Psychological insights into the research and practice of embodied conversational agents, chatbots and social assistive robots: a systematic meta-review

ABSTRACT

This study presents a systematic literature search and narrative meta-review of the current state of research on conversational agents (CAs), including embodied CAs, chatbots, and social assistive robots (SARs). The investigation identifies 1,830 academic articles, of which 315 articles satisfied the inclusion criteria for the review. Systematic reviews across various fields are reported, including mental disorders, neurodevelopmental disorders, dementia/cognitive impairment, other medical conditions, elderly support, health promotion, mental health, education, industrial applications, agent characteristics, and robot characteristics. The study highlights challenges in current CA research, such as the scarcity of high-quality comparative studies and the acceptance of CAs by users and caregivers, particularly in elderly support. The article also categorises ethical discussions into nine elements: privacy, safety, innovation, user acceptance, psychological attachment, care philosophy, evaluation, social systems compatibility, and rule development. It also offers insights into the development of future guidelines. The role of CAs in fostering human relationships through their conversational function is emphasised to provide guidance for subsequent CA research and social implementation. As advancements in CA technology and research continue to progress, there is an increasing demand for sophisticated psychological investigations addressing relationships, emotions, and the self.

KEYWORDS:

• Large language models
• artificial intelligence
• interactive robots
• ethics
• education
• mental health

1. Introduction

Research on conversational agents (CA) has made a remarkable progress. Chatbots are investigated and utilised in various areas, including healthcare, education, and industries, and increasing since 2016 (Adamopoulou and Moussiades 2020). The research and utilisation of social assistive robots (SARs) are also spreading, especially in nursing and dementia patient care.

The use of computers and the exchange of information and its effects between humans and computers, which are the basis of CA research, have long been studied in psychology in the context of computer-based test interpretations (Harris 1987), computer-mediated communication (CMC) (Wright et al. 2003), human–computer interaction (HCI) (Ghani and Deshpande 1994), and so on. As the CA technology advances, and intelligent computer agents become more human-like in their behaviour, the importance of psychology in CA research is growing. A discrepancy in the level of trust displayed toward robots by 3-year-olds and 7-year-olds has been recently attributed to differences in the development of attachment formation and theory of mind (Di Dio et al. 2020). The utility of artificial intelligence (AI) and computer agents in forecasting, identifying, and remedying mental health issues has been investigated (D’Alfonso 2020). The similarity of results obtained from questionnaire evaluations conducted by chatbots and human evaluators (Caballer et al. 2020) and the inferiority of chatbot communication in relational aspects compared to human communication have been documented (Lew and Walther 2023).

To date, a substantial body of research in psychology pertaining to CAs can be considered an extrapolation of psychological studies on human beings. However, the rapid progression of CA technology has ushered in a new epoch in related areas of research. It is imperative to assess the veracity of established psychological concepts, phenomena, and theories in the context of interactions between humans and computer agents as well as between computer agents alone. Consequently, the scope of psychology is broadening, and psychological research on CAs portends to furnish valuable insights for the construction of socially-integrated agents, as well as the development of various psychological theories.

The research area on CA is interdisciplinary and diverse, making it difficult to get a complete picture. Although several general reviews have been reported (Lambert et al. 2020; Romero, Casadevante, and Montoro 2020; Xu et al. 2021), to the best of our knowledge, no meta-review has yet been conducted. Therefore, this study aims to conduct a systematic literature search and a narrative review of the systematic reviews on CA. By conducting a meta-review with respect to high-quality systematic reviews, we believe that we can capture the overall picture of CA research, identify its current state, future challenges, and aspects in which psychologists should be particularly involved.

1.1. Key perspectives

With the development of the information and communication technology (ICT), research on computer-mediated communication (CMC) between intelligent agents, including humans, has developed. In particular, the interface design that enables seamless or intuitive communication between humans and machines and the associated human and machine capabilities, cognitive abilities, and ways of interacting with the world have been studied in the context of HCI (Ess 2008). The area of HCI that deals with the interaction of intelligent agents that behave like humans with autonomy, adaptability, and persona, unlike conventional information systems, is called the human–agent interaction (HAI). Intelligent agent systems are usually virtual entities that exist within the screens of smartphones, personal computers (PCs), and other devices. Those with humanoid or non-humanoid externals are called embodied agents. Agents that simulate or emulate human conversation are called CAs.

While agents focus on the agency of individuals or systems, those with machine-like entities are called robots. The application areas of the human–robot interaction (HRI) include the human monitoring of work robots, human manipulation of robotic tasks in hazardous environments inaccessible to humans or automated machines with human passengers, and robots for social interaction with humans (Sheridan 2016). SARs are robots that assist users in social interaction and primarily relevant in the third context.

A chatbot is a feature that allows communication between agents via chat. It may also refer to the agent itself with such a feature. Chatbots usually focus on voice or text communication and may also present images. Voice chatbots, often referred to as voice assistants, can also be known as virtual assistants (e.g. Google Assistant, Siri, Cortana, and Watson). Early chatbots, such as ELIZA, PARRY, and ALICE, relied on simple parsing, keyword searching, and pattern matching techniques to process input from user speech and generate responses using manually created rules (Lokman and Ameedeen 2018). These are called retrieval- or rule-based chatbots, which can be used in domain-specific applications due to their limited response flexibility. In recent years, chatbots that use AI, machine learning (ML), and natural language processing (NLP) techniques have been developed to understand the intent of input and generate responses according to algorithms that convert input into responses learned in advance (Chow, Sanders, and Li 2023a). These chatbots are called generative-based or AI chatbots that are expected to be more flexible in their responses and more human-like. By contrast, they have the disadvantage of possibly generating undesirable responses. Accordingly, a hybrid type combining retrieval and generative was proposed (Lokman and Ameedeen 2018). Furthermore, chatbots employing ML and NLP are not exclusively generative. Some chatbots utilise ML and NLP solely for intent recognition and entity extraction, relying on predefined responses (Agarwal and Wadhwa 2020).

Retrieval-based chatbots are often used as closed-domain chatbots to address issues and provide information. Meanwhile, generative-based and hybrid chatbots are effective as open-domain, conversational chatbots. In recent years, NLP and ML technologies have been used to address the challenges encountered by chatbots. The development of the NLP and ML techniques, sentiment analysis, and emotion estimation has seen growing expectations for chatbots that can achieve natural conversations in the open domain and aim to build relationships with users. However, many challenges remain, including semantics, consistency, and interactiveness (Huang, Zhu, and Gao 2020).

ECAs, chatbots, and SARs share similarities. ECAs are frequently equipped with a chat function, yet they are distinguished by their embodied and gestural features. Unembodied agents are often called chatbots, but chatbots are not necessarily agents. For example, a retrieval-based Q&A chatbot with low autonomy and adaptivity and no clear persona is not called an agent. Among SARs, telepresence robots are neither agents nor chatbots. Pet-type SARs may or may not be called agents depending on the degree of anthropomorphism. This study examines the social dimensions of ECA, chatbots, and SAR, collectively referred to as CA.

The integration of virtual reality (VR) and augmented reality (AR) with CA is a potential area for the future. With the development of HAI research, knowledge on multimodal human–agent interaction, such as face and gesture, speech, and conversational content recognition, has been accumulated (Chollet et al. 2009). Multimodal HAI is inferred to become more familiar in the future, not only through interaction with chatbots and SARs in real-life situations, but also through VR and AR.

As the technology spreads throughout the society, the ethical aspects of CA are also being actively discussed. In the old days, the Three Laws of Robotics were proposed in a science fiction novel (Asimov 1950). These laws state that an autonomous robot should follow human instructions and protect its own self-preservation and human safety (Asimov 1950). In the real world, the Japanese government proposed seven Social Principles of Human-Centric AI: human-centeredness, education and literacy; privacy; safety; fair competition; fairness, accountability and transparency; and innovation (Japanese Council for Social Principles of Human-centric AI 2019). In addition, a list of values that public services provide through chatbots (Makasi et al. 2020) and 12 considerations for CA based on a report by the US National Academy of Medicine were presented (McGreevey, Hanson, and Koppel 2020).

CAs are already in practical use in the fields of education, health, entertainment, information retrieval, business Q&A, e-commerce, and assistants and conversation partners in daily life (Lambert et al. 2020; Shawar and Atwell 2007). However, most of them employ retrieval-based chatbots and remain in closed-domain, task-oriented conversations. In this current situation, it is particularly meaningful to systematically organise in which domains these CA technologies have been developed and which CAs involve open-domain conversations and relationship-building techniques at a level that can withstand practical application. Making comparisons across domains clarifies the domain and the kind of research that should be conducted in the future. Furthermore, issues that should be addressed across disciplines, such as ethical guidelines, can also be organised. In particular, the development of CA research for the support of the elderly and dementia care is a topic that should be of interest to developed countries, which are suffering from an aging population and a shortage of nursing care personnel. Telehealth using ICT devices, such as eHealth, digital health, and mobile health, as well as medical automation and efficiency using ML and AI have become widespread in recent years; CA has played a vital role in this trend (Chow 2021; Hauser-Ulrich et al. 2020; Kretzschmar et al. 2019). However, CA is still in its developmental stage. Moreover, there is an insufficient differentiation between the use of CA for psychiatric medicine and the promotion of the mental and physical health of non-clinical populations. From this perspective, this study aims to contribute to the organisation of directions for future research, development, and practical application, especially in the field of psychology.

