Demystifying the impact of artificial intelligence word-of-mouth (AI-WOM) and knowledge management on social media engagement: comparing Vietnam and China

Over the past few years, the features of artificial intelligence (AI) have seen amazing breakthroughs in managing increasingly complicated tasks; thus, they have significantly reshaped the way consumers interact in the digital space (Moussawi et al., 2022). As a dominant channel for communication and information exchange, social media enables users to engage in knowledge-management and community-building activities (Santini et al., 2020). AI word-of-mouth (AI WOM) has evolved into a key driver of transformation in digital communities, shaping consumer perceptions and fostering engagement (Tassiello et al., 2024). It is explained as oral communication between customers and technology in a parasocial setting, with comments resulting from AI development (Tassiello et al., 2024). Social media platforms leveraging AI-generated content provide unique opportunities for personalized interaction and real-time knowledge exchange (Pataranutaporn et al., 2021). By exploring AI capabilities for generating and distributing content, AI WOM drives consumer involvement and strengthens online communities. Notably, Meta AI is a new chatbot available in the chatbox and search bar on Facebook and Instagram that is designed to learn, create content and connect while interacting with users, contributing to AI WOM by improving user engagement and interaction (Ward, 2024). As AI continues to reshape social media landscapes, the urgency to explore its impact on consumer engagement becomes increasingly evident. According to the explanation presented above, the importance of introducing AI WOM into research is growing, as existing academic research on how consumers interact with this aspect is still in their early phases.

AI has significantly transformed social media interactions, with chatbots playing a crucial role in enhancing customer engagement (Hari et al., 2021; Gołąb-Andrzejak, 2022; Bakkouri et al., 2022). Amidst the increasing reliance on AI-driven consumer interactions, research on AI-based chatbots and their impact on consumer engagement remains underdeveloped. Traditional WOM has been extensively analyzed in marketing literature (Engel et al., 1969; Mason, 2008; Bartschat et al., 2021; East, 2023). However, AI WOM, characterized by its automated and data-driven nature, presents unique dynamics that warrant dedicated investigation. This new form of WOM activates distinct AI cognitive traits, including perceived animacy, perceived intelligence, anthropomorphism, algorithm interpretability, perceived unbiasedness and perceived AI expertise. At the same time, social platforms increasingly function as knowledge management (KM) systems, where consumers continuously acquire, share and apply information (Al-Sharafi et al., 2022). The role of knowledge acquisition, sharing and application has become important in explaining how value is cocreated through interactions with AI-powered agents (Chandra and Rahman, 2023). Understanding how AI-WOM stimulates these knowledge processes is essential because AI-generated messages can accelerate learning, guide decision-making and encourage users to contribute knowledge to the community.

These factors ultimately influence consumer social media engagement (CSME), which has evolved from simple content consumption to more complex behaviors such as learning, socializing, sharing and codeveloping. Modern users interact not only with other humans but also with AI agents; therefore, the determinants of engagement are now more multidimensional and technologically driven (Hlee et al., 2022). This change indicates that engagement on social media can no longer be explained only by human-to-human communication, as AI now plays an active role in shaping how users process information and participate online. Clarifying how AI-WOM and knowledge processes drive these different engagement forms fills a critical theoretical gap.

While prior studies have explored AI WOM and algorithmic persuasion in influencing consumer decision-making, most have primarily focused on AI-generated recommendations and automated decision-making rather than the dynamic and interactive nature of chatbot-facilitated engagement (Mogaji and Jain, 2024; Zarouali et al., 2022). Existing research has examined AI credibility, trust and transparency (Lee, 2018; Grimmelikhuijsen, 2022), but it has not comprehensively addressed how specific chatbot attributes – such as perceived animacy, intelligence, anthropomorphism, algorithmic interpretability and unbiasedness – shape consumer engagement and knowledge-management behaviors. For businesses aiming to leverage AI-driven conversations effectively, recognizing how AI WOM impacts customer interaction on social platforms is of great importance. This gap is critical as AI chatbots are not merely recommendation tools but interactive agents that adapt in real time, fostering deeper consumer involvement (Balakrishnan and Dwivedi, 2021). Furthermore, KM theories, while widely applied in organizational settings (Cabrera and Cabrera, 2005; King, 2009), have not been sufficiently integrated into AI-driven consumer engagement research. Prior studies have yet to examine how AI WOM, facilitated through chatbots, enhances consumer learning, social interactions and content cocreation (Adam et al., 2020; Kot and Leszczyński, 2022). Unlike traditional electronic word-of-mouth (eWOM), AI WOM introduces complexities in how consumers perceive AI expertise, algorithmic transparency and decision-making autonomy, influencing their involvement with AI-generated insights (Hong et al., 2023). Filling this gap is important to advance both academic understanding and real-world application as our study integrates consumer psychology with AI engagement models, offering a framework to understand how chatbot interactions influence KM and trust. Nguyen et al. (2021) pointed out the necessity for further exploration in future studies to uncover additional dimensions of AI quality to better understand its impact on customer experiences and engagement. Furthermore, cultural differences significantly affect consumer responses to AI-driven interactions. Vietnam and China, despite sharing certain cultural similarities, exhibit distinct attitudes toward technology adoption and digital communication. Vietnam’s younger generation is rapidly integrating AI-driven applications into daily life, while China has a more mature and regulated AI ecosystem with widespread chatbot utilization in various sectors (Liu et al., 2023). However, the extent to which AI WOM impacts consumer engagement in these two markets remains underexplored. This research tackles the existing gaps by investigating how AI chatbots on social media enhance consumer engagement among Vietnamese and Chinese undergraduate students. While prior studies have largely relied on single-country samples, this research provides a comparative perspective, shedding light on how cultural factors influence consumer responses to AI-driven interactions.

This study addresses critical gaps by identifying key drivers of AI WOM and analyzing their effects on customer engagement, thereby providing both theoretical insights and practical implications for optimizing AI-powered interactions.

This research strives to provide an original perspective by assessing how AI WOM affects customer engagement in the context of social media, specifically in Vietnam and China. Our objectives are threefold:

1. to assess how key chatbot attributes – perceived animacy, intelligence, anthropomorphism, algorithm interpretability, unbiasedness and AI expertise – influence AI WOM;

2. to delineate the impact pathway of AI WOM by analyzing its effect on three core KM processes: knowledge acquisition, knowledge sharing and knowledge application, and determine the role of these knowledge processes in shaping consumer engagement on social media; and

3. to conduct a cross-country comparison to uncover how cultural and technological factors moderate these relationships.

By integrating the algorithmic persuasion framework (APF) with KM theories, this research advances the theoretical discourse on AI-driven consumer interactions while offering actionable insights for marketers, AI developers and business strategists seeking to optimize AI-powered engagement strategies. The comparative analysis of Vietnam and China reveals critical nuances; for instance, the stronger effect of algorithm interpretability in China contrasts with the greater importance of perceived animacy in Vietnam. These findings challenge the universal applicability of single-market AI engagement models and provide a framework for understanding AI adoption across diverse technological and cultural ecosystems.

This study is grounded in the APF developed by Zarouali et al. (2022), The APF explains how algorithm-mediated communication influences user behavior through a circular process with five components: input, algorithm, persuasion attempt, persuasion process and persuasion effects. These components form a feedback loop in which user responses generate new data that in turn shape subsequent algorithmic decisions (Zarouali et al., 2022; Julier, 2017; Vasudevan, 2020). By integrating traditional persuasion theories with personalization models such as the personalization process model of Vesanen and Raulas (2006), the APF highlights the cyclical and data-driven nature of digital persuasion. This makes it suitable for examining the interaction between AI WOM and consumer behavior in digital environments. Prior studies have applied the APF to audience coping behavior in targeted advertising (Chen et al., 2024) and to consumer responses to data-driven advertising (Noort et al., 2024), yet it has not been used to study consumer perceptions of AI digital assistants.

In this research, perceptions of AI chatbots including perceived animacy, perceived intelligence, perceived anthropomorphism, perceived algorithm interpretability, perceived unbiasedness and perceived AI expertise are treated as inputs that influence the persuasive attempts of AI chatbots and shape AI WOM. AI WOM in turn is expected to affect consumers social media engagement and individual knowledge management processes knowledge sharing, knowledge acquisition and knowledge application. The iterative structure of the APF reflects how user perceptions of AI chatbots, for example, perceived intelligence or anthropomorphism affect engagement with AI-generated content, which then becomes new input for future algorithmic outputs (Zarouali et al., 2022).

