Imitation of cultural practices is ubiquitous in humans and often involves faithful copying of intransitive (i.e., non-object directed) gestures and societal norms which play a crucial role in human cumulative cultural evolution. Apart from learning these directly from a tutor, humans often learn passively as third-party observers from the interactions of two or more individuals. Whether third-party imitation has evolved outside humans remains unknown. In the current study, we investigated whether undomesticated blue-throated macaws (Ara glaucogularis) could imitate in a third-party setting. A naïve test group (N = 6) passively observed a conspecific demonstrator performing rare intransitive actions in response to specific human gestural commands. Directly afterwards, the observer received the same gestural commands and performance-contingent rewards. An equally naïve control group (N = 5) was tested correspondingly, in the absence of third-party demonstrations. The test group learned more target actions (mean = 4.16 versus mean = 2.2) in response to the specific commands, significantly faster and performed them more accurately than the control group. The test group also spontaneously imitated some of the actions even before they received any gestural commands or rewards. Our findings show that third-party imitation, even for intransitive actions, exists outside humans, allowing for rapid adaption to group specific behaviours and possibly cultural conventions in parrots.

Human cumulative culture is largely underpinned by the human ability to imitate along with other social learning processes (e.g., emulation) and pedagogy, that enable faithful transmission of cultural knowledge through generations. Heyes defined imitation as when an observer "learns and reproduces 'something' about the topography of the model's body movements " whereas in other kinds of social learning, an observer learns about the goal of action (goal emulation) or location and characteristics of objects (local and stimulus enhancement). Human children develop the tendency to imitate an adult who is directly demonstrating an action to the child ('second-person imitation'-A learns from B), from the first year of their life. They also develop third-party imitation (A learns from B and C interacting) ability only towards the end of their second year perhaps linked to the development of an understanding for self-other equivalence. Third-party imitation entails the replication of an act in the appropriate interactive context after passively observing the act in the interaction of two individuals. Thus, in third-party imitation, the observer can also acquire knowledge (e.g. normative knowledge) about the interaction itself, i.e. how to behave in a certain interactive situation. For example, in some hunter-gatherer societies, the socially appropriate conduct expected of a particular gender in specific situations is acquired through passive observational learning. Thus, third-party imitation is an important skill that has been reported only in humans and the evolution of this ability remains largely unexplored. A comparative approach investigating third-party imitation in distantly and non-distantly related non-human species promises to shed light on whether this ability evolved in other animal taxa and its mechanistic underpinnings.

Imitation of a single demonstrator ('second-person' or 'second-party' imitation') has been reported in a few non-human species, e.g., in two-target task studies in which observers show significantly more target observed acts than non-observed ones. Apart from whether imitation happens in second-party or third-party settings, it is further distinguished between whether the copied body movements or actions of the model are already present within the behavioural repertoire of the species, which is referred to as contextual imitation or whether the imitated actions are completely novel, which is called production learning or (-by other scientists-) "true imitation. Thus, third-party imitation can be contextual (actions are familiar but not produced in an interactive context before) or focus on production learning (novel or improbable actions), with actions demonstrated through the interaction of two individuals in the presence of a passive observer.

Considering that imitation studies involving object-directed-actions (e.g., two-target task studies) often cannot exclude alternative social learning mechanisms such as goal emulation, or stimulus enhancement, evidence for true or contextual imitation can be obtained most convincingly in studies that focus on intransitive actions, i.e. actions that do not involve objects or goals. Here, only the topography of body movements can be the target of social learning. Animal studies of imitation of intransitive actions remain rare and restricted to second-party settings. Maybe surprisingly, given that gestural communication is ubiquitous in non-human primate species, gestures -- often in the form of intransitive actions -- are typically either biologically predisposed responses or ritualized interactions but do not seem to be acquired through imitation (neither through second-party nor third-party imitation). Concerning imitation of intransitive actions, Moore reported 'second-party' imitation of intransitive gestures in an African grey parrot (Psittacus erithacus). The experimenters interacted directly with the parrot several times a day, performing certain stereotyped body movements. Nevertheless, over the course of five years, this parrot produced 10 actions using head, beak, wings and feet and was claimed to have imitated intransitive actions of the human demonstrators. The other few studies that have reported imitation of intransitive actions in animals have used the "Do-as-I-do" (DAID) or "Do-as-the-other-does" (DATOD) paradigm, in which the animals are actively trained by the experimenter to pay attention to and copy either a human or a conspecific demonstrator. All these studies were conducted in a second-party setting (a single model demonstrated the actions) and thus did not address third-party imitation.

