Introduction

A typical situation on a dog walk is that while two dogs are interacting, their owners like to interpret the behavior demonstrated by their dogs in terms of meaning and function. The owners often disagree with each other, as one may see a playful interaction whereas the other perceives it as rather aggressive. Even extensively trained humans, like behavioural scientists who observe human or animal interactions professionally, sometimes have difficulty to define and agree about the context in which a social interaction takes place (e.g. playful or aggressive) and then also understanding the meaning of behavioral signals [1, 2].

In order to ensure an optimal reaction in a given situation, however, it may be more important to predict the outcome of a situation rather than to define its context, per se. Using the example with the two dogs, determining to what extent the interaction is playful or aggressive is less critical than predicting whether or not the dogs are going to attack and harm each other in the next seconds. To make reliable predictions about future events and to minimize the costs of surprise, humans apply knowledge acquired during previous experiences to new situations (see [3, 4]). As accurate predictions usually lead to appropriate action choices (e.g. avoid risks or identify and react to danger, recognize beneficial chances or reflect mistakes, see [4–6], they have a beneficial effect on survival. In other words, predicting the outcome of a situation is highly adaptive.

Predictions about others’ intentional mental states and future actions are crucial during social interactions. To attribute intentional states to others, humans need to understand actions as being goal-directed. This ability, which is also important to predict outcomes of actions, emerges at a young age. Children around the age of six months are already able to perform goal-directed gazing [7, 8], a skill which continues to develop over the years. For example, Cannon & Woodward (2012) found that, in their first year, infants are already developing an understanding of goal-directed actions to make online predictions of others’ next actions [9]. They look predictively to an object that another person had previously reached, even when the object’s location had changed. For adults it was shown that within complex scenes, viewers tend to focus their attention on the most meaningful and task-relevant information [10], represented by a predictive gazing pattern (e.g. predicting the bounce point of a tennis match with direct fixation [11]). These findings suggest that human subjects rely more on predictive gazing (by producing eye movements similar to those produced when performing the task) than on reactive gazing (i.e. analysing the action) when observing a task [12], showing how important predictions of actions are in everyday life. However, so far most studies have focused on how humans predict the outcome of complex social interactions among members of their own species (but see [13]). But it would also be highly adaptive to predict the behaviour of other species, in particular when there is a risk of injury (see also dog example above).

While this gap in the literature still exists, much is known about how humans define the social context of a given situation and from it infer the meaning of the behaviours demonstrated by the subjects involved in the interaction. A typical approach for behavioural scientists is to establish ethograms by coding a huge number of interactions and their outcomes to interpret behaviors [14]. When those ethograms are established for a given animal, it is possible to identify social contexts by recognizing typical behaviours such as gestures. For example, thorough research has made it possible for experts to unambiguously recognize playful contexts in chimpanzees by detecting typical play gestures [15–17], affiliative contexts in ravens marked by the offering of non-edible items [18] and playful contexts in dogs characterised by their typical play bow [19].

However, some types of interactions are more difficult than others for non-experts to assign context to. Whereas it seems generally easy for humans to interpret a situation based on auditory signals in primates [20] and in dogs [21], the picture is much more complex for visual signals.

For example, most people are quite unskilled in interpreting the emotional states of chimpanzees [22], likely because chimpanzees have facial expressions that functionally differ from those of humans, despite the obvious physical similarities [23, 24]. Also, humans generally have problems interpreting the body language of dogs [22, 25–31]. Such interpretive decisions also depend on the kind of context. For example, Amici and colleagues (2019) found that six-year-old kids performed above chance level in interpreting dog emotions. Specifically, kids were better at detecting aggressive expressions than fearful and sad expressions. Indeed, from an evolutionary point of view, humans should predict aggressive outcomes better than playful and neutral ones (e.g. [4, 6]), because knowing when another individual of any species is aggressive might provide a clear selective advantage as one can react accordingly and avoid a potentially deadly encounter.