To closely examine people’s exposure to AI-based chatbots for their well-being and engagement, the theory of KM is applied to our research. KM is “an essential field for practice and research in information systems from information economics, strategic management, organizational culture” to AI (Baskerville and Dulipovici, 2015). The application of KM to our research was in light of two main reasons. First, KM has emerged as important to technological development (Bell DeTienne and Jackson, 2001; Maria Mårtensson, 2000). Similarly, Blair (2002) described “knowledge management” as a strategy developed to manage the fast-paced expansion of information. Second, KM has not only been applied to business organizations and human resource managers (Benson, 1977), but has also affected people’s lives (Maria Mårtensson, 2000). Because KM can be used for individuals, we used three factors: knowledge sharing, knowledge acquisition and knowledge application. While knowledge acquisition refers to “the processes that use existing knowledge and capture new knowledge” (Lee and Choi, 2007) and knowledge sharing refers to “the process of disseminating various resources among the individuals taking part in specific activities” (Al-Emran and Teo, 2019), knowledge application refers to “the process that enables the individuals to access the knowledge smoothly via the existing efficient storage and retrieval techniques” (Arpaci et al., 2020).

AI WOM is a distinct phenomenon that takes place in a hybrid, parasocial environment, where consumers perceive technology as having a human-like social presence. While eWOM involves human-generated opinions shared asynchronously through digital platforms (Hennig-Thurau et al., 2004; Ismagilova et al., 2020; Petrescu et al., 2023), AI WOM operates in real-time, triggered by consumer requests or contextual cues and delivers personalized recommendations through advanced AI and machine learning algorithms (Tassiello et al., 2024). Unlike eWOM, which is grounded in human authenticity, AIWOM reflects algorithmic agency and socially persuasive, bidirectional exchanges that evoke perceptions of social presence and emotional responsiveness (Huang and Rust, 2021; Hennighausen et al., 2025). While neither Huang and Rust (2021) nor Hennighausen et al. (2025) explicitly use the term AIWOM, their conceptualizations of feeling AI and AI-mediated communication jointly illustrate how AI can act as a social and emotional communicator in digital service interactions. Building upon these foundations, this study defines AIWOM as the communicative effect of real time, AI-mediated interactions between users and chatbots integrated into social media or e-commerce platforms.

Social media has become central to modern life and shapes communication, entertainment and consumption (Habibi et al., 2014; Strauss and Frost, 2016; Zaglia, 2013). Firms use social media strategically to influence preferences and create long-term relationships (Li et al., 2021). Consumer engagement, which includes transactional and nontransactional interactions affects brand performance through interactive communication and cocreated experiences (Cambra-Fierro et al., 2013; Garg et al., 2020; Jaakkola and Alexander, 2014).

Several multidimensional models of social media engagement have been proposed. Abdullah and Siraj (2018) identify seven subprocesses: learning, socializing, sharing, cocreating, advocating, criticizing and reviling. This study adopts four key dimensions from the online brand community engagement model proposed by Brodie et al. (2013): codeveloping, sharing, socializing and learning. Codeveloping refers to participation in improving offerings through idea generation, codesign or collaborative production (Vargo and Lusch, 2007; Jaakkola and Alexander, 2014; Bendapudi and Leone, 2003; Vargo and Lusch, 2007). Sharing captures the exchange of knowledge, experiences and content that influences others’ perceptions (Abdullah and Siraj, 2018; Brodie et al., 2013; Jaakkola and Alexander, 2014). Socializing refers to interactive exchanges where consumers develop shared norms, attitudes and language with peers or brands (Longmore, 1998; Brodie et al., 2013; Chu and Sung, 2015). Learning involves the acquisition of knowledge and skills that inform purchasing and consumption decisions and deepen understanding of the market (Brodie et al., 2013; Abdullah and Siraj, 2018). These four dimensions are used to operationalize consumers social media engagement.

Animacy gives objects a lifelike quality, primarily connected to their movement and visual texture, rather than resembling human form (Gao et al., 2019; Heider and Simmel, 1944). Animacy can be seen as a crucial element in recognizing AI chatbots as social entities that foster users’ sense of social presence (Jin and Youn, 2022). In this paper, perceived animacy refers to how lifelike users perceive AI, based on traits such as responsiveness, intentionality and social presence. While research on animacy has traditionally been rooted in psychology, its application has increasingly shifted toward information systems, robots and AI studies in recent years (Bartneck et al., 2009). The previous research has proven that AI systems designed with human-like voices and behaviors increase social presence, making interactions more engaging and enjoyable (Nass and Gong, 2000; Nass and Lee, 2001; Nielsen et al., 2015). This heightened engagement often results in users sharing their positive experiences, effectively generating AI WOM. Therefore, we propose:

Perceived intelligence refers to how autonomous and adaptive an AI system appears (Moussawi et al., 2022). It has become a key identity of AI-powered systems (Ogiela and Ko, 2018; Russell and Norvig, 2010) and is expected to enhance intelligence through algorithms and data predictions (Akter et al., 2019). Recent research highlights perceived intelligence as a unique trait of AI assistants, shaping positive user attitudes (Zhang et al., 2024). Critical elements such as knowledge, sensibility and responsibility play a significant role in defining intelligence (Bartneck et al., 2009; Kiesler et al., 2008; Parise et al., 1996). However, the relationship between AI assistants and AI WOM remains underexplored. AI WOM provides personalized recommendations but is still a novel concept lacking in-depth study (Tassiello et al., 2024). Given that perceived intelligence fosters familiarity and trust, we propose:

Perceived anthropomorphism relates to how users perceive nonhuman agents, such as AI or digital assistants, as possessing humanlike traits (Qiu and Benbasat, 2009; Moussawi and Koufaris, 2019). This concept involves assigning human traits, such as emotions, cognitive abilities and social behaviors, to AI systems, which can enhance user trust and emotional connection (Mourey et al., 2017; Touré-Tillery and McGill, 2015).

Prior studies have extensively explored anthropomorphism in robotics and marketing, highlighting its role in fostering positive user attitudes and behaviors (Bartneck et al., 2009; Hudson et al., 2016). Research indicates that, within the context of AI-driven digital assistants, perceived anthropomorphism significantly affects positive emotions, trust and adoption intentions (Maduku et al., 2024; Yen and Chiang, 2020; Pillai and Sivathanu, 2020). Within the APF, perceived anthropomorphism acts as an input that influences AI’s persuasive attempts, shaping AI WOM outcomes. However, there remains a gap in exploring how this input affects AI WOM within APF’s iterative feedback loop. Specifically, can humanlike traits in AI systems, which promote user interaction, also foster more frequent and organic WOM behaviors? From the light of the APF and prior findings, the following hypothesis is proposed:

According to Doshi-Velez and Kim (2018), algorithm interpretability refers to the capacity to offer explanations in human-interpretable terms. Conversely, interpretability relates to the level at which users can comprehend the outcome of a decision (Kim et al., 2020) or the ability to understand and conclude about the model output (Chakraborty et al., 2017). Applications of algorithmic interpretability vary depending on the domain. For instance, AI systems in finance can be more reliable by improving their interpretability (Hong et al., 2023). In addition, doctors can better determine problems and assess a model’s credibility if it is capable of understanding its output in the medical field (Hakkoum et al., 2022). As Shin (2020) and Chen (2024) have shown that the more that customers make algorithmic systems more comprehensible, the more likely they will be inclined to appreciate its value and will be motivated to transmit WOM about their interaction. Furthermore, this influences AI WOM behavior and increases the community’s acceptance and popularity of AI systems (Tassiello et al., 2024). Instead of being baffled by intricate algorithms, they boldly communicate their views, influencing public opinion and accelerating the use of AI (Lee, 2018). Thus, perceived algorithm interpretability not only improves user understanding but also amplifies AI WOM by motivating users to communicate their experiences. For these reasons, the authors propose hypotheses:

Bias in AI chatbots encompasses various aspects, including ethical, cultural and commercial considerations (Ray, 2023). In this paper, the authors define perceived unbiasedness as the belief that an AI chatbot’s recommendations or information are free from commercial bias and solely serve the user’s benefit. In previous research, perception of AI credibility is considered as having an impact on the outcomes of the conversations, such as trust in content (Zhou and Lu, 2024; Yen and Chiang, 2020) or intention to follow AI’s recommendation (Khan and Mishra, 2023). When an AI chatbot is perceived as biased, customers may question the information provided and lose trust, leading to decreased support, intention to follow recommendations or share WOM (Dietvorst et al., 2015; Cheng et al., 2021; Chau et al., 2013; Verma et al., 2023). Conversely, when the AI chatbot is viewed as a fair and objective source, consumers are more likely to trust and disseminate information, regardless of the product type (Davenport et al., 2020; Lee et al., 2010). Therefore, the authors propose:

Perception of AI expertise is based on how users perceive the technology’s competence, knowledge and overall capabilities Manser Payne and O’Brien (2024). AI-powered systems demonstrate expertise through data processing, predictive analytics and problem-solving capabilities, enhancing user trust and engagement (Bendig and Bräunche, 2024). Perceived expertise plays a role in influencing risk, trust and perceived usefulness, ultimately contributing to the intention to use (Ramrath et al., 2024, p. 275).

