The model suggests that when users encounter new technologies, they assess both perceived usefulness and perceived ease of use to determine whether to adopt them early (Davis, 1989). These two psychological factors are key determinants of user acceptance. Perceived ease of use refers to the degree to which an individual believes that using a particular technology would be free of effort. In contrast, perceived usefulness indicates the extent to which users believe the technology can enhance their job performance (Yoon and Lee, 2021).

Scholars have extended the application of the TAM in e-commerce by integrating the unique characteristics of online shopping environments with the concept of consumer trust. Empirical studies have validated several models that combine e-commerce trust with technology acceptance frameworks. For instance, Gefen and Straub (2003) proposed a model that integrates the TAM with theories of trust formation in e-commerce. He argued that external factors in e-commerce influence perceived usefulness and perceived ease of use, which in turn affect consumer trust. This trust subsequently shapes consumers’ attitudes and behaviors toward online shopping.

Artificial intelligence virtual anchors are digital avatars created through AI technology that integrate voice systems and motion recognition to replace human anchors (Shi and Yue, 2023). Advanced AI capabilities and extensive speech databases enable these virtual anchors to perform anchor duties effectively (Kaplan and Haenlein, 2019), thereby driving significant industry transformation. In live streaming, they are gradually replacing human anchors (Santhoshkumar et al., 2023). AI virtual anchors offer lower investment costs and greater stability than human anchors, and are not constrained by broadcast duration or location (Niu et al., 2023). From a service perspective, Dong et al. (2023) notes that virtual anchors possess appeal and novelty, naturally generating traffic through topic engagement, thereby boosting livestream performance. From an industry development perspective, AI virtual anchors can address the “anchor dilemma” faced by human live streamers in e-commerce. On the one hand, human anchors may face reputational damage; on the other hand, the uneven quality of human anchors makes it difficult for e-commerce companies to recruit suitable candidates. It highlights a key direction for the future development of e-commerce live streaming (Zhong et al., 2024).

Intelligence is the fundamental attribute of virtual humans, endowed by AI technology (Slota et al., 2023). Virtual humans perceive their internal and external environments through intelligent technologies and execute actions with inherent intelligence (Orozco et al., 2011). Their intelligent behaviors reflect psychological states and simulate emotional characteristics such as mood (Vinayagamoorthy et al., 2006), enabling them to interact with their surroundings while possessing feedback-generating capabilities that influence them (Herrero and Antonio, 2005). As virtual human intelligence continues to evolve, new economic production models are emerging. As representatives of AI, virtual humans have become labor subjects participating in the division of labor, enhancing productivity and presenting new opportunities for comprehensive human-society development (Su and Guo, 2019). To better serve users, developers integrate more human-like intelligent behaviors into virtual human programs through algorithms (Li, 2003), thereby improving their intelligence for seamless interaction in application scenarios (Pan et al., 2007). Zhong et al. (2025) conducted grounded research revealing that the intelligence of e-commerce virtual anchors comprises four dimensions: guidance intelligence, recognition intelligence, analysis intelligence, and feedback intelligence. The study developed corresponding measurement scales and validated their scientific validity through empirical research.

Trust forms the foundation of exchange behaviors (Granovetter, 1973). In e-commerce contexts, trust refers to consumers’ expectations regarding the performance perceptions of products and services (Le, 2015). Gu (2022) posits that consumer trust constitutes the core competitive advantage in livestreaming commerce, where trust in anchors and products significantly influences purchasing decisions (Wang, 2023). With AI technology increasingly integrated into various interactive scenarios, scholars have expanded trust research to the human-machine relationship level, exploring the logic behind users’ trust in AI. Reyes et al. (2025) found through game experiments that visualizing the uncertainty of AI predictions can effectively enhance user trust. Zarifis and Cheng (2024), focusing on financial consulting scenarios, proposed that trust-building in generative AI requires adaptation to problem types—AI’s human-like interaction features can strengthen trust when addressing ambiguous financial questions.

