Understanding the Continuance Intention for Artificial Intelligence News Anchor: Based on the Expectation Confirmation Theory

Abstract

The Metaverse accelerates the development of the meta-human industry and human-AI interactions in both traditional media outlets and online platforms. As a typical application of meta-human, artificial intelligence (AI) news anchors have been gradually utilized for program reports instead of newscasters in China. In this paper, through the lens of expectation confirmation theory, we establish a conceptual model consisting of perceived anthropomorphism (ANT), perceived intelligence (PI), perceived attractiveness (PA), perceived novelty (PN), information quality (IQ), confirmation of expectation (CE), trust (TRU), and satisfaction (SAT) to explore continuous intention (CI) of watching news reported by AI anchors among online users. By leveraging on a sample of 598 eligible questionnaires, the partial least square structural equation model is employed and the results show that the holistic continuing intention for AI news anchor is positive but not robust. Further analysis indicates that SAT, PI, and TRU can predict CI directly, meanwhile CE, ANT, and PA associate with CI through the mediation of satisfaction. In addition, trust and satisfaction serve as serial mediators between IQ and CI. There is no direct relationship between CE & CI, ANT & CI, and PN & SAT. Nevertheless, user gender and previous experience can moderate the relationships of ANT & CI and PN & SAT, respectively. It can be seen that the proposed model can explain 80.1% of the variance in CI. The implications are intended to provide references for further commercialization of AI news anchors.

1. Introduction

The Metaverse has become the most promising prophecy of contemporary technology. As the prominent constitute of the Metaverse, meta-human has applied in the fields of education, retail, healthcare, gaming, automobiles, etc., while the commercial value has been constantly tapped and released. Specifically, the virtual human had a market value of 11.3 billion dollars in 2021 around the world and is expected to contribute 440.3 billion dollars by 2031, which reveals a strong growth momentum [1]. Among various types of meta-human, virtual anchors have the ability of human-AI interactions to replace reporters to deliver the news, which can reach the widest users and are regarded as the most active field of the meta-human industry in China [2]. The market size of virtual anchors has been valued at 300 million CNY in 2021 and is projected to reach 2.5 billion CNY by 2024 in China [3], demonstrating great potential for future development. Practically, AI news anchors become a typical application of virtual anchors and the representative products in China are shown in

Table 1. Examples of representative AI news anchors in China.

Release Time	Name	Institution
November 2018	Xin Xiaohao	Xin Hua News Agency, Beijing, China
May 2019	Guo Guo	People’s Daily, Beijing, China
October 2021	Time Xiaoni	Beijing Radio & Television Station, Beijing, China
March 2019	Xiao Qing	IFLYTEK Co.,Ltd., Hefei, China

.

The prior AI news anchors’ research mainly focused on retrospection of the origin and historical development [4], application analysis in the industry [5], comparative studies with human announcers [6], or philosophical reflections about their values and worries [7,8], which are primarily subjective conceptual discussions from the perspective of media organizations. Recently, several studies have examined users’ current acceptance and perceptions about the AI entities empirically [9,10,11]. However, long-term attention reflects individual sustained concern and future decisions in the long run. It remains to be unknown about users’ continuous watching intention and long-term social acceptance of the cutting-edge technology. With the consistently expanding market scale of AI news anchors, objective data examining users’ long-term watching intention would be more meaningful for the AI agents.

Based on expectation confirmation model (ECM) and characteristics of AI news anchors, we investigate users’ long-term attention to AI news anchors by conducting an online survey empirically and try to unpack the underlying mechanisms that may give rise to such long-term attention through the key variable of continuance intention (CI), which reflects an individual’s sustainable intention to continue adopting new application. Therefore, we established a conceptual model focusing on users’ continuance intention that is tailored to AI news anchors’ current development state and tested the model. The research questions (RQs) are presented as follows:

• RQ1: Are users willing to continue watching videos of AI news anchors?
• RQ2: What are the direct factors that influence continuance intention for AI news anchors and how do they affect users’ continuance intention?
• RQ3: What are the indirect factors that impact continuance intention for AI news anchors and their influencing mechanisms?

The contributions of this research are summarized as follows: Firstly, our study tentatively contributes to the research intersection of AI news anchors, virtual agents acceptance, and post-adoption studies theoretically by testing a conceptual model incorporated ECM and extended variables depicted the features of AI news anchors. Secondly, we have found that perceived intelligence and trust could impact continuance intention directly. Perceived anthropomorphism exerts a positive influence on continuance intention via the mediation of satisfaction indirectly but impacts continuance intention negatively and directly among female users. Moreover, the direct association between perceived novelty and satisfaction was not supported but was moderated by previous consumption on AI news anchors’ video. Finally, based on these new findings, practical suggestions for media outlets and practitioners are provided to facilitate the long-term success of AI news anchors.

The rest of the article proceeds as follows: Section 2 describes the literature about AI news anchors and expectation confirmation model. Section 3 illustrates the research hypotheses in the proposed model. Section 4 shows the online survey procedure and demographic information. Section 5 presents the data analysis results. Section 6 reveals the findings and discussions. Section 7 concludes and elaborates the implications and limitations of this study.

2. Literature Review

2.1. Meta-human and AI News Anchor

The meta-human can be understood as a digital image created by various converging technologies that exists solely on the screen of the display device. They generally possess human-like physical characteristics and personality traits. They can imitate humans in expression and action, perceive and interact with the external environment using human intellect [12,13]. In essence, they are virtual avatars that resemble human beings in appearance and behavior and are designed to interact with people. Currently, popular products of meta-human in China include virtual idols, virtual employees, and virtual anchors [14], while the industry overseas pays more attention to virtual influencers (VIs), virtual content generation, and virtual humanized assistants/advisors. From hand-painted depictions to computer-generated imagery, from 2D graphic to 3D modeling, the virtual digital human becomes more humanized to create intimate, caring, and immersive interactive experiences. It can be said that the evolutionary logic of the virtual digital human is to highlight human subjectivity and improve users’ experience in human-AI interaction.

As a significant application for meta-human, AI news anchors are generative AI entities that can substitute or supplement human anchors to broadcast the programs in an uninterrupted and multilingual way, which have been widely used in TV programs, live streaming, brand marketing, film, and entertainment. As an emerging digital medium, AI anchors gradually penetrate into the process of social communication. Studies related to the user’s acceptance and usage intention of human-like agents are summarised in

Table 2. Main elements influencing individual acceptance of virtual digital human.

