Concept-based imitation training in an Asian elephant (Elephas maximus) using the do-as-I-do paradigm

1 Introduction

Due to their remarkable intelligence, complex social structure, and long lifespans, Elephantidae are among the most studied species on the planet (Chusyd, 2021; Plotnik, 2022). They exhibit advanced social learning by observing and interacting within their groups, transcending reliance on personal experiences (Bates et al., 2025). This capability facilitates the transmission of specialised foraging methods and communicative vocalisations, contributing to the formation of distinct cultural behaviours within elephant populations (McComb et al., 2001; Bates et al., 2025). Their sophisticated social structures and strong familial bonds further enhance their learning processes (Hedwig et al., 2021; Bates et al., 2025). Given the importance of social learning and cultural transmission in elephant societies, developing training and enrichment approaches that promote these opportunities may provide important cognitive and welfare benefits for elephants under human care.

Over the past decade, a growing body of research has demonstrated a positive relationship between cognitive challenge and improved welfare outcomes for animals housed in human care (Rose and Lewton, 2024). Cognitive enrichment and training have been associated with increased behavioural diversity, reduced stereotypic behaviours, and improved indicators of psychological well-being across a range of taxa (Alligood et al., 2017; Fernandez, 2022; Ward et al., 2024). Despite these advances, current approaches may still underestimate the breadth and complexity of cognitive challenges that animals routinely engage with in the wild (Doyle et al., 2024).

For elephants living in human care, a persistent welfare challenge remains the mismatch between evolved cognitive demands and the comparatively predictable conditions of human care (Doyle et al., 2024). Contemporary welfare science increasingly recognises that meeting elephants’ physical needs alone is insufficient; (Doyle et al., 2024; Veissier et al., 2024) positive welfare requires opportunities for behavioural choice, agency, and engagement with cognitively meaningful challenges (Ghimire et al., 2024).

Evidence from welfare assessments indicates that elephants in human care commonly express compromised welfare through indicators such as stereotypic behaviour and altered stress physiology, both of which are sensitive to management conditions (Greco et al., 2016). Importantly, epidemiological studies demonstrate that these indicators covary with modifiable aspects of husbandry, including social environment and daily management practices, highlighting the potential for targeted interventions to improve welfare outcomes (Greco et al., 2016). Cognitive enrichment and training are increasingly framed as welfare tools, capable of promoting positive affective states by increasing behavioural diversity, controllability, and engagement (Hall et al., 2021; Rose and Lewton, 2024).

Concept-based training represents an extension of traditional training approaches by shifting from isolated motor behaviours to rule-based cognition that can generalise across contexts (Jaakkola et al., 2010; Zentall, 2014). Training the concept of imitation is particularly relevant for elephants, given the central role of social learning in elephant behaviour and adaptation (Bates et al., 2025). Experimental evidence demonstrates that elephants in human care can use social information to acquire novel tasks, supporting the relevance of socially mediated learning opportunities in human care (Greco et al., 2013). From a welfare perspective, imitation-concept training may therefore function as a form of elevated cognitive care, reinstating socially meaningful learning processes that are otherwise limited in their environment and aligning elephant management with contemporary welfare frameworks that emphasise agency, challenge, and positive experiences (Plotnik and Jacobson, 2022; Rose and Lewton, 2024).

Imitation, a form of social learning, is a complex and multifaceted process and can be categorized into four primary types: intransitive, transitive, contextual, and production (Whiten and Mesoudi, 2008; Subiaul, 2016). Intransitive imitation involves replicating actions devoid of object manipulation, such as gestures or simple movements, whilst transitive imitation encompasses actions performed with objects (Subiaul, 2016). Contextual imitation demands consideration of the action’s situational context and purpose (Over and Carpenter, 2012; Tennie et al., 2012). Lastly, production imitation refers to the acquisition of entirely new actions or skills through observation and repetition, allowing an individual to learn behaviours that were not previously part of its repertoire (Heyes, 2012; Whiten, 2017).

A pivotal method for studying imitation in animals is the “do-as-I-do” task, initially introduced by Hayes and Hayes (1952). This procedure involves an animal observing a specific behaviour performed by a human or trained conspecific, followed by a request for the animal to reproduce the demonstrated behaviour. The request is typically communicated through a pre-established cue or symbolic gesture. Successful replication of the demonstrated action by the animal, as opposed to an alternative behaviour, serves as evidence of intransitive imitation (Haldar et al., 2024).