Initial trust—the foundational form of trust—shapes future relationships between communicators (McKnight et al., 2002). This initial assessment occurs during anchors’ first contact with audiences and is based on an evaluation of their comprehensive capabilities (Shi and Li, 2021). As the first step in trust-building, it establishes an enduring foundation for sustained relationships (Audun et al., 2006). In e-commerce, establishing initial trust is critical to consumer decisions and platform sustainability. The virtual nature of transactions introduces heightened uncertainty and credit risks, making initial trust acquisition particularly vital for digital commerce (Salo and Karjaluoto, 2007). Research on initial trust in AI interaction scenarios further reveals unique formation mechanisms. Yanzeng et al. (2025) found that bubble coloring designs in chatbots enhance information processing fluency, strengthening users’ initial trust in AI and, consequently, increasing their willingness to disclose themselves. Dao et al. (2025), using Vietnamese Gen Z consumers as the sample, found that in medical AI contexts, the “traditional self” in individual self-concept negatively affects initial trust, whereas the “modern self” positively influences it. That initial trust directly promotes willingness to adopt AI.

In management studies, the “value co-creation” approach manifests through consumer experiences, where businesses create value by providing exceptional consumer interactions (Prahalad and Ramaswamy, 2004). Throughout this process, consumers act as value creators while enterprises serve as co-creators through service engagement (Grönroos, 2008). Yi and Gong (2013) categorized consumer value co-creation into two dimensions: participation behavior and citizenship behavior. Participation behavior includes information seeking (collecting external information to meet needs), information sharing (discussing known information with others), compliance behavior (adhering to rules and instructions), and interpersonal interaction (engaging with others). Citizenship behavior encompasses feedback (providing suggestions for long-term development), advocacy (promoting products/services), altruism (helping peers), and tolerance (accepting others’ mistakes).

In psychology, “fear” is considered one of the most primal human emotions, arising from our self-protective mechanisms in response to unfamiliar environments or potential threats (Wang and Kong, 2023). Technophobia analyzes the thoughts or behaviors of users who develop negative emotions, such as anxiety and fear, towards computers and thus destroy or resist them (Bin and Yan, 2020). As research progressed, scholars broadened the construct from computer-specific anxiety to encompass apprehension toward emerging technologies in general (Aksoy et al., 2020).

Technophobia towards AI is relatively common among people. In an online survey conducted by the British Science Association, 60% of respondents believed AI would reduce employment opportunities within a decade, and 36% viewed AI development as a threat to humanity’s long-term survival (Oh et al., 2017). Scholarly research on various AI systems identifies two primary sources of technophobia: First, the perceived greater-than-anticipated job displacement—where AI’s superior productivity might replace significant numbers of workers (Mills, 1997)—has fueled fears of being replaced by AI (Xu and Song, 2022). Second, uncertainties surrounding continuous technological advances, particularly the fear that AI could attain autonomous consciousness and pose existential risks to humanity (Han and Zhao, 2020).

This study conducts independent testing of the four sub-dimensions of e-commerce virtual anchors’ intelligence (guidance intelligence, recognition intelligence, analysis intelligence, and feedback intelligence), based on the following core rationale: First, the sub-dimension classification is grounded in robust empirical evidence and literature support. Zhong et al. (2025) employed grounded research methods to deeply immerse in real-world human-computer interaction scenarios of e-commerce virtual anchors, systematically refining and validating these four core sub-dimensions. It confirms that the classification accurately aligns with consumers ‘perception of virtual anchors’ intelligence performance, representing empirical induction from practical scenarios rather than subjective categorization. It provides a fundamental theoretical basis for testing the impact of each dimension separately.

Secondly, the functional orientations and operational pathways of each sub-dimension exhibit fundamental differences. Guidance intelligence emphasizes proactive service adaptation; recognition intelligence focuses on precise demand capture; analysis intelligence prioritizes problem-solving efficiency; while feedback intelligence concentrates on interactive response quality. Merging these dimensions into a higher-order factor would only validate the influence of “overall intelligence,” failing to reveal the differentiated impacts of the various intelligence dimensions on perceptual variables and initial trust, and thereby losing the ability to identify core influence pathways accurately.

Finally, the independent testing aligns with the study’s practical value orientation. One of the core objectives of this research is to provide actionable optimization strategies for relevant enterprises. By separately evaluating the effectiveness of each sub-dimension, it directly clarifies the optimization priorities for businesses. This precise guidance value is unattainable with high-order factor models, which can only demonstrate “overall effectiveness of intelligence” but fail to address the core question: “which types of intelligent capabilities enterprises should prioritize for optimization.”

