they are part of everyday training: the dog's name, eye contact, a particular tone (how dogs learn from people and other dogs).

Dogs have also been shown to integrate human facial and vocal emotional cues (Albuquerque et al., 2016), and they seek such information in uncertain situations. Faced with an ambiguous object, a fan, 83% of dogs looked referentially to their owner, and the owner approaching or retreating had a much stronger effect than voice and facial expression alone (Merola, Prato-Previde & Marshall-Pescini, 2012b). In a related study, 62% of dogs alternated their gaze between object and informant when the informant was the owner, and 52% with a stranger; regulation by the emotional message was clearer with the owner (Merola, Prato-Previde & Marshall-Pescini, 2012a).

These findings show the channel through which unintended cues act: a dog facing uncertainty looks to a person, and movement had a particularly strong effect (how dogs use human information in uncertain situations). The finding comes from one laboratory and one stimulus, so its generality is limited.

4. Do Owners' Beliefs Steer Dogs in Cognition Experiments?

4.1 The Problem for Dog Cognition Research

In many dog cognition experiments, the owner holds the dog at the start of a trial or stands nearby, in a position from which the dog can read small body cues. If owners systematically influenced their dogs, much of the gesture comprehension literature would be open to doubt.

4.2 The Vienna Pointing Study

A study at the Clever Dog Lab in Vienna tested this with a two-choice object-choice task of 20 trials per dog, in which an experimenter pointed to one of two cups while the owner held the dog (Schmidjell, Range, Huber & Virányi, 2012). In the first experiment, 69 pet dogs were assigned to four groups. Blindfolded owners could not see the point. In the enhancement group, owners saw the point and were told the dog should follow it. In the decrease group, owners saw a point to the empty cup and were told the dog should ignore it. In the no-pointing group, there was no point, but the owner knew where the food was.

Dogs in all three pointing groups chose above chance, and the belief manipulation made no measurable difference. In the no-pointing group, where only the owner had the information, dogs chose at chance level. Owner knowledge alone did not guide the dogs (Schmidjell et al., 2012).

The second experiment, with 31 dogs, asked what happens when owners try to help. When owners actively directed their dog, the dogs succeeded above chance, with 10 of 16 succeeding individually. When owners directed the dog while the experimenter pointed the other way, the owners' influence failed, with 2 of 15 dogs succeeding individually, and the groups differed significantly (Schmidjell et al., 2012).

4.3 The Budapest Study

An independent laboratory in Budapest reached a converging conclusion with different conditions (Hegedüs, Bálint, Miklósi & Pongrácz, 2013). In a two-way object-choice task with a momentary distal point, one group of owners wore opaque glasses. In a "passive Clever Hans" group, owners were promised a gift and an entry in a "smartest dogs registry" if their dog made no errors. In an "active Clever Hans" group, owners were told to help by gently pushing the dog toward the correct side. A further control group completed the design; its exact procedure is not described here.

The authors concluded that the experimenter's momentary point remained a reliable cue for adult companion dogs even when owners gave voluntary or involuntary cues, and that owner presence in short visual communication studies may not necessarily distort results (Hegedüs et al., 2013). Sample size and detailed statistics are not reported here, so the comparison with the Vienna study is limited to the direction of the main conclusion.

4.4 What These Results Mean

Two laboratories did not find a measurable Clever Hans effect from owners' beliefs in a short pointing task. That is reassuring for this paradigm, with limits. The point was a clear, salient signal; where no such signal exists, unintended cues may matter more. The Vienna authors noted that their results cannot easily be generalized to other dog-owner pairs, that owner behavior was not measured because dogs may react to cues observers cannot detect, and that a possible effect of the pointing experimenter, who knew the correct cup, was not investigated (Schmidjell et al., 2012).

The fourth point matters: in the original case, it was the questioner who supplied the answer.

5. When Visible Human Behavior Overrides the Dog's Own Information

5.1 Choosing Less Because the Owner Did

Visible behavior is a different matter. In a food choice task with 54 pet dogs, dogs choosing independently between a large and a small quantity of food chose the larger one significantly more often (Prato-Previde, Marshall-Pescini & Valsecchi, 2008). After watching their owner favor the small quantity, they chose the large quantity significantly less often. When both options were equally small and the owner favored one, conformity was higher than when the quantities differed.

The authors concluded that owners can lead dogs to counterproductive choices. Because the owners acted deliberately, this is not a Clever Hans effect in the narrow sense, but it shows that a visible human choice can override the dog's own perception, most strongly when that perception does not clearly favor one option.