1.2. Research questions

Based on the abovementioned perspectives, this study elucidates the following research questions (RQs) through a narrative meta-review of systematic reviews:

RQ1. In what domains has research on ECA, chatbots and SAR progressed?

RQ2. What is currently known from research on ECA, chatbots and SAR in each domain?

RQ3. What are the future challenges for the research and social implementation of ECA, chatbots and SAR in each domain?

2. Materials and methods

2.1. Literature search

This review was conducted in accordance with the PRISMA 2020 statement (Page et al. 2021), where relevant. The first author was responsible for the overall procedure. The coauthors approved the literature inclusion/exclusion and classification. Five major databases of scholarly articles in psychology and information technology were searched: MEDLINE (PubMed), ScienceDirect (Elsevier), EBSCO host (Academic Search Complete, Business Source Complete, Library, Information Science & Technology Abstracts, MasterFILE Complete etc.), Web of Science (Clarivate) and ACM Guide to Computing Literature (ACM Digital Library). All databases were searched on July 29, 2023 to extract literature containing the following keywords in the title, abstract, or a set of keywords: (review OR synthesis) AND (‘chat-bot*’ OR chatbot* OR ‘agent interaction*’ OR ‘conversational agent*’ OR ‘conversational robot*’ OR ‘conversational avatar*’ OR ‘conversational character*’ OR ‘conversational assistant*’ OR ‘conversation* agent*’ OR ‘conversation* robot*’ OR ‘conversation* avatar*’ OR ‘conversation* character*’ OR ‘conversation* assistant*’ OR ‘virtual agent*’ OR ‘virtual robot*’ OR ‘virtual avatar*’ OR ‘virtual character*’ OR ‘virtual assistant*’ OR ‘relational agent*’ OR ‘relational robot*’ OR ‘relational avatar*’ OR ‘relational character*’ OR ‘relational assistant*’ OR ‘interactive agent*’ OR ‘interactive robot*’ OR ‘interactive avatar*’ OR ‘interactive character*’ OR ‘interactive assistant*’ OR ‘interaction* agent*’ OR ‘interaction* robot*’ OR ‘interaction* avatar*’ OR ‘interaction* character*’ OR ‘interaction* assistant*’ OR ‘social agent*’ OR ‘social robot*’ OR ‘social avatar*’ OR ‘social character*’ OR ‘social assistant*’ OR ‘socially assistive agent*’ OR ‘socially assistive robot*’ OR ‘socially assistive avatar*’ OR ‘socially assistive character*’ OR ‘socially assistive assistant*’ OR ‘socially assist agent*’ OR ‘socially assist robot*’ OR ‘socially assist avatar*’ OR ‘socially assist character*’ OR ‘communication agent*’ OR ‘communication robot*’ OR ‘communication avatar*’ OR ‘communication character*’ OR ‘virtual human*’ OR ‘virtual counselor*’ OR ‘virtual consultant*’ OR ‘virtual mentor*’ OR ‘virtual coach*’ OR ‘virtual therapist*’ OR ‘personal assistant*’). The filters used were as follows: academic journal, academic article, and written in English. The search results were consolidated in EndNote^TM X8.2. The duplicates were removed and then transferred to Rayyan for making decisions (Ouzzani et al. 2016).

2.2. Eligibility criteria

The retrieved literature was selected for inclusion or exclusion according to the following criteria:

1. The literature must be a peer-reviewed academic journal article.

2. The literature must be a review article on embodied conversational agent (ECA), chatbot, or SAR.

3. The article must have undergone a systematic literature search.

4. The search target must be literature, not applications or robots.

5. The text must be written in English.

Conference proceedings and dissertations were not considered. The characterisation of a review as pertaining to ECAs, chatbots, or SARs was determined through the presence of articles referencing these entities in the primary literature analyzed within the reviewed articles. For example, a paper was included if it examined chatbots and included in a review on eHealth, but excluded if it used CA for social science research or simulation methods. We also excluded reviews that discussed CA application, but primarily focused on linguistics or deep learning algorithms. We excluded human-operated avatars as virtual counsellors or coaches because these are not CA. Articles on agents that are not CAs, such as those on multi-agent systems that did not include conversational objectives, were also excluded. The criteria for conducting a comprehensive literature search mandate clear specification of the databases searched and the search parameters. Thus, we excluded reviews that merely claimed to have performed a literature review without specifying the databases and keywords employed, and those that failed to report the number or provide a synopsis of the papers used. Review articles that searched applications, agents, patents, or materials rather than literature were also excluded. The review type was not specified. Many narrative reviews and meta-analyses were included. Subjects included both healthy individuals and patients, from children to the elderly. All accepted review articles included at least one article addressing an interaction with CA as an intervention or an exposure. However, comparisons and outcomes were established in some articles, but not in others.

2.3. Information extraction and classification

The information extracted ranged from the number of retrieved references, number of included articles adopted, and year of publication of the included articles. As a basic policy, we focused on reviews targeting the elderly, who have been the focus of particular attention in recent years and extracted them as ‘elderly assistance,’ including those in the medical field. However, the review on ‘dementia/cognitive impairment,’ which has a considerable number of references, was included as a single category, despite some references on subjects other than the elderly. Support for people with psychiatric disorders and mental health care for people without disorders are often discussed together. However, significant differences exist between medical and non-medical care (e.g. application of the Pharmaceutical and Medical Device Act); hence, we tried to distinguish them as much as possible in this study and used the separate categories of ‘psychiatric disorders’ and ‘mental health.’ Other categories were defined as those for which a certain number of references with a similar content could be identified.

3. Results

3.1. Overview of the included articles

The search yielded 1,830 articles, excluding duplicates. Of these, 1,515 articles that showed wrong publication type and research topic, were not a systematic and literature review, or had a wrong language were excluded. The remaining 315 articles were selected for review (Figure 1). The eligible articles were classified into 11 categories based on the research theme, with allowed overlap. Twenty-nine references overlapped into two or three categories.

Figure 1. PRISMA chart for this study. The literature search yielded 1,279 articles, of which 1,074 were screened out of 205 duplicates; 893 were excluded; and 181 were included in the review.

For RQ1, Figure 2 shows a detailed distribution of each category by year of publication. Systematic reviews on CA were published as early as 2012. In 2019 and 2021, the number of reported systematic reviews greatly increased from six papers in 2018 (seven in the number of categories) to 30 in 2019 (36 in the number of categories) and 69 in 2021 (75 in the number of categories). In July 2023, the number of papers reached 70 (79 in the number of categories). The current figure is rapidly approaching the aggregate number of publications recorded in 2022.

Figure 2. Distribution of the categories by year of publication

3.2. Category-specific findings

For RQ2 and RQ3, the systematic review findings for each study area are summarised below.

3.2.1. Mental disorders

Figure 3 displays a list of the systematic reviews on mental disorders. Twenty-five reviews were identified. The composition of the contents included eHealth/mHealth (Denecke, Schmid, and Nüssli 2022; Gooding and Kariotis 2021; Gual-Montolio et al. 2022; Hoermann et al. 2017; Scholten, Kelders, and Van Gemert-Pijnen 2017; Toh et al. 2022), psychiatry (Pacheco-Lorenzo et al. 2021; Vaidyam et al. 2019; Vaidyam, Linggonegoro, and Torous 2021), chatbots (Abd-Alrazaq et al. 2019; Abd-Alrazaq et al. 2021; Ahmed et al. 2023; Lim et al. 2022b; Ogilvie, Prescott, and Carson 2022; Omarov, Narynov, and Zhumanov 2023), and VR (Chard and van Zalk 2022), excluding nine reviews combined with other disciplines.

Figure 3. List of the systematic reviews of conversational agent studies on mental disorders. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

CA interventions for mental disorders exhibit promise in the areas of motivation, behavioural monitoring (Hoermann et al. 2017), depression and social isolation reduction (Lim et al. 2022b; Toh et al. 2022), addiction treatment (Ogilvie, Prescott, and Carson 2022), psychoeducation (Vaidyam et al. 2019), detection and diagnosis (Pacheco-Lorenzo et al. 2021; Vaidyam, Linggonegoro, and Torous 2021), and VR integration (Chard and van Zalk 2022). However, there is a lack of high-quality research (Denecke, Schmid, and Nüssli 2022), and there is no obvious evidence of effectiveness. Additionally, its practical application presents numerous obstacles, including ethical and legal issues (Gooding and Kariotis 2021; Scholten, Kelders, and Van Gemert-Pijnen 2017), as well as the establishment and maintenance of natural responses and connections in open conversations (Abd-Alrazaq et al. 2021).