Source	Setting	Method and Participants	Involved Variables	Main Arguments and Findings
Tan (2020) [9]	AI news an- chor	Mixed, interviews with 6 experts and online survey with 458 young adults (aged between 18–35) in China.	Relative advantage, compatibility, observability, social influence, perceived risk, hedonic motivation, personal innovation, behavioral intention, and usage behavior.	- Relative advantage, observability, social influence, hedonic motivation, and personal innovation positively associate with behavioral intention. - Observability, social influence, hedonic motivation, and behavioral intention positively associate with usage behavior.
Wang et al. (2021) [20]	AI news an- chor	Mixed, interviews with 16 news program audience and online survey with a sample of 418 questionn- aires in China.	Perceived novelty, perceived intelligent advantage, perce- ived credibility, perceived appearance anthropomor- phism, perceived agency, attitude, and acceptance.	- Perceived novelty, perceived credibility, perceived appearance anthropomorphism, and perceived agency can positively predict attitude and acceptance. - Attitude can positively predict acceptance.
Xue et al. (2022) [11]	AI news an- chors	Quantitative, 2 online ex- periments conducted in China, experiment 1 and experiment 2 recruited 200 participants respec- tively.	Appearance, gender, voi- ce, perceived attractive- ness, inherent impression of traditional news anchors, and watching intention.	- The appearance, gender, and voice of AI news anchors are all significantly related to perceived attractiveness. - Non-humanoid female AI news anchors with anthropomorphic voices are perceived as most attractive among audiences. - The audience’s inherent impression of traditional news anchors negatively regulates the relationship between perceived attractiveness and watching intention. - The appearance, gender, and voice of AI news anchors influence watching intention through the mediation of perceived attractiveness.
Xiang et al. (2023) [22]	VIs on so- cial media platform	Fuzzy-set qualitative com- parative analysis (FsQCA), conducted in China, cases included 36 virtual humans on RED.	Appearance image, per- sona positioning, under- lying technology, interac- tive application, and accep- tance.	- Persona positioning is the core element to improve public acceptance. - No single elements are sufficient and necessary to generate public acceptance of VIs. - Persona positioning and appearance image are two modes that can improve public acceptance for VIs, while underlying technology and interactive application are important in both configurations.
Wang (2023) [23]	Virtual hosts in variety sh- ow	Mixed, interviews with 3 university students, online survey with 537 young adults (aged between 18–25), and data analysis using bullet subtitles (“danmu”) from video websites, conducted in China.	Perceived novelty, perceived intelligent advantage, perce- ived credibility, perceived appearance anthropomor- phism, perceived agency, attitude, and acceptance.	- Perceived novelty, perceived appearance anthropomorphism, perceived agency pos- itively impact attitude, while perceived intelligent advantage negatively impacts attitude.
Lu et al. (2021) [19]	Virtual You- tubers on You- tube	Qualitative, semi-structured interviews with 21 dedica- ted viewers of virtual You- tubers in China.	None.	- Virtual Youtubers’ appearance and persona, the opportunity to interact with favorite anime characters, strong interests in long series of anime, curiosity, seeking relaxation and experiencing a sense of community are motivations to watch virtual Youtubers’ live streaming.
Lou et al. (2022) [24]	VIs on so- cial media platform	Qualitative, semi- structured interviews with 26 followers of VIs from Singapore.	None.	- There are six primary motivations for users following VIs: novelty, information seeking about VIs, entertainment, surveillance of VIs’ daily life, esthetics, resonance in values and sense of belonging to a community.
Choudhry et al. (2022) [21]	VIs on so- cial media platform	Qualitative, semi- structured interviews with 30 participants from 16 countries who currently follow the selected VIs on social media.	None.	- Content-driven interest, the novelty of VIs, high-frequency interaction with VIs, and visual attractiveness are primary reasons behind user involvement with VIs. - People follow non-human VIs and animated VIs for their niche domains rather than aesthetic value.
Wortelboer (2022) [18]	VIs on so- cial media platform	Qualitative, semi- structured interviews with 29 participants from 14 countries who currently follow the selected VIs on social media.	None.	- Drivers for first-time engagement with VIs: social media engagement, visibility of VIs, hard to distinguish the real identities of VIs, and curiosity for VIs. - Drivers for long-term engagement with VIs: recognition of the true identities of VIs, positive attitudes for VIs, information seeking about VIs, entertainment, and parasocial relationship. - Determinants that drive professionals to engage are significantly different from those who are personally involved.
Taglinger et al. (2023) [25]	Digital assis- tants in on- line stores	Quantitative, online sur- vey with a sample of 174 questionnaires in German.	Performance expectancy, effort expectancy, social influence, hedonic motivation, habit, trust, and behavioral intention.	- Performance expectancy and habit can positively influence behavioral intention.
Sestino et al. (2023) [15]	Digital-based healthcare ser- vices deliver- ed by doctors’ avatars in the metaverse	Quantitative, an interna- tional experiment with a sample of 689 participants from Europe, Northern America, Southern Ame- rica, Asia, Africa.	Perceived anthropomor- phism, emotional recep- tivity, and intention to use.	- Higher-level of human-like interactions (manipulated as virtual avatars in the metaverse) exert an indirect effect on individuals’ intention to use such digital services via the mediation of perceived anthropomorphism. - Such effect only exists among individuals with higher levels of emotional receptivity.
Philip et al. (2020) [16]	Virtual med- ical agent	Quantitative, 2 experiments, participants are outpatients from the Sleep Clinic at the University Hospital of Bordeaux (France), experiment 1 and experiment 2 recruited 179 and 139 participants respectively.	User characteristics (gen- der, age, education, and suspected sleep disorders), usability, satisfaction, bene- volence, credibility, engage- ment, and acceptance.	- The virtual medical agent is found to be more acceptable among older and less- educated patients. - Higher level of trust and acceptance will lead to higher degree of engagement.
Dupuy et al. (2021) [17]	Virtual compa- nion to deliver personalized advice for sleep problems	Quantitative, experiment with a sample of 3479 questionnaires.	Age, gender, education, initial severity of insomnia complaints, familiarity with technologies, length of interaction, pandemic context, trustworthiness, and acceptance.	- Individual’s age, education level, fami- liarity with the technology, trustworthi- ness and length of interaction significant- ly affect user’s acceptance of the virtual companion.