This study establishes the core mediating role of “initial trust” between perceived usefulness and perceived ease of use and value co-creation behavior. This framework serves as the key logical pivot in the model, as demonstrated below: From the perspective of theoretical transmission logic, the TAM fundamentally operates through “perceived variables → attitude/behavior” transformation (Davis, 1989). In human-computer interaction contexts, trust serves as the essential bridge connecting cognitive evaluation to behavioral decision-making (McKnight et al., 2002). While perceived usefulness reflects consumers ‘assessment of e-commerce virtual anchors’ “practical value” and perceived ease of use evaluates “operational convenience,” these evaluations do not directly trigger value co-creation. Consumers must first establish initial trust in the anchors’ reliability and service capabilities through these perceptions, which then eliminates interaction concerns and facilitates the transition from “cognitive recognition” to “participation behavior.” Therefore, the core transmission path in this study is: E-commerce virtual anchor intelligence → Perceived usefulness/Perceived ease of use → Initial trust → Value co-creation behavior. Given the unique nature of the research scenario, in which e-commerce virtual anchors function as “dynamic AI interaction agents” lacking offline physical support, the mediating role of initial trust is particularly critical. In first-time contact scenarios, consumers lack understanding of virtual anchors’ service quality and response reliability. While perceived usefulness and perceived ease of use can lower the “technical acceptance threshold,” initial trust further reduces “uncertainty in virtual transactions,” motivating consumers to engage in participation behaviors such as liking, asking questions, and even proactively promoting or reporting flaws. Existing research confirms that in AI-driven human-computer interactions, trust serves as the core mediator transforming perceived value into behavioral intention (Kaplan and Haenlein, 2019). Without this intermediary, the explanatory power of perceived variables on co-creation behaviors would significantly weaken. From a model rigor perspective, this study clarifies the mediating role of initial trust, avoiding the logical leap that “perceived usefulness/perceived ease of use directly impacts co-creation.” In prior studies on e-commerce trust, Gefen and Straub (2003) established that perceived variables influence consumer behavior through trust, while Hampton-Sosa and Koufaris (2005) further highlighted the pivotal mediating role of initial trust in first-contact scenarios. This research builds upon this theoretical framework to align variable relationships with the practical logic of consumer decision-making.

A total of 400 formal questionnaires were distributed in this study: 200 through online channels and 200 through offline channels. Researchers excluded questionnaires with incorrect responses to screening items and those suspected of being completed carelessly (e.g., selecting the same option for six consecutive questions or containing blank items). Ultimately, 337 valid questionnaires were retained, with an effective response rate of 84.25%.

Table 1 presents the demographic characteristics of the participants. The sampling conformed to the audience characteristics of AI virtual anchors’ live streaming rooms.

This study conducted descriptive statistics on 337 valid samples using SPSS 20.0, with the results shown in Table 2. It indicates that all item standard deviations were less than 2, and the absolute values of skewness and kurtosis were all below 2—both meet the criteria for normal distribution. This result demonstrates that the sample data follow a normal distribution and are suitable for subsequent statistical tests.

As shown in Table 3, the correlation coefficients among latent variables all remained within the standard threshold of 0.7, indicating relatively strong overall correlations. While these results provide preliminary predictive support for subsequent hypothesis testing, correlation analysis alone cannot conclusively determine quantitative dependencies between variables. Therefore, structural equation modeling should be employed in this study to validate the specific relationships among these variables further.

As the study’s data originate from self-reported questionnaires, it may be susceptible to standard-method bias due to a single data source. To mitigate this, we employed Harman’s widely recognized single-factor test for validation. The test’s core logic is that if significant standard method bias exists, the variance across all measurement items would be highly concentrated in a single common factor. Typically, a variance explanation ratio of less than 40% for the first common factor is considered a non-biased result.

The specific test operation is as follows: all the measurement items in the questionnaire are carried out for non-rotated exploratory factor analysis, the common factor is extracted by the principal component analysis (the extraction standard is eigenvalue>1), the orthogonal rotation is carried out by the maximum variation method, the test results are based on the total variation analysis, see Table 4.