3.2.2. Neurodevelopmental disorders

Figure 4 presents a list of systematic reviews on neurodevelopmental disorders. A total of 21 reviews were identified. With the exception of four reviews that covered multiple disciplines, the composition of contents included the use of SARs (Alabdulkareem, Alhakbani, and Al-Nafjan 2022; Aresti-Bartolome and Garcia-Zapirain 2014; Islam, Hasan, and Deowan 2023; Kouroupa et al. 2022; Lorenzo et al. 2021; Pennisi et al. 2016; Saleh, Hanapiah, and Hashim 2021; Salimi, Jenabi, and Bashirian 2021; Valencia et al. 2019), special needs education (Papakostas et al. 2021; Perez, Burgos, and Rodriguez 2021), autism spectrum disorder (ASD) social skills (Maddalon et al. 2023; Sani-Bozkurt and Bozkus-Genc 2021; Syriopoulou-Delli and Gkiolnta 2020), emotional communication in ASD (Cano et al. 2021), motor rehabilitation in ASD (Jouaiti and Henaff 2019), and support for adults with intellectual disabilities (MacHale, Ffrench, and McGuire 2023).

Figure 4. List of the systematic reviews of conversational agent studies on neurodevelopmental disorders. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

SARs have been extensively studied for neurodevelopmental disorders, and meta-analyses have shown improvement in social and communication skills (Kouroupa et al. 2022). While the initial focus was on training using generic tools (Aresti-Bartolome and Garcia-Zapirain 2014), the focus gradually shifted to the elements of companion robots and entertainment agents with an awareness of the user experience (Sani-Bozkurt and Bozkus-Genc 2021; Syriopoulou-Delli and Gkiolnta 2020). The direction is expected to increase the motivation for learning by promoting positive emotions, such as joy and fun, and enhance social skills through natural interactions (Valencia et al. 2019). Future work is needed to improve the quality of communication and understanding of emotions and appropriate response skills of SARs and CAs to define and standardise the training methods for therapists, effectively use robots in therapy, and examine the long-term effects of their use in schools and homes (Pennisi et al. 2016; Salimi, Jenabi, and Bashirian 2021).

3.2.3. Dementia/cognitive impairment

Figure 5 lists the systematic reviews on dementia/cognitive impairment. Thirty-four reviews were identified. The composition of the contents included general care and treatment (Cho et al. 2023; Choukou et al. 2023; Hirt et al. 2021; Hsieh et al. 2023; Hung et al. 2019; Koh, Felding, et al., 2021; Lu et al. 2021; Su et al. 2022), caregiving and relationships (Budak et al. 2021; Dada et al. 2022; Ghafurian, Hoey, and Dautenhahn 2021; Hoel, Feunou, and Wolf-Ostermann 2021; Jones et al. 2020; Kruse et al. 2020; Moyle et al. 2017), acceptability (Felding et al. 2023; Whelan et al. 2018; Yu et al. 2022), pet SARs (Kang et al. 2020; Koh, Ang, et al., 2021), cognitive training (Yuan et al. 2021), design(Rampioni et al. 2021), spoken dialogue (Russo et al. 2019), and mild cognitive impairment or amnesia (Boumans et al. 2022; Mancioppi et al. 2019), excluding nine reviews combined with other disciplines.

Figure 5. List of the systematic reviews of conversational agent studies on dementia/cognitive impairment. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

SARs for dementia/cognitive impairment can be divided into the general nursing and treatment and home care contexts. The effects of SARs in general nursing and treatment are varied in effectiveness across measures, with meta-analyses showing effects in agitation, depression, anxiety, and happiness, but not in QOL, or arousal (Cho et al. 2023; Hsieh et al. 2023; Lu et al. 2021). Home care research focuses on the treatment and promotion of engagement and companion SARs, health guidance, assistance with activities of daily living, and initial assessment (Dada et al. 2022; Hoel, Feunou, and Wolf-Ostermann 2021; Kruse et al. 2020). The initial results in both contexts are promising, despite concerns about hygiene and practical (cost and cognitive load) and ethical (pros and cons of attachment formation, lie feedback, privacy, etc.) aspects (Budak et al. 2021; Dada et al. 2022; Ghafurian, Hoey, and Dautenhahn 2021; Hirt et al. 2021; Hung et al. 2019; Koh, Felding, et al., 2021). In addition, research focuses on specific themes, such as acceptability, design, and spoken dialogue. Involving parties in design, creating personalised agents, and enhancing communication skills, including spoken dialogue, can promote acceptability (Rampioni et al. 2021; Russo et al. 2019; Whelan et al. 2018). Acceptance by healthcare providers, caregivers, and cohabitants is also an important factor (Kruse et al. 2020). A series of high-quality effectiveness studies involving system use in daily life and in routine medical care situations must be conducted (Yu et al. 2022; Yuan et al. 2021).

3.2.4. Other medical conditions

Figure 6 depicts a list of other systematic reviews on medicine. Fifty-two reviews were identified. The composition of the contents included general medicine and ethics (Boada, Maestre, and Genís 2021; Chattopadhyay et al. 2020; Chew and Achananuparp 2022; Ciecierski-Holmes et al. 2022; Geoghegan et al. 2021; May and Denecke 2021; Mikkonen et al. 2023; Tudor Car et al. 2020), children (Dawe et al. 2019; Lewis et al. 2021; Littler et al. 2021; Moerman, van der Heide, and Heerink 2019; Triantafyllidis et al. 2023; Trost et al. 2019), chatbots (Federici et al. 2020; Kim et al. 2023; Mirbabaie, Stieglitz, and Frick 2021; Safi et al. 2020), motor assistance (Basteris et al. 2014; Tamburella et al. 2022), COVID-19 (Almalki and Azeez 2020; Amiri and Karahanna 2022; Mbunge et al. 2022; Pallavicini et al. 2022; White, Martin, and White 2022; Xia and Nan 2023), chronic disease (Bin Sawad et al. 2022; Lyzwinski, Elgendi, and Menon 2023; Orpin et al. 2023; Pernencar, Saboia, and Dias 2022; Schachner, Keller, and Wangenheim 2020; Shetty et al. 2022), cancer (Bibault et al. 2019; Shin and Choi 2020; Wang et al. 2023), speech therapy(Chen et al. 2016), HIV(McGuire et al. 2021), cerebral palsy (Malik et al. 2016), tuerculosis management (Needamangalam Balaji et al. 2022), pregnancy and maternity care (Chua et al. 2023; Silveira et al. 2023), emergency and disaster management (Lillywhite and Wolbring 2023), and dental care (Alhaidry et al. 2023), excluding nine reviews combined with other disciplines.

Figure 6. List of the systematic reviews of conversational agent studies on other medical conditions. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

CA is used in other medical care, starting with stroke (Basteris et al. 2014), hearing impairment, dysarthria, aphasia (Chen et al. 2016), cerebral palsy (Malik et al. 2016), management of cancer, hypertension, asthma, orthopedic interventions, ureteroscopy, varicocele (Geoghegan et al. 2021), NICU and pediatric conditions (Lewis et al. 2021; Littler et al. 2021), COVID-19 (Almalki and Azeez 2020; Amiri and Karahanna 2022; Mbunge et al. 2022), HIV (McGuire et al. 2021), treatment and monitoring of various diseases, health services support, and patient education. CA and chatbots have been employed for information provision and disease management, while SARs have been used for rehabilitation, exercise therapy, and childcare. Overall, the expectations for acceptance and effectiveness are high, but research is in its early stages and should have high-quality effectiveness based on standardised assessments and real-world clinical applications, as well as in addressing ethical issues, including privacy and sociopolitical aspects (Boada, Maestre, and Genís 2021; Chattopadhyay et al. 2020; Chew and Achananuparp 2022; May and Denecke 2021). Meta-analysis has so far provided no clear evidence on the efficacy of CA in medicine (Mikkonen et al. 2023; Shetty et al. 2022).

3.2.5. Elderly support

Figure 7 lists the systematic reviews on elderly support. Thirty-two reviews were identified. The composition of the contents included general support for the elderly (Abdi et al. 2018; Asgharian, Panchea, and Ferland 2022; Frennert and Ostlund 2014; Kachouie et al. 2014; Winkler et al. 2023), caregiving (Abbott et al. 2019; Bemelmans et al. 2012; Gibelli et al. 2021), acceptability (Kachaturoff et al. 2021; Loveys et al. 2022; Vandemeulebroucke, de Casterlé, and Gastmans 2018), effectiveness (Lee et al. 2022; Pu et al. 2019; Støre, Beckman, and Jakobsson 2022; Vandemeulebroucke, Dzi, and Gastmans 2021), coaching (Bevilacqua et al. 2020; El Kamali et al. 2020; Stara et al. 2020), psychiatric symptoms (Chen, Jones, and Moyle 2018; Dunham et al. 2021; Kulpa, Rahman, and Vahia 2021; Salvemini et al. 2019), and loneliness (Gasteiger, Loveys, et al., 2021; Latikka et al. 2021), excluding eight reviews combined with other disciplines.