But even within the human species, we cannot be sure that we all interpret the emotional signals of the individuals involved in a situation the same way. Whereas it has long been assumed that human emotions are universal [32], more recent studies have shown that there are huge cultural differences in how emotions are expressed, interpreted, and labelled [33–38].

Overall, these studies provide evidence that humans can selectively attend to the emotional context of a social situation in order to assess it and make predictions about possible future outcomes. However, results from the above-mentioned studies do not allow reliable conclusions about our ability to predict intra- and interspecific social interactions in dynamic natural settings to be drawn. In the current study, we therefore aimed at testing the ability of healthy adult participants to assess various social situations between members of three species. For this purpose, we presented short video clips of three different real-life interactions (aggressive, neutral and playful) of dyads of dogs, monkeys (macaques) and human children. We then asked participants to either categorize the context of the situation–i.e. whether it was aggressive, neutral or playful–or to predict the outcome of the observed interactions. Dogs were chosen to assess interspecific interactions because they have been living closely with humans for up to 30,000 years and have undergone a special domestication process [39]. Macaques were chosen because of their relatedness as a primate species.

We hypothesized that humans would perform above chance level in predicting the outcome of social interactions across all species, as this improves the ability to react accordingly and economizes cognitive effort, which is beneficial for survival. Secondly, we hypothesized that participants would perform better at predicting the outcome of an interaction compared to categorizing the context, as the first task is more important for survival and therefore more adaptive than the second. Thirdly, we hypothesized that performance in assessing social interactions would vary depending on context. In particular, we expected that participants would perform better at rating the context and predicting the outcome of aggressive than they would for playful and neutral situations, as this is highly adaptive. Finally, we hypothesized that humans would be better at assessing social interactions between members of their own species due to inherent intra-species advantage, i.e. that we interact more frequently with other humans than we do with other species.

Methods

Results

Fig 1 presents the average context categorization scores and Fig 2 presents the average predictions of outcomes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Depicts the average correct responses for the context identifications for each species and context in group 1.

https://doi.org/10.1371/journal.pone.0277783.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Depicts the average performance of the participants for the estimation of the future outcome of the interactions for each species and context in group 2.

https://doi.org/10.1371/journal.pone.0277783.g002

Participants performed similarly at predicting the outcome and categorizing the context (main effect of type of task: F(1, 90) = 8.70e-4, p = 0.977; η² = 9.67e-6). However, there was a significant 3-way interaction of species, context and type of task (F(4,360) = 11.160, p < .001; η² = 0.110), indicating that the performance of participants in the two groups differed depending on species and outcome. Furthermore, there was a significant interaction between species and type of task, indicating that participants in the two groups showed performed differently depending on the species (F(2, 180) = 21.40, p < .001; η² = 0.192). However, context did not impact participants’ performance, as there was no significant interaction between type of task and context (F(2,180) = 0.943, p = 0.392; η² = 0.110). Moreover, we found a significant interaction between species and context (F(4,360) = 29.40, p < .001; η² = 0.246), demonstrating that participants performed differently for the three species and this effect depended on the context. Finally, performance was different for the three species (main effect of species: F(2,180) = 17.44, p < .001; η² = 0.162), with all participants performing better with primates than dogs. Performance was also different for the three main contexts (main effect of context: F(2,180) = 66.51, p < .001; η² = 0.425).

For the comparisons against chance level and the pairwise comparison we first consider the two types of task (tested in the groups) separately. Then we present the relevant comparisons between groups.

Discussion

The participants in our study performed above chance level by predicting social situations in humans, dogs and monkeys. Presented with a choice of three outcomes, participants selected the correct one in 50–80% of interactions—as our first hypothesis predicted. Thus, for most of the nine situations participants were able to predict the outcome of the social interactions for all the species.