. There mainly exist four categories of elements that impacting individual’s adoption towards virtual digital human, i.e., personal characteristics (e.g., gender, age, usage experience, and professional background) [15,16,17,18], physical characteristics of digital human (e.g., appearance, voice) [11,19], users’ perceptions of virtual human (e.g., novelty, entertainment, aesthetic value, anthropomorphism, and intelligence) [19,20,21], and users’ affective evaluations of digital human (e.g., attitude, satisfaction, and trust) [16,18].

2.2. Expectation Confirmation Model

Devoted to understand users’ continuance intention for AI news anchors, the expectation confirmation model (ECM) proposed by Bhattacherjee is selected as the theoretical framework for this study. ECM was developed from the expectation disconfirmation theory (EDT) put forward by Oliver. The EDT assesses user satisfaction by comparing the expectation before using a certain product with the disconfirmation that arises after the actual use, which is extensively applied in marketing discipline to explain consumer behavior [26,27]. Different from EDT, ECM concentrates on the discipline of information systems and investigates individual post-adoption perception from users’ personal experiences and overall appraisal. Key components in ECM incorporate perceived usefulness, confirmation, satisfaction, and continuance intention. Briefly, ECM posits that (1) satisfaction is determined by perceived usefulness and confirmation jointly; (2) continuance intention is affected by a combination of satisfaction and perceived usefulness; and (3) perceived usefulness is evaluated by confirmation of expectation resulting from previous usage experience [28]. Since the birth of ECM, the theory has been broadly employed in a wide range of areas such as communication services [29], online learning [10], and mobile applications [30], etc., to predict individual continuance intention, which breaks out of the old pattern followed in technology acceptance studies that only pay attention to initial acceptance. In the process of social adoption for AI news anchors, we believe that the users will subconsciously compare their cognitive thoughts before and after watching the videos of AI news anchors and thus generate individual satisfaction. We thereby retain three variables in ECM, namely the confirmation (disconfirmation), satisfaction, and continuance intention, and combine them with the core characteristics of AI news anchors such as intelligence and anthropomorphism to establish the theoretical model, aiming to clarify the mechanism underlying users’ willingness to continue watching AI news anchors.

3. Research Model and Hypotheses Development

Based on the expectation confirmation model (ECM), the proposed model consists of continuance intention (CI), satisfaction (SAT), confirmation of expectation (CE), trust (TRU), perceived anthropomorphism (ANT), perceived intelligence (PI), perceived novelty (PN), perceived attractiveness (PA), and information quality (IQ), which are defined in

Table 3. Definitions of the constructs.

Construct	Definition	Reference
CI	Users’ willingness to continue watching news programs hosted by AI anchors.	Liao et al. [31]
SAT	Users’ overall emotive state resulting from watching videos of AI news anchors.	Bhattacherjee [32]
CE	The extent to which users’ initial expectations about AI news anchors have been met after they watch videos of AI news anchors.	Oliver [26]
TRU	Users’ cognitive trust and emotional trust for AI news anchors.	Glikson et al. [33]
ANT	Users’ perception of AI news anchors’ human-like traits.	Kühne et al. [34]
PI	Users’ cognition of AI news anchors’ competence and performance.	Bartneck et al. [35]
PN	Users’ evaluation of AI news anchors’ novelty degree.	Wells et al. [36]
PA	Users’ perception of whether the appearance of AI news anchors is pleasant or not.	Filieri et al. [37]
IQ	Users’ assessment of the quality of news content broadcasted by AI anchors.	Abbasi et al. [38]

. Hypotheses in the theoretical model are composed of both original hypotheses from ECM (H1&H2a) and new hypotheses developed from the extant literature rather than ECM (H2b-H8).

3.1. Continuance Intention, Satisfaction, and Confirmation of Expectation

Continuance intention, as introduced by Bhattacherjee, refers to an individual’s sustained intention to continue using a particular information technology (IT) based on their prior usage experience [28]. Unlike initial acceptance, continuance intention evaluates usage intention in the long run and is regarded as an effective means of assessing the long-term success of IT [39]. We select it as the outcome variable to explore users’ acceptance of the large-scale and continuous commercialization of AI news anchors.

Confirmation of expectation, or disconfirmation of expectation, is captured by the cognitive dissonance generated from an individual’s comparison of pre-usage expectation to the observed post-usage performance of an IT. Through second-hand information, such as vendor claims or media reports, individuals generally develop certain expectations for an IT before using it, while after actual consumption, their cognitive appraisals can be modified by the observed IT performance from their personal experiences, which illustrates the rationale of confirmation (disconfirmation) of expectation [40]. When the post-usage experience exceeds the original expectations, it is manifested as positive (dis)confirmation; when the post-usage experience fails to meet the original expectations, it is manifested as negative (dis)confirmation; when the IT performs just at the level of the original expectations, it is a confirmation of expectation. The terms confirmation and disconfirmation are two sides of the same concept. In essence, there is no substantial difference between them. Meanwhile in research related to ECM, they are also used interchangeably [28,32,41]. Therefore, we uniformly use the term confirmation of expectation to refer to the confirmation/disconfirmation phenomenon in the expectation confirmation model for the sake of convenience.

Satisfaction occurs after the psychological process of confirmation of expectation, which can be viewed as the affective response derived from the prior IT consumption experience [32] (e.g., the work performance of AI anchors, the quality of news programs). When the actual usage experience meets or exceeds the initial expectations, satisfaction is manifested as pleasurable and positive feelings. And vice versa, when the actual usage experience does not meet the expectations of users, satisfaction is manifested as unpleasant and negative feelings.