AI WOM, which involves AI-generated recommendations and insights, is an emerging phenomenon influencing consumer decision-making. Although AI WOM can improve trust and decision confidence, its credibility remains an open question (Tassiello et al., 2024). Given that perceived AI expertise enhances reliability, it is likely to strengthen AIWOM’s persuasiveness and effectiveness. Based on this rationale, the authors propose:

Although these processes may evolve dynamically in real-world AI-mediated environments, the present study models AIWOM as an upstream informational and communicative stimulus that precedes knowledge-related processes. This ordering is theoretically grounded in two streams of literature. First, recent work conceptualizes AIWOM as a distinct communication phenomenon emerging from interactions between consumers and AI-enabled systems, suggesting that AI-generated recommendations and related communication can serve as externally available informational inputs for subsequent user responses (Tassiello et al., 2024). Second, knowledge and information processing research suggests that exposure to externally available information typically precedes downstream cognitive activities such as acquisition, interpretation and use, rather than following them (Alavi and Leidner, 2001; Eveland, 2001). On this basis, the present study treats AIWOM as an antecedent to knowledge acquisition and subsequent downstream outcomes.

At the same time, this study acknowledges that alternative causal structures are plausible. In AI-mediated environments, users who acquire or apply knowledge may subsequently become more likely to rely on, reproduce or reinforce AI-related WOM through repeated interactions with algorithmic systems. Recent research on human AI feedback loops suggests that interaction with AI can recursively shape subsequent judgments and behaviors, indicating that reciprocal reinforcement may emerge over time rather than in a single linear sequence (Sun et al., 2023; Glickman and Sharot, 2024). Nevertheless, the current ordering is retained because it is more consistent with the present study’s cross-sectional design, the theoretical role assigned to AIWOM as an external communication input and the parsimonious objective of isolating how AI-related perceptions are translated into later knowledge and engagement outcomes.

AI WOM provides personalized, data-driven feedback that supports knowledge acquisition, causal understanding and decision-making (Zhai et al., 2021). Tailored recommendations from AI systems can increase confidence in purchase decisions and reinforce cognitive engagement (Grewal et al., 2020). AI WOM also encourages sharing of experiences, insights and AI-generated content within digital communities (Hennig-Thurau et al., 2004) and promotes socializing by stimulating conversations and strengthening community ties (Brodie et al., 2013; Shahzad et al., 2024. Furthermore, AI WOM can support codeveloping when users provide feedback that helps firms refine products and services (Brodie et al., 2013). Although these research are highly relevant, the role of AI WOM in customers’ social media engagement remains underexplored. With the relations to the four dimensions of customers’ social media engagement of AI WOM, the author proposed:

AI WOM can be understood as the spread of AI-based suggestions, reviews and collective feedback that influence attitudes and behavior (Zhang et al., 2022). As AI becomes more embedded in daily interactions, AI WOM becomes a key source shaping user knowledge. Compared with offline WOM that depends on human opinions, AI WOM relies on data-based knowledge and large-scale information aggregation to provide relevant information at appropriate times (Pihlaja et al., 2017). Prior research shows that WOM affects attitudes, trust and purchase intention (Lin and Lu, 2010; Agag and El-Masry, 2016; Li and Jaharuddin, 2021; Qi and Kuik, 2022). Knowledge acquisition is a central outcome in AI-mediated interactions, because users absorb, process and extend knowledge through contact with AI (Jo, 2023). AI learning systems and personalized recommendation tools enhance user effectiveness, decision-making and perceived usefulness by supporting knowledge acquisition (Jo and Park, 2023; Jo and Park, 2024; Belle, 2023). The authors proposed the hypothesis:

Knowledge sharing is a vital process that enhances organizational capabilities, particularly by strengthening absorptive capacity (Lee et al., 2016). Research suggests that AI WOM can serve as a form of knowledge dissemination, enabling users to share their experiences and insights regarding AI functionalities, performance and best practices (Tassiello et al., 2024). Prior studies on eWOM indicate that online discussions and recommendations contribute to collective knowledge sharing, as they provide valuable information that aids decision-making (Cheung and Thadani, 2012). Similarly, AI WOM can facilitate knowledge-sharing behaviors by motivating users to discuss AI-related experiences and provide solutions to AI-related challenges (Freire, 2023). Furthermore, social exchange theory, as introduced by Blau (2017), posits that individuals participate in knowledge sharing when they perceive reciprocal benefits, such as gaining valuable AI insights from others. Given this theoretical foundation, we propose the following hypothesis:

Knowledge application refers to the process through which individuals and organizations can effectively access and use knowledge through efficient storage and retrieval systems (Alavi and Leidner, 2001). It functions as a core element of KM, where individuals or organizations translate information into actionable insights. AI WOM significantly contributes to enabling knowledge application by providing users with tailored, data-driven recommendations that enhance their ability to act on knowledge effectively (Duan et al., 2019). When AI WOM is explainable and transparent, users perceive the recommendations as more reliable and credible, thereby increasing their confidence in applying the knowledge obtained from these AI-driven suggestions (Chen et al., 2023). However, knowledge application is not solely dependent on information availability but also on users’ ability to process and integrate AI WOM into their cognitive frameworks. Cognitive load theory (Sweller, 1994) suggests that an overload of AI-generated recommendations may hinder knowledge application by overwhelming users with excessive or conflicting information. Thus, we propose the following hypothesis:

Knowledge application is described as the technique of using obtained knowledge in making choices, tackling issues or performing duties (Alavi and Leidner, 2001). McLean and Osei-Frimpong (2019) stated that “users who can apply knowledge gleaned from an AI system would perceive it as more beneficial, enhancing the utility derived from its use.” Effective knowledge applications can lead to higher user satisfaction (Nguyen and Malik, 2021; Wang et al., 2021; Malik et al., 2021). In addition, people who put their knowledge into practice also frequently actively share their experiences, engage with the social and cocreate material (He et al., 2019). Similarly, applying knowledge can help consumers feel more confident in discussions, thereby promoting sustainable engagement on digital interfaces (Hollebeek and Macky, 2019). Nevertheless, there is no research that has looked into how knowledge application affects customer involvement on social media. The question of whether knowledge application can greatly improve customer engagement on social media platforms is brought up by the fact that involvement includes learning, sharing, socializing and codeveloping. For these reasons, the authors propose:

Knowledge acquisitions have been proven to positively and significantly affect customer involvement in various activities of the company (Dahiyat et al., 2012). Previous research also highlights that customers must obtain relevant skills and knowledge to engage effectively with brands (Hollebeek et al., 2019; Hibbert et al., 2012). Behnam et al. (2021) further suggest that knowledge acquisition deepens customers’ understanding of a brand, enhancing their ability to participate in brand interactions. This indicates that learning, as a core dimension of engagement, is significantly influenced by the extent to which customers actively seek and acquire knowledge. After gaining valuable knowledge, consumers may be more inclined to share information with others (Liu and Liu, 2008). While previous studies suggest that this willingness can vary across different contexts and subjects, this research focuses on examining the direct interaction between users and AI chatbots implemented on social media platforms. For instance, when users interact with AI-driven chatbots to learn about a product or service, they may feel more confident in discussing their insights with peers, contributing to knowledge exchange within online communities. This aligns with findings that acquiring knowledge plays a crucial role in successful socialization or codevelopment (Ostroff and Kozlowski, 2006; Sjödin, 2018). Therefore, the authors propose:

Knowledge sharing is known to be vital in fostering learning by enhancing sustainable use of technologies like chatbots, and fosters socializing by creating a collaborative climate that encourages frequent interaction and reciprocity (Al-Sharafi et al., 2022; Feiz et al., 2017; Lin et al., 2009). Similarly, knowledge sharing is a fundamental aspect of codeveloping, as individuals “assist the organization by sharing ideas, experience and information for augmenting the service offering” (Helander et al., 2024). In addition, when individuals share knowledge through AI chatbots, they are likely to extend this behavior to sharing information, experiences and opinions with others in real-world interactions. However, despite these established relationships, no study has explored how knowledge sharing affects consumer engagement in the context of social media. Given that engagement encompasses learning, sharing, socializing and codeveloping, it raises the question: Can knowledge sharing significantly enhance consumer engagement on social media platforms? In accordance with this reasoning, the authors propose:

A cross-sectional online survey was used as the primary data collection method. The survey instrument was developed based on validated scales from prior literature. A pretest with 20 participants was performed to examine and confirm the comprehensibility of the questions. All scales were initially created in English, with certain items modified for better understanding. We then adapted and translated the scales into Vietnamese and Chinese using a translation and back-translation procedure. A total of 892 participants from Vietnam and China made up the final sample, collected through an online survey using nonprobability sampling, with snowball sampling with screening-based eligibility criteria as the specific method used. This method was deemed appropriate to access populations of social media users who have experience with AI chatbots. The survey was distributed through social media channels (e.g. Facebook and WeChat) and university student networks in both countries to initiate the sampling chains.

