Introduction

In the last two decades, a large body of research has established that dogs (Canis familiaris) show sophisticated social cognitive abilities in interspecific interactions with humans (Hare and Tomasello 2005; Reid 2009; Udell and Wynne 2008; Takaoka 2009). Dogs can exploit a range of behavioral cues from humans, with some of their abilities paralleling social cognitive skills of humans (e.g., Adachi et al. 2007; Hare and Tomasello 2005; Takaoka et al. 2013). For example, Adachi et al. (2007) reported that dogs looked longer at visual images of people when these images did not match the voice of someone calling them, suggesting that dogs have cross-modal representations of particular humans. Using the same expectancy-violation paradigm, Takaoka et al. (2013) demonstrated that dogs categorize human gender using visual and auditory information.

Several experiments suggest that domestic dogs are particularly sensitive to human pointing gestures (Bräuer et al. 2006; Hare and Tomasello 1999; Miklósi et al. 1998). In several of these studies, the so-called object choice task was used to test sensitivity to human communicative gestures. In this task, a human experimenter first hides a reward in one of the two identical containers out of view of the subject. The experimenter then provides the subject with a cue such as orienting, leaning, or pointing toward the baited container, and the subject is allowed to make a choice. Dogs can use various human gestures including pointing, gazing, bowing, nodding, and eye glance to find hidden rewards.

Dogs’ responsiveness to human communicative cues appears very strong; they will even try to use misleading signals given by humans. Szetei et al. (2003) showed that in a two-alternative object choice task, dogs chose an empty container indicated by the experimenter even after they saw or sniffed the baited container. In a related study, Prato-Previde et al. (2008) showed that dogs tended to choose a plate containing a smaller quantity of food if their owners misled them by showing a strong preference for that plate rather than another with a much larger quantity of food. Even a stranger can mislead dogs in this way (Marshall-Pescini et al. 2011b).

Although dogs respond positively to human communicative gestures, they do not follow pointing gestures automatically. Petter et al. (2009) found that, after intensive training, dogs learned to approach a cooperative human who always pointed at a baited container (honest pointing) more often than a deceptive human who always pointed at an empty container (dishonest pointing). Kundey et al. (2010) expanded Petter et al.’s (2009) study using three kinds of gestures: static pointing, momentary pointing, and standing behind a container. They found that dogs had difficulty inhibiting approach directed by static pointing in comparison with the other gestures. Kundey et al. (2010) further explored whether additional training would improve dogs’ ability to inhibit following dishonest static pointing. Dogs did learn to do this, but only if the reward was visible at choice.

Although the above results suggest that dogs have difficulty inhibiting approaching a location indicated by misleading pointing, it is not clear how dogs evaluate the human who provides misleading pointing gestures. Prior deceptive behavior by a particular person is an important clue to that individual’s future trustworthiness (Vanderbilt et al. 2011). Kundey et al. (2010) did not include a condition in which the experimenter provided correct information after providing incorrect information. That is to say, in Petter et al. (2009), dogs may have interpreted the dishonest pointing as a simple discriminative cue to avoid that container. Another possibility is that they inferred something about the behavioral or personal characteristics of the misleading pointer. However, little work has been done on dogs’ judgments of particular people’s traits, and whether they can use this information flexibly when subsequently responding toward those people.

Human children monitor the past accuracy of informants and use this information when they decide which informant to trust (Corriveau and Harris 2009; Harris and Corriveau 2011). They prefer to accept information from someone who has proved accurate rather than someone who has made errors when naming objects (Koenig et al. 2004). Gaze-following behavior is also influenced by children’s previous experience with the looker; they follow the gaze of a reliable looker more than that of an unreliable looker (Chow et al. 2008). This tendency to trust some informants more than others is called “selective trust.”