Satisfaction (SAT) and confirmation of expectation (CE) both play a central role when forming continuance intention (CI). Derived from the ECM, the interplay between the three constructs was ascertained in Bhattacherjee’s study [28], namely, CE positively influences SAT, and SAT then positively impacts CI. The causal chain has been demonstrated over a wide range of IT continuance contexts. For example, Li et al. found that for chatbot service, CE could positively affect CI through the mediating effect of SAT [42]. Similar conclusions have also been advocated under the contexts like smart fitness wearables [43], online shopping [41], and social networking sites [29]. In addition, studies also investigated the relationship between CE and CI. For example, Venkatesh et al. validated that positive disconfirmation would lead to lower continuous usage intention employing the polynomial model [44]. However, Liao et al. have concluded that in the context of online games, positive disconfirmation had a significant positive relationship with CI [31]. We hypothesize:

Hypothesis 1 (H1). 

Users’ satisfaction with AI news anchors is positively associated with their CI for AI news anchors.

Hypothesis 2 (H2a). 

Users’ extent of confirmation is positively associated with their satisfaction with AI news anchors.

Hypothesis 2 (H2b). 

Users’ extent of confirmation is positively associated with their CI for AI news anchors.

3.2. Trust

Prior research suggests that the concept of trusting beliefs in the field of artificial intelligence have two dimensions [33], one commonly addressed is referred to as cognitive trust, which is embedded in individuals’ rational and deliberative evaluation of the trustworthiness of evidence [33,45]. Relevant clues that may contribute to cognitive trust toward AI news anchors include high-quality information, objective and unbiased attitude, professional competence, etc. Another less commonly addressed is known as emotional trust. It can be conceived as the extent to which one feels secure and comfortable to obtain news via AI anchors [46]. Emotional trust generally exists in interpersonal relationships for the emotional bonds between individuals [47]. Considering the similarity between the interactions of humans and AI anchors and interpersonal relationships, we incorporate emotional trust to examine the irrational factors in human-AI trust relationships. We thus conceptualize trust for AI news anchors as a combination of both cognitive trust and emotional trust.

When it comes to the adoption of AI news anchors, trust serves as the important foundation for successful interaction between humans and AI anchors and had been validated as a critical predictor of usage intention or CI in similar research contexts such as chatbots [48], information system [49], fake news spreading [45], and social media influencer [50]. Moreover, studies have also confirmed that trust can affect CI through the mediation of satisfaction. For example, Wang et al. pointed out satisfaction acted as a mediating variable between platform trust and continuous usage intention in the context of a crowdsourcing competition platform [49], Likewise, consumer trust was proved to be significantly correlated with their satisfaction of e-commerce [51]. Therefore, we hypothesize:

Hypothesis 3 (H3a). 

Users’ extent of trust is positively associated with their CI for AI news anchors.

Hypothesis 3 (H3b). 

Users’ extent of trust is positively associated with their satisfaction with AI news anchors.

3.3. Perceived Anthropomorphism

As one of the most salient variables in human-robot interaction research, perceived anthropomorphism denotes an individual’s perception of human-like characteristics of a non-human entity [34]. The measurement for anthropomorphism is multifaceted, which includes but not limited to physical characteristics such as appearance, movement, facial expression, voice, and gesture [35,52], psychological characteristics such as intention, emotion, and cognition [53], and social cues such as name, gender, communication style, and personality [54,55]. According to the current development of AI news anchors, we mainly assess anthropomorphism from aspects of appearance, body movement, facial expression, voice, hosting style, and personality traits.

Despite the uncanny effect, most research still indicates that a higher level of anthropomorphism can trigger positive feelings about AI [56]. For example, Jang et al. identified that anthropomorphism had a positive effect on satisfaction for voice shopping via smart speaker [57]. To go further step, due to the humanoid features of AI anchors, such as unique names, avatars, and hosting styles, users may get induced to the characters exhibited by AI news anchors, thus forming a stronger willingness to continue watching. As an instance, in a comparable research setting, Balakrishnan et al. reported that perceived anthropomorphism had a significant and positive effect on users’ CI towards service robots [58]. Besides, as one of the most notable features of AI news anchors, the higher the degree users’ perception of anthropomorphism is, the more likely it is to surpass their original expectations for AI anchors, which ultimately leads to a greater extent of confirmation. To illustrate, Moussawi et al. suggested that perceived anthropomorphism could exert significant influence on CE regarding personal intelligent agents [59]. Hence, we hypothesize:

Hypothesis 4 (H4a). 

Users’ perception of anthropomorphism about AI news anchors is positively associated with their satisfaction with AI news anchors.

Hypothesis 4 (H4b). 

Users’ perception of anthropomorphism about AI news anchors is positively associated with their CI for AI news anchors.

Hypothesis 4 (H4c). 

Users’ perception of anthropomorphism about AI news anchors is positively associated with their extent of confirmation of AI news anchors.

3.4. Perceived Intelligence

Perceived Intelligence (PI) mostly depends on individual perception of the performance and competence of an AI-powered system [35]. Generally speaking, when intelligent entities exhibit qualities like problem-solving, goal-achieving, reasoning, learning, and adaptability to the environment, people shall perceive intelligence [60]. For AI news anchors, we evaluate PI from dimensions of professional competence, degree of intelligence, professional skill, sensibility, and agency [35,61,62,63].

As one of the most prominent features of AI-powered IT, perception of intelligence helps to generate more positive cognition in human-AI interactions [64]. In particular, Moussawi et al. revealed that a higher degree of PI could result in a higher degree of trust in personal intelligent agents [65]. While in the long run, functionality will be crucial when making further decisions. To elaborate, Balakrishnan et al. verified that PI would affect people’s CI for chatbot services [58]. Furthermore, intelligence as another core trait of AI news anchors has also been found to have a notable impact on CE [54,59]. In sum, we hypothesize:

Hypothesis 5 (H5a). 

Users’ perception of intelligence about AI news anchors is positively associated with their trust in AI news anchors.

Hypothesis 5 (H5b). 

Users’ perception of intelligence about AI news anchors is positively associated with their CI for AI news anchors.

Hypothesis 5 (H5c). 

Users’ perception of intelligence about AI news anchors is positively associated with their extent of confirmation of AI news anchors.