Evidence of imitation using the Do-As-I-Do paradigm has been documented across a diverse range of species, including domestic dogs (Canis lupus familiaris), non-human primates, cetaceans, and birds (Voelkl and Huber, 2000; Heyes and Saggerson, 2002; Whiten et al., 2004; Byrne and Tanner, 2006; Kuczaj and Yeater, 2006; Topál et al., 2006; Jaakkola et al., 2010). These studies demonstrate that animals can acquire behaviours by observing and replicating the actions of a demonstrator, sometimes using multiple sensory modalities to interpret demonstrated behaviours. Given the role of social learning in elephant behaviour and adaptation (Bates et al., 2025), it is therefore reasonable to expect that elephants may be capable of imitation-based learning, and that providing opportunities for such learning in human care may offer important cognitive and welfare benefits. However, formal imitation paradigms have received little scientific investigation in elephants. Consequently, the potential for imitation-based concept training to function as a form of cognitive enrichment and welfare tool in elephants remains largely unexplored.

This study therefore aimed to investigate whether an Asian elephant could learn and generalise the concept of imitation using the Do-As-I-Do paradigm, in which an individual is required to replicate the behaviour of a demonstrator following a generalised copy cue. In addition, the study explored the potential role of imitation-based concept training as a form of cognitive challenge that may expand learning opportunities, facilitate socially mediated learning, and contribute to cognitive enrichment and welfare in elephants under human care.

2 Materials and methods

2.1 Subjects and testing environment

The subjects of this study were two 10-year-old male Asian elephants, both of whom were born at Dublin Zoo, Ireland into a herd of seven. The two elephants, who have resided together since birth, were transferred to Sydney Zoo in October 2020 from Dublin Zoo and have remained the only elephants at the facility since February 2022.

At Sydney Zoo, the elephants occupied a shared area of 6,000 square meters, which included two on-display habitats, two raceways providing access to three outdoor stalls, and three indoor stalls. The elephants were housed together 24 hours per day, only being separated for small periods during some specific training. The elephants were trained using positive reinforcement methods with primary rewards being sweet potato, apple, corn, bamboo, teff and oaten hay and they were fed according to their standard daily routine. Tactile from their trainers was used as a secondary reinforcer a times.

Prior to arrival at Sydney Zoo, the elephants had limited prior conditioning, none of which were concept-based behaviours. Whilst at Sydney Zoo, both elephants had participated in several concept-based behaviour studies that focused on olfactory and memory recall behaviours, none of which have been published. The elephants learned the “imitate” behaviour concurrently and in this study the same animal served as the imitator (the subject required to reproduce the demonstrated behaviour) and demonstrator (the individual performing the target behaviour).

2.2 Behaviour set-up and materials

The “Lower Raceway” was selected as the area to set up the materials needed for the study due to its even, flat surface, which facilitated the erection of a visual barrier (Figure 1). The Lower Raceway measured 4 meters in width and 20 meters in length. A large sheet of plywood, measuring 2.4 meters in height, 1.2 meters in width, and 10 millimetres in thickness, was utilized to create a visual barrier. This barrier was secured to the fence bollard with nylon rope in two sections. All training sessions were conducted within this designated space, while the elephants retained free access to their habitats and alternative food sources, including grass and hay.

Two trainers positioned themselves on either side of the visual barrier, approximately 2 meters apart, each working with one elephant placed in front of them (Figure 2). This arrangement ensured that each elephant could only see the trainer directly in front of them and the adjacent elephant, while the visual barrier effectively blocked the view of the opposing trainer. Trainers and observers would routinely check visual access between each repetition to ensure the imitator remained blind to any visual cues.

2.3 General procedure

The experimental procedure and training methodology used in this study were closely modelled on previously published Do-As-I-Do imitation paradigms (Jaakkola et al., 2010; Zamorano-Abramson, 2019). These studies employed comparable shaping strategies, cueing procedures, and criteria for determining successful imitation, in which a demonstrator performs a target behaviour and the subject is subsequently given a copy cue to reproduce that behaviour, with reinforcement contingent on accurate replication.