5.2 Owners and Strangers

A follow-up study with two experiments, one varying food quantity and one palatability, had either the owner or an unfamiliar friendly person indicate the less advantageous option (Marshall-Pescini, Prato-Previde & Valsecchi, 2011). The dogs conformed to the human's indication even though it led them to the worse option, and owner and stranger did not differ in their influence. An unfamiliar experimenter, examiner or judge is therefore not automatically a neutral presence.

5.3 The A-not-B Error

In an A-not-B task, an object is hidden repeatedly at location A and then, in full view, at location B. Searching at A after seeing the object hidden at B is a perseverative error. In a study comparing dogs, human-reared wolves and 10-month-old infants, dogs made more perseverative errors when the hiding was done in an ostensive-communicative way than in a non-communicative or non-social way: 75% of trials, compared with 39% and 17% (Topál, Gergely, Erdőhegyi, Csibra & Miklósi, 2009). Human-reared wolves did not show this context dependence.

The original interpretation was that dogs, like infants, are especially sensitive to human communicative signals. That interpretation is disputed. Published comments proposed simpler explanations, in particular local enhancement, meaning that the human's actions at location A drew the dog's attention there. In their response, the original authors acknowledged that attentional processes and local enhancement are involved, and argued that dogs' behavior is affected by a combination of factors, including a specific susceptibility to human communicative signals (Topál, Miklósi, Sümegi & Kis, 2010).

5.4 A Replication

A joint team from Budapest and Vienna replicated the task without motor search in the A trials and without "sham-baiting" in the B trials (Kis et al., 2012). The tendency to err after ostensive-communicative hiding was replicated, and neither an inability to inhibit a rewarded motor response nor deficits in working memory or attention explained the errors. The error diminished when location B was ostensively enhanced. The authors concluded that such tasks show how susceptibility to human social signals could mislead dogs.

The replication is not fully independent of the original group, and the mechanism remains debated. The result, however, is clear: how a person behaves during a task can override the dog's own observation.

5.5 Linking the Findings

In short pointing tasks with a clear experimenter cue, owners' expectations alone did not measurably shift dogs' choices. When people visibly chose, demonstrated or communicated, dogs' choices shifted, sometimes against their own interest and perception. The practical risk lies less in what a handler secretly believes than in what a handler visibly does, often without realizing it.

6. Interpretation Bias: The "Guilty Look"

6.1 The Classic Experiment

The "guilty look" is the most familiar everyday example of how human beliefs shape what people see in dogs. In an experiment with 14 dogs, the owner forbade the dog to eat a treat and left the room (Horowitz, 2009). The experimenter then manipulated whether the dog ate it, and the owner was told, accurately or falsely, whether the dog had obeyed, and greeted or scolded the dog accordingly.

The "guilty look" behaviors did not differ according to whether the dog had actually disobeyed. They increased when the owner scolded and were most pronounced in dogs that had obeyed but were scolded anyway (Horowitz, 2009). The display followed the owner's behavior, which followed the owner's belief, rather than the dog's own action.

6.2 Replication and Owner Judgments

An independent study replicated the pattern. Dogs that had broken a rule were no more likely to show the display than dogs that had not, the display was essentially absent without a scolding owner, and owners were no better than chance at judging from greeting behavior whether their dog had transgressed (Hecht, Miklósi & Gácsi, 2012). A further study likewise found that owners could not tell from greeting behavior whether the dog had transgressed (Ostojić et al., 2015).

6.3 Why This Is a Clever Hans Problem

The guilty look combines both directions from section 2: the owner's belief changes the owner's behavior, the dog responds, and the owner reads the response as confirmation of the belief that produced it (what the research on the guilty look shows). It is therefore not a classic Clever Hans effect in the narrow sense, but a related example of how human behavior can alter a dog's response and how human expectations can then shape the interpretation of that response.

The Horowitz study is small, and these experiments say nothing about whether dogs can experience anything like guilt. They do show that the behavior commonly taken as evidence of guilt is a poor indicator of what the dog did, and that a dog's behavior in the presence of a person who expects something is not a neutral readout of its state (why behavior is not a direct readout of emotion).

7. Detection Dogs and Handler Effects

7.1 Why Detection Work Is Vulnerable

In scent detection, the Clever Hans problem has large practical consequences, because an alert can lead to a search, an arrest or a medical decision. The situation resembles the original case: a dog works close to a person who may know or suspect where the target is and who interprets the dog's behavior. The dog's olfactory abilities are not in question (how olfaction shapes dog behavior). The question is whether a particular alert reflects the odor or the handler.