To mitigate the risk of overrepresentation from a single group, the initial “seed” participants were deliberately selected to represent a range of ages and educational backgrounds. Participants were encouraged to share the survey within their networks, but no specific quotas or incentives for chain-referral were used, which helped maintain organic dissemination. Two screening questions were used to identify target respondents: “Have you ever used AI chatbots on social media (i.e. Meta AI on Facebook and Instagram)?” and “Based on the explanation above, would you be interested in continuing to use AI chatbots on social media or trying them out if you had the chance?” Participants who answered “Yes” to all questions proceeded with the survey, which took approximately 20 min. Sample characteristics and data screening were conducted using SMARTPLS 4 software for exploratory factor analysis. Following the removal of extreme values, data from 761 respondents were retained to test the proposed model (see Figure 1), leading to an effective rate of 85.31%. The sample was balanced by gender, with 370 females (48.6%) and 391 males (51.4%). Respondents’ ages ranged from below 18 to above 45 years. Specifically, 5% were under 18, 38% were 18–25, 25% were 26–30, 18.7% were 31–35, 7.5% were 36–40, 3.5% were 40–45 and 2.4% were over 45 years old. Regarding education profile, 16.8% were high school students or below, 20.1% took college/vocational training, 50.3% possessed a bachelor’s degree, 11.2% held a master’s degree and 1.6% attained a doctoral degree. Finally, in terms of income, the highest proportion is 53.6%, which is the percentage of respondents with income under US$500, followed by 23.1% with income from US$500 to US$1,000 per month, 10.9% with income from US$1,001 to US$1,500 per month, 7.5% with income from US$1,501 to US$2,000 per month, 3.4% with income from US$2,001 to US$2,500 per month, 0.8% with income from US$2,501 to US$3,000 per month and lastly, 0.7% earning over US$3,000 per month (see Table 1).

To validate the theoretical model (see Figure 1), measurement scales were modified from previous studies on perceived animacy; perceived intelligence; perceived anthropomorphism; perceived algorithm interpretability; perceived unbiasedness; perceived AI expertise; codeveloping, sharing, socializing and learning as four dimensions of consumer’s social media engagement; AI WOM; knowledge application; knowledge acquisition; and knowledge sharing. All items were modified to fit the research background. Afterward, the authors translated items from English to Vietnamese and Chinese to ensure respondents’ understandability (Harkness et al., 2004). A five-point Likert scale was used for participants’ responses, offering options from “strongly disagree” to “strongly agree.”

Perceived animacy, perceived intelligence and perceived anthropomorphism were measured using six, five and five items, respectively, all derived from Bartneck et al. (2009). Perceived algorithm interpretability was measured with three items from Köhler et al. (2011), while perceived unbiasedness was measured with three items from Benbasat and Wang (2005). Perceived AI expertise was measured with four items from Manser Payne and O’Brien (2024). CSME included four dimensions: Codeveloping, sharing, socializing and learning. Codeveloping, sharing and socializing were each assessed using three items, while learning was evaluated with four items, all adapted from Brodie et al. (2013) and Lim and Jiang (2020). AI WOM was measured with four items based on Le et al. (2024). The scale for knowledge acquisition was measured with five items from Al-Emran and Teo (2019). Knowledge sharing was also gauged with five items, sourced from Al-Emran and Teo (2019) and Arpaci et al. (2020). Knowledge application was measured with four items from Arpaci et al. (2020). Full items are included in the  Appendix.

The partial least squares structural equation modeling (PLS-SEM) approach, implemented using SmartPLS4 software (Ringle et al., 2023), was used to evaluate the theoretical framework (see Figure 1) and hypotheses proposed in this study. PLS-SEM is a robust analytical technique that offers a balance between explanation and prediction (Lim et al., 2023), making it particularly suitable for this research. PLS-SEM, in contrast to conventional techniques, does not depend on rigid assumptions about data distribution, making it possible to estimate complex models involving multiple indicators, constructs and structural pathways (Hair et al., 2019). This effectiveness extends to its ability to address complex research designs that include moderators and reflective-formative higher-order constructs (Becker et al., 2022; Cheah et al., 2023). To assess the overall model fit, the bootstrapping method was used, which evaluates R² values and the significance of relationships among constructs, in line with the guidelines provided by Dijkstra and Henseler (2015). Given the cross-country nature of this research, a multigroup analysis (MGA) was also conducted to compare the structural relationships between the Vietnamese and Chinese samples. After considering these advantages, the proposed model was tested using SmartPLS4 software, providing a comprehensive evaluation of the research framework (see Figure 1).

The correlations among variables from a shared data collection method impose challenges for subsequent data analysis (Podsakoff et al., 2012). Potential common method bias (CMB) could arise in our research because data were gathered at a single time point. While procedural controls are conducted prior to data collection (ex ante), statistical controls are applied afterward (ex post). Procedural controls eliminate the risk of CMB by requiring a scrupulous questionnaire design. Such design is obtained by assistance of many researchers: Adaptive questionnaires to respondents are recommended to draw their attention throughout the test (Kmetty and Stefkovicsb, 2021); clear and straightforward questions are commonly used, indicating respondents’ exposure to AI chatbot (Podsakoff et al., 2012). In addition, questions were designed to motivate respondents through concise and straightforward language, indicating the respondents’s importance to the question (Podsakoff et al., 2012). Finally, breaks between groups of questions are designed with a view to helping respondents avoid “straight-line response” (Herzog and Bachman, 1981) which implies a situation when respondents start to answer questions with little difference. Next, the Harman’s single-factor test was conducted to assess the presence of CMB in our research. The test results showed that a single component explained 38.275% of the variance. CMB may become a concern if a single factor accounts for more than 50% of the variance (Fuller et al., 2016). Because the first unrotated factor in our study accounted for less than this critical threshold, CMB was not considered to be a major concern. Furthermore, the full collinearity test (Kock and Lynn, 2024) revealed that the variance inflation factor (VIF) scores ranged from 1.289 to 2.361, all of which were under the 3.33 cutoff. Therefore, CMB did not significantly influence our study.

The findings (refer to Table 2) demonstrate that the outer loadings for all items exceed the recommended threshold of 0.5 (Hair et al., 2021). Composite reliability (CR) values surpass 0.7, aligning with the 0.6 benchmark suggested by Marcoulides (1998) and the 0.7 threshold proposed by Henseler and Sarstedt (2013), Hair et al. (2021) and Bagozzi and Yi (1988). In addition, the average variance extracted (AVE) for all items is above 0.5, indicating a positive correlation among the variables (Rasoolimanesh et al., 2016; Hair et al., 2021). CR values also exceed the recommended benchmark of 0.7 (Hair et al., 2021). In this research, Cronbach’s alpha (CA) ranges from 0.700 to 0.937, further reinforcing the validity of the variables (Hair et al., 2021). To validate the model, heterotrait–monotrait ratio of correlation (HTMT) was used to evaluate the discriminant validity. As shown in Table 3, the majority of HTMT values are below the 0.9 threshold suggested by Gold et al. (2001) and Henseler et al. (2015). Moreover, the Fornell–Larcker criterion is satisfied, as the square root of each construct’s AVE is greater than its correlations with other constructs (Fornell and Larcker, 1981). Thus, confirming the discriminant validity of the latent variable measures.

The assessment of collinearity revealed that the VIF values within the outer and inner model below the critical threshold of 5.0, suggesting that collinearity issues are not significant (Hair et al., 2019). Following this, the structural model was assessed using the R-squared (R²) coefficient, which ranged from 0.358 to 0.636, signifying moderate to good explanatory power. In addition, the model displayed a good alignment with the sample data, as reflected by a normed fit index (NFI) exceeding 0.7 and a standardized root mean square residual (SRMR) of 0.052, which is smaller than the cutoff value 0.08 (Hu and Bentler, 1999), confirming the model’s robustness, as presented in Table 4.

We further analyze the correlations among latent variables (Table 5). The examination of Table 5 indicates that the most robust connection exists between knowledge sharing and knowledge application (0.740), showing that greater knowledge sharing is closely linked to an enhanced ability to apply knowledge effectively. This relationship underscores the necessity of creating an environment that supports collaborative knowledge sharing, where individuals actively exchange information and insights, leading to improved knowledge utilization and organizational performance.