Human infants begin to understand and produce pointing at an early stage in their first year of life (Behne et al. 2012). Pointing is such a conventional communicative gesture that young children have difficulty interpreting it in a novel, unconventional way. Couillard and Woodward (1999) investigated young children’s ability to comprehend deceptive pointing. An experimenter provided a misleading cue about the location of a reward by pointing at the container including no reward. Children were found to have difficulty rejecting the deceptive advice. In Vanderbilt et al. (2011), children first observed the pointer helping finders locate a hidden reward (helper trial) or tricking finders into looking for the reward in a wrong location (tricker trial). In the choice test that followed, 5-year-olds trusted the helper’s advice more often than the tricker’s, suggesting that young children inferred whether the pointer was reliable or not from observing his/her behaviors, and used this knowledge to predict the pointer’s future behavior. In the same study, 3- and 4-year-olds showed poorer performance than 5-year-olds, suggesting that selective trust develops over time.

At present, we know little about whether dogs can track the reliability of a person and use the information to adjust their behavior flexibly. Marshall-Pescini et al. (2011a) reported that dogs distinguished between generous and selfish actors after observing their food-related interactions with a beggar. When the beggar tapped the generous actor on the arm, the latter gave a small piece of cereal to the beggar, whereas the selfish actor refused to do so. When dogs were then allowed to move freely, they preferentially approached the generous actor. This study may suggest that dogs evaluate characteristics of actors from third-party interactions, but a simple association between the generous donor and the reward cannot be ruled out. In Kundey et al. (2011), dogs observed interactions between a recipient and two demonstrators, who both showed a treat to the recipient but then either gave it to her or not. Subsequently, when dogs were allowed to choose one of the demonstrators, they showed a strong preference for the giver. Again, however, this preference may result from the association between the giver and the reward. Furthermore, dogs are likely to have experienced similar activities with their owners and others during their daily lives, learning about features of people who give them food; this may generalize to the generous demonstrator in experimental situations.

In contrast, most studies on children’s selective trust have been conducted within less familiar contexts, including word learning (Harris and Corriveau 2011; Koenig et al. 2004) and comprehension of deceptive pointing (Heyman et al. 2013; Vanderbilt et al. 2011). In Harris and Corriveau (2011), the unreliable informant named the familiar object incorrectly, for example, by saying “That’s a cup” when presented with a ball. In Vanderbilt et al. (2011), the unreliable pointer provided young children with misleading pointing to an empty container.

Misleading pointing should be an unfamiliar gesture for dogs, because dog owners normally point toward locations they want the dog to look at or approach. No research has examined how dogs evaluate humans who provide misleading cues. Furthermore, it is unclear whether dogs can infer how reliable someone is based on their own direct interactions with that person, and generalize the knowledge to predict the person’s subsequent behavior. To this end, we investigated how dogs would react to honest pointing gestures after observing the same human pointing misleadingly in an object choice task. We tested dogs in three phases. Phase 1 was a regular object choice task in which an experimenter pointed at the correct (baited) container. In Phase 2, the same experimenter deliberately pointed at the empty container after explicitly showing the dogs which container was baited. Phase 3 was identical to Phase 1. Our question was whether dogs would again respond positively to the pointing gesture even after they observed misleading pointing in Phase 2. If dogs judge someone as unreliable after receiving a misleading cue, they might transfer this evaluation to a subsequent situation, in which case they should be less likely to follow the same experimenter’s pointing in Phase 3 than in Phase 1. Alternatively, if dogs do not use their experience in this way, or if they simply do not draw inferences about the reliability of the experimenter, they should continue to follow the experimenter’s pointing gestures even after receiving inaccurate information.

Experiment 1

Experiment 1 tested the basic proposition that dogs would cease to follow the points of an informant who proved untrustworthy.