7.2 Handler Beliefs and Alert Locations

The most direct test in working dogs comes from a study with 18 drug and/or explosive detection dog and handler teams (Lit, Schweitzer & Oberbauer, 2011). Each team completed two sets of four brief search scenarios. Handlers were falsely told that in two conditions a paper marker indicated target scent, which tested human influence; two conditions contained decoy food or toy scents, which tested influence from the dog's side. The conditions were a control, a paper marker, a decoy scent, and a marker at a decoy. None contained any target scent, so every alert was incorrect.

There were 225 incorrect responses, with no differences in mean responses across conditions. There were more correct searches, meaning searches without an alert, in conditions without markers. Within the marked conditions, handlers reported that dogs alerted more at the marked locations than elsewhere. The authors concluded that human more than dog influences affected alert locations (Lit et al., 2011).

Without target odor, the study cannot say how beliefs affect detection of real targets, and it cannot separate dogs reacting to handler cues from handlers calling alerts on ambiguous behavior. With 18 teams, the sample is modest, yet it is the strongest direct evidence that handler expectations change detection outcomes.

7.3 Blinding as the Methodological Standard

A methodological review of canine olfactory detection research treats handler and experimenter expectancy as a major threat to validity, since handlers may move differently when they know where targets are (Lazarowski et al., 2020). Single-blind testing, in which only the handler is unaware of the target, is most common; double-blind testing, in which nobody present and able to interact with the dog knows the answer, is methodologically an especially important standard for controlling expectancy effects and more closely approximating operational conditions. Studies have reported performance declines once double-blind testing was introduced, and handler-induced false alerts are a distinct source of error.

A systematic review of 54 scent detection studies found that testing methods were heterogeneous and susceptible to potential bias, with no accepted quality standards (Johnen, Heuwieser & Fischer-Tenhagen, 2017). The authors derived a best-practice standard covering target odor, task, set-up, samples, test design, trainer and training, and breeds.

7.4 When Training Success Does Not Survive Blinding

A study on prostate cancer detection shows what can happen when rigorous testing follows promising training (Elliker et al., 2014). Ten dogs were trained in two stages, and three reached the second stage. In a double-blind test with unfamiliar urine samples, two dogs had sensitivities of 0.13 and 0.25, no better than chance. The authors suggested that the dogs may have memorized individual samples during training.

This is not primarily a handler cueing result, but it shows that performance with familiar samples during training can overestimate performance under blind conditions with new samples.

7.5 Removing the Human From the Decision

Another approach removes the handler from the decision altogether. In a proof-of-concept study, an automated three-port olfactometer randomized the target position by computer, detected a 4-second nose hold with an infrared sensor and delivered the reward through an automated feeder (Aviles-Rosa, Gallegos, Prada-Tiedemann & Hall, 2021). The experimenter could not see which port held the target, so testing was double-blind.

Twelve naive mixed-breed shelter and rescue dogs, tested in two cohorts, all learned the task within 23 days and all exceeded 85% accuracy in their final sessions, with values between 85.00 ± 0.00% and 98.75 ± 1.25%. With the odor lines disconnected, accuracy fell to 5.00 ± 3.37%, which showed that the dogs were not using the apparatus or other non-odor cues. The false-alert proportion never exceeded 20% during training (Aviles-Rosa et al., 2021).

The authors named the limitations: laboratory use only, lower ecological validity than traditional searches, a cost of about USD 5,000, programming requirements and variable odor concentration. The dogs were a research population, not operational detection dogs.

7.6 Objective Measures of the Dog's Response

A further approach makes reading the dog less dependent on the handler. In 200 coded videos of 10 detection dogs on a carousel apparatus, sniffing was significantly shorter for true negatives than for other response types, and true negatives usually involved one sniffing episode, the others commonly two (Concha et al., 2014). The authors named a handler's failure to identify a positive alert as one source of false negatives. Such measures have so far been tested only in small samples and a controlled set-up.

7.7 Field Accuracy Is Context Specific

In a study of 164 police dogs in Polish police training and testing environments, with 1,219 search tests, dogs needed on average 64 seconds to indicate; 87.7% of indications were correct, 5.3% were false, and in 7.0% of tests the target was not found within 10 minutes (Jezierski et al., 2014). Accuracy fell to 63.5% outdoors and 57.9% in cars. During police examination trials, dogs made more false alerts and fewer correct indications, and searches took longer than in the final stage of training.