A significant correlation also exists between knowledge acquisition and knowledge sharing (0.727), highlighting the strong link between sharing information and the ability to gain new knowledge effectively. This implies that individuals who engage more frequently in knowledge-sharing activities are better positioned to enhance their learning abilities and absorb new concepts. The researchers also identified a strong link between knowledge acquisition and knowledge application (0.719), highlighting how these two capabilities reinforce each other. In addition, a significant correlation is found between AI expertise and AI WOM (0.702), suggesting that the more sophisticated and capable an AI system is, the more actively it engages in online discussions with users. Furthermore, the strong link between perceived animacy and perceived intelligence (0.695) indicates that AI systems perceived as more lifelike tend to be regarded as more intelligent. These findings emphasize the interplay between knowledge utilization, AI capabilities and human perceptions in shaping interactions with intelligent systems.

To evaluate the hypotheses, t-test analysis and path coefficient p-values were used to examine the statistical significance of intervariable relationships (Table 6). To strengthen the reliability of the results, the authors applied a nonparametric approach by using the bootstrap technique to assess the accuracy of the hypotheses (Becker et al., 2022). The findings indicate that most hypotheses are supported, as reflected by t-test values exceeding 2.3 and p-values below 0.05 (Rasoolimanesh et al., 2016).

The hypothesis findings indicate that H1 and H3 were supported for Vietnamese consumers. Specifically, perceived animacy positively influenced AI WOM (H1: β = 0.187, p = 0.000), and perceived anthropomorphism also had a significant positive effect on AI WOM (H3: β = 0.217, p = 0.000). However, H2 was dismissed, as perceived intelligence did not significantly impact AI WOM (H2: β = 0.044, p > 0.05). The hypothesis findings indicate that perceived algorithm interpretability, perceived unbiasedness and perceived AI expertise both have a positive influence on AI WOM among Vietnamese consumers, supporting H4 (H4: β = 0.089, p < 0.05), H5 (H5: β = 0.161, p < 0.01) and H6 (H6: β = 0.224, p < 0.000), respectively. About the relationship between AI WOM and consumers’ social media engagement, H7 (H7: β = 0.187, p < 0.000) confirms that the two hypothesized objects have a positive mutual effect. Similarly, AI WOM significantly influences knowledge acquisition (H8: β = 0.561, p < 0.000), knowledge sharing (H9: 0.539, p < 0.000) and knowledge application (H10: β = 0.537, p < 0.000), supporting all three hypotheses. With p-values of 0.09 and 0.09, hypotheses H11 and H12 are supported. In Vietnam, the relationship between knowledge sharing and consumers’ social media engagement is statistically significant (H13: β = 0.167 and p < 0.007), indicating a positive effect of knowledge sharing activities on customer engagement.

For Chinese consumers, the hypothesis testing results revealed that H1 and H3 were supported. Perceived anthropomorphism positively influenced AI WOM (H1: β = 0.129, p < 0.05), and perceived animacy also had a significant effect (H3: β = 0.086, p < 0.05). However, similar to Vietnamese consumers, H2 was rejected, as perceived intelligence did not significantly impact AI WOM (H2: β = −0.024, p > 0.05). Similar to Vietnam, Chinese AI WOM are positively impacted by perceived algorithm interpretability, perceived unbiasedness and perceived AI expertise, hence, supporting H4 (H4: β = 0.292, p < 0.000), H5 (H5: β = 0.175, p < 0.000) and H6 (H6: β = 0.350, p < 0.000), respectively. While H7 (H7: β = 0.388, p = 0.000) confirms a positive relationship between AI WOM and Chinese consumers’ social media engagement, H8 and H9 confirmed the significant effect of AI WOM on knowledge acquisition and knowledge sharing, as evidenced by H8: β = 0.821, p = 0.000 and H9: β = 0.789, p < 0.000. In addition, AI WOM has a direct and positive influence on knowledge application (H10: β = 0.746, p = 0.000), supporting the hypothesis H10. Similarly, the hypotheses H11 and H12 were also supported (both p-values = 0.000), indicating that as knowledge acquisition and application increase, social media engagement among consumers also rises. In China, the relationship between knowledge sharing and consumers’ social media engagement is not significant (p = 0.339). This finding implies that knowledge-sharing efforts do not significantly drive customer engagement on social media platforms in this context.

In addition, as presented in Table 7, all VIF values for the inner model relationships are well below the conservative threshold of 5. The highest VIF recorded is 2.803 (for KS → CSME), while the values for the six perception constructs (PAN, PI, PA, AI, UR and AIE) range from 1.859 to 2.568. This indicates that while some degree of correlation among these independent variables exists as is common in perceptual and behavioral research, it does not reach a level that would compromise the reliability of the path estimates. Consequently, the concern regarding potential multicollinearity among the six perception variables is statistically addressed; the observed correlations are within acceptable limits and do not constitute harmful multicollinearity. Therefore, it can be concluded that the structural model in this study was not adversely affected by multicollinearity, ensuring the robustness and interpretative validity of the hypothesized relationships tested.

Before conducting cross-country comparisons, this study assessed measurement invariance between the Vietnamese and Chinese samples using the measurement invariance of composite models (MICOM) procedure (Henseler et al., 2016). First, configural invariance was established, as identical measurement models, indicators, data treatment and algorithm settings were applied across the two national samples, and acceptable model fit was observed for both groups. Second, compositional invariance was evaluated using a permutation-based approach with 5,000 resamples and two-tailed testing. The original correlations of the composite scores were close to 1, and the corresponding 95% confidence intervals included the value of 1 for all constructs. In addition, all permutation p-values exceeded the 0.05 threshold, indicating that compositional invariance was established across the Vietnamese and Chinese samples. Third, the permutation-based confidence intervals for mean and variance differences included zero for all constructs, suggesting no statistically significant differences in composite means or variances between the two groups. MICOM results confirm configural and compositional invariance and therefore establish partial measurement invariance, which is sufficient for conducting cross-group comparisons of structural path coefficients (Henseler et al., 2015). An MGA was conducted to examine whether the structural relationships differ between Vietnam and China. As shown in Table 8, the results indicate several statistically significant differences between the two countries. In particular, the effect of AI on AIWOM differs significantly across the two groups. Significant differences are also observed in the effects of AIWOM on CSME, knowledge acquisition, knowledge sharing and knowledge application. Furthermore, the relationships between knowledge acquisition, knowledge sharing, knowledge application and CSME differ significantly between Vietnam and China. In contrast, no significant cross-country differences are found for the effects of perceived anthropomorphism, perceived intelligence, perceived animacy, unbiased recommendations and AI expertise on AIWOM. This suggests that the process through which AIWOM is formed is largely similar in the two national contexts. An examination of the direction of the path coefficient differences shows that most AIWOM-related effects are stronger in China than in Vietnam. However, the effects of knowledge acquisition and knowledge sharing on CSME are stronger in Vietnam. Overall, these findings suggest that AI-driven WOM plays a more prominent role in shaping consumer engagement in China, whereas knowledge-related processes are relatively more important in Vietnam.

This study builds on the APF and KM to investigate how AI WOM relates to CSME in Vietnam and China. It explores how users perceive, interpret and respond to AI-driven recommendations within social media environments, offering insights into the dynamics of human–AI interaction. By conducting a detailed analysis, this research expands the understanding of AI-driven consumer behavior, addresses gaps in existing studies and offers valuable insights into how AI influences social media engagement.