Yet, our data does not support our second hypothesis, as participants predicting the outcome of a social interaction in general did not perform better than participants who had categorized the context. Only in neutral human interactions did participants perform better at predicting outcomes than categorizing context. This shows that contexts can indeed be ambiguous, particularly for human interactions [see also 38]. It furthermore suggests that predicting the outcome of a given situation can sometimes be more constructive than simply categorizing its affective content. Surprisingly, there was one situation in which participants categorized the context better than predicting its outcome–i.e. in neutral situations with monkeys. Here, the performance in categorizing the context was very high–over 90%, whereas outcome predictions were about 60%. Thus, categorizing the context can be useful in general, but depending on the situation and the species, either categorizing or predicting may be more accurate. Obviously, these two tasks are not entirely independent of each other. When participants make a decision about the potential outcome of a situation, it is very likely that they also consider the social context, and when their task is to rate the context, they might also do it by partly taking into account its outcome. However, sometimes humans seem to consider context and outcome independently. In an everyday situation they might for example only speculate about the outcome without thinking how to define the context. It might also be important in what order the two tasks are tested experimentally (see S1 File). To investigate that question further, future studies should concentrate on one context and one species to exclude possible interaction effects between these factors.

Future studies should also take into account individual differences. Participants in our study that were especially skilful in predicting the outcome for one species or one context, also performed well for the other species or in the other contexts. Thus, there are people who are skilled predictors of social interactions in general–independent from species and context. Interestingly, we did not find such a correlation for rating the context, also indicating that contexts can indeed be ambiguous.

Our third hypothesis was that participants would be overall better at assessing aggressive situations than playful and neutral ones, independently of the species. We did not find evidence in our data to support this hypothesis. In contrast, participants performed poorly when assessing dogs`aggressive behaviour. In particular, they rated aggressive contexts among dogs at chance level, and they predicted outcomes below chance level. They also assessed aggressive interactions in dogs worse than playful and neutral ones. Thus, dogs`aggressive behaviour is not well-recognized. In addition, participants were unable to predict what could potentially occur next. Furthermore, other studies have shown that humans perform surprisingly poorly at detecting anxiety and aggression in dogs [28, 40, 41], but see also [22]. This is most likely the reason for the relative frequency of reported biting incidents [42, 43], as humans fail to notice dogs`displacement and appeasement behaviors before an attack [44]. A possible method of preventing severe biting incidents could entail that prospective dog owners are better educated about dog behaviour before adopting, as it has also been found that owning a dog does not improve the ability to assess dog behaviour [13, 45].

Interestingly, participants in our study also underestimated human aggressions. Participants performed below chance level at assessing the context, and also failed to reliably predict the outcome of aggressive interactions, performing worse than with playful and neutral contexts. It is possible that humans are biased to assume good intentions from other humans and from “man’s best friend”, sometimes preventing us from recognizing aggressive situations in these species.

In general, participant performance depended on both species and context. Thus, we have to reject our fourth hypothesis that humans will generally be better at assessing social interactions between members our own species than of other animal species. In the following, we consider the three contexts separately. For the playful context, participants assessed the situation better for dogs and humans than for monkeys. It is possible, that dogs`playful faces might be easier for humans to understand than the playful faces of other primates [46]. Also, other studies have found that humans more easily recognize positive dog emotions [13, 25, 28, 30, 31]. This could potentially be a consequence of co-domestication, meaning that through the process of domestication, humans became better at assessing when a dog is playing, and is therefore non-threatening. However, in other studies we found no evidence that humans have an innate tendency to understand dogs, in particular when it comes to aggressive interactions [13, 22], see also above.

Also, in contrast to our fourth hypothesis, participants were not better at assessing the social interactions of humans than of monkeys and dogs. (This was only the case for predicting the outcome in the neutral context.) Thus, we did not find a clear difference between species.