Trained detection dogs therefore perform well above chance, but their accuracy varies with breed, training level, drug type and environment, and drops under examination conditions. The study did not report handler blinding, so it does not show handler cueing; it shows that an accuracy figure means little without its conditions.

8. Gestures Versus Words in Training

8.1 When Gesture and Word Disagree

In everyday training, the most common unintended cue is a visible body movement the handler does not count as part of the cue: bending forward slightly with "down", stepping back with "come". The dog may learn the movement rather than the word.

A study with water rescue dogs, trained equally on gestural and verbal requests, tested this (D'Aniello, Scandurra, Alterisio, Valsecchi & Prato-Previde, 2016). Owners asked for "sit", "lie down", "stay" and "come" by gesture only, by word only, or with an incongruent combination. Gestures were followed significantly better than words. Females responded more to gestures, while males responded better to words than females did. When gesture and word conflicted, the dogs mostly followed the gesture, except when a verbal "come" was combined with a gestural "stay" while the owner moved away. The authors concluded that gestures are more salient, but that context, here the motivation to stay close to a departing owner, also matters.

8.2 Familiar and Unfamiliar Handlers

A second study with working dogs compared congruent and conflicting gestural and verbal signals from the owner or a stranger across four obedience tasks (Scandurra et al., 2017). The requested behavior was less likely when a stranger gave verbal signals, and latencies increased with unfamiliar handlers using vocal signals. With conflicting signals, dogs followed a stranger's gestures more readily than the stranger's words, while owners' verbal commands were more reliable.

A verbal cue is tied to a particular voice and context, whereas a gesture looks similar whoever makes it, so a dog may follow a stranger's gestures but not the stranger's words (how dogs generalize learned cues).

8.3 Limits of the Gesture Findings

Both studies used specialized working dog populations, and their sample sizes are not reported here. The finding cannot be turned into a rule that "dogs always follow body language": even among the water rescue dogs, responses depended on sex and on the specific cue combination. Which signal dominates for a given dog probably depends on its training history and on which signal has most reliably predicted reinforcement.

In practice: It is common for a dog to appear to "know" a verbal cue that turns out to be controlled by a body movement the handler makes at the same time, such as a slight forward lean before "down" or a turn of the shoulders before "come". This is an observation from training practice, not a finding from controlled studies, but it fits the experimental results on the salience of gestures.

9. Designing Against Clever Hans

9.1 Blinding

Pfungst's method still defines the basic control: compare performance when the people present know the answer with performance when they do not. In a single-blind design, the handler does not know the answer, but someone else present may. In a double-blind design, nobody who can interact with or observe the dog during the trial knows it. In detection research, double-blind testing is methodologically an especially important standard for controlling expectancy effects and approximating real operational conditions (Lazarowski et al., 2020).

Someone has to prepare the trial and must then stay out of the dog's sight, and the person scoring the behavior must also be blind, otherwise observer bias can enter. Video coding by observers unaware of the condition is one solution (how behavioral assessments are designed and scored).

9.2 Automation and No-Target Controls

Automation removes people from the decisive moment, as in the automated line-up described above (Aviles-Rosa et al., 2021). Just as important was its control with disconnected odor lines. Such controls test whether the dog can solve the task without the information it is supposed to use; if performance stays high, the dog is using something else. The handler-belief study used the same logic by including no target odor at all, which made every alert informative about non-odor influences (Lit et al., 2011). Both follow Pfungst: remove the information the animal is supposed to use, and see what remains.

9.3 Measuring the Human, Not Only the Dog

Few studies measure what the human actually does. The Vienna study did not measure owner behavior, because dogs may respond to cues observers cannot detect (Schmidjell et al., 2012). A null result therefore means that no effect on the dog was found, not that no cues were given.

9.4 Replication Across Laboratories

Clever Hans effects are only one reason why single studies can overestimate effects. A methodological paper on comparative cognition argued that published effect sizes are inflated and that limited species availability, behavioral variation between sites and small samples make many claims practically unfalsifiable (Farrar, Boeckle & Clayton, 2020). The authors proposed formal assessment of replicability, better statistics and collaborative infrastructure across laboratories. A review of 285 dog cognition articles recommended replicating studies where knowledge rests on small study sets or small samples, testing more diverse populations and paying more attention to olfaction (Bensky, Gosling & Sinn, 2013).