Third-party imitation, i.e., imitation of a demonstrator interacting with another individual, has been systematically tested in human children. Here the subjects passively observed an interaction between two individuals and could reproduce the relevant action when tested in the same interactive context. However, in non-human animals, only one single study has tested third-party imitation, namely a study on dogs (Canis familiaris), a domesticated species. Here, a trained dog performed one of two intransitive actions in response to verbal commands of an experimenter, while a task-naïve, observer dog (i.e., the "third-party") watched the interaction. Subsequently, it was tested whether the observer dog would show the actions demonstrated by the trained conspecific using the same verbal commands. The dogs failed to show evidence of third-party imitation in this study. Yet, even if some dogs had been successful, the obtained results could have been confounded by domestication effects in dogs. Tennie and colleagues had tested unenculturated chimpanzees for contextual imitation in a comparable third-party setting, where a trained chimpanzee performed an arbitrary action to obtain food from an experimenter (social context) while the naïve observers watched the interaction. Except for one chimpanzee, the rest failed to imitate the arbitrary action in this study, consistent with previous findings in the species, where naïve chimpanzees were unable to reproduce novel begging gestures (production learning) demonstrated by a model to obtain food from an experimenter. In a different set of studies, Pepperberg and colleagues, used a method called "model/rival technique" in which two trainers interacted with each other while a single African grey parrot observed their interaction. However, this technique was employed as a general tool for training the parrot to learn object labels by vocal imitation rather than for specifically investigating its ability to learn via third-party imitation. Moreover, the grey parrot received direct interactive training concurrently and hence the effect of learning as a third party cannot be teased apart retrospectively from these experiments. Thus, third-party imitation in non-human and non-domesticated animals has not been studied specifically, even though it is likely to exist, and a comparative approach promises insights into the evolution of third-party imitation.

We hypothesize, that third-party imitation should most likely prevail in social animals in which social behavioural repertoire (including those mediated through vocal and gestural communication) varies within and across populations. For example, third-party imitation may be observed in animal societies with fission-fusion dynamics which brings about frequent changes in group composition or in matrilocal/patrilocal societies, where one sex of the species disperses to join unrelated groups. Here, cultural learning about group-typical social behaviours including idiosyncratic behaviours might aid group cohesion and some might be required for integration into a group.

In this study, we aim to test third-party imitation in parrots, focusing on contextual imitation of intransitive actions, i.e., imitation of actions that are present in the repertoire of the species but occur infrequently and demonstrated through the interaction between a human experimenter and a conspecific. As indicated by the aforementioned studies parrots are particularly interesting candidates for studying third-party imitation. Parrots characteristically live in fission-fusion societies in which learning group-typical interactive social behaviours, such as gestures or coordinated (synchronised) movements by social observation, may be important for group cohesion. Parrots are among the very few animal taxa possessing vocal imitation ability tool-making skills foraging cultures and vocal dialects. They have been also known to show gestural communication. Perhaps most importantly, parrots have previously been successful in second-party imitation of transitive and intransitive actions and they appear to learn from a human third party- setting as a previous study suggests but it remains untested whether they are capable of learning via third-party imitation.

To convincingly investigate a non-human species' capacity for third-party imitation, it is essential to ensure the following conditions: (1) demonstrations must not be overtly directed at the subjects using human-specific cues (e.g., ostensive signals like calling or eye gaze), to ensure passive observation; (2) subjects should not be human-enculturated (as with domestic dogs), in order to maintain ecological validity; and most importantly (3) test subjects (observers) must have no prior training in either the target actions or imitation/copying actions itself (as seen in DAID paradigms), to avoid confounding effects of automatic imitation (sensorimotor experience gained through trained actions) or second-party imitation (if DAID trained, animals may have a predisposition to copy actions and ignore the interactive context).

We complied with all points raised above and operationally defined third-party imitation as copying rare target actions, matching the model's action topography upon observing them in a specific interactive social context, i.e., after observing the behavioural response of a conspecific demonstrator to a human gesture (here). To ensure the actions used to test for contextual imitation (rather than 'true imitation' or 'production learning') were sufficiently rare, we established a baseline by recording the natural frequency/occurrence of the target actions and through post-hoc analysis. In the experiment, a test group of (N = 6) hand-raised blue-throated macaws that were habituated to receiving food from human hands, but were not trained to perform specific behaviours on command and were never intentionally trained to imitate, observed a trained conspecific demonstrator performing five intransitive target actions (lift leg, fluff, spin, vocal and flap wings), one at a time, in a session. The order of actions demonstrated was randomised across subjects. Each target action was performed in response to specific hand commands from the experimenter (see Fig. 1A). Subjects in the test group observed these demonstrations passively from an adjacent room separated by a transparent plexiglass (see Fig. 1B). Following these demonstrations (i.e., after 3-5 s), a second experimenter (who had been passively present in the adjacent room together with the subject) gave the same gestural commands to the subject (see Supplementary Videos S1-5), testing whether subjects would express the demonstrated target actions within a timeframe of 12 (± 2) seconds of response time, standardized arbitrarily to match the assumed attention span of a macaw (based on observation of experimenters). A control group (N = 5) was exposed to the same gestural commands, yet without ever observing any target action linked to the commands (see Supplementary Video S6-7). We rewarded the subjects in both test and control groups for performing the target action with the respective command. Overall, this study aimed to investigate whether macaws learned to reproduce intransitive target actions in the correct interactive context (response to correct gestural command) more accurately and faster, after observing a model perform them in a third-party setting than in the absence of demonstrations.