Figure 7. List of the systematic reviews of conversational agent studies on elderly support. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

CA for elderly support has been studied for a long time. The following challenges were identified in 2014: influence of cultural background, appearance or aesthetics and user expectations; difficulties in home use; subject bias (more women, fewer studies with stakeholders, such as caregivers); considerations for person-centered care and development; lack of reproducibility studies and insufficient sample size; low validity of scales and data; studies of acceptability and role of robots; and conflict between technical determinism and social construction (Frennert and Ostlund 2014; Kachouie et al. 2014). Since then, not all of these challenges had been resolved, especially the ethical and legal aspects, which have not yet been fully discussed (Abdi et al. 2018; Gibelli et al. 2021; Vandemeulebroucke, de Casterlé, and Gastmans 2018; Vandemeulebroucke, Dzi, and Gastmans 2021). Although improvements in psychosocial aspects, such as mood, loneliness, social connectedness, communication, and engagement, as well as reductions in stress indicators are expected (Bemelmans et al. 2012), a meta-analytic evaluation found only a reduction in agitation of pet-type SARs (Abbott et al. 2019; Lee et al. 2022). The poor quality of research (Pu et al. 2019) and problems of acceptability (Abbott et al. 2019) are pointed out as reasons for the unclear effectiveness of CA and SAR. The perceptions and attitudes of caregivers and the elderly themselves are important. Guidelines and support systems must also be developed for home use. In general, elderly support has been the subject of much research on SARs, especially pet-based SAR. In contrast, CA has been extensively studied with respect to coaching for behaviour change with older adults (El Kamali et al. 2020; Stara et al. 2020). Interventions for loneliness and other psychiatric symptoms in older adults with SARs have also been studied, but high-quality studies and robust evidence are lacking (Gasteiger, Loveys, et al., 2021; Salvemini et al. 2019). Future studies should compare safety, acceptability, cost-effectiveness, and impact on caregivers with existing treatments (Dunham et al. 2021; Winkler et al. 2023).

3.2.6. Health promotion

Figure 8 provides a list of the systematic reviews on health promotion. Thirty were identified. The composition of the contents included health promotion in general (Denecke and May 2023; Kennedy et al. 2012; Milne-Ives et al. 2020; Montenegro, Costa, and Righi 2019; Sezgin et al. 2020; Shan, Ji, Xie, Qian, et al., 2022; ter Stal et al. 2020), chatbots (Abd-Alrazaq, Safi, et al., 2020; Aggarwal et al. 2023; Gabarron et al. 2020; Han et al. 2023; Oh et al. 2021; Singh et al. 2023; Whittaker, Dobson, and Garner 2022; Zhang et al. 2020), coaching (Kramer et al. 2020; Tsiouris et al. 2020), personalisation (Chew 2022; Kocaballi et al. 2019), and physical activity and weight loss (Luo et al. 2021; Noh et al. 2023), excluding nine reviews combined with other disciplines.

Figure 8. List of the systematic reviews of conversational agent studies on health promotion. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

CA for health promotion has generally been well evaluated in its early stages, with 41 studies recognised as early as 2012 and RCTs showing the effects of CA on behaviour change and acceptability (Kennedy et al. 2012). Evaluation has been particularly high for acceptability and user experience (Milne-Ives et al. 2020). The target population ranges from children to the elderly. The modalities of conversation vary from text, voice, gesture, and multimodal (Montenegro, Costa, and Righi 2019; Sezgin et al. 2020). The technologies used also include convolutional neural networks, Wizard of Oz methods, Markov chains and reinforcement learning, rule-based expert system models, NLP (e.g. bag-of-words), and various acoustic models (Montenegro, Costa, and Righi 2019). The majority of interventions are multicomponent, and behaviour change techniques, such as shaping knowledge, self-belief, repetition and substitution, feedback and monitoring, goals and planning, antecedents, natural consequences, comparison of behaviour, and identification, have been used (Sezgin et al. 2020). More CA and chatbots have been studied when compared to SARs. However, chatbot studies on diet and weight change have been limited, and positive results have been obtained regarding increased physical activity (Han et al. 2023; Oh et al. 2021). Although the quality of research is limited, and no definitive conclusions have been drawn (Gabarron et al. 2020; Zhang et al. 2020), recent meta-analyses show positive effects of chatbots on increased physical activity and improved diet and sleep (Singh et al. 2023). Standard models must be established for designs that include empathy with emotional expression and relational behaviour (Kennedy et al. 2012; ter Stal et al. 2020). Communication skills also present challenges (Milne-Ives et al. 2020). The integration of chatbots and coaching with automated patient monitoring and real-time assessment (Tsiouris et al. 2020) and the development of theory-based personalisation (Denecke and May 2023; Kocaballi et al. 2019) are also expected to increase the effectiveness of CA for health promotion.

3.2.7. Mental health

Figure 9 lists the systematic reviews on mental health. Twenty-three reviews were identified. The composition of the contents included mental health support by SARs (Robinson, Cottier, and Kavanagh 2019; Scoglio et al. 2019), CA for suicide prevention (Martínez-Miranda 2017), CA of online mental health interventions for the youth (Shan, Ji, Xie, Li, et al., 2022; Zhou et al. 2021), mental health during the COVID-19 pandemic (Wilson and Marasoiu 2022; Zhang and Smith 2020), SAR for child mental health (Kabacinska, Prescott, and Robillard 2021), and CA for promoting resilience (Pusey, Wong, and Rappa 2020), excluding 14 reviews combined with other disciplines.

Figure 9. List of the systematic reviews of conversational agent studies on mental health. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

The number of systematic reviews of CA studies in the mental health area is limited. Existing reviews suggest that SARs target the elderly and children (Kabacinska, Prescott, and Robillard 2021; Robinson, Cottier, and Kavanagh 2019; Scoglio et al. 2019), while CAs and chatbots tend to be used for suicide prevention (Martínez-Miranda 2017), youth mental health interventions (Zhou et al. 2021), and mental health care during the COVID-19 pandemic (Zhang and Smith 2020). CA is often used as a self-help tool available online, while chatbots are utilised as tele psychological support. While there have been systematic reviews of coaching for behaviour change in health promotion and non-psychiatric areas of medicine, no systematic reviews have yet been reported on coaching for promoting wellbeing in the mental health area with non-clinical groups, such as healthy children, workers, and the elderly.

3.2.8. Education

Figure 10 lists the systematic reviews on education. Forty reviews were identified. The composition of the contents included the overall use and design of AI and SARs in education (Armando, Ochs, and Régner 2022; Bonaiuti et al. 2022; Chen, Chen, and Lin 2020; Dai and Ke 2022; Davis, Park, and Vincent 2021; Martha and Santoso 2019; Mejbri et al. 2022; Woo et al. 2021), language education (Ao and Yu 2022; Bibauw, François, and Desmet 2019; Klimova and Seraj 2023; Klímová et al. 2023; Neumann 2020; van den Berghe 2022; van den Berghe et al. 2019), chatbots (Deng and Yu 2023; Huang, Hew, and Fryer 2022; Hwang and Chang 2021; Jeon, Lee, and Choi 2023; Kuhail et al. 2023; Lin and Yu 2023; Lo 2023; Lo and Hew 2023; Mohammad et al. 2023; Okonkwo and Ade-Ibijola 2021; Perez, Daradoumis, and Puig 2020; Sallam 2023; Urquiza-Yllescas et al. 2022; Wollny et al. 2021; Wu and Yu 2023; Zhai and Wibowo 2022; Zhang, Zou, and Cheng 2023), special needs education (Cinquin, Guitton, and Sauzéon 2019), science, technology, engineering, and mathematics (STEM) education (Papadopoulos et al. 2020; Zhan et al. 2022), VR applications (Pirker and Dengel 2021), nursing or medical education (Buchanan et al. 2021; Nizhenkovska et al. 2022), music education (Martinez-Roig, Cazorla, and Faubel 2023), and teacher education in low- and middle-income countries (Hennessy et al. 2022).