Based on the total variance results in Table 4, the single-factor analysis by Harman yields the following conclusions: First, nine common factors with eigenvalues greater than one were identified, indicating that no single factor dominates all item variance and preliminarily excluding the possibility of severe standard-method bias. Second, the initial variance explained by the first common factor (32.106%) is significantly below the 40% threshold, suggesting that item variance is not concentrated in a single factor and standard method bias has minimal interference. Third, the cumulative variance explained by the nine common factors reaches 76.508%, demonstrating complete factor extraction and statistically valid results. In summary, the data in this study show no significant standard-method bias, providing a solid foundation for subsequent empirical analysis.

The KMO value for this study is 0.924, which exceeds the conventional minimum threshold of 0.7. Meanwhile, the approximate chi-square value of Bartlett’s sphericity test reaches 9706.383, with corresponding degrees of freedom at 741, and the significance level (sig) is 0.000, lower than the critical test value of 0.05. This study conducted commonality tests on the sample data, with all item commonalities exceeding 0.6. The above test results indicate that the questionnaire data collected in this study exhibit good construct validity of Broussonetia papyrifera and are suitable for subsequent factor analysis.

The CITC and reliability analysis results of this study are shown in Table 5. The Cronbach’s α values for all scales are above 0.7, while the item-specific CITC values remain below the 0.35 threshold. Notably, removing any single item would lower Cronbach’s α below the original scale’s level, indicating good reliability and high internal consistency.

The AVE and combined reliability results of this study are shown in Table 6. The standardized factor loadings of all nine latent-variable items were >0.5, with AVE values exceeding 0.5 and CR values above 0.7, indicating good representativeness of the latent variables and satisfactory research convergent validity.

This study analyzes discriminant validity using the AVE values for each latent variable, with the results presented in Table 7. There were significant correlations among the nine variables (p < 0.01), with all pairwise correlation coefficients being smaller than the square root of AVE, indicating that the variables exhibit both correlation and independence, demonstrating good discriminant validity and meeting the research requirements.

In this study, the research model with nine variables was constructed by using AMOS24.0 software, as shown in Figure 2:

The test results of the goodness-of-fit indicators for the research model are as follows: X²/df = 1.284 < 3; RMSEA = 0.029 < 0.08; GFI = 0.889 > 0.8; NFI = 0.913 > 0.9; CFI = 0.979 > 0.9; IFI = 0.979 > 0.9; TLI = 0.980 > 0.9. All the above indicator results show good fitness. Based on the goodness-of-fit indicator test results, the research model has a good fit, and no modifications are required.

This study tested the hypothesis relationship H1-12 and verified the model path in Figure 2. The specific test results are shown in Table 8.

Based on the analysis of the above table, this study arrives at the following research hypothesis conclusions:

Among H1-12, only H12, which posits that consumers’ initial trust positively influences consumer citizenship behavior, is not supported; all other hypotheses are validated.

This study employed the Process-Bootstraping method (5,000 bootstrap samples) in SmartPLS4, using technophobia as a moderating variable to examine its effects across two groups (perceived ease of use/perceived usefulness and initial trust). Following Hair et al. (2010) criterion (t > 1.96 or p < 0.05), the results are shown in Table 9. The moderation effect of technophobia on perceived usefulness initial trust was not significant (H14 not supported), while the moderation effect on perceived ease of use-initial trust was statistically significant (H13 supported).

To examine the statistically significant moderating effect, we analyzed the model using the Process-Bootstraping method in SmartPLS4. The simple slope results are shown in Figure 3. Participants with low technophobia (at-1SD) demonstrated stronger predictive power than those with high technophobia (at +1 SD), indicating that technophobia weakens the predictive effect of perceived ease of use on initial trust. Therefore, Hypothesis H13 is supported.

This study confirmed the influence pathways and structural models of human–computer interaction intelligence embodied in AI virtual anchors within the context of technophobia. Of the 14 hypotheses, 12 received support, whereas two did not.