Methods

Subjects

Thirty-four pet dogs (mean = 4.32 years, SD = 3.28, 10 males and 24 females) and their owners participated. The subjects included 2 Airedale terriers, 1 Border collie, 5 Chihuahuas, 1 Golden retriever, 2 Miniature dachshunds, 11 Pomeranians, 1 Shetland sheepdog, 2 Toy poodles, 1 Wire-haired fox terrier, 3 Japanese spitzes, and 5 Mongrels. Their participation was voluntary. The owners were given a detailed explanation of the test procedure before signing an informed consent form. The owners remained blind to the study’s purpose to exclude Clever Hans effects, and they were forbidden from pointing or making related gestures throughout the experiment.

Apparatus

Two identical plastic opaque containers were used. Each dog’s favorite food, such as dog biscuit or jerky, was used as a reward. To control olfactory cues, we hid a piece of the same food on the underside of the bottom of each container; this food was not visible to the dog.

Procedure

We used a two-alternative object choice task. The procedure consisted of pre-training and three test phases. The experiment took place in Kyoto University Kokoro laboratory room, an unfamiliar location for the dogs.

Pre-training

Preliminary training served to accustom the dog to the experimental situation. Two containers were positioned upside down, 1 m apart from each other. An experimenter sat approximately 30 cm back from and equidistant from the containers. The owner was instructed to restrain the dog lightly at about 1.5 m from the containers by holding the dog’s leash or harness. The experimenter called the dog’s name and visibly hid a piece of food under one of the containers. The experimenter then lifted both containers simultaneously and said “OK, let him/her go,” to ask the owner to release the dog. The trial was repeated up to four times (two trials with the reward at each side) until the dog came up to the baited container without hesitation.

Test

Phase 1 Neither dog nor owner observed the baiting process: The owner covered the dog’s eyes with his/her hand or the dog and the owner faced away during the process. They were allowed to see the containers once baiting was completed. The experimenter asked the owner to close his/her eyes during the next pointing cue to avoid any Clever Hans effect. The experimenter called the subject’s name to draw its attention. As the dog watched, the experimenter extended her arm with her index finger extended to about 5 cm from the baited container (momentary proximal pointing) three times. The experimenter continued looking at the dog’s face during pointing. After the pointing gesture, the experimenter placed her hands on her lap, bent her head to look down centrally between the containers, and said “OK, let him/her go,” to ask the owner to release the dog. The experimenter maintained her position until the dog chose. The experimenter upturned the first container the dog approached to within 10 cm and gave the reward to the dog if it was there. If the dog chose the unbaited container, the experimenter repeatedly showed the dog that it was empty. The dog was allowed to choose only one container. The procedure was then repeated with the baited location reversed; this second trial completed Phase 1.

Phase 2 The procedure was identical to Phase 1 except for the following: First, the dog was shown the content of both containers (food or nothing). After baiting one container, the experimenter called the dog’s name to get its attention and lifted both containers simultaneously to deliberately show which container had the food. The experimenter continued this, sometimes by tapping the containers until she confirmed that the dog visually checked both containers. Second, the experimenter then pointed at the unbaited container, instead of the baited one. Thus, the dog would find the food only by choosing the container not pointed at by the experimenter. Another trial was then conducted with the baited location counterbalanced.

Phase 3 The procedure was identical to Phase 1. Two trials were run with the baited location counterbalanced.

Thus, six test trials in total were given to each dog. No side was baited more than twice consecutively. We asked whether dogs would follow the experimenter’s pointing gesture even after the misleading pointing by the experimenter in Phase 2. If a dog failed to choose a container within 30 s, a “no choice” response was recorded and the next trial started. All trials were videotaped for later analysis (30 frames/s).

Data analysis

We initially coded dogs’ choices during the test trials. Because the choices were unambiguous, the same experimenter subsequently confirmed the choices from videotape. Within-observer reliability was 100 %. For each phase, we categorized dogs that twice chose the container pointed to by the experimenter as “Obedient choosers”. Dogs showing any other choice pattern—following the pointing cue in zero trials or one trial only—were recorded as “Non-obedient choosers”. We calculated the numbers of “Obedient choosers” and “Non-obedient choosers”, and used binomial tests to determine whether the proportion of Obedient choosers was above chance, which was .25, in each phase. We also used a McNemar test to examine whether a significant proportion of Obedient choosers in Phase 1 switched to Non-obedient choosers in Phase 3, analyzing data from only those dogs that completed both trials of Phase 2.