First, this research reveals and confirms antecedents affecting AI WOM. The research indicates a strong positive association between perceived animacy and AI WOM for both Vietnamese (H1: β = 0.187, p < 0.000, t = 3.595) and Chinese consumers (H1: β = 0.086, p < 0.05, t = 2.154). This implies that users who perceive AI as more lifelike are more willing to engage in WOM discussions about AI. This finding supports earlier studies, which emphasize the influence of perceived animacy on shaping user attitudes toward intelligent virtual assistants (Priya and Sharma, 2023). Notably, the effect size is larger for Vietnamese consumers, implying that animacy plays a more pronounced role in driving AI WOM in Vietnam compared to China. However, H2 is rejected, as perceived intelligence does not significantly influence AI WOM in either market, as indicated by nonsignificant results for Vietnamese consumers (H2: β = 0.044, p > 0.05, t = 0.808). Similarly, in China, this correlation lacked statistical significance (H2: β = −0.024, p > 0.05, t = 0.557). This contradicts previous research suggesting a link between AI intelligence and user attitudes (Priya and Sharma, 2023). One potential reason for this divergence may lie in the fact that, while intelligence enhances AI functionality, it does not necessarily evoke emotional engagement strong enough to encourage AI WOM. In addition, perceived anthropomorphism shows a strong positive correlation with AI WOM in both Vietnam (H3: β = 0.217, p < 0.000, t = 3.733) and China (H3: β = 0.129, p < 0.05, t = 3.413). This indicates that the more human-like AI appears, the more users are inclined to talk about it. These findings are consistent with Priya and Sharma (2023), who emphasized the impact of anthropomorphism on AI attitudes. Next, the relationship between perceived algorithm interpretability and AI WOM exhibited similar patterns among Vietnamese and Chinese consumers. Specifically, the results indicate a significant effect in both Vietnam (H4: β = 0.089, p < 0.05, t = 2.042) and China (H4: β = 0.292, p < 0.000, t = 6.190), with the impact being notably stronger in the Chinese market. This implies that higher algorithm interpretability enhances consumers’ trust and willingness to share AI-related experiences, with the stronger effect in China potentially stemming from greater AI adoption and familiarity in digital environments. These results align with the work of Hong et al. (2023), suggesting that enhancing algorithm interpretability can contribute to greater trust and reliability in AI-driven financial systems. Next, this research highlights the substantial positive influence of perceived unbiasedness on AI WOM, with all p-values <0.05 for both surveys conducted in Vietnam (H5: β = 0.161, p < 0.01, t = 3.310) and China (H5: β = 0.175, p < 0.000, t = 4.114). This result aligns with the article by Verma et al. (2023), which suggested a positive relationship between eWOM credibility and the intention to use eWOM messages when making purchasing decisions. Our findings reveal a significant positive correlation between AI expertise and AI WOM in both Vietnam (H6: β = 0.224, p < 0.000, t = 4.255) and China (β = 0.350, p = 0.000, t = 6.860). Users who perceive AI systems as competent and knowledgeable are more likely to engage in AI-related WOM and are inclined to share information and experiences related to these technologies. The strength of this relationship between the two countries suggests that Chinese users may be more influenced by AI expertise when forming and sharing AI WOM. This observation aligns with the conceptual framework proposed by Tassiello et al. (2024), who introduced the notion of AIWOM to describe the evolving consumer behaviors in response to AI-driven innovations. Further corroborating our findings, recent research by Wang et al. (2024) proposed a comprehensive scale for measuring customer experiences with AI-enabled products, identifying key dimensions such as classification, delegation, data capture and social and anthropomorphic experiences. These dimensions collectively enhance the perceived expertise of AI chatbots, thereby fostering greater consumer engagement and positive WOM dissemination.

Second, this research establishes the role of AI WOM in shaping consumer engagement on social media. The finding reveals a significant positive relationship between AI WOM and consumers’ social media engagement for both Chinese (H7: β = 0.388, p < 0.000, t = 7.183) and Vietnamese (H7: β = 0.270, p < 0.000, t = 3.895) consumers. This suggests that AI WOM significantly influences how actively users participate in social media in both markets, albeit with a stronger effect in China. This relationship aligns with previous research, which suggests that AI WOM fosters user discussions and interactions in online platforms by enhancing trust and perceived credibility (Simay et al., 2022). It is writ large that the weaker effect in Vietnam suggests that AI WOM alone may not be enough to drive engagement, as users may require additional social validation or influencer involvement.

Third, this study underscores the influence of AI WOM on knowledge-related behaviors. The findings confirm a significantly positive relationship between AI WOM and knowledge acquisition for both Chinese (H8: β = 0.821, p < 0.000, t = 38.051) and Vietnamese (H8: β = 0.561, p < 0.000, t = 12.061) consumers. This suggests that AI WOM effectively enhances users’ ability to acquire new information, with a notably stronger effect in China. In line with previous findings, this study confirms AI’s significant influence on the dissemination and accessibility of knowledge (Ngo et al., 2024). The stronger impact in China could be attributed to higher AI penetration and a greater reliance on AI-driven search and recommendation systems, which facilitate faster and more efficient knowledge acquisition. In addition, the results reveal a significantly positive relationship between AI WOM and knowledge sharing in both China (H9: β = 0.789, p < 0.000, t = 28.899) and Vietnam (H9: β = 0.539, p < 0.000, t = 12.712). This aligns with prior research highlighting the role of trust in digital interactions and knowledge dissemination (Yaqub and Al-Sabban, 2023). However, limited research has previously examined this specific relationship. Furthermore, AI WOM has a significantly positive impact on knowledge application for both Chinese (H10: β = 0.746, p < 0.000, t = 23.524) and Vietnamese (H10: β = 0.537, p < 0.000, t = 11.934) consumers. This implies that AI WOM enhances users’ ability to apply acquired information, with a stronger effect observed in China. This study is consistent with existing literature, emphasizing AI’s role in personalized knowledge dissemination (Owoseni et al., 2024). The stronger effect observed in the Chinese context may be attributed to greater exposure to AI-powered educational technologies and digital literacy initiatives, which facilitate knowledge retention and application.

This study examines the association between knowledge application and consumers’ social media engagement, which includes a range of interactive behaviors such as learning, socializing, sharing and codeveloping. For Chinese consumers, this correlation is particularly strong (H11: β = 0.675, p < 0.000, t = 8.645), showing that knowledge-applicators are more involved in social media interaction in all its forms. This result is consistent with earlier studies that highlight how knowledge applications can promote engagement in digital environments (Haque and Abulaish, 2025). Similarly, the effect size is significantly smaller for Vietnamese consumers, even though the relationship is still significant and positive (H11: β = 0.176, p < 0.01, t = 2.613). The outcome could be explained, in part, by the fact that cultural variations may clarify this difference, as the digital engagement and cooperation activities might be more deeply embedded in Chinese social media culture. For Vietnamese consumers, knowledge acquisition was found to be positively associated with consumers’ social media engagement (H12: β = 0.177, p < 0.05, t = 2.620). Vietnamese populations displayed increased acquisition in knowledge practices with greater user’ involvement on social platforms (Zeng et al., 2022). In contrast, this correlation is negative in China (H12: β = −0.229, p < 0.000, t = 3.808), implying that higher social media engagement is associated with a decrease in knowledge acquisition among certain consumer groups. In addition, the relationship between knowledge sharing and customers’ social media engagement demonstrates varying levels of significance across the two studied contexts, Vietnam and China. In Vietnam, the results indicate a positive and significant relationship (H13: β = 0.167, p < 0.01, t = 2.681). This suggests that knowledge-sharing activities contribute to increased customers’ social media engagement, aligning with earlier research that stresses the value of knowledge exchange in fostering online interactions and participation (Molinillo et al., 2019; Xu et al., 2024). Conversely, in the case of Chinese samples, the relationship between knowledge sharing and customers’ social media engagement does not show statistical significance (H13: β = −0.076, p > 0.05, t = 0.957). This result implies that knowledge-sharing behaviors do not have a strong influence on social media engagement within this context. A possible explanation for this finding is that Chinese consumers may engage with social media primarily for entertainment, brand-related content or influencer-driven interactions rather than for knowledge exchange. This aligns with prior research suggesting that social media engagement in China is often influenced by interactive and entertainment-focused content, as knowledge sharing may sometimes carry implicit motives. This research emphasizes the value of a contextual approach in applying knowledge-sharing strategies to enhance customer engagement. While knowledge sharing effectively drives engagement in Vietnam, entertainment-based or influencer-led content may be more influential in the Chinese market.

Finally, the cross-country comparison provides important insights into how national context shapes the mechanisms through which AI attributes and AI-driven WOM influence consumer engagement. The findings indicate that the impacts of perceived animacy, perceived intelligence, perceived anthropomorphism, unbiased recommendations, perceived accountability and AI expertise on AIWOM are largely comparable across the two countries, suggesting a shared evaluation of fundamental AI characteristics. The similar impact of perceived unbiasedness on AI WOM in Vietnam and China suggests this is a foundational AI attribute universally valued across both markets. Regardless of cultural or infrastructural differences, users share a common need for fair, impartial AI recommendations. When AI is perceived as unbiased, it reinforces user trust, encouraging them to share positive experiences (AI WOM) as a natural social behavior. The comparable effect sizes across both countries further support the notion that unbiasedness is a core, cross-cultural value in assessing AI credibility. Regarding AI expertise, the results show similar positive effects in both countries, however, the effect on AI WOM is stronger among Chinese users. In a market where AI is deeply integrated into “super-apps” like WeChat or Alipay and is used for a wide range of high-stakes decisions (investment, health care and logistics), the perceived expertise of AI becomes a key metric of its value. As Chinese consumers interact more frequently with sophisticated AI applications, perceptions of AI competence and technical expertise may become more salient drivers of trust and information diffusion. Consequently, users who perceive AI systems as highly capable are more likely to disseminate AI-generated recommendations within their networks. In Vietnam, although AI expertise remains an important factor, the relatively earlier stage of AI adoption may lead consumers to rely on a broader set of cues, such as social influence or platform reputation, when deciding whether to share AI-related information.