A total of 56 training sessions were conducted over a six-month period. Sessions occurred one to two times per day between 11:00 and 15:00 h, across three consecutive days each week. This schedule was selected to maintain trainer consistency while allowing sufficient recovery and consolidation time between training days. Each session lasted approximately 10–15 minutes and involved the same pair of elephants with their respective primary trainers. Two additional trainers were present during all sessions and served as independent observers. Across all 56 imitation training sessions, the elephants participated voluntarily for the full duration of each session, and behavioural measures recorded during sessions (including response latency, task engagement, and absence of distraction) were consistent with sustained participation.

All sessions were video recorded using an iPhone 11 to allow for post-session review and verification of behavioural responses. Procedural controls were implemented throughout the study to prevent inadvertent cueing. These measures included regular checks between repetitions by trainers and observers to confirm that the imitating elephant had no visual access to the trainer working with the demonstrator elephant, supplemented by review of recorded footage. Initial verification of visual occlusion was conducted by presenting high-value food items in the cueing location in front of the demonstrator elephant; the absence of any response from the imitator elephant indicated that the visual barrier was effective.

The study was organised into three distinct phases (Table 1), with progression between phases contingent upon performance criteria. Advancement occurred once the imitator’s responses stabilised and only a small number of non-systematic, random errors were observed, indicating reliable comprehension of the task demands at each stage. All behaviours selected for this study were intransitive and could be performed while the animal remained stationary, which was important for maintaining clarity of the demonstration and minimising potential confounding effects of positional or spatial cues. Phase 1 behaviours consisted of well-established behaviours within the animals’ repertoire. Phase 2 introduced additional behaviours that also did not require changes in location but included behaviours with a shorter reinforcement history. Phase 3 behaviours consisted of combinations of behaviours from Phases 1 and 2, which had not previously been performed together, allowing assessment of imitation beyond previously trained responses. In this phase, the demonstrator was instructed to perform both tasks simultaneously before the imitator received the copy signal.

Table 1. List of model behaviours and their description.

All behaviours chosen were intransitive and did non-object orientated actions.

2.4 Phase 1 (training phase)

The imitate signal was trained using a combination of verbal and visual cues that were already familiar to both elephants. To establish that the imitator was attending to the demonstrator rather than responding to the trainer alone, the trainer working with the demonstrator initially delivered a verbal cue concurrently with the copy signal presented to the imitator. This pairing was designed to prompt the imitator to seek behavioural information from the demonstrator elephant rather than relying solely on the trainer’s cues. Over successive training sessions, the demonstrator’s verbal cue was progressively faded and replaced with a purely visual signal by their trainer. This signal remained the same throughout the study. The only stimulus that changed for the imitating elephant was the actions being performed by the demonstrator elephant.

Throughout this shaping procedure, the trainer of the imitator withheld the copy signal until the demonstrator was engaged in one of three trained behaviours. By the end of this phase, the imitator reliably waited for the demonstrator’s response before executing its own behaviour, indicating that the copy signal had become functionally linked to the demonstrator’s actions rather than to direct trainer prompting.

2.5 Phase 2 (testing familiar conditioned behaviours)

The goal of this phase was to determine if the imitator elephant could generalise the imitate signal with other well-known behaviours and successfully imitate them. The set-up for the sessions in Phase 2 remained the same as Phase 1. The 7 behaviours (Table 2) used in this phase were all known behaviours to the imitator, however the four additional behaviours (EF, FL, HD, HN) had never been associated with the copy behaviour. The sessions in Phase 2 were made of behaviours from Phase 1 with the new behaviours randomly interjected. To begin, only one new behaviour was added in each training session before multiple new behaviours could be introduced during sessions.

Table 2. Performance accuracy and statistical outcomes for imitation trials across behaviours and phases.

Importantly, the imitator was required to reproduce these newly associated behaviours using only the established copy signal. Unlike the initial shaping phase, no additional verbal or gestural cues would be provided to indicate which behaviour should be performed. Successful imitation of these behaviours therefore required the imitator to rely on the demonstrator’s actions in conjunction with the imitate cue.

2.6 Phase 3 (testing novel combination behaviours)

This phase replicated the trial setup from Phase 2. In this third phase, four combinations of behaviours were presented to the demonstrator animal. These behaviours had never been requested simultaneously from either elephant, resulting in a novel combination behaviour. The demonstrator elephant was instructed to perform both behaviours at the same time, after which the imitator elephant received the imitate signal and was asked to execute the behaviour.