A small single-laboratory study, run by people who know the hypothesis and often the correct answer, is exactly where unintended cues and inflated effects are hardest to detect. Until a finding has been replicated across sites, it is best treated as promising rather than settled (what is known about canine cognition).

10. What Follows for Practice

10.1 Unintended Cues in Everyday Training

The research suggests that the most relevant cues in everyday training are not hidden beliefs but visible behavior: body orientation, gaze, movement and timing. Dogs track whether a person is attending (Call et al., 2003; Schwab & Huber, 2006), respond to the owner's approach or retreat more than to voice and face alone (Merola et al., 2012b), and tended to follow gestures over words when the two conflicted (D'Aniello et al., 2016).

Two patterns are common: a body movement that precedes or accompanies the verbal cue, and a glance at the correct object or location in scent games, retrieves or object discrimination. Both carry information the handler does not count as part of the cue.

In practice: In scent games and object searches, handlers who know where an item is hidden often turn their head or body slightly toward it, slow down near it or change their tone as the dog approaches. A simple safeguard is to let another person hide the item so that the handler does not know where it is. This is an observation from practice, not a result from a controlled study.

10.2 Testing Whether the Dog Responds to the Word

In practice: To check whether a dog responds to the verbal cue alone, the handler stands still with arms relaxed, looks at a fixed point above the dog or wears sunglasses, and gives the word once in a neutral voice. Cues can be mixed in random order so the dog cannot predict the next one from the sequence. For object names, the objects can be placed out of the handler's view so the handler cannot look at the right one. A video recording lets someone else check whether a movement slipped in. These steps come from training practice and adapt the logic of blind testing; they are not a validated test procedure.

If performance drops under these conditions, that is information, not failure: part of what the dog learned was the handler's body, and training can fade that cue while the word becomes predictive.

10.3 Evaluating Claims Such as "My Dog Understands Words"

Claims about dogs' understanding of words, numbers or concepts are especially vulnerable, because the person making the claim usually knows the answer and is present during the demonstration (what is known about word learning in dogs). A video of a dog fetching the named toy from a pile is not evidence of word comprehension if the owner could see the toys and knew which one was named.

The minimum standard follows from Pfungst: the person giving the cue must not know the answer or must not be visible to the dog, and the person judging the response must not know which answer was expected. Only then does a correct response show that the dog solved the task without human information.

10.4 Keeping the Evidence in Proportion

Not every success is a Clever Hans effect. Owners' beliefs did not measurably change dogs' choices in two pointing studies (Schmidjell et al., 2012; Hegedüs et al., 2013), detection dogs in field tests performed far above chance (Jezierski et al., 2014), and dogs in an automated, double-blind set-up reached high accuracy without a human in the loop (Aviles-Rosa et al., 2021).

What the evidence supports is narrower: visible human behavior can shift dogs' choices (Prato-Previde et al., 2008; Marshall-Pescini et al., 2011; Kis et al., 2012), handler expectations can shift where alerts are called (Lit et al., 2011), and training success can fail under double-blind testing (Elliker et al., 2014). The right response is not suspicion of dogs' abilities, but tests that show which information the dog actually uses.

11. Summary at a Glance

The Clever Hans effect describes behavior guided by unintended cues from people who know or expect the correct answer. In the original case, a horse was correct in 90 to 100% of responses when the questioner knew the answer and in at most 10% when the questioner did not; the cue was a slight upward head movement the questioners did not know they were making (Pfungst, 1911).

Dogs are highly sensitive to human gestures, attention and emotional signals (Hare et al., 2002; Call et al., 2003; Gácsi et al., 2004; Schwab & Huber, 2006; Merola et al., 2012a, 2012b). In two independent pointing studies, owners' beliefs alone did not measurably change dogs' choices (Schmidjell et al., 2012; Hegedüs et al., 2013), whereas visible human choices shifted them, with strangers as influential as owners (Prato-Previde et al., 2008; Marshall-Pescini et al., 2011). Communicative behavior can also induce search errors, although the mechanism is disputed (Topál et al., 2009, 2010; Kis et al., 2012).

The "guilty look" followed owners' scolding rather than dogs' actions (Horowitz, 2009; Hecht et al., 2012; Ostojić et al., 2015). In detection work, handlers' false beliefs changed where alerts were called (Lit et al., 2011), double-blind testing is an especially important methodological control for expectancy effects (Lazarowski et al., 2020; Johnen et al., 2017), training success did not survive it in one study (Elliker et al., 2014), automation can remove the human from the decision (Aviles-Rosa et al., 2021), and field accuracy depends strongly on context (Jezierski et al., 2014). In training, gestures often dominated words, especially with unfamiliar handlers (D'Aniello et al., 2016; Scandurra et al., 2017).