We measured latency from a frame-by-frame analysis of the video recordings, defining latency as from the moment the dog was released until it made a choice. A second observer, blind to the purpose of the study, scored a randomly selected sample of trials (20 %). Pearson’s correlation between the two observers’ scores was high (r = 0.92). We used one-way repeated-measures ANOVA to determine whether latency to choose differed in the three phases. Only trials in which dogs made a choice were included in the latency analysis.

To assess behavioral differences in the three phases, from video recordings we also analyzed gaze alternation, defined as looking between a person (the experimenter or the owner) and either container during the choice time. As relatively few dogs showed gaze alternation, we counted those dogs that showed the behavior at least once.

All experiments in this paper adhered to the ethical standards of Kyoto University and were approved by the Animal Experiment Committee of the Graduate School of Letters, Kyoto University.

Results

Eight dogs were excluded from the analysis because they failed to complete pre-training. We analyzed the behavior of all dogs that completed two trials in each phase.

Latency

In the latency analysis (see Fig. 1a), we excluded three dogs that made no choice in any trial of Phases 1 or 3. The ANOVA showed a marginally significant difference in latency across phases (F 2,44 = 3.01, P = .06). Post hoc pairwise comparison with Bonferroni correction revealed a significant difference between Phases 2 and 3 (P = .05), but not between Phases 1 and 2, or 1 and 3. As seen in Fig. 1a, the results suggest that dogs were generally slower to make a choice in Phase 3.

Fig. 1
figure 1

Mean latency (±SEM) to make a choice (a Experiment 1, b Experiment 2)

Choice

The numbers of Obedient and Non-obedient choosers are shown in Fig. 2a. Some of the Non-obedient choosers followed the experimenter’s pointing in only one of the two trials. As preliminary analysis revealed no difference between dogs that chose obediently on trials 1 and 2, we pooled the trials. In Phase 1, 14 of 24 dogs (58 %) were Obedient choosers (binomial test: P < .01, chance = .25), compared to only 6 of 25 dogs (24 %) in Phase 2, (binomial test: P = .62), and only 2 of 16 dogs (13 %) in Phase 3 (binomial test: P = .94). Thus, the proportion of Obedient choosers was significantly above chance only in Phase 1. The difference in the proportion of Obedient choosers and Non-obedient choosers between Phases 1 and 3 was significant (McNemar test: P < .01). These results indicate that most dogs followed the experimenter’s initially reliable pointing cue to find the food in Phase 1. However, after being exposed to misleading pointing by the same experimenter in Phase 2, fewer dogs relied on the experimenter’s pointing in Phase 3, despite it being the only available cue.

Fig. 2
figure 2

Percentage of dogs which followed the pointing (a Experiment 1, b Experiment 2). Black represents dogs following the experimenter’s pointing in both trials (obedient chooser). Gray represents dogs following in one trial. White represents dogs choosing the container which the experimenter did not point to in both trials

Note that the number of dogs completing each phase decreased progressively. Two dogs in Phase 1, one dog in Phase 2, and ten dogs in Phase 3 failed to make a choice within 30 s (no choice). In Phase 3, six of the ten dogs that did not choose showed gaze alternation at least once; after release, they sat equidistant between owner and experimenter and alternated gaze among the owner, the experimenter, and the containers. No dog showed gaze alternation in either Phase 1 or Phase 2.

Discussion

This experiment investigated whether dogs would follow an experimenter’s initially reliable pointing gesture after being exposed to misleading pointing by the experimenter in Phase 2. Although the procedure in Phases 1 and 3 was identical, most dogs followed the experimenter’s pointing in the former, but not the latter. This result suggests that not only can dogs use human communicative cues to find hidden food, but also they modify their behavior based on recent experience, an effect attributable to observing the experimenter’s misleading pointing in Phase 2. In other words, dogs were sensitive to the reliability of the human who gave the cues and their evaluation influenced their subsequent behavior.