3 Data coding and analysis

3.1 Data coding

The data were coded manually from video recordings of each trial. Each trial was recorded in terms of the target behaviour performed by the demonstrator elephant and the subsequent response of the subject elephant. Correct responses were scored 1, incorrect 0. To ensure consistency in coding, two trained observers independently scored each trial, with discrepancies resolved through discussion and agreement. For behaviours tested in smaller sample sizes (e.g., EF, FL, HN), trials were classified as either correct (subject imitated the behaviour) or incorrect (subject did not imitate the behaviour or performed an alternate behaviour). Data for behaviours performed in combination (e.g., Combination 1, Combination 2, Combination 3, Combination 4) were similarly coded, with correct responses representing instances where the subject successfully imitated both components of the compound behaviour.

3.2 Analysis

All statistical analyses were conducted using “R” (R Core Team, 2024), with binomial tests performed using the base function to determine whether the observed number of correct responses deviated significantly from chance expectation. For each phase of testing, the chance level of correct responses was calculated based on the number of behaviours available for selection in each trial.

• Phase 1: Three behaviours were available for selection, resulting in a chance level of 0.33.
• Phase 2: Seven behaviours were available for selection, resulting in a chance level of 0.14.
• Phase 3: As in Phase 2, seven behaviours were available for selection, resulting in a chance level of 0.14.

Each behaviour was tested for significance against the corresponding chance level using the binomial test for one-tailed probabilities (greater), assuming that performance significantly exceeding chance would indicate the presence of an imitative response. The p-values were computed for each behaviour, with significant results considered for p-values less than 0.05.

For Phase 3, where combination behaviours were included, separate analyses were conducted to assess performance on individual components of the compound behaviour, and results were interpreted accordingly.

4 Results

4.1 Phase 1

In Phase 1, the subject was tested on three behaviours, Head Shake, Ear Touch, and Trunk Raise (Table 2). Accuracy rates for these behaviours were between 60.9% and 80.0%. Binomial tests against a chance probability of 0.33 indicated that all behaviours were performed significantly above chance.

4.2 Phase 2

In Phase 2, the task was expanded to seven possible behaviours (Table 2). Accuracy remained high, between 80.6% and 94.4% for the three behaviours that carried over from phase one. The four additional behaviours introduced in this phase (Ear Flap, Foot Lift, Head Nod and Head Down) had accuracy rates between 54.4% and 100%. All seven behaviours in Phase 2 were performed above chance.

4.3 Phase 3

Performance remained consistently high on the three core behaviours with accuracy rates between 87.1% and 90.3%. The four additional actions from Phase 2 showed similarly strong imitation accuracy with results between 80.0% and 100% during this phase.

Four novel combination actions were assessed to test conceptual imitation beyond single action copying (Table 2). The elephant had a success rate of 75-100% across all combinations of behaviours.

5 Discussion

The aim of this study was to investigate whether an Asian elephant could learn and generalise the concept of imitation using the Do-As-I-Do paradigm and to explore whether imitation-based concept training could function as a form of cognitive enrichment. The results demonstrated that the subject elephant successfully imitated familiar behaviours, generalised the imitation cue to additional behaviours, and replicated combination behaviours, with performance significantly above chance across all phases. Similar findings have been reported in other species trained using Do-As-I-Do paradigms. For example, domestic dogs have been shown to reproduce both single actions and action sequences following a generalised imitation cue (Topál et al., 2006), while bottlenose dolphins have demonstrated the ability to copy both familiar and novel behaviours on command (Kuczaj and Yeater, 2006; Jaakkola et al., 2010). Studies in non-human primates have similarly shown the capacity to imitate demonstrated actions and learn socially through observation (Voelkl and Huber, 2000; Whiten et al., 2004; Byrne and Tanner, 2006), and evidence of imitation has also been documented in birds under controlled experimental conditions (Heyes and Saggerson, 2002). The ability of the subject elephant to imitate both familiar behaviours and novel combinations suggests that performance was unlikely to be explained solely by stimulus–response learning and is more consistent with concept-based learning and imitation described in other species (Zentall, 2006; Heyes, 2012).

The findings of this study are particularly relevant when considered in the context of elephant social learning. The ability of the subject elephant to reproduce behaviours demonstrated by a conspecific suggests that imitation may function as a mechanism for socially mediated learning. Elephant societies rely heavily on knowledge transmission between individuals, and disruption to these pathways can negatively affect behaviour, social cohesion, and survival outcomes (Bates et al., 2025). Facilitating imitation-based learning in elephants under human care may therefore provide a means of supporting socially relevant learning processes.