However, significant differences emerge in the effect of algorithm interpretability on AIWOM, with a stronger influence observed in China. One plausible explanation relates to the higher maturity of China’s digital ecosystem and the widespread integration of algorithmic systems in everyday services such as e-commerce, fintech and content recommendation platforms. Chinese consumers are therefore more accustomed to interacting with algorithm-driven interfaces and may place greater emphasis on understanding how AI systems generate recommendations. In such contexts, interpretability functions as an important trust-building mechanism that encourages users to share AI-generated information with others. In contrast, Vietnamese consumers may still rely more heavily on interpersonal validation and social cues when evaluating digital recommendations, which reduces the relative importance of algorithm transparency in triggering AI WOM.

In addition, the effects of AIWOM on knowledge acquisition, knowledge sharing, knowledge application and CSME are significantly stronger in China, highlighting the more central role of AIWOM in translating AI perceptions into engagement outcomes in this context. This interpretation can be explained through the affordance theory and an information adoption lens. From an affordance perspective, AIWOM is more likely to lead to knowledge application when platform environments make information easier to interpret, personalize and act upon through functions such as responsiveness, recommendation integration and seamless movement from content exposure to platform-based action. From an information adoption perspective, AI-generated content is more likely to be applied when users perceive it as actionable, useful and easy to incorporate into ongoing digital practices. Country-level indicators provide contextual support for why this conversion may be stronger in China. In the 2025 Network Readiness Index, China ranks 24th out of 127 economies with a score of 65.74, whereas Vietnam ranks 40th with a score of 56.00, indicating stronger overall network readiness in China. The same 2025 country profiles also show higher scores for China in the technology pillar (64.37 vs 52.02) and the people pillar (67.83 vs 47.25), which is consistent with a context in which AI-related content may be more readily translated into practical use. Although these indicators do not establish causality in the present model, they do suggest that AIWOM in China may circulate within a more digitally integrated environment, making its conversion into knowledge application more likely (Network Readiness Index, 2025a, 2025b). Previous literature has also argued that differences in technological infrastructure, institutional support and digital experience can shape individuals’ trust in algorithmic systems and their willingness to rely on AI-based information, leading to cross-national variations in behavioral outcomes (Omrani et al., 2022). Knowledge application shows a much stronger relationship with CSME in China, while knowledge acquisition and knowledge sharing have stronger effects in Vietnam. One plausible explanation is that Chinese users may be more accustomed to converting AI-related information into action within highly integrated digital environments. In such contexts, once knowledge is applied, it is more likely to produce visible engagement behaviors, which helps explain why knowledge application shows a much stronger effect on CSME in China. By contrast, the stronger effect of knowledge acquisition in Vietnam suggests that engagement may develop through a more learning-centered process. Vietnamese users may be more likely to participate after first acquiring and processing useful information, rather than moving directly from AI-mediated exposure to action. This implies that the Chinese pattern is more consistent with engagement driven by enacted knowledge, whereas the Vietnamese pattern reflects engagement driven by prior learning and sense making. From a comparative perspective, the stronger role of knowledge application in China was broadly plausible, whereas the stronger role of knowledge acquisition in Vietnam is better interpreted as an exploratory insight that points to a more gradual knowledge to engagement pathway in that context (Alavi and Leidner, 2001; Network Readiness Index, 2025a, 2025b). Knowledge sharing shows a significant positive effect on CSME in Vietnam, while this relationship is not significant in China. One possible explanation is that Vietnamese social media users tend to rely more on peer interaction and community-driven exchanges when engaging online. Hence, sharing experiences and information may therefore stimulate discussion and participation within digital communities. In contrast, Chinese users may engage with social media more through entertainment-oriented content, influencers or platform-driven features, which may reduce the influence of peer knowledge exchange on their engagement behaviors. Overall, these findings emphasize the importance of national context in understanding how AI attributes and AIWOM jointly influence consumer engagement.

To strengthen the comparative contribution of this research, we distinguish between the cross-country differences that were theoretically anticipated and those that emerged as exploratory findings. Theoretically, we anticipated that perceived algorithm interpretability and the overall influence of AI WOM would be significantly stronger in China. This expectation was based on China’s more mature and regulated AI ecosystem, where users have higher AI penetration and familiarity with algorithmic logic. In contrast, the finding that knowledge sharing significantly drives engagement in Vietnam but remains nonsignificant in China emerged as an exploratory insight, suggesting that Vietnamese users use social media for collaborative exchange while Chinese users may prioritize entertainment or influencer-led interactions.

The current research contributes to the literature on AI perception by extending it in several key areas. First, drawing on our model was advanced based on APF, which broadens others perceptions of AI, a novel approach that uniquely positions this study within the broader academic dialogue on AI WOM and consumer engagement (Zarouali et al., 2022). This study uses APF with AI WOM and CSME, establishing a robust foundation for examining how algorithmic recommendations and perceived AI attributes influence digital interaction. Second, this study is among little empirical research examining the direct relationship between AI WOM toward CSME, and uncovers the unique role of AI WOM in influencing consumer behavior on social media platforms. Unlike traditional WOM, AI WOM is characterized by its reliance on AI-generated content, which may differ in credibility, engagement and influence. This distinction opens up new avenues for understanding the mechanisms through which AI WOM interactions shape online engagement. Third, we extended the explanatory power of the AI WOM by incorporating perceived animacy, perceived intelligence, perceived anthropomorphism, perceived algorithm interpretability, perceived unbiasedness and perceived AI expertise into the research model, providing a detailed examination of their effects (Balakrishnan and Dwivedi, 2021; Hong et al., 2023). Our study presents a detailed framework for understanding how users’ perceptions of AI capabilities impact their readiness to engage in AI WOM. These attributes are crucial in differentiating AI WOM from traditional human-driven WOM, offering fresh insights into the technological and psychological factors underlying consumer engagement with AI systems. Fourth, the study broadens the existing literature by focusing on the pivotal role of KM, comprising knowledge sharing, knowledge application and knowledge acquisition. While past research has predominantly examined the direct influence of AI-generated recommendations and opinions on consumer behavior (Guerra-Tamez et al., 2024; Mogaji and Jain, 2024), this study introduces a more nuanced approach by integrating KM processes as an essential bridge linking AI WOM to engagement outcomes. Fifth, this study broadens the scope of KM theories into the realm of AI-driven consumer interactions. Traditionally, KM has been examined in organizational and human-based contexts (Cabrera and Cabrera, 2005; King, 2009), with limited exploration of its role in AI communications. By demonstrating how AI WOM facilitates knowledge sharing, application and acquisition, this research expands the theoretical boundaries of KM by positioning AI as an active agent in knowledge dissemination and utilization. Sixth, this research further contributes to theoretical insights by examining how knowledge acquisition, application and sharing relate to CSME in the realm of AI WOM. While much of the previous literature has extensively explored consumer engagement on social media (Brodie et al., 2013; Hollebeek et al., 2014), the role of AI chatbots as facilitators of knowledge exchange in this domain remains underexplored. Our study highlights that knowledge acquisition from AI chatbots fosters learning-oriented engagement on social media. Unlike traditional WOM, where consumers rely on human-generated information, AI-generated recommendations and explanations enhance consumers’ ability to absorb and process new information, leading to more active participation in discussions. We contribute to the literature by emphasizing that knowledge application influences socializing behaviors in online communities. Consumers who successfully apply AI-driven insights in their personal or professional contexts are more prone to engaging in discussions, sharing experiences and seeking validation from their peers. This study extends theoretical perspectives on knowledge sharing by demonstrating its role in content-sharing and codevelopment activities. AI chatbots enable consumers to become knowledge contributors, facilitating the cocreation of user-generated content. Unlike conventional social media engagement, where knowledge flows from human experts or influencers, AI-powered knowledge sharing introduces a dynamic, bidirectional exchange that strengthens online community participation.

Moreover, this study advances theoretical discourse by extending AI ethics and transparency beyond consumer trust to encompass their broader impact on CSME. Prior research has largely framed perceived unbiasedness and algorithm interpretability as factors that shape consumer trust in AI-driven decisions (Lee, 2018; Grimmelikhuijsen, 2022). However, this study argues that these factors do not merely influence trust but also drive consumer engagement by shaping their interactions with AI-generated content. Specifically, when AI is perceived as transparent and unbiased, consumers are more likely to actively engage in social media platforms through four distinct mechanisms: learning, socializing, sharing and codeveloping.