Dogs took longer to make a choice in Phase 3 compared to earlier phases. This reflects ambivalence about the experimenter-given cue following misleading cues in Phase 2. Gaze alternation, observed only in Phase 3, also suggested hesitation about obeying the experimenter’s pointing after prior misinformation. Gaze alternation by dogs has been interpreted as a request for human intervention in the case of being unable to solve a task by themselves (Miklósi et al. 2003). Together, these results strongly suggest that dogs reacted differently to the experimenter in Phases 1 and 3.

However, it could be argued that fatigue or diminishing motivation to participate in the task across phases affected the dogs’ behavior, especially in view of the fact that their obedience was not rewarded in Phase 2. We conducted a second experiment to address this issue.

Experiment 2

The procedure was identical to Experiment 1 except that first experimenter left the room after Phase 2, to be replaced by another experimenter who provided the pointing cue in Phase 3. The prediction was that if the dogs’ behavioral change in Experiment 1 was because they made an inference about the reliability of the first experimenter, they should follow the pointing gesture by the novel experimenter as they did in Phase 1. Alternatively, if dogs had simply lost motivation to participate in the task across phases, fewer dogs should follow the pointing gesture by the novel experimenter.

Methods

Subjects

Thirty-one pet dogs (mean = 4.32 years, SD = 2.78, 17 males and 14 females) and their owners participated. The subjects included 1 Airedale terrier, 1 Border collie, 2 Chihuahuas, 2 Welsh Corgi Penbrokes, 2 Flat-coated retrievers, 4 Labrador retrievers, 1 Lakeland terrier, 4 Miniature dachshunds, 1 Pitbull terrier, 1 Pomeranian, 1 German Shepherd, 2 Shibas, 1 Toy poodle, 2 White shepherds, 1 English springer spaniel, 1 West highland white terrier, 1 Cavalier King Charles spaniel, 1 Sealyham terrier, and 2 mongrels.

Apparatus and procedure

The apparatus and procedure were identical to those in Experiment 1, except that the first experimenter left the experimental room after Phase 2 and another experimenter of the same gender entered the room. It took about 10 s for the change. In Phase 3, the second experimenter gave the pointing gesture in the same way as the first experimenter in Experiment 1.

Data analysis

Coding of the videos and analyses were the same as in Experiment 1. Interobserver reliability of the latency measure based on 20 % of the trials was high (Pearson’s correlation: r = .96). Additionally, to examine whether dogs responded differently to the experimenters between Experiments 1 and 2, we used a Fisher exact test to compare the proportion of the dogs that followed the pointing in both trials and the dogs that did not in Phases 1 and 3. For this analysis, we excluded dogs that followed pointing in only one of the two trials, because these dogs chose the same side in both trials, possibly reflecting a simple side bias. Finally, we used a generalized linear mixed model (GLMM) analysis with a binomial distribution and a logit link function to examine the effect of different factors on the dogs’ choices. The dependent variable was choice in each phase (2: dogs followed pointing in both trials, 1: dogs followed pointing in 1 trial, 0: dogs never followed pointing). As fixed factors, we included experiment (Experiments 1 and 2), phase (Phases 1 and 3), and the interaction between the two. Each dogs’ identity was entered as a random factor. To find the best (most parsimonious) model with the lowest number of factors, we input all variables that were likely to affect dogs’ choices and excluded variables that failed to contribute to decrease Akaike’s information criterion (AIC) of the model in a stepwise manner. The analysis was run on R version 3.1.1 using the glmmML function included in the glmmML package.

Results and discussion

Five dogs were excluded from the analysis because they were failed to complete pre-training.