3.5. Perceived Novelty

Novelty reflects an individual intrinsic motivation to seek freshness and different experiences when adopting new things. When people find an IT innovation to be new, interesting, and different from existing technology, they may feel excitement and curiosity, so as to perceive the novelty [66]. The pursuit of novelty is related to people’s desire for more choices, which can evoke pleasant emotions and experiences, and is reckoned as an influential factor when making choices [36]. Talukdar et al. suggested that perceived novelty had a significant positive correlation with consumers’ satisfaction with VR devices [66]. In a similar vein, Kim et al. also empirically validated that perceived novelty had a positive impact on users’ satisfaction in the area of service robots [67]. Given that an AI news anchor is a brand new interactive tool for users to access news information, we hypothesize:

Hypothesis 6 (H6). 

Users’ perception of novelty about AI news anchors is positively associated with their satisfaction with AI news anchors.

3.6. Perceived Attractiveness

Perceived attractiveness describes the attractiveness of the communicator’s face, appearance, and image, which can be defined as the individual’s perception of whether the physical appearance of AI anchors is pleasant or not [37]. Existing research has demonstrated that PA plays an essential role in forming impressions and changing attitudes, which is commonly used in advertising and marketing studies to measure the effectiveness of communicators such as brand endorsers and social influencers [68]. Wiedmann et al. concluded that influencers’ attractiveness was significantly correlated with consumers’ satisfaction towards luxury brands [69]. Li et al. reported a positive relationship between attractiveness and image satisfaction among Weibo users in the area of influencer marketing [70]. Whereas this study posits that the pleasant external image of AI anchors can evoke pleasant and positive emotions among users. We thereby hypothesize:

Hypothesis 7 (H7). 

Users’ perception of attractiveness about AI news anchors is positively associated with their satisfaction with AI news anchors.

3.7. Information Quality

Information quality (IQ) implies users’ evaluation of the quality and characteristics of the content broadcasted by AI anchors, which is a multi-dimensional variable. Under the context of AI news anchors, this study focuses on the credibility, accuracy, informativeness, and clarity of information [71]. While the format of news programs may change over time, the primary purpose of watching news programs remains to obtain information. Therefore, we introduce information quality to measure the content attributes of AI anchors. Currently, research has demonstrated that IQ is closely related to trust [72]. Le et al. suggested that a higher level of information quality of vloggers’ videos could lead to stronger trust in electronic word-of-mouth [73]. McKnight et al. also ascertained the significant and positive association between IQ and trust regarding e-commerce trade [74]. We hypothesize:

Hypothesis 8 (H8). 

The information quality of news broadcasted by AI anchors is positively associated with users’ trust in AI anchors.

In conclusion, the proposed model and assumptions are depicted in Figure 1.

4. Method

4.1. Stimuli

Considering the fact that AI news anchors have not yet been popularized among the public, to ensure respondents have an intuitive understanding of AI news anchors, we provided an introduction to AI news anchors and a 3 min and 30 s video of AI anchors delivering news at the beginning of the questionnaire, meanwhile set a countdown of 3 min and 30 s on the same page to guarantee the quality of filling behavior. By extensively looking through videos of AI news anchors, related literature, news and information, industry reports, and ranking lists, we chose hyper-realistic 3D meta-human “Xiao C” from China Media Group (CMG) and “Shen Xiaoya” from Shanghai Media Group (SMG) as representatives of virtual anchors. The snapshots of the AI anchors are shown in Figure 2. Next, we edited their programs “C+ Moment of the Two Sessions” (presented by “Xiao C”) and “Good Morning Metaverse” (presented by “Shen Xiaoya”) into a total of 3 min and 30 s video as the stimulus material. We selected them as the stimulus materials mainly because: (1) “Xiao C” and ”Shen Xiaoya“ rank the first and third place respectively in China meta-human Influence Index Report in 2022 [14]. 1 What is more, the two virtual anchors are from CMG and SMG, respectively. As they cater to both nationwide and local audiences, their impacts span a wide range of areas; (2) “C+ Moment of the Two Sessions” and ”Good Morning Metaverse” were launched in March 2022 and January 2023, respectively, which are relatively up-to-date and can reflect the latest technical progress; (3) Extant work showed that non-humanoid female AI news anchors with anthropomorphic voices are more attractive [11], plus given the uncanny effect, two female virtual anchors with natural voices and relatively lower levels of anthropomorphism were selected as research materials to ensure that participants could fully experience differences between AI news anchors, but will not result in divergence of subsequent choices due to excessive differences between virtual anchors.

4.2. Procedure

To examine users’ watching intention for AI news anchors, we conducted an online survey from 25 March 2023 to 13 April 2023 by means of convenient sampling, which ultimately collected 692 questionnaires. We collected data in two channels: (1) Posted recruiting posters and links of the questionnaire on social media platforms (i.e., Wechat, QQ, Douban, Weibo, RED) (n = 646); (2) Recruited participants on Douban and Weibo, respondents who passed the quality check could get 2 CNY as a reward (n = 46). To control the quality of the collected sample, the following questionnaires were excluded as invalid: (1) the response time is too short (<300 s) or too long (>2100 s) (n = 19); (2) all items in the scale are exactly the same (n = 75). After screening, a total of 598 valid questionnaires remained, with a recovery rate of 86.4%. According to the rules of thumb for the structural equation model, the sample size should be larger than 10 times the maximum measurement equation or maximum structural equation, and a sample of 598 was sufficient for stable estimates [75]. In terms of academic ethics, we added a consent form in the introduction of the questionnaire. The use of data was approved by all respondents. Respondents’ privacy and collected data were strictly protected during the whole research process.

4.3. Participants

Of the 598 participants, 299 (50%) are male and 299 (50%) are female. Most respondents are well-educated, with 450 people (75.3%) having a bachelor’s degree or above (including current students). The age distribution is somehow balanced, with the largest proportion of people aged 18–25 years old (34.6%), followed by those aged 41–50 years old (23.9%) (n = 143) and those aged 31–40 years old (n = 134, 22.4%). The respondents are from 27 provinces (n = 595) in China, Europe (n = 2), and Australia (n = 1). 179 people have watched videos of AI news anchors before (29.9%), most of them never pay attention to the AI-powered agents (n = 224, 37.5%) or occasionally look through related information/watch related videos (n = 307, 51.3%). 352 participants reported that they knew a bit about AI news anchors (58.9%). 208 respondents (34.8%) reported that they had never heard of AI news anchors.