This study found that Hypothesis H12 (“Consumers’ initial trust positively affects consumers’ citizenship behavior”) was not supported. One explanation is that initial trust represents the earliest stage of trust formation (McKnight et al., 2002), dominated by preliminary expectations rather than deep participation. In contrast, consumers’ citizenship behavior entails discretionary actions aimed at improving product or service quality (Gruen, 1995). First, from the perspective of the match between trust levels and behavioral thresholds, initial trust is essentially “cognitive-level trust” and falls under the category of in-role behavior, which can only reduce the psychological barriers to shallow participation behaviors such as asking questions and leaving comments. In contrast, citizenship behavior belongs to “affective-behavioral deep co-creation” and is an extra-role behavior, which requires support from higher-level trust such as emotional connection and identity recognition (Bove et al., 2009). The transition of trust from the cognitive to the affective level cannot be achieved merely through initial contact. Second, when AI virtual anchors first appeared, their limited intelligence generated negative press coverage. Van Der Meer and Brosius (2024) empirically demonstrated that such negative news reduces audiences’ willingness to share information and increases their skepticism. Consequently, even when consumers believed AI virtual anchors could serve them, they remained reluctant to endorse them and showed low tolerance for potential shortcomings.

This study also found that Hypothesis H14 (“Technophobia negatively moderates the relationship between perceived usefulness and consumers’ initial trust”) was not supported. From the perspective of the dimensional specificity of technophobia, technophobia encompasses sub-dimensions such as fear of employment substitution, fear of operational complexity, and fear of technological out-of-control. A plausible reason is that technophobia primarily reflects individuals’ fears of job displacement by new technologies, especially by AI (Jia, 2019). Although AI virtual anchors possess utilitarian capabilities that could threaten human anchors, the sampled consumers may not have perceived their own employment to be at risk; as a result, technophobia did not interfere with the formation of initial trust. Perceived usefulness corresponds to the instrumental value of virtual anchors and has a very weak correlation with consumers’ perceived employment threats, whereas perceived ease of use is directly linked to the fear of operational complexity. This determines that technophobia cannot interfere with the “perceived usefulness-initial trust” link. Meanwhile, the formation of perceived usefulness is based on the objective functional values of AI virtual anchors, such as 24*7 service and precise product matching. Such values are irreplaceable and represent rational value judgments, which are not easily disturbed by subjective emotions like technophobia. In contrast, perceived ease of use depends on the subject’s subjective operational experience. Factors such as age and educational background may create difficulties in learning and using new technologies, leading to a sense of rejection toward new technologies among subjects (Ilona et al., 2018). Moreover, technophobia is also associated with the difficulties and uncertainties of accessing new technologies and new things (Liu, 2011), rather than the actual utility of new technologies and new things. Thus, consumers can acknowledge the usefulness of AI virtual anchors while simultaneously experiencing unease about potentially complex operations; this logic explains why Hypothesis H13, which posits an adverse moderating effect of technophobia on the relationship between perceived ease of use and consumers’ initial trust, was supported.

The core theoretical contribution of this study focuses on the trust formation mechanism and value co-creation pathways in human-computer interaction scenarios from the perspective of technophobia, particularly highlighting the unique moderating role of technophobia in the TAM-trust relationship. This innovative perspective provides a new entry point for research in related fields, specifically manifested in the following three aspects:

First, it extends the applicability and depth of the TAM. This study extends it to the context of “dynamic human-computer interaction agents” such as AI virtual anchors, systematically validating the complete transmission chain of “e-commerce virtual anchor intelligence → perceived usefulness/perceived ease of use → initial trust → value co-creation behavior.” By subdividing intelligence into four independent dimensions—guidance, recognition, analysis, and feedback—the study reveals the differential impact of these dimensions on perceptual variables, thereby enhancing the TAM model’s explanatory power at a finer granularity in AI scenarios.

Second, it clarifies the boundary moderating value of technophobia and enriches its application scenarios. This study is the first to empirically demonstrate that technophobia is not a generalized negative psychological factor that interferes with trust formation but rather specifically moderates the critical boundary condition of “perceived ease of use → initial trust.” By triggering consumer concerns about operational complexity and behavioral hesitation, technophobia weakens the conversion efficiency of ease-of-use perceptions into initial trust, while having no significant impact on the “perceived usefulness → initial trust” transmission. This finding clarifies the mechanism of technophobia in human-computer interaction trust formation, fills a research gap in technophobia theory within the context of e-commerce virtual anchors, and provides a core theoretical explanation for understanding “why some consumers recognize the value of e-commerce virtual anchors but are reluctant to establish trust.”