12. Research Gaps and Critical Appraisal

12.1 Small Samples and Single Laboratories

Many key studies are small: one horse (Pfungst, 1911), 14 dogs in the guilty look experiment (Horowitz, 2009), 12 in the attention study (Call et al., 2003), 16 in the home study (Schwab & Huber, 2006), 12 in the automated line-up (Aviles-Rosa et al., 2021), 18 teams in the handler-belief study (Lit et al., 2011), and 10 dogs in the cancer detection study, of which 3 reached the second stage (Elliker et al., 2014). Several findings come from one laboratory, and the A-not-B replication included members of the original team (Kis et al., 2012). The field as a whole has been criticized for inflated effect sizes and small samples (Farrar et al., 2020; Bensky et al., 2013).

12.2 Null Results Are Task Specific

The reassuring results on owner beliefs come from short two-choice pointing tasks with a salient experimenter cue (Schmidjell et al., 2012; Hegedüs et al., 2013). They do not show that owners have no influence in tasks without a clear cue, in longer sessions, when the dog's own information is ambiguous, or in assessments outside the laboratory. The pointing experimenter's own influence was not investigated.

12.3 The Human Side Is Rarely Measured

Few studies record the gaze, posture, movement and timing of the people involved and relate them to the dog's responses, so it is unclear which human behaviors carry information and whether people can learn to suppress them. The handler-belief study could not separate dog responses to handler cues from handlers misreading dogs (Lit et al., 2011).

12.4 Specialized Populations

The gesture studies used water rescue and working dogs (D'Aniello et al., 2016; Scandurra et al., 2017); the detection studies used detection teams, police dogs from one national system, or naive shelter and rescue dogs in a laboratory (Lit et al., 2011; Jezierski et al., 2014; Aviles-Rosa et al., 2021); the cognition studies used volunteer pet dogs. Transfer to family dogs is plausible for the mechanisms but uncertain for effect sizes.

12.5 Mechanisms Remain Contested

For the A-not-B error, the relative contribution of communicative sensitivity and simpler attentional processes is still debated, and the original authors describe a combination of factors (Topál et al., 2010). For social referencing, it is unclear how far results with a single ambiguous object generalize (Merola et al., 2012b). Human odor affected dogs' stress behavior and cognitive bias (d'Aniello et al., 2018; Parr-Cortes et al., 2024), but whether such cues contribute to Clever Hans effects in cognitive or detection tasks has not been tested.

12.6 Reporting of Blinding

Scent detection testing lacked accepted quality standards (Johnen et al., 2017), and the police dog field study did not report handler blinding (Jezierski et al., 2014). Reporting who knew what, who scored responses and whether no-target controls were used would make the risk of cueing easier to judge.

13. Conclusion

The Clever Hans case identified a mechanism that applies with particular force to dogs: an animal highly attentive to people can solve a task by reading those who know the answer. Dogs track human attention, gestures and movement closely, visible human behavior can override their own information, and in detection work handler expectations changed where alerts were called.

Yet owners' beliefs alone did not measurably steer dogs in two controlled pointing studies, and dogs in double-blind set-ups and field tests performed well above chance. The lesson of Clever Hans is methodological rather than skeptical: whenever it matters what a dog has learned, the test must be designed so that the people present cannot supply the answer.

Key Insights (Takeaways)

• The Clever Hans effect describes behavior guided by unintended cues from people who know the answer; Hans was correct in 90 to 100% of responses when the questioner knew the answer and in at most 10% when not (Pfungst, 1911).

• Dogs closely track human attention, gaze, body orientation and movement (Call et al., 2003; Schwab & Huber, 2006; Merola et al., 2012b).

• Owners' beliefs alone did not measurably change dogs' choices in two pointing studies, while visible human choices did shift dogs (Schmidjell et al., 2012; Hegedüs et al., 2013; Prato-Previde et al., 2008).

• The "guilty look" follows the owner's scolding, not the dog's action (Horowitz, 2009; Hecht et al., 2012).

• Without any target odor, handlers' false beliefs changed where alerts were called (Lit et al., 2011); double-blind testing is an especially important methodological control for expectancy effects in detection research (Lazarowski et al., 2020).