4.4. Measures

Referring to Rahman’s study [76], we employed a six-point scale to measure the proposed variables (1 = “strongly disagree”, 6 = “strongly agree”). This format forces the participants to give either a positive response or a negative response, avoiding the neutral attitude that Asians often hold [76]. The detailed questionnaire items and sources are shown in

Table A1. Constructs and items.

Items	Source
Perceived Novelty (PN)
PN1: For me, watching videos of AI anchors is a novel experience.	Wang, Wu, and Wang, 2021 [4]; Dang, 2020 [95]; Self-developed.
PN2: I find AI news anchors introduced a novel perspective to my views about hosts.
PN3: I think AI news anchors are new and refreshing.
PN4: I feel the news programs hosted by AI anchors are novel, intriguing, and flexible. ∗
Perceived Attractiveness (PA)
PA1: I think AI anchors are good-looking.	McCroskey & McCain, 1974 [96].
PA2: I feel that AI news anchors are groomed in a decent and elegant way.
PA3: I like the physical appearance of AI anchors.
PA4: Generally speaking, I consider AI anchors to be attractive.
Perceived Intelligence (PI)
PI1: I feel AI anchors are qualified for the hosting and broadcasting of news programs.	Bartneck, Kulić, Croft et al. 2009 [35]; Balakrishnan & Dwivedi, 2021 [61]; Zhu, Li, Nie et al. 2021 [62]; Self developed.
PI2: I think AI anchors can report news in an efficient and intelligent way (e.g., AI anchors can quickly convert text into audio, can switch hosting scenes at will).
PI3: I believe AI anchors have the professional skills required to host a news program.
PI4: I consider AI anchors to have good skills in hosting.
PI5: I feel that I can resonate with the news delivered by AI anchors. §
PI6: I feel AI anchors reply sensibly and manner well when interacting with people.
PI7: I find AI anchors to be stable in their work and available for 24/7 broadcasting. ∗ ℓ
Perceived Anthropomorphism (ANT)
ANT1: I think the physical appearances of AI anchors are human-like.	Bartneck, Kulić, Croft et al. 2009 [35]; Moussawi, Koufaris, & Benbunan-Fich, 2022 [54]; Self-developed.
ANT2: I feel the voices of AI anchors are natural.
ANT3: I consider AI news anchors to have an amiable tone of voice.
ANT4: I find that AI anchors have elegant and smooth body movements when hosting programs.
ANT5: I find the facial expressions of AI anchors vivid and natural when hosting programs.
ANT6: I think AI anchors have their own hosting styles. ∗
ANT7: I feel AI anchors are friendly. ℓ
Confirmation Expectation (CE)
CE1: I feel AI anchors are smoother than I expected. ∗	Bhattacherjee, 2001 [28]; Self developed.
CE2: I feel the physical appearances of AI anchors are more vivid and lifelike than I expected. ∗
CE3: I find AI anchors’ job performances are better than I expected. ∗
CE4: Overall, most of my expectations of AI news anchors were confirmed.
Information Quality (IQ)
IQ1: I believe the news information presented by AI anchors is reliable.	DeLone & McLean, 1992 [71].
IQ2: I think the news information presented by the AI anchors is accurate and there is no slip of the tongue.
IQ3: I feel the news information presented by AI anchors is informative, enriching my knowledge and cognition.
IQ4: I find the AI anchors host the news program with standard pronunciation and clear articulation, and the voice is easy to understand.
Trust (TRU)
TRU1: I believe AI news anchors are reliable. ℘	Komiak & Benbasat, 2006 [46].
TRU2: I believe AI anchors report the news without bias.
TRU3: I feel AI news anchors have integrity and honesty, delivering the news content faithfully.
TRU4: I feel secure about accessing news via AI news anchors.
TRU5: I feel comfortable accessing news via AI news anchors.
TRU6: I feel content about accessing news via AI news anchors.
Satisfaction (SAT)
SAT1: I think it is a wise choice to use AI anchors in news programs.	Bhattacherjee & Lin, 2015 [32]; Isaac et al. 2019 [97]; Li, Lee, Emokpae et al. 2021 [42].
SAT2: I find it pleasing to watch videos of AI anchors. †
SAT3: I feel satisfied with the work performances of AI anchors.
SAT4: I like AI news anchors.
SAT5: In general, the news programs hosted by AI anchors are satisfactory. †
Continuance Intention (CI)
CI1: I will continue watching news programs hosted by AI anchors.	Bhattacherjee, 2001 [28]; Ashfaq, Yun, Yu et al. 2020 [98].
CI2: The probability that I re-watch news programs hosted by AI anchors is high in the future.
CI3: Compared with traditional news programs, I prefer to watch news programs with AI anchors.
CI4: If I have a chance, I will recommend AI news anchors to others. †
CI5: I will try to use AI anchors to access news information in the future.

.

5. Data Analysis Results

In this study, partial least squares structural equation modeling (PLS-SEM) was used by SmartPLS 3, which aims to maximize the explanatory capacity (i.e., $R^2$) of the endogenous latent variables and is suitable for theoretical prediction and exploratory research [77].

5.1. Measurement Model

First, the measurement model was examined by the PLS algorithm in SmartPLS 3, with maximum iterations and stop criterion setting at 1000 and ${10}^{-7}$, respectively. Indicators of acceptable reliability and validity include [78,79,80]: (1) factor loading > 0.7, (2) Cronbach’s alpha ($\alpha$) and composite reliability (CR) > 0.7, (3) average variance extracted (AVE) > 0.5, (4) the square root of AVE value of each construct > correlation coefficients between the given construct and other constructs (Fornell–Larcker criterion).

Table 4. Construct reliability and validity.

Construct	Item	Loadings	Cronbach’s $\mathit{\alpha }$	CR	AVE
ANT	6	0.782–0.882	0.917	0.935	0.707
CE	4	0.858–0.901	0.906	0.934	0.780
CI	4	0.893–0.939	0.937	0.955	0.842
IQ	4	0.819–0.863	0.858	0.903	0.700
PA	4	0.823–0.932	0.904	0.933	0.777
PI	5	0.807–0.858	0.891	0.920	0.696
PN	4	0.793–0.895	0.872	0.911	0.720
SAT	3	0.915–0.918	0.905	0.941	0.841
TRU	5	0.811–0.893	0.916	0.937	0.749

compares the reliability and convergent validity of each variable. It can be seen that the factor loadings of all items range from 0.782 to 0.939 > 0.7, demonstrating that the questionnaire items have good indicator reliability. The values of the Cronbach $\alpha$ and CR range from 0.858 to 0.937 and 0.903 to 0.955 respectively, all of which are greater than 0.7, indicating that the constructs have stable internal consistency. The convergent validity is satisfactory, with AVE values ranging from 0.696 to 0.842 > 0.5. For discriminant validity, we employed the Fornell-Larcker criterion. Results are presented in

Table 5. Discriminant validity (Fornell–Larcker criterion).