Figure 10. List of the systematic reviews of conversational agent studies on education. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

In the education domain, humanoid SARs (Woo et al. 2021), ECAs capable of nonverbal communication (Martha and Santoso 2019), and chatbots specialised for specific functions (e.g. Q&A, language learning, and mentoring) have been studied (Huang, Hew, and Fryer 2022; Hwang and Chang 2021; Jeon, Lee, and Choi 2023; Wollny et al. 2021). The educational contents studied are language education (Bibauw, François, and Desmet 2019), medical, nursing, and STEM education (Papadopoulos et al. 2020; Zhan et al. 2022), music education (Martinez-Roig, Cazorla, and Faubel 2023), special needs education (Cinquin, Guitton, and Sauzéon 2019), and teacher education (Hennessy et al. 2022). CA plays an educational role, either alone or with teachers, and demonstrates advantages, such as content integration, rapid access, motivation and engagement, multiple users, and immediate support (Chen, Chen, and Lin 2020; Huang, Hew, and Fryer 2022; Okonkwo and Ade-Ibijola 2021). Meta-analysis showed a positive impact of chatbots on learning outcomes, especially in higher education for short periods of time (Wu and Yu 2023). Practical challenges include the lack of distinct educational effects partly due to the novelty effect, lack of ethical discussion, negative attitudes of those involved, and lack of guidelines for robot maintenance, supervision, and utilisation (Hwang and Chang 2021; Perez, Daradoumis, and Puig 2020; Woo et al. 2021). Empirical studies, particularly in elementary and secondary education, longitudinal studies addressing novelty effects, efficacy comparisons with other IT-based education, and studies on the impact of teacher perception and support are all required to answer research questions (Davis, Park, and Vincent 2021; Huang, Hew, and Fryer 2022).

3.2.9. Industrial applications

Figure 11 presents a list of the systematic reviews on industrial applications. Thirty-four reviews were identified. The composition of the contents included stakeholder communications (Bakkouri, Raki, and Belgnaoui 2022; Bălan 2023; da Silva Oliveira and Chimenti 2021; Gopinath and Kasilingam 2023; Jenneboer, Herrando, and Constantinides 2022; Kambur and Yildirim 2023; Lim et al. 2022a; Ling et al. 2021; Mariani, Hashemi, and Wirtz 2023; Misischia, Poecze, and Strauss 2022; Nicolescu and Tudorache 2022; Pentina et al. 2023; Pillarisetty and Mishra 2022; Ramesh and Chawla 2022; Shah et al. 2023; Syvänen and Valentini 2020; Van Pinxteren, Pluymaekers, and Lemmink 2020), use in hospital (Berdahl et al. 2022; Gonzalez-Gonzalez, Violant-Holz, and Gil-Iranzo 2021; Haubold, Obst, and Bielefeldt 2020; Scholten, Vissenberg, and Heerink 2016), service or social robots (Blaurock et al. 2022a; 2022b; De Keyser and Kunz 2022; Lajante, Remisch, and Dorofeev 2023; Oruma et al. 2022), ML (Bavaresco et al. 2020), sex industry (González-González, Gil-Iranzo, and Paderewski-Rodríguez 2020), apparel industry (Landim et al. 2021), autmobile industry (Lee and Jeon 2022), museum or library (Gasteiger, Hellou, et al., 2021; Sylaiou and Fidas 2022; Yan, Zhao, and Mazumdar 2023), and food aid (Martin et al. 2022).

Figure 11. List of the systematic reviews of conversational agent studies on industrial applications. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

Research on the industrial applications of CA can be classified into five categories as follows: communication among stakeholders (Syvänen and Valentini 2020), healthcare operations (Haubold, Obst, and Bielefeldt 2020), service or social robots (De Keyser and Kunz 2022), ML applications (Bavaresco et al. 2020), and industry-specific agents. For communication among stakeholders, expectations are particularly high for chatbots, which can exhibit various affordances (i.e. affective, cognitive, functional, relational, and material) to enhance UX and customer loyalty. Meta-analysis indicated that chatbot playfulness, attitude, usefulness, facilitating conditions, and social influence are important factors influencing user adoption (Gopinath and Kasilingam 2023). In healthcare organisations, research focuses not only on improving the quality of care and services, but also on how to effectively incorporate CAs into the operations of healthcare organisations (Gonzalez-Gonzalez, Violant-Holz, and Gil-Iranzo 2021). For service robots, the design of appearance and agents for enhancing customer satisfaction is being studied (Blaurock et al. 2022b). The range of CA applications, including ML, is wide and includes commercial transactions, such as pre-sales, sales execution, and post-sales customer care (Bavaresco et al. 2020), fashion (Landim et al. 2021), sex industry (González-González, Gil-Iranzo, and Paderewski-Rodríguez 2020), museums (Gasteiger, Hellou, et al., 2021), and food aid (Martin et al. 2022). Future challenges must take an interdisciplinary approach, including service management and communication science, based on meso (organisation) and macro (society) level considerations and micro level agent characteristics for industrial applications and promote utilisation to increase customer satisfaction and trust (Syvänen and Valentini 2020). In addition, amidst some over-expectations, the challenge is how to incorporate CA and its management into actual service delivery processes to create value considering ethical and legal aspects (De Keyser and Kunz 2022).

3.2.10. Agent characteristics

Figure 12 lists the systematic reviews on agent characteristics. Thirty-three reviews were identified. The composition of the contents included CA features and design in general (Abrantes et al. 2023; Alsheddi and Alhenaki 2022; Arsenyan, Mirowska, and Piepenbrink 2023; Eiris and Gheisari 2017; Hatfield et al. 2022; Kelly, Kaye, and Oviedo-Trespalacios 2023; Loveys et al. 2020), chatbots (Agarwal, Agarwal, and Gupta 2022; Ahmed et al. 2022; Albites-Tapia et al. 2022; Bilquise, Ibrahim, and Shaalan 2022; Caldarini, Jaf, and McGarry 2022; Kavaz, Puig, and Rodriguez 2023; Lee, Ju, and Lee 2023; Lin, Huang, and Yang 2023; Mariciuc 2022; Miklosik, Evans, and Qureshi 2021; Mohamad Suhaili, Salim, and Jambli 2021; Phiri and Munoriyarwa 2023; Rapp, Curti, and Boldi 2021; Silva and Canedo 2022; J. Wang et al., 2021; Yang et al. 2023), social cues (Chaturvedi et al. 2023; Feine et al. 2019; Liew and Tan 2021), nonverbal communication (Tubin, Mazuco Rodriguez, and de Marchi 2022; Wang and Ruiz 2021), and trust or social presence (de Barcelos Silva et al. 2020; Felnhofer et al. 2023; Rheu et al. 2021), excluding two reviews combined with other disciplines.

Figure 12. List of the systematic reviews of conversational agent studies on agent characteristics. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

Research on CA features has been conducted as architecture, engineering, and construction research on associated costs, site management, scheduling, system evaluation and analysis, collaboration and communication, safety, and education (Eiris and Gheisari 2017). Future challenges include the integration of new emerging hardware and software technologies for interacting with and displaying virtual humans and interaction among humans, avatars, and agent systems and devices that integrate VR and AR (Eiris and Gheisari 2017).

Chatbot research focuses on high-level statistical performance and system development and testing, including deep and reinforcement learning architectures (J. Wang et al., 2021). Compared to the evaluation of some features, chatbot recognition, and acceptance, less research has been conducted on the FAQs, troubleshooting, recruiting, chatbot relationships, trust, promotion, health, security, user classification, purchasing, personalisation, and surveys (Miklosik, Evans, and Qureshi 2021). Research on social cues (Liew and Tan 2021) and nonverbal communication (Tubin, Mazuco Rodriguez, and de Marchi 2022) is underway, but it is not at a stage where standardised perceptions or definitive evidence can be established (Loveys et al. 2020; Wang and Ruiz 2021). One meta-analysis suggested that in immersive VR, human-operated avatars exhibit a higher social presence than computer-controlled agents (Felnhofer et al. 2023).

3.2.11. Robot characteristics

Figure 13 lists the systematic reviews on robot characteristics. Twenty-seven reviews were identified. The composition of the contents included robot features in general or specific robots (Aitsam, Davies, and Di Nuovo 2022; Amirova et al. 2021; Borboni et al. 2023; Coronado, Itadera, and Ramirez-Alpizar 2023; Damholdt et al. 2023; Lambert et al. 2020; Lundgren et al. 2022; Maroto-Gomez et al. 2023), emotion recognition and emotional expression (Cavallo et al. 2018; de Wit, Vogt, and Krahmer 2023; Song and Luximon 2020; Y. Y. Wang et al., 2021), relationships with children (de Jong et al. 2019; Filippini and Merla 2023; Pashevich 2022; van Straten, Peter, and Kühne 2020), usability and user experience (David, Thérouanne, and Milhabet 2022; Jung, Lazaro, and Yun 2021; Naneva et al. 2020; Ringwald et al. 2023; Shourmasti et al. 2021), privacy (Lutz, Schottler, and Hoffmann 2019), personality influence (Mou et al. 2020), persuasion (Liu, Tetteroo, and Markopoulos 2022), and programming (Coronado et al. 2020), excluding two reviews combined with other disciplines.

Figure 13. List of the systematic reviews of conversational agent studies on robot characteristics. The Y-axis shows the authors and the number of papers adopted in order of publication year. The X-axis depicts the earliest and latest publication years of the adopted papers.