• Gestures often dominated conflicting words, especially with unfamiliar handlers, but not every dog always follows body language (D'Aniello et al., 2016; Scandurra et al., 2017).

• Claims such as "my dog understands words" need tests in which neither the cue giver nor the judge knows the correct answer.

References

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

Aviles-Rosa, E. O., Gallegos, S. F., Prada-Tiedemann, P. A., & Hall, N. J. (2021). An automated canine line-up for detection dog research. Frontiers in Veterinary Science, 8, 775381. https://doi.org/10.3389/fvets.2021.775381

Bensky, M. K., Gosling, S. D., & Sinn, D. L. (2013). The world from a dog's point of view: A review and synthesis of dog cognition research. Advances in the Study of Behavior, 45, 209–406. https://doi.org/10.1016/B978-0-12-407186-5.00005-7

Call, J., Bräuer, J., Kaminski, J., & Tomasello, M. (2003). Domestic dogs (Canis familiaris) are sensitive to the attentional state of humans. Journal of Comparative Psychology, 117(3), 257–263. https://doi.org/10.1037/0735-7036.117.3.257

Concha, A., Mills, D. S., Feugier, A., Zulch, H., Guest, C., Harris, R., & Pike, T. W. (2014). Using sniffing behavior to differentiate true negative from false negative responses in trained scent-detection dogs. Chemical Senses, 39(9), 749–754. https://doi.org/10.1093/chemse/bju045

D'Aniello, B., Scandurra, A., Alterisio, A., Valsecchi, P., & Prato-Previde, E. (2016). The importance of gestural communication: A study of human–dog communication using incongruent information. Animal Cognition, 19(6), 1231–1235. https://doi.org/10.1007/s10071-016-1010-5

d'Aniello, B., Semin, G. R., Alterisio, A., Aria, M., & Scandurra, A. (2018). Interspecies transmission of emotional information via chemosignals: From humans to dogs (Canis lupus familiaris). Animal Cognition, 21(1), 67–78. https://doi.org/10.1007/s10071-017-1139-x

Elliker, K. R., Sommerville, B. A., Broom, D. M., Neal, D. E., Armstrong, S., & Williams, H. C. (2014). Key considerations for the experimental training and evaluation of cancer odour detection dogs: Lessons learnt from a double-blind, controlled trial of prostate cancer detection. BMC Urology, 14, 22. https://doi.org/10.1186/1471-2490-14-22

Farrar, B. G., Boeckle, M., & Clayton, N. S. (2020). Replications in comparative cognition: What should we expect and how can we improve? Animal Behavior and Cognition, 7(1), 1–22. https://doi.org/10.26451/abc.07.01.02.2020

Gácsi, M., Miklósi, Á., Varga, O., Topál, J., & Csányi, V. (2004). Are readers of our face readers of our minds? Dogs (Canis familiaris) show situation-dependent recognition of human's attention. Animal Cognition, 7(3), 144–153. https://doi.org/10.1007/s10071-003-0205-8

Hare, B., Brown, M., Williamson, C., & Tomasello, M. (2002). The domestication of social cognition in dogs. Science, 298(5598), 1634–1636. https://doi.org/10.1126/science.1072702

Hecht, J., Miklósi, Á., & Gácsi, M. (2012). Behavioral assessment and owner perceptions of behaviors associated with guilt in dogs. Applied Animal Behaviour Science, 139(1–2), 134–142. https://doi.org/10.1016/j.applanim.2012.02.015

Hegedüs, D., Bálint, A., Miklósi, Á., & Pongrácz, P. (2013). Owners fail to influence the choices of dogs in a two-choice, visual pointing task. Behaviour, 150(3–4), 427–443. https://doi.org/10.1163/1568539X-00003060

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

Jezierski, T., Adamkiewicz, E., Walczak, M., Sobczyńska, M., Górecka-Bruzda, A., Ensminger, J., & Papet, E. (2014). Efficacy of drug detection by fully-trained police dogs varies by breed, training level, type of drug and search environment. Forensic Science International, 237, 112–118. https://doi.org/10.1016/j.forsciint.2014.01.013

Johnen, D., Heuwieser, W., & Fischer-Tenhagen, C. (2017). An approach to identify bias in scent detection dog testing. Applied Animal Behaviour Science, 189, 1–12. https://doi.org/10.1016/j.applanim.2017.01.001