	ANT	CE	CI	IQ	PA	PI	PN	SAT	TRU
ANT	0.841
CE	0.789	0.883
CI	0.731	0.738	0.917
IQ	0.640	0.738	0.641	0.837
PA	0.784	0.718	0.717	0.568	0.881
PI	0.791	0.693	0.696	0.680	0.747	0.834
PN	0.649	0.646	0.631	0.536	0.748	0.638	0.849
SAT	0.775	0.772	0.894	0.667	0.720	0.710	0.632	0.917
TRU	0.723	0.746	0.774	0.792	0.668	0.668	0.614	0.816	0.865

. We can see that the square root of AVE values of all constructs (in bold on the diagonal) are greater than the inter-construct Pearson correlations (off-diagonal values), in which discriminant validity is confirmed.

5.2. Structural Model

Next, we further evaluated the structural model via bootstrapping for a two-tailed test (with settings of 5000 sub-samples and significance level at 5%) and blindfolding at the default setting.

Table 6. Hypotheses testing.

Hypothesis	Path	Path Coefficient	T-Statistic	${\mathit{f}}^2$	p-Value	95%CI a	Result
H1	SAT → CI	0.716	15.215	0.663	0.000	[0.624, 0.808]	Supported
H2a	CE → SAT	0.183	3.336	0.041	0.001	[0.079, 0.291]	Supported
H2b	CE → CI	0.066	1.688	0.007	0.091	[−0.010, 0.143]	Not
H3a	TRU → CI	0.087	2.060	0.012	0.039	[0.006, 0.170]	Supported
H3b	TRU → SAT	0.437	9.924	0.300	0.000	[0.348, 0.523]	Supported
H4a	ANT → SAT	0.211	3.841	0.049	0.000	[0.102, 0.312]	Supported
H4b	ANT → CI	−0.016	0.332	0.000	0.740	[−0.113, 0.078]	Not
H4c	ANT → CE	0.642	15.436	0.422	0.000	[0.561, 0.726]	Supported
H5a	PI → TRU	0.240	6.017	0.091	0.000	[0.163, 0.319]	Supported
H5b	PI → CI	0.096	2.500	0.017	0.012	[0.022, 0.174]	Supported
H5c	PI → CE	0.185	3.707	0.035	0.000	[0.080, 0.276]	Supported
H6	PN → SAT	0.022	0.584	0.001	0.559	[−0.056, 0.095]	Not
H7	PA → SAT	0.114	2.555	0.015	0.011	[0.022, 0.198]	Supported
H8	IQ → TRU	0.629	15.898	0.622	0.000	[0.546, 0.700]	Supported

Note: ${}^a$ 95%CI is short for confidence interval computed at the 95% level.

demonstrates the results of hypotheses testing. As we can see from the graph, SAT positively affects CI ($\beta$ = 0.716, p < 0.001, 95%CI = [0.624, 0.808]), so H1 was verified. For CE, it is shown to have a positive correlation with SAT ($\beta$ = 0.183, p < 0.01, 95%CI = [0.079, 0.291]) but not directly related to CI ($\beta$ = 0.066, p = 0.091, 95%CI = [−0.010, 0.143]), so H2a was validated but H2b was not. For TRU, the data analysis shows that it can positively predict CI ($\beta$ = 0.087, p < 0.05, 95%CI = [0.006, 0.170]) and SAT ($\beta$ = 0.437, p < 0.001, 95%CI = [0.348, 0.523]), and therefore, H3a and H3b were supported. H4a hypothesized that ANT and SAT had a positive and significant relationship, which was verified by the calculated results ($\beta$ = 0.211, p < 0.001, 95%CI = [0.102, 0.312]). H4b proposed there was a direct and positive association between ANT and CI, but was not supported ($\beta$ = −0.016, p = 0.740, 95%CI = [−0.113, 0.078]). H4c was confirmed for that higher ANT would lead to stronger CE ($\beta$ = 0.642, p < 0.001, 95%CI = [0.561, 0.726]). For PI, it positively predicted TRU ($\beta$ = 0.240, p < 0.001, 95%CI = [0.163, 0.319]), CI ($\beta$ = 0.096, p < 0.05, 95%CI = [0.022, 0.174]), and CE ($\beta$ = 0.185, p < 0.001, 95%CI = [0.080, 0.276]), thus H5a, H5b, and H5c were ascertained. H6 presumed that PN and SAT were positively and significantly correlated, but the results did not support the assumption ($\beta$ = 0.022, p = 0.559, 95%CI = [−0.056, 0.095]). H7 predicted that PA and SAT had a positive and significant correlation, and the results were consistent with the assumption ($\beta$ = 0.114, p < 0.05, 95%CI = [0.022, 0.198]). H8 hypothesized that IQ positively and significantly influenced TRU and the results were supportive ($\beta$ = 0.629, p < 0.001, 95%CI = [0.546, 0.700]). To sum up, 11 of 14 assumptions were verified by the surveyed data.

In Table 6, the effect size ($f^2$) is also presented. The $f^2$ value is used to measure the effects of the independent variable on the dependent variable in each pair of paths, with 0.02, 0.15, and 0.35 representing weak, medium, and strong effect sizes respectively [81]. The results show that the effect sizes of SAT on CI, ANT on CE, and IQ on TRU are substantial. Trust has a medium effect size on satisfaction, while the effect sizes of CE on SAT, ANT on SAT, PI on TRU, and PI on CE are relatively weak.

To further assess the effectiveness of the proposed model, the coefficient of determination ($R^2$) and predictive relevance ($Q^2$) are adopted to evaluate the predictive and explanatory capacity of the model, as shown in

Table 7. Values of $R^2$ and $Q^2$.