Latency

Three dogs that failed to choose in either trial of Phase 3 were excluded from the analysis of latency (Fig. 1b). A one-way repeated-measures ANOVA revealed no significant difference in latency across phases (F 2,44 = 1.79, P = .18), suggesting that dogs showed no hesitation to follow the second experimenter’s pointing cue despite the first experimenter providing misleading cues earlier.

Choice

The numbers of Obedient and Non-obedient choosers are shown in Fig. 2b. Preliminary analysis revealed no difference between dogs that chose obediently on trials 1 and 2, so the trials were pooled. In Phase 1, 16 of 26 dogs (62 %) were Obedient choosers (binomial test: P < .01, chance = .25), compared to 3 of 23 dogs (13 %) in Phase 2 (binomial test: P = .95), and 7 of 18 dogs (39 %) in Phase 3 (binomial test: P = .13). In Phase 3, relatively more dogs followed pointing in both trials, but the number of Obedient choosers was significantly above chance only in Phase 1. However, in contrast to Experiment 1, the proportion of Obedient choosers and Non-obedient choosers between Phases 1 and 3 was not statistically different (McNemar test: P = .12). The number of dogs completing both trials decreased across phases; 0, 3, and 8 dogs failed to make a choice within 30 s in Phases 1–3, respectively. However, no dogs showed gaze alternation.

We compared the dogs’ choices between Experiments 1 and 2. The only factor remaining on the best GLMM model was type of the interaction between experiment and phase. We presented the effects of the factor in the best model (Table 1). In Phase 1, Experiment 1, 14 dogs followed pointing and 1 dog failed, whereas in Phase 1, Experiment 2, 16 dogs followed pointing and 2 dogs failed. Thus, in Phase 1, the proportion of dogs that followed the pointing and the proportion of dogs that failed was not significantly different between Experiments 1 and 2 (Fisher exact test: P = 1). In contrast, in Phase 3, Experiment 1, 2 dogs followed the pointing and 5 dogs failed, whereas in Phase 3, Experiment 2, 7 dogs followed the pointing and 1 dog failed. The proportion of dogs that followed the pointing or not was significantly different between Experiments 1 and 2 (Fisher exact test: P = .04). These results indicate that the dogs in Experiment 2 followed the pointing in Phase 3, in contrast to the dogs in Experiment 1.

Table 1 GLMM parameter estimate in Experiment 1 and Experiment 2

The above results rule out simple lowered motivation for the task across phases. Dogs used the pointing cue from the novel experimenter to locate the food. Combined with those from Experiment 1, these results confirm that not only can dogs use human communicative cues to find hidden food, they can also change their behavior adaptively depending on their preceding experience with cues.

General discussion

This study investigated whether dogs followed human pointing gestures after observing the human giving them a misleading cue. In Experiment 1, most dogs followed the experimenter’s pointing in Phase 1, but the proportion doing so declined markedly in Phase 3, although the task was identical. In Experiment 2, dogs used a pointing gesture by a novel experimenter in Phase 3 even after experiencing the first experimenter’s misleading gesture, suggesting that the reluctance to follow the experimenter’s pointing in Phase 3 of Experiment 1 was not simply due to decreased motivation. We believe these results provide strong support for the finding of Experiment 1.

One possible objection to our conclusion is that dogs might have learned to avoid the container to which the first experimenter pointed in Phase 2. However, as there were only two trials in each phase, it seems unlikely that the dogs could have learned such a relation between cues and outcomes. In fact, in Petter et al. (2009), dogs required many trials to learn to respond differently to a cooperator and a deceiver. In their study, in cooperator trials, the experimenter always pointed at the baited container, whereas in deceiver trials another experimenter always pointed at an empty container. Although dogs initially tended to obey cues even when being misled, they eventually learned to approach the cooperator more often than the deceiver.

Might the delay to change experimenters in Experiment 2 account for dogs’ responses in Phase 3? We believe it unlikely that dogs completely lost their memory within that short time. In support of our view, Fujita et al. (2012) demonstrated that dogs can maintain information about particular experiences for more than 10 min, during which they walked around outside with no expectation of a memory test that followed (i.e., incidental memory).