In addition, this study extends the APF and KM literature by comparing AI WOM and social media engagement across Vietnam and China – two culturally and economically distinct markets. Prior research has predominantly focused on Western contexts or single-country settings (Priya and Sharma, 2023; Simay et al., 2022), leaving an unexplored gap regarding how AI WOM operates in emerging Asian economies. By revealing divergent effects (e.g. perceived animacy’s stronger impact in Vietnam vs algorithm interpretability’s dominance in China), this research draws attention to the critical role of cultural and technological contexts in shaping AI interactions. For instance, China’s advanced AI adoption may explain its stronger AI WOM-engagement link, whereas Vietnam’s reliance on social validation highlights nuanced behavioral drivers. These contrasts challenge universal assumptions about AI adoption, suggesting that contextual factors (e.g. digital literacy and AI familiarity) critically moderate theoretical relationships. This dual-country analysis not only fills a geographic gap but also pioneers a comparative framework for future studies in underrepresented regions.

Regarding practical implications, this research offers valuable insights for various stakeholders, including customers on social media, marketers, managers and AI developers. Understanding the importance of AI-driven consumers’ perception in shaping AI WOM, as well as the importance of AI WOM in driving KM and social media engagement enables these groups to optimize AI-driven strategies, enhance user experience and maximize engagement in both Vietnam and China.

For social media users, the study underscores the significance of AI-generated recommendations in shaping their engagement and online interactions. Customers who perceive AI as lifelike and human-like tend to be more engaged in AI WOM, sharing their experiences and offering recommendations to other users. This implies that users should be aware of how AI systems curate their content and recommendations, ensuring they remain critical consumers of information rather than passive recipients. In addition, as AI-driven recommendations significantly impact perceived informativeness and credibility, users should develop strategies to balance AI-generated content with self-awareness to make more informed decisions. By understanding how AI operates, users can leverage these recommendations effectively for entertainment, shopping or educational purposes while avoiding potential manipulation or filter bubbles.

For marketers, the study underscores the necessity of leveraging AI WOM to design more effective AI-driven marketing strategies. First, the strong positive link between perceived animacy and AI WOM implies that marketers should focus on creating AI systems that exhibit lifelike qualities. For instance, incorporating conversational tones, empathetic responses and interactive features into AI chatbots or recommendation systems can make them more engaging and relatable to users. Second, the significant impact of anthropomorphism on AI WOM indicates that marketers should humanize AI interactions by using avatars, personalized greetings and culturally relevant language. Through enhancing users’ emotional connection with AI systems, this approach encourages the sharing of their experiences with others. Third, the study highlights how the interpretability and impartiality of algorithms are crucial in nurturing trust and facilitating engagement. Marketers should prioritize transparency by clearly communicating how AI algorithms work and ensuring that recommendations are free from bias. For example, providing users with explanations for why certain content is recommended or allowing them to customize their preferences can enhance trust and satisfaction. In China, where algorithm interpretability has a particularly strong impact, marketers should invest in user-friendly interfaces that simplify AI processes and build confidence in the system’s reliability. By addressing these factors, marketers can create AI-driven campaigns that attract users and encourage them to share positive experiences, ultimately amplifying brand reach and credibility.

From a managerial perspective, the study emphasizes how AI WOM influences knowledge-related behaviors, and how effective KM enhances social media engagement, providing a foundation for implementing impactful AI strategies. Recognizing the potential of AI-driven recommendations can help managers enhance users’ learning and information-sharing experiences. For instance, in Vietnam, where knowledge sharing significantly influences social media engagement, managers can develop AI systems that facilitate collaborative learning and content cocreation. In China, where knowledge application has a stronger impact, managers should focus on creating AI tools that help users apply information in practical ways, such as through interactive tutorials or personalized learning paths. By aligning AI strategies with these knowledge-related behaviors, managers can foster deeper engagement and loyalty among users.

For AI developers and designers, this study provides valuable guidance on optimizing AI WOM to enhance user perception and engagement. The results indicate that AI-generated recommendations should be designed to appear more human-like and engaging to encourage interaction. Developers should integrate natural language processing techniques to refine AI-generated messages, making them more conversational and contextually relevant. In addition, improving the personalization capabilities of AI models is crucial for delivering highly relevant recommendations. This can be achieved by using machine learning algorithms that assess user behavior, sentiment and real-time interests to refine content suggestions. Furthermore, AI designers should prioritize creating transparent AI systems that help users to comprehend the rationale behind specific recommendations. Providing users with customization options, such as allowing them to adjust AI-generated recommendations based on their interests, can significantly improve engagement and trust. Another key consideration is the ethical aspect of AI development – ensuring that AI WOM does not promote misinformation or manipulate users into making biased decisions. Developers should implement safeguards to filter out low-quality or misleading content while promoting transparency in the AI decision-making process. As a result, AI-generated recommendations can contribute to a more genuine and impactful user experience.

While this study highlights the effect of AI WOM on customer engagement within social media contexts, several limitations should be considered to improve its theoretical and practical contributions. First, the study is based exclusively on quantitative survey data, which may introduce self-reporting biases and limit the ability to fully explore user perceptions. Future studies could benefit from adopting mixed-method approaches, such as qualitative interviews or experimental studies, to strengthen causal explanations. Second, the study is based on the APF but does not explore other relevant models, such as human–computer interaction or the technology acceptance model. By integrating these perspectives, a deeper understanding of AI-driven engagement can be achieved. Third, the study focuses only on Vietnam and China, which restricts the generalizability of its findings. Future research should examine different cultural and technological settings to identify cross-cultural differences. Fourth, the sample for this study was notably skewed toward young people with lower incomes who were heavy social media users. This may limit how well the findings reflect older adults or business users. To improve the generalizability of future work, researchers should prioritize more diverse and representative sampling methods. In addition, the survey only included people who had already used AI chatbots. While this helped ensure that participants could evaluate the topic, it may also mean that the results lean toward more favorable perceptions and higher engagement than what might be found in the general population. Including both users with and without prior experience in future studies would allow for a more balanced view of how AI WOM influences engagement. Next, by integrating AI WOM, KM and social media engagement, this study establishes a comprehensive theoretical foundation for further investigation on AI-driven consumer behavior. Future studies could further explore how different AI-generated knowledge types (e.g. explicit vs tacit) impact engagement and how personalized AI WOM influences consumer trust and decision-making. A further limitation concerns the directional ordering between AIWOM and knowledge-related processes. Although the present study positions AIWOM as an upstream communication stimulus that facilitates subsequent knowledge acquisition and downstream outcomes, alternative structures are also plausible in AI-mediated environments characterized by continuous interaction and feedback loops. For example, users who acquire, apply or share AI-related knowledge may subsequently become more likely to rely on AI-generated recommendations or to generate further AIWOM over time. Recent evidence suggests that human AI interactions can operate recursively, such that exposure to AI output may shape later cognition and behavior in ways that reinforce subsequent interaction patterns (Glickman and Sharot, 2024). Future research could therefore examine dynamic or reciprocal models using longitudinal, panel, experience sampling or cross-lagged designs to better capture the coevolving relationship between AIWOM and knowledge processes. Finally, this study treats AI WOM as a single concept, without distinguishing between AI-generated recommendations and conversational AI interactions. Future studies should differentiate these types of AI engagement to develop a more detailed theoretical framework. Addressing these limitations will improve the study’s methodological rigor, theoretical depth and broader relevance across industries and user groups. A further limitation of this study relates to the operationalization of the AI-WOM construct. Although AI-WOM is conceptually positioned as a distinct, algorithmically mediated and interactive form of communication, its measurement in this study is adapted from existing WOM and eWOM scales. While this approach is consistent with prior research in emerging domains and ensures initial content validity, it may not fully capture the unique characteristics of AI-WOM, such as real-time interactivity, perceived algorithmic agency, personalization and parasocial dynamics inherent in human–AI communication. As a result, the current operationalization may only partially reflect the multidimensional nature of AI-WOM. Given the novelty of the construct, future research is encouraged to develop and validate a dedicated AI-WOM measurement scale using rigorous scale development procedures (e.g. Churchill, 1979; MacKenzie et al., 2011). Such efforts should aim to identify and empirically validate the underlying dimensions of AI-WOM, potentially incorporating aspects such as perceived autonomy, responsiveness, cocreation and algorithmic transparency. Developing a context-specific and psychometrically robust scale would not only enhance construct validity but also enable more precise theorization and empirical testing of AI-driven communication phenomena in future studies.

This research used a survey method to collect the data. All procedures performed in this study involving human participants were conducted ethically according to the ethical standards of the Institutional and/or National Research Committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. Informed consent was obtained from all individual survey participants included in the study.

This research does not include and disclose any case reports of individuals participating in the study.