Regarding robot features, the superiority of real-world SARs over on-screen ECAs and chatbots and the promise of mixed-reality robotics have been reported (Lambert et al. 2020). However, many challenges exist as regards improving the quality of research to be addressed, including the development of methods for evaluating human–robot interactions, continued interest in robots without novelty effects, examining satisfaction with social interactions over time, studies in settings other than assisted living and medical stay programmes, and studies of large samples (Amirova et al. 2021; Lambert et al. 2020). For social implementation, issues like acceptability (Song and Luximon 2020), emotion recognition and expression (Cavallo et al. 2018), usability and UX-related responses (Shourmasti et al. 2021), child-specific validation issues (de Jong et al. 2019), and ethical and legal issues around privacy (Lutz, Schottler, and Hoffmann 2019), remain unresolved. With the progression of technology, robots have exhibited enhanced information processing capabilities. Improvements in visual semantic comprehension and emotion recognition have been observed (Borboni et al. 2023; Lundgren et al. 2022), elevating anticipations for collaborative robotics and industrial applications of SARs in general.

3.2.12. Composite domains

Mental disorders had overlaps with neurodevelopmental disorders (Provoost et al. 2017), other medical conditions (Bérubé et al. 2021), mental health (Abd­Alrazaq et al. 2020; Bendig et al. 2019; Gaffney, Mansell, and Tai 2019; He et al. 2023; Jabir et al. 2023), neurodevelopmental disorders and dementia/cognitive impairment (Guemghar et al. 2022), and neurodevelopmental disorders and other medical conditions (de Filippis et al. 2020). Most CA applications in clinical psychology are in the development and testing stages and have not yet reached a routine practice evaluation or application (Provoost et al. 2017). The CA is also used for the behavioural support and health monitoring for cancer, diabetes, heart failure, hearing impairment, asthma, Parkinson’s disease, dementia, ASD, intellectual disability, and depression. However, only a few studies have been conducted, and no conclusive evidence has been presented on their usefulness (Bérubé et al. 2021). Weak evidence based on a meta-analysis has been found for CA, including effectiveness in improving depression, distress, stress, anxiety, and fear of heights, although high-quality RCTs are scarce.

Overlaps between mental health and health promotion (Curtis et al. 2021; Giansanti 2023; Laranjo et al. 2018), and between mental health, health promotion and other medical care were also observed (Denecke and May 2022; Esterwood and Robert 2021; Li et al. 2023; Martinengo et al. 2022; Pereira and Díaz 2019; Sadasivan et al. 2023). The impact of matching patient and robot personalities on outcomes has been reported (Esterwood and Robert 2021). However, longitudinal studies for determining the optimal design of CAs to promote mental and physical health are lacking (Curtis et al. 2021). Meta-analysis on CA in the healthcare domain indicates moderate or higher acceptability and effects on physical functioning, healthy lifestyle, mental health, and psychosocial outcomes (Li et al. 2023).

Dementia/cognitive impairment and elderly support overlapped in seven studies (Alnajjar et al. 2019; Asl et al. 2022; de Araujo et al. 2021; Fardeau, Senghor, and Racine 2023; Góngora Alonso et al. 2019; He, He, and Liu 2022; Huq, Maskeliūnas, and Damaševičius 2022). The differences in the acceptability of SARs with and without diagnosis (Góngora Alonso et al. 2019), effectiveness of using robots in computerised cognitive training for older adults and those with mild cognitive impairment (Alnajjar et al. 2019), and poor validity of SAR studies for depressive symptoms in elderly residents of long-term care facilities were identified (de Araujo et al. 2021). Additionally, dementia/cognitive impairment overlapped with other medical conditions, and a review of CA studies focusing on the rehabilitation of adults with neurological diseases uncovered seven prototypes (Hocking et al. 2023). There were no efficacy trials conducted and the existing literature exhibited heterogeneity, indicating that the research is still in its preliminary stages.

One overlap between agent and robot features was found (Oliveira et al. 2021; Park and Whang 2022). Empathic ECAs and SARs evolved to be a domain-independent multimodality processing both emotion and cognition and are emotionally expressive, adjusting empathy to situations and relationships (Park and Whang 2022). About half of the studies reported positive effects of interaction with ECAs and SARs on prosociality (Oliveira et al. 2021). In addition, an overlap between neurodevelopmental disorders and elderly support was also found (Even et al. 2022). CA has the potential to improve the QOL for older adults with intellectual disabilities, contingent upon thoughtful consideration of their learning requirements, existing capabilities, and privacy issues.

4. Discussion

A meta-review of the systematic reviews on the ECA, chatbot, and SAR provided a comprehensive picture of the results and challenges of research unique to each of the 11 research domains and those common to these domains. Below is a more detailed discussion of the results of the review along each RQ.

4.1. Overall trends and differences among the domains

For RQ1, we discerned overarching trends in CA research and identified distinct differences across domains. The number of reports of systematic reviews increased in 2019 and 2021, which may be due to the development of ML, deep learning, and NLP technologies and the growing societal expectations and perceptions of AI (Arnold et al. 2019; Kalyanathaya, Akila, and Rajesh 2019). This change is reflected in the increasing number of more specific keywords (e.g. chatbots and deep learning) in research topics for industrial applications (Bavaresco et al. 2020). Specifically, an increasing number of CAs are utilising ML, deep learning, and NLP techniques to comprehend natural language input and equipping them with human-like response capabilities (Almalki and Azeez 2020; Amiri and Karahanna 2022; Chew and Achananuparp 2022; Kovacek and Chow 2021; Laranjo et al. 2018; Mbunge et al. 2022; Perez, Daradoumis, and Puig 2020; Siddique and Chow 2021; J. Wang et al., 2021). Although there are probably some expectations ahead of time, the number of review papers on CA is increasing as the technology develops. It is a research area that is developing every day right now and expected to develop further in the future.

For the 11 areas covered herein, the largest number of systematic reviews reported to date is 52 for other medical conditions, followed by 40 for education, 34 for dementia/cognitive impairment, industrial applications and agent characteristics. Although the high number of reviews related to the elderly and medical care is noticeable, there have also been many reviews conducted on non-clinical groups, such as health promotion and education. However, the existing research on mental health is blurry in its distinction of psychiatric disorders, dementia/cognitive impairment, neurodevelopmental disorders, and other adjacent areas. Imagine the large social demand for medical care and elderly support. Research on these areas has been conducted in the lead. Research on non-clinical groups, such as suicide prevention, youth mental health, mental health of the general population during the COVID-19 pandemic, and mental health of middle-aged and older adults, including workers, and research aimed at promoting wellbeing are highly important. Therefore, research on mental health that is not related to medical care is an area where further growth can be expected in the future.

Interestingly, the main type of CA studied in each domain differs. Figure 14 roughly shows the relationship of ECA, chatbot, and SAR and each domain. ECAs and chatbots are mostly studied in mental disorders and health promotion, while chatbots and SARs are mostly studied in education, industrial applications, and other medical conditions. Meanwhile, ECAs and SARs are mostly studied in mental health, dementia/cognitive impairment, elderly support, and neurodevelopmental disorders.

Figure 14. Image of the proximity of the ECA, chatbot, and SAR to each practice domain.

ECAs and chatbots excel in gathering and providing information, delivering dialogue-based mental and behavioural change support, and communicating among stakeholders. As a software reality installed on websites or on PCs and tablets, ECAs often guide a structured programme, communicate with the user, and provide information as appropriate, while chatbots as chat functions in chat applications on mobile devices or on websites often provide dialogue-based information, assist in self-exploration, and support and record self-monitoring. ECAs have been used to diagnose mental disorders (Vaidyam, Linggonegoro, and Torous 2021), eHealth/mHealth (Denecke, Schmid, and Nüssli 2022), support behaviour change for health promotion (Sezgin et al. 2020), and virtual coaching (Kramer et al. 2020). Meanwhile, chatbots are used to provide psychotherapy for mental illness (Lim et al. 2022b), promote physical activity to improve health (Oh et al. 2021), and communicate among stakeholders in the industrial domain (Syvänen and Valentini 2020).

In contrast, SARs are characterised by the presence of realities and excel at engaging with subjects who have difficulty with text- or voice-only communication, such as the elderly and children. SARs have been used in service robots in industrial applications (e.g. De Keyser and Kunz 2022), in areas involving physical activity outside the home, such as rehabilitation in medicine (e.g. Tamburella et al. 2022); in other medical conditions, neurodevelopmental disorders, education, and other settings involving children (e.g. Chen, Chen, and Lin 2020; Dawe et al. 2019; Lorenzo et al. 2021); and in areas involving the elderly, such as elderly support and dementia/cognitive impairment (e.g. Lu et al. 2021; Vandemeulebroucke, Dzi, and Gastmans 2021).

While the differences in the CA types primarily studied between the domains may be reasonable, excessive bias would limit the user choice. For example, limiting SARs for the elderly to pet types is expected to lead to resistance by some users (Koh, Ang, et al., 2021). For children, SARs are effective; however, considering the cost, an ECA or a chatbot may be more convenient for some purposes. Aside from developing effective CA research in each domain, considering expanding user options is also important.