Studies in human children suggest another possible explanation. Pet dogs are often treated as family members, so infants and dogs share a similar social environment. Such environmental conditions offer the possibility for the development of similar behavioral traits (Lakatos et al. 2009). Infants begin to understand others’ pointing at an early age (Itakura and Tanaka 1998; Lakatos et al. 2009), and puppies can also use pointing to find hidden food (Gácsi et al. 2009; Hare et al. 2002; Riedel et al. 2008). Furthermore, young children appraise the reliability of informants and use this knowledge to decide which informant to trust subsequently (Corriveau and Harris 2009; Koenig et al. 2004). Whereas children categorize an unfamiliar informant as trustworthy by default, this is reversed if an informant provides the children with incorrect information (Vanderbilt et al. 2011). The present study suggests that dogs engage in a similar form of appraisal. Dogs readily used the pointing cue in Phase 1, suggesting a default trust of information provided by a human. This trust was challenged when the pointer provided the dogs with incorrect information in Phase 2; in Phase 3, the dogs’ response to the cue by the same experimenter was quite different, with fewer dogs following the pointing cue. The facts that dogs took longer to choose in Phase 3 compared to other phases in Experiment 1 and that gaze alternation by dogs occurred only in Phase 3 of Experiment 1 further indicate ambivalence about cues from an experimenter who previously provided inaccurate information.

This research represents a step forward in understanding dogs’ ability to infer human traits based on their prior experience with the humans. Previous studies reported that dogs had difficulty inhibiting an approach to a location cued by a misleading point (Kundey et al. 2010; Petter et al. 2009), but none examined how dogs actually judge the human who provided the misleading information. Here, we showed that dogs drew an inference about the reliability of the misleading human and used this inference to change their response to the informant.

Young children make inferences about reliability not only from direct interactions (Chow et al. 2008) but also from observations of the third-party interaction (Vanderbilt et al. 2011). In Vanderbilt et al. (2011), children watched a video in which a reliable informant pointed at a correct container to help a finder to locate a hidden reward, but an unreliable informant pointed at an empty container to trick a finder into looking for a reward there. After watching the video, more 5-year-olds trusted pointing by the reliable informant than the other. Marshall-Pescini et al. (2011a) showed that dogs preferred a generous donor over a selfish donor after observing their food-sharing interaction with a beggar. However, as noted in the Introduction, we cannot exclude the possibility that dogs may have learned the features of a person who is willing to give them food in daily interactions. Nitzschner et al. (2012) designed a task that did not involve food. In their second experiment, dogs witnessed two interactions as a bystander. The “nice” experimenter behaved in a friendly manner with another dog, but the “ignoring” experimenter walked through the experimental area without interacting with another dog. After the observation, the subjects showed no preference for the nice experimenter, suggesting that dogs may not use indirect experiences to evaluate humans. Failure to use the third-party interactions may be because the human–dog encounter was not strong enough to keep the dog’s attention. More studies are needed on dogs’ evaluations of third-party interactions.

A final interesting question concerns the extent to which dogs generalize the reliability of an informant. In the present study, the same communicative gesture was used across the three phases. Would dogs use their inference about the reliability of an informant in different situations? In obedience training, dogs might refuse to obey commands given by someone who previously gave misleading information. It would also be interesting to study the duration of selective trust in particular individuals. This could be tested by postponing Phase 3 for several hours, or even days.

In sum, we have shown that although dogs readily follow pointing by an unfamiliar person to locate hidden food, they change their behavior if they recently experience misleading pointing by that person. The results suggest that dogs, like human children, make inferences about the reliability of an informant from her prior behavior and generalize this inference to predict the person’s behavior in the future situations. Thus, dogs have problem-solving skills that are functionally similar to those observed in human children. To adapt to human society, dogs would benefit from the ability to monitor people’s reliability to predict their future behaviors.