To minimise the potential bias associated with reinforcement, in the form of trainer interaction or food, additional opportunities to obtain these were always provided at other times during the daily routine. This was intended to reduce the likelihood that participation in imitation sessions was driven solely by access to reinforcement or trainer interaction. Anecdotally, the imitation sessions consistently showed high levels of participation and engagement, suggesting that the learning task itself may have been reinforcing. However, it is important to note that this interpretation is limited by the lack of direct experimental controls isolating reinforcement variables, and participation was not systematically compared under conditions where alternative reinforcement opportunities were simultaneously available. This increased engagement may be associated with the novelty and cognitive challenge of the task, which is consistent with previous research indicating that cognitively demanding and novel enrichment activities can enhance engagement and behavioural diversity in animals under human care. For example, Clark (2011) demonstrated that increased cognitive challenge in captive primates led to higher levels of engagement and reduced abnormal behaviours, while Melfi (2013) highlighted that training itself can function as enrichment when it provides animals with opportunities to engage in problem-solving and learning. Similarly, Fernandez (2022) emphasised that structured training and cognitive tasks can increase behavioural diversity and promote positive welfare states, and more recent work has shown that enrichment strategies incorporating cognitive challenge are associated with improved welfare outcomes across species housed in human care (Rose and Lewton, 2024; Mellor et al., 2020). Nevertheless, no direct measures of welfare outcomes were collected in this study, and therefore the potential relationship between imitation-based training and cognitive welfare remains speculative. Future studies incorporating controlled comparisons and longitudinal welfare assessments would be required to determine whether imitation-based training has measurable impacts on welfare.

It is important to consider the limitations of this study, particularly the small sample size, with a single subject acting as the imitator. As a result, the findings cannot be generalised across all elephants in different facilities, and individual variation in cognition, temperament, and learning ability may have influenced performance. The consistently high accuracy observed across all phases may therefore reflect individual-specific capabilities rather than species-level capacity, limiting the robustness of broader conclusions. Prior experience of the subject elephant must also be taken into account when interpreting the results. The imitator elephant had previously participated in multiple concept-based behaviour studies and was therefore familiar with shaping procedures, experimental setups, and rule-based learning tasks. Previous research across species has demonstrated that prior training experience and familiarity with experimental procedures can influence performance in cognitive tasks (Griffin et al., 2015). The subject’s previous experience with concept-based learning may therefore have facilitated more rapid acquisition and generalisation of the imitation concept, potentially inflating performance relative to individuals without such experience. While the findings provide valid evidence that imitation learning can occur in an elephant under these conditions, the robustness of the results is constrained by the potential influence of individual learning history and prior training exposure. Future research involving individuals with different training histories would be valuable to determine the extent to which prior concept learning experience influences imitation learning performance in elephants. This could include comparative studies between elephants with and without prior concept-based training, as well as longitudinal studies examining how imitation learning develops over time. Expanding this work across different social groupings and management conditions would also provide insight into how imitation-based training can be applied as a practical tool to enhance cognitive enrichment and welfare outcomes in elephants under human care.

Despite these limitations, this study provides the first documented application of imitation-based concept training using the Do-As-I-Do paradigm in elephants and demonstrates the potential for imitation training to increase the diversity of learning opportunities available to elephants under human care. By providing opportunities for socially mediated learning and cognitively challenging tasks, imitation training may represent a valuable tool for cognitive enrichment and welfare in elephants managed in human care.

6 Conclusion

Collectively, the results of this study provide evidence that the imitating elephant consistently engaged in imitative responses that significantly exceeded levels predicted by chance, even as the complexity of the task increased over time. Throughout the various phases of the experiment, the imitating demonstrated the ability to replicate a diverse range of behaviours, from familiar actions to novel, previously unobserved combinations. This consistency in performance indicates that responses were unlikely to be driven solely by rote memorisation or fixed associations between individual behaviours and specific cues or rewards and is consistent with the interpretation that the subject applied the imitation cue beyond individual trained actions.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Used to help interpret the P value and assist with some sentences.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Paul Rose, University of Exeter, United Kingdom

Reviewed by: Ashley Edes, Saint Louis Zoo, United States

Ian Hickey, Chester Zoo, United Kingdom