Third, it refines the mechanism by which initial trust differentially affects value co-creation. This study finds that initial trust can only significantly drive consumer participation behaviors (e.g., low-threshold interactions such as asking questions or leaving comments), but cannot trigger citizenship behaviors (e.g., proactive advocacy or defect feedback, which require deeper participation). It confirms an essential difference in the depth of trust required for the two dimensions of value co-creation—citizenship behaviors demand sustained and profound trust accumulation, which cannot be supported solely by the initial trust formed during first contact. This conclusion enhances the theoretical understanding of the relationship between trust and value co-creation, providing an important reference for future research to distinguish the behavioral effects of trust at different levels.

Based on the research findings, this study provides actionable, practical strategies for e-commerce enterprises, live streaming platforms, and regulatory authorities, focusing on three core directions: “Optimizing virtual anchor AI functions, alleviating consumer technophobia, and tiered guidance for value co-creation behavior.”

First, E-commerce Enterprises: Prioritize upgrading core AI capabilities to enhance consumer perception and experience precisely. E-commerce enterprises need not pursue comprehensive AI upgrades; instead, they should focus on dimensions with greater impact on consumer perception variables, such as recognition intelligence and feedback intelligence. Strengthen demand recognition algorithms to accurately interpret consumer inquiries, optimize human-like feedback phrasing and response speed to reduce operational concerns, and efficiently boost initial consumer trust through targeted investments.

Second, E-commerce Enterprises, Live Streaming Platforms, and Regulatory Authorities: Collaborate to alleviate consumer technophobia and lower the entry barrier for first-time users. E-commerce enterprises may adopt a hybrid live streaming model featuring “human anchors + virtual anchors” as a transitional approach. Live streaming platforms can create simplified interaction guides (e.g., “One-Click Question Tutorials”) and organize experiential activities, such as virtual anchor sales competitions. Regulatory authorities can disseminate information on the principles of virtual anchor technology and its security mechanisms through official channels. This tripartite cooperation aims to mitigate consumer fear of unfamiliar technology and operational hesitations.

Third, E-commerce Enterprises and Live Streaming Platforms: Implement tiered guidance for consumer value co-creation that aligns with consumers’ initial trust levels. During the initial trust phase, e-commerce enterprises and live streaming platforms should focus on low-barrier participation behaviors (e.g., asking questions, commenting, liking) by using interactive prompts and coupon incentives to encourage participation. As consumers build sustained trust, deeper citizenship behaviors—such as defect feedback and proactive promotion—can be encouraged through point rewards and exclusive benefits, thereby avoiding consumer resistance to excessive demands.

First, there is a lack of tracking analysis on the evolution of trust after the first contact. This study relied on a cross-sectional design that captures only static data at the “first-contact” moment; it does not trace how initial trust evolves into sustained trust. Because the mechanism linking first contact to long-term behavior remains unclear, future research should adopt longitudinal designs to map the trajectory from initial to sustained trust. Such studies could track the same consumer cohorts, examine how sustained trust shapes value co-creation behaviors, and provide empirical guidance for the long-term deployment of AI virtual anchors.

Second, other potential interfering variables in the first-contact scenario have not been considered. This study introduced technophobia as a moderating variable, but individual characteristics of consumers (such as AI technology involvement and online shopping experience) and external characteristics of e-commerce virtual anchors (such as anthropomorphism degree and voice style) that may significantly affect the formation of initial trust during the first contact were not taken into account. In future research, the researchers will incorporate key interfering variables in the first-contact scenario to improve the theoretical model and enrich its interpretive dimensions.

Finally, the measurement of technophobia did not distinguish between the differences between “mild concerns” and “severe resistance.” The present study treated technophobia as a single construct and did not distinguish between “mild concerns” (e.g., worry about operational unfamiliarity) and “severe resistance” (e.g., outright aversion to AI). The moderating effect of perceived ease of use on initial trust may vary slightly across these levels, suggesting that the current understanding of technophobia’s moderating mechanisms could be refined. Future research should disaggregate technophobia by severity and test level-specific moderating effects, enabling e-commerce companies to design tailored fear-relief strategies for consumers with differing degrees of technophobia.