One could argue that our clips were not representative of species behaviors, or that the perceived intensity of the emotional contexts might differ across clips, or that some clips may not have enough context-specific information to be successfully assessed by participants. However, the clips were carefully selected with effort to eliminate biases towards any species. Indeed, each clip was assessed correctly either for context or outcome by at least 3 to 23 participants, indicating that the clips had sufficient context-specific information to be successfully assessed, but they were not self-explaining (see S1 File for details). For example, the performance was poor for every aggressive dog clip, indicating that the clips were comparable to each other and that humans are not very good in assessing aggressive dog interactions. We believe that the advantages of our decision to use clips of real social interactions outweigh the disadvantages of possibly different intensity of the emotional situations. Future studies should control for the apparent strength of the emotional context by using criteria; for example, whether there is physical contact between subjects or not, see [14].

One open question that emerges from the current study is how exactly humans`assessment of a social situation works. A recent study used eye-tracking of human participants while they watched video clips with naturalistic emotional expressions from dogs or humans. The study found that, although emotional states of dogs are more evident in bodily cues, humans focused mainly on the head of the subjects in the video [47]. Another study found that when only the facial expressions of primates and dogs are presented, participants predominantly attend to human eyes and animal mouths [48]. In particular, for judging great apes it would be helpful, if human observers would rather rely on vocalisations, as apes express aggressive intent mainly through vocalisations [49]. Future studies should use questionnaires to find out which features participants use to assess such a social interaction, and more importantly, to find ways to improve the performance of participants [13, 45, 46, 48].

The skill of perceiving and anticipating goal-directed actions does not seem to be uniquely human. Great apes, for example, are able to make online goal-based predictions about the actions of other individuals. Kano and Call (2014) tested great apes with an eye-tracking experiment in which the apes were familiarized with movie clips showing a human hand reaching to grasp one of two objects. In the test event the objects’ locations were swapped and the hand made an incomplete reach between the objects. The apes predictively looked at the familiarized goal object rather than the familiarized location when viewing the hand action in the test event. Thus, similar to human children, great apes make online goal-based predictions about the actions of other individuals [9, 50]. To compare the prediction ability of humans and non-human primates, Myowa-Yamakoshi, Scola and Hirata (2012) examined anticipatory eye movements in human infants, adults, and chimpanzees. The experiment showed that chimpanzees anticipate action goals in the same way as humans do, even though the chimpanzees used different methods to scan goal-directed actions [51]. Furthermore, there is evidence that dogs can perceive human actions as goal directed [52, 53]. Recent experiments have also used eye-tracking on animals to find out how they observe emotional expressions in other species. Dogs, for example, focus mainly on body signals when viewing emotional expressions of both conspecifics and humans [47]. Rhesus macaques explore human and monkey faces differently than chimpanzee and dog faces, indicating that their experience with the observed species might play a role [48]. Thus, animals are also able to assess natural social interactions in their own and other species, although they might sometimes use different mechanisms than humans.

To summarize, we show that humans perform above chance level at assessing natural social interactions in humans, in dogs and in monkeys–even without prior special experiences with the non-human species [13]. How well humans predict outcomes and categorize situations depends on the species and on the context. Humans are not better at assessing social interactions of their own species compared to other species. Surprisingly, participants perform particularly poorly when assessing dogs`aggressive behaviour. Thus, aggressive behaviours in dogs are not well recognized and the outcomes of aggressive interactions between dogs are hard for the average person to predict.

Supporting information

Acknowledgments

First, we want to thank dog owners and parents of the children who recorded clips and gave permission to use them, we also thank Katrin Schumann for her help to get access to the clips. For help with data collection in Jena we are grateful to Rico Stecher. Thank you also to the students of the Project Seminar Social Cognition in Domestic Dogs 2018 at FSU Jena, who did a pilot study for this project. We are grateful to Moritz Doerfler for Coding and Sasha Donnier for commenting on the manuscript. Finally, we want to thank Andrew Zeilstra for proof reading.