4.2. Requirement for ethical contemplations

In the context of RQ2, ethical considerations are actively explored across numerous domains, leading to a growing body of insights. The examination of ethical considerations is an indispensable aspect in the research, development, and societal integration of CAs. This review has demonstrated that ethical considerations and legal framework are extensively discussed, with a few exceptions in certain domains. Some studies have noted the lack of comprehensive ethical discussions in the medical domain (e.g. Vandemeulebroucke, Dzi, and Gastmans 2021), in contrast to the educational domain, where ethical and safety considerations have been discussed but not incorporated into the assessment of learning efficacy (Woo et al. 2021). Ethical considerations and legislation have been the focus of extensive discussion, particularly in the realms of medicine, dementia/cognitive impairment, and education. Based on the discussions inherent to the reviewed CA academic research, ethical considerations can be categorised into the following nine categories: privacy, safety, innovation, user acceptance, psychological attachment, care philosophy, evaluation, social systems compatibility, and rule development.

Privacy is generally an important ethical aspect of technology use, and CA is no exception. Most especially when using new technologies, such as multichannel data sensor integration and cloud computing, it is important to disclose information and obtain consent regarding privacy protection and the treatment of personal information (Cavallo et al. 2018). If the system is based on ML, what information should be used as data for training and how much of it must be disclosed should be thoroughly discussed, including the users’ views. Meanwhile, the excessive protection of information may hinder fair competition and technological innovation. When considering privacy, the viewpoints of promoting technological innovation, convenience, and industrial application must be included while assuming protection and respect for the individual.

Safety is of paramount importance in ethical consideration. The deployment of technology necessitates the evaluation of anonymity and psychological safety of user groups or social contexts (Gooding and Kariotis 2021). SARs must be free of physical hazards and adhere to hygiene standards. The possibility of addiction behaviours, akin to internet gaming disorder, particularly with ECAs and chatbots, must be thoroughly examined. As with any type of human interaction, the use of information technology will never be entirely risk-free. The potential for risk must be unambiguously identified, and the locus of responsibility must be clearly delineated to guarantee safety.

Innovation is a matter of CA development. Ethical development includes the transparency and accountability of algorithms, user-centered development, user involvement in design, improvement design and quality of support, and preparation of a framework for development and implementation. One important argument for ethical development is technical determinism versus social construction (Frennert and Ostlund 2014). Technical determinism is the idea that technological development leads to social development. The user’s role is passive and does not influence equipment usage. By contrast, in social construction, technology and society mutually shape a robot. Robot outcomes are determined by human choices, and social conditions influence technological development. The user-participatory design is an important element in advancing social construction.

User acceptance is related to usability, UX, familiarity, and convenience. Not all users have positive attitudes toward SARs, especially in dementia/cognitive impairment and elderly support (Abbott et al. 2019; Koh, Ang, et al., 2021). Acceptance by caregivers, nurses, teachers, and other stakeholders is also an important factor (Gibelli et al. 2021; Kruse et al. 2020). From an ethical perspective, the conception must be palatable to all, not merely to those who are capable of utilising, comprehending, or operating it.

Psychological attachment is a matter of attachment formation with CA and robots. It is a particularly important factor in medical care involving dementia/cognitive impairment and children. People with dementia/cognitive impairment are at risk of misidentifying animal-type SARs as real animals or becoming overly dependent on SARs (Koh, Ang, et al., 2021). Children also form attachments to CAs more strongly than adults. Consequently, they may feel a strong sense of loss when SARs malfunction or are retrieved after an experiment (Kabacinska, Prescott, and Robillard 2021). In operating CAs, users should be aware of the risks associated with excessive attachment formation and should be alerted to the dangers that come with their usage.

Care philosophy is the aspect concerning the principles and value standards that underlie activities, such as technology use, care, and education. CA development has called for a rethinking of the nature of social activities, including care, nursing, medicine, education, and business. Respect for the supported and users and respect for individual lives and ideas are very important aspects. Suitable and good relationships between human beings and CAs must be designed on this basis. There are currently only a limited number of cases in which CAs replace specialists, often in an auxiliary or complementary role. In implementing CA, one must be aware of the balance among strategic, ethical, and human-oriented aspects (Haubold, Obst, and Bielefeldt 2020). Patients and consumers could be depersonalised by excessive commoditization and techno-utilitarianism (Gooding and Kariotis 2021). A philosophical anthropology from the perspective of inter subjectivity holds significant relevance (Boada, Maestre, and Genís 2021). The impact of the emergence of robots and their replacement of humans in the society must be examined in terms of freedom, responsibility, happiness, care, and justice. For example, a false positive feedback by CA is fraught with ethical issues in terms of the philosophy of assistance (Yuan et al. 2021). Alternatively, prejudice against the use of CAs and the impact of being cared for by CAs on the user’s sense of existence must be fully considered.

Appropriate evaluation is another important ethical perspective requiring a systematic and comprehensive assessment of the consequences of CA use (Haubold, Obst, and Bielefeldt 2020). The evaluation includes the monitoring of the long-term effects on patients, users, supporters, staff, work teams, and organisations. Identifying the efficacy and side effects and establishing trust in the society are critical for the ethical social implementation.

Compatibility with social systems is an aspect that looks at the impact of sociopolitical structures on the way care and education are provided, the lives of individuals, and the development of technology rather than just the one-to-one relationship between supporters and recipients (Boada, Maestre, and Genís 2021). Ethics, jurisprudence, and sociopolitical establishments are interdependent and cannot subsist in isolation from one another. In reference to CA, the examination must contemplate the imbalances linked with the digital divide and corresponding societal frameworks and regulations.

Finally, the development of rules is a crucial aspect in ethical considerations (Okonkwo and Ade-Ibijola 2021). Establishing clear regulations for the deployment of ECAs, chatbots, and SARs, codifying ethical principles, and circumscribing their utilisation in specific circumstances can be efficacious not only in terms of ethical regards but also through the establishment of explicit rules. Specifically, the principle of informed consent must be meticulously executed for individuals with dementia/cognitive impairment, the elderly, those with neurodevelopmental conditions, and minors based on their individual circumstances. Additionally, as the range of activities that CAs can perform autonomously increases, it will be necessary to devise detailed rules and restrictions for their application in each domain.

4.3. Lack of comparative studies and addressing acceptability

Pertaining to RQ3, the current challenges in CA research include the paucity of high-quality comparative studies and the imperative to address acceptability. At this stage, the domains lack high-quality comparative studies with respect to CA. According to meta-analyses, chatbots reduce the symptoms of depression (Lim et al. 2022b), distress, stress, and fear of heights (Abd-Alrazaq, Rababeh, et al., 2020), while pet-type SARs reduce agitation in the elderly, including dementia (Abbott et al. 2019; Lu et al. 2021), and have overall medical benefits (Chattopadhyay et al. 2020). However, many challenges exist, including the heterogeneity of primary studies and a few comparisons with active controls. The meta-analyses for robot characteristics were not identified in this review. Although some effects have been reported in the RCTs, high-quality comparative studies with large sample sizes, standardised effect measures, verification in everyday situations outside the laboratory, elimination of novelty effects, and verification of long-term effects are required. Specifically, studies assessing effectiveness should control for factors impacting outcomes, including age, gender, personality, acceptability, and individualisation.

One factor influencing effectiveness is acceptability. Acceptability is an important factor not only in terms of ethics, but also in terms of expanding effectiveness and social implementation. Acceptability is often an issue, especially in initiatives targeting the elderly. This is not only the acceptability of the users themselves, but also the attitude of caregivers and nurses (Koh, Felding, et al., 2021). According to this review, older adults with dementia or cognitive impairment are more aware of the need for SARs and more accepting than those without (Góngora Alonso et al. 2019). In terms of the cognitive load, older adults without cognitive problems seem to feel more comfortable using SARs (Kachaturoff et al. 2021). Moreover, the lower perception of need seems to result in poor acceptance in a non-clinical group. In contrast, acceptance is generally good in the context of long-term care (Loveys et al. 2022). Even if they have not been diagnosed, older adults themselves are likely to be more accepting if they are aware of the specific purpose and need for their care (e.g. nursing care). However, not all care recipients seem to be positive about CA (Vandemeulebroucke, de Casterlé, and Gastmans 2018); thus, devising an introduction and being cautious are important.

4.4. Quality of conversational functioning

Also related to RQ3, enhancing conversational quality emerges as another pivotal challenge that CA research should confront in the forthcoming periods. With the development of the NLP and emotion estimation technologies, the ability of CAs to generate responses has increased. CAs are now required to do more than simply provide information suitable for input. They must now build a relationship with the user as an agent and assist the user by acting as a companion and an entertainer. In the future, CAs must be reliable and dependable, as well as accessible and acceptable. They must exude a presence that people want to be with and be involved with. To achieve this, they must be able to resolve the several challenges that exist in terms of conversational function.