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

Kis, A., Topál, J., Gácsi, M., Range, F., Huber, L., Miklósi, Á., & Virányi, Z. (2012). Does the A-not-B error in adult pet dogs indicate sensitivity to human communication? Animal Cognition, 15(4), 737–743. https://doi.org/10.1007/s10071-012-0481-2

Lazarowski, L., Krichbaum, S., DeGreeff, L. E., Simon, A., Singletary, M., Angle, C., & Waggoner, L. P. (2020). Methodological considerations in canine olfactory detection research. Frontiers in Veterinary Science, 7, 408. https://doi.org/10.3389/fvets.2020.00408

Lit, L., Schweitzer, J. B., & Oberbauer, A. M. (2011). Handler beliefs affect scent detection dog outcomes. Animal Cognition, 14(3), 387–394. https://doi.org/10.1007/s10071-010-0373-2

Marshall-Pescini, S., Prato-Previde, E., & Valsecchi, P. (2011). Are dogs (Canis familiaris) misled more by their owners than by strangers in a food choice task? Animal Cognition, 14(1), 137–142. https://doi.org/10.1007/s10071-010-0340-y

Merola, I., Prato-Previde, E., & Marshall-Pescini, S. (2012a). Dogs' social referencing towards owners and strangers. PLOS ONE, 7(10), e47653. https://doi.org/10.1371/journal.pone.0047653

Merola, I., Prato-Previde, E., & Marshall-Pescini, S. (2012b). Social referencing in dog-owner dyads? Animal Cognition, 15(2), 175–185. https://doi.org/10.1007/s10071-011-0443-0

Ostojić, L., Tkalčić, M., & Clayton, N. S. (2015). Are owners' reports of their dogs' "guilty look" influenced by the dogs' action and evidence of the misdeed? Behavioural Processes, 111, 97–100. https://doi.org/10.1016/j.beproc.2014.12.010

Parr-Cortes, Z., Müller, C. T., Talas, L., Mendl, M., Guest, C., & Rooney, N. J. (2024). The odour of an unfamiliar stressed or relaxed person affects dogs' responses to a cognitive bias test. Scientific Reports, 14(1), 15843. https://doi.org/10.1038/s41598-024-66147-1

Pfungst, O. (1911). Clever Hans (the horse of Mr. von Osten): A contribution to experimental animal and human psychology (C. L. Rahn, Trans.). Henry Holt.

Prato-Previde, E., Marshall-Pescini, S., & Valsecchi, P. (2008). Is your choice my choice? The owners' effect on pet dogs' (Canis lupus familiaris) performance in a food choice task. Animal Cognition, 11(1), 167–174. https://doi.org/10.1007/s10071-007-0102-7

Range, F., Heucke, S. L., Gruber, C., Konz, A., Huber, L., & Virányi, Z. (2009). The effect of ostensive cues on dogs' performance in a manipulative social learning task. Applied Animal Behaviour Science, 120(3–4), 170–178. https://doi.org/10.1016/j.applanim.2009.05.012

Scandurra, A., Alterisio, A., Marinelli, L., Mongillo, P., Semin, G. R., & D'Aniello, B. (2017). Effectiveness of verbal and gestural signals and familiarity with signal-senders on the performance of working dogs. Applied Animal Behaviour Science, 191, 78–83. https://doi.org/10.1016/j.applanim.2017.02.003

Schmidjell, T., Range, F., Huber, L., & Virányi, Z. (2012). Do owners have a Clever Hans effect on dogs? Results of a pointing study. Frontiers in Psychology, 3, 558. https://doi.org/10.3389/fpsyg.2012.00558

Schwab, C., & Huber, L. (2006). Obey or not obey? Dogs (Canis familiaris) behave differently in response to attentional states of their owners. Journal of Comparative Psychology, 120(3), 169–175. https://doi.org/10.1037/0735-7036.120.3.169

Topál, J., Gergely, G., Erdőhegyi, Á., Csibra, G., & Miklósi, Á. (2009). Differential sensitivity to human communication in dogs, wolves, and human infants. Science, 325(5945), 1269–1272. https://doi.org/10.1126/science.1176960

Topál, J., Miklósi, Á., Sümegi, Z., & Kis, A. (2010). Response to comments on "Differential sensitivity to human communication in dogs, wolves, and human infants." Science, 329(5988), 142. https://doi.org/10.1126/science.1184152

Udell, M. A. R., Dorey, N. R., & Wynne, C. D. L. (2008). Wolves outperform dogs in following human social cues. Animal Behaviour, 76(6), 1767–1773. https://doi.org/10.1016/j.anbehav.2008.07.028