The first challenge is basic communication skills, especially in the care of people with dementia and the elderly. User-friendly communication reduces the cognitive load and increases acceptance (Russo et al. 2019). Basic communication skills are also important in health promotion as a basis for relationship building (Milne-Ives et al. 2020). In the educational domain, the integration of natural verbal and nonverbal communication, movement, and character art is required (Martha and Santoso 2019). Nonverbal communication has particularly been the subject of conflicting reports as regards its benefits and inaccuracies in implementation (Wang and Ruiz 2021) and is an area for further study.

Second, CAs must accurately assess the user state, especially their emotional state, to tailor responses to the user state. The relevant challenges include the real-time assessment of emotional states in psychiatry (Scholten, Kelders, and Van Gemert-Pijnen 2017) and the assessment of learners’ emotional states and motivational responses in educational treatment and education (Martha and Santoso 2019). The emotional expression of CAs is also an important issue in addition to user emotion recognition. Studies on the appropriate design regarding appropriate verbal and nonverbal communication styles, expression of empathy and emotions, and personality are particularly needed when working with patients with ASD and children (Cano et al. 2021). The implementation of the appropriate emotional expression is also a challenge with regard to suicide prevention in the mental health domain (Martínez-Miranda 2017).

Empathy and relationship building are becoming increasingly important as conversational functioning. In health promotion, empathy and relationship-building behaviours are important in addition to basic communication skills (Kennedy et al. 2012). Establishing a consensus on the CA design for health promotion is also a challenge (ter Stal et al. 2020). In the industrial domain, affective and relational affordances promoting user engagement are recognised as the key features (Syvänen and Valentini 2020). Empathy and anthropomorphism are important in other medical conditions, especially in working with children (Chew and Achananuparp 2022; van Straten, Peter, and Kühne 2020). The impact of anthropomorphic versus biomorphic and personality differences on a SAR’s ability to empathise is unclear and needs theory-based testing (Park and Whang 2022). Furthermore, research on the human–agent relationships based on phenomena and systematized theories identified in psychology (e.g. social inhibition and facilitation, social learning theory, social exchange theory, and the persona effect) will help identify the conditions under which CAs build and maintain productive relationships with their users and would be able to clarify (Scholten, Kelders, and Van Gemert-Pijnen 2017).

Conversationally competent CAs can elicit trust from human users, a crucial factor for their successful implementation in elderly support programmes (El Kamali et al. 2020). In industrial settings, chatbots have demonstrated their potential to enhance customer satisfaction and trust by exhibiting human-like behaviours (Jenneboer, Herrando, and Constantinides 2022). Future studies should explore the influence of emotional expressions on building trust and their interplay with situational factors such as importance and urgency of decisions (Song and Luximon 2020). Another research area pertains to the differential impact of robots versus agents on human trust, with the physicality of robots being a well-established determinant (Lambert et al. 2020). However, SARs encounter the obstacle of being highly specialised and potentially limited in their ability to provide emotional or cognitive support (Park and Whang 2022). A systematic examination of theories regarding human-CA trust, including agent personas, persona matching, and conversational functionality, is also an area ripe for exploration (Davis, Park, and Vincent 2021; Liew and Tan 2021; Song and Luximon 2020).

The Chat Generative Pre-Trained Transformer (ChatGPT) developed by OpenAI has garnered significant interest due to its advanced communication abilities. Beyond producing human-like responses in English, ChatGPT also performs proficiently in other languages, such as Japanese, without notable degradation in comprehension or output quality. This systematic review encompasses its applications in neurodevelopmental disorders (Maddalon et al. 2023), other medical domains (Alhaidry et al. 2023; Mirbabaie, Stieglitz, and Frick 2021), and educational fields (Sallam 2023). While ChatGPT is a promising technology, it faces challenges such as lack of specialised knowledge, outdated information, output accuracy, and model uncertainties (Chow, Sanders, and Li 2023b). Nonetheless, the evolution of large language models (LLMs), such as GPT, holds substantial promise for enhancing the conversational capabilities of computer assistants in the future. While autonomous psychotherapy delivery by CA therapists, as opposed to relying solely on human assistance, demands intricate processing of conversational tasks (Grodniewicz and Hohol 2023), the advancement of LLMs may pave the way for achieving this objective.

4.5. Suggestions for future research and social implementation

As described above, CA research has been increasing in recent years, with ECA, chatbot, and SAR research progressing in each domain and accumulating knowledge on ethics, effectiveness, acceptability, and conversational functioning. The research challenges are summarised here. First, high-quality comparative studies are lacking in many areas. The verification of effectiveness is particularly needed for loneliness, other psychiatric symptoms, positive aspects (e.g. QOL), and wellbeing, for which effectiveness has not yet been clarified. More research on mental health, especially in a non-clinical group of young adults, and on neurodevelopmental disorders compared to research on older adults would help correct the gap between the domains. Second, in conducting comparative studies, standardised assessment methods for the key outcomes have not yet been established in several domains. Third, verbal and nonverbal communication skills that can meet the expectations of a companion robot and an entertainment agent must be acquired. The required communication skills include natural responses in open dialogues, which enhance emotion recognition and expression, empathy, and other behaviours for building and maintaining relationships. The ability to interact via voice and text must be promoted. The fourth challenge is to promote acceptability. Further improving UX and establishing theory-based methods of personalisation are near-term issues. Fifth is the challenge of integration with other new technologies. The integration of chatbots and SARs with patient monitoring, real-time evaluation, and VR and AR technologies is required to create new assistive technologies and user experiences. The final challenge is the development of an interdisciplinary approach that includes service management and communication sciences. Until now, interdisciplinary efforts are being made in engineering, medicine, social work, psychology, and law. The incorporation of disciplines in the areas of practice and service is expected to broaden the scope of application, especially for industrial applications.

The implications for social implementation that can be drawn from this review are presented here. First is the formulation of ethical guidelines for development and implementation. In this review, we proposed nine elements that should be included in the ethical guidelines based on the previous studies. Based on these elements, clarifying the ethical guidelines that should be addressed by the industry will help ensure that new technologies will be used in a safer manner and to enrich the society. Second, a standard CA model for each domain would be established. The establishment of a standard model will give users an indication of how much they should expect from the technology. Third is the creation of standard implementation guidelines for each domain that should provide guidance on the effective implementation of CA in the field, including how to choose a CA, the maintenance required, and the necessary management structure. Fourth, training materials for professionals who use or collaborate with CAs, such as medical professionals, caregivers, and teachers, should be created. Some fields (e.g. nursing education) already provide training for using the latest technology. This type of training will be required in any professional training programme. The development of such materials will help eliminate prejudice, excessive expectations, and misconceptions about CA use. Finally, explanatory materials for users and their families must be prepared.

4.6. Limitations

This meta-review provided a comprehensive and informative understanding of the current status and challenges of CA research, but it has limitations. First, this review only included systematic review articles. Although systematic reviews aggregated the findings of each domain, and the large number of systematic reviews partly reflected the progress level of research in each domain, some aspects may not have been captured because we did not look at primary research. For example, we did not focus on kindergarten initiatives; however, research on storytelling with SARs in kindergarten is recognised at the primary research level (Tavernise and Bertacchini 2016). Second,, we did not address in-depth research on ML and NLP in related areas because we focused on CA. However, the development of basic technologies, such as the diagnosis and detection of mental disorders (Thieme, Belgrave, and Doherty 2020) and the analysis of dialogue in psychotherapy (Aafjes-van Doorn et al. 2021), is an important and indispensable foundation for CA research. Finally, the quality of conference proceedings differs across disciplines. By excluding them from this review, we may have overlooked valuable articles published within such proceedings.

5. Conclusion

This study presented a meta-review of systematic reviews on CA to identify the research findings and challenges in each major research area. Overall, we found a gradual increase in the reporting of systematic reviews from 2018 to 2019 and 2020 to 2021, with medical care, elderly support, dementia/cognitive impairment, health promotion, and education as the most frequently reported areas. By contrast, fewer systematic reviews were reported on neurodevelopmental disorders and mental health. The use of ECAs, chatbots, and SARs is differentiated among the domains. Despite the disparities among various domains, the discourse on ethics, in its entirety, is advancing. The nine aspects of privacy, safety, innovation, user acceptance, psychological attachment, care philosophy, evaluation, compatibility with social systems, and development of rules must be further discussed, and the guidelines must be compiled. Research is still in its early stages, and although meta-analyses have been conducted in some domains, high-quality comparative studies are still lacking. Efficacy studies that consider acceptability are needed in areas related to the elderly. CAs now particularly need to improve the quality of conversational functioning. With technological development, CAs can accurately understand the user’s condition and generate personalised responses in a human-like manner. A multifunctional CA is required to serve as a companion and an entertainer based on basic communicative competence, recognition and expression of emotions, and trust built on empathy and relationship building while realising care, education, information provision, and information gathering functions. Accordingly, a proposal for research and social implementation was made based on these findings.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Available by contacting the corresponding author.

Additional information

Funding

This work was supported by the Ritsumeikan Global Innovation Research Organization (R-GIRO) of Ritsumeikan University, Japan.