Construct	${\mathit{R}}^2$	${\mathit{R}}^2$ Adjusted	${\mathit{Q}}^2$
CI	0.812	0.810	0.677
SAT	0.758	0.756	0.631
TRU	0.659	0.658	0.490
CE	0.635	0.634	0.490

. The $R^2$ metric attempts to gauge the collective impacts of exogenous variables on a given endogenous variable, which can be categorized as weak, moderate, and substantial predictive accuracy based on values of 0.25, 0.5, and 0.75, respectively [82]. The $Q^2$ statistic is used to measure predictive relevance. A $Q^2$ value greater than zero indicates that the associated exogenous constructs have predictive relevance for a particular endogenous latent variable [78]. In the present study, TRU, SAT, and PI can predict 80.1% of the variance in CI, and combined effects of TRU, CE, PA, and ANT account for 75.6% of the variance in SAT, both of which exhibit strong predictive accuracy. While 65.8% of the variance in trust can be explained by PI and IQ, 63.4% of the variance in CE can be explained by PI and ANT. The adjusted $R^2$ for both TRU and CE remain at a moderate level of predictive accuracy. In addition, the $Q^2$ values of all endogenous latent variables are all above zero, suggesting that the relevant paths are predictive for each endogenous latent variable. In sum, the structural model with path coefficients and the explained variance of each endogenous construct are illustrated in Figure 3.

5.3. Mediation and Moderation Analysis

To further clarify the relationships between the constructs and understand why H2b, H4b, and H6 were not verified, we conducted mediation and moderation analysis. PLS-SEM measures the strength of mediating effect via the metric of variance accounted for (VAF), which is calculated by the ratio of indirect effects to total effects. A VAF larger than 80% represents a complete mediation effect, and 20% < VAF < 80% represents a partial mediation effect [83].

Table 8. Mediation analysis.

Indirect Path	Indirect Effect	p-Value	Total Effects	p-Value	VAF
ANT → SAT → CI	0.151	0.000	0.262	0.000	0.576
CE → SAT → CI	0.131	0.002	0.197	0.001	0.665
IQ → TRU → SAT → CI	0.197	0.000	0.252	0.000	0.782
PA → SAT → CI	0.082	0.012	0.082	0.012	1.000

displays the mediated paths with relatively strong mediating effects in this study (VAF > 50%). From which it can be seen that although neither ANT nor CE is directly and significantly associated with CI (H4b and H2b), they both positively affect CI through partial mediation of satisfaction (${\beta }_{ANT}$ = 0.151, $p_{ANT}$ < 0.001; ${\beta }_{CE}$ = 0.131, $p_{CE}$ < 0.01). In addition, SAT serves as a full mediator between PA and CI ($\beta$ = 0.082, p < 0.05). IQ indirectly impacts CI through a partial mediation of trust and satisfaction ($\beta$ = 0.197, p < 0.001).

Referring to the study [84], the researchers conducted a multi-group analysis (MGA). MGA is primarily used to assess the path differences between different groups of categorical moderator variables, which is a useful tool for analyzing moderation effects in PLS-SEM [85]. Specifically, we set maximum iterations and bootstrapping sub-samples to 1000 and 5000 respectively, and divided the six moderators (i.e., gender, age, education, attention to AI news anchors, knowledge about AI news anchors, and previous consumption of AI news anchors’ video) into two groups. The calculated results are presented in

Table 9. Multi-group analysis.

Path	Moderating Variable and Path Coefficient among Different Groups		Path Coefficient-Diff
	gender
	male (n = 299)	female (n = 299)	(male-female)
ANT → CI	0.101	−0.130 *	0.231 *
	previous consumption of AI news anchors’ video
	used to watch (n = 179)	never (n = 419)	(used to watch-never)
PN → SAT	0.157 *	−0.019	0.176 *
ANT → CI	−0.079	0.014	−0.093
CE → CI	0.188 *	0.024	0.165
PI → CI	0.111	0.093	0.018
TRU → CI	0.045	0.087	−0.042
	attention to AI news anchors
	high attention (n = 67)	low attention (n = 531)	(high-low attention)
ANT → CI	0.138	−0.028	0.166
CE → CI	0.029	0.071	−0.042
PI → CI	0.207	0.090 *	0.116
TRU → CI	0.084	0.087 *	−0.003
	knowledge about AI news anchors
	know well (n = 38)	know little (n = 560)	(know well-know little)
ANT → CI	0.054	−0.009	0.063
CE → CI	−0.159	0.070	−0.229
PI → CI	0.316	0.081 *	0.235
TRU → CI	−0.015	0.095 *	−0.109

Note: 1. * p < 0.05. 2. “High attention” represents participants who look through AI news an-chors’ videos/information from time to time or frequently, while “low attention” consists of participants who look through AI news anchors’ videos/information occasionally or ones who never pay attention to relevant videos/information. 3. “Know well” includes respon-dents who chose “know very well about AI news anchors” or ones who chose “have a good understanding about AI news anchors”, while “know little” includes respondents who chose “know a bit about AI news anchors” or “never heard of AI news anchors”.

. The results show that the influence of ANT on CI (H4b) is moderated by gender ($\Delta \beta$ = 0.231, p = 0.017 < 0.05). Only in the female group ($\beta$ = −0.130, p = 0.046 < 0.05) dose ANT negatively predict CI, whereas this effect is not significant in the male group ($\beta$ = 0.101, p = 0.135). Similarly, the relationship between PN and SAT (H6) is moderated by previous consumption of AI news anchors’ video ($\Delta \beta$ = 0.157, p = 0.034 < 0.05), which asks whether participants have previously watched AI news anchors’ videos or not. H6 only shows significance in the group that had previously watched the AI news anchor videos ($\beta$ = 0.157, p = 0.024 < 0.05), and statistically insignificant for those who had never watched before ($\beta$ = −0.019, p = 0.662). Besides, age, education, attention to AI news anchors, and knowledge about AI news anchors, were not validated as moderating variables. Nevertheless, it is worth noticing that for attention to AI news anchors and knowledge about AI news anchors, the sample size of “high attention” (n = 67) and “know well” (n = 38) is small (see Table 9 for details). These groups of participants are likely to have sufficient exposure to AI news anchors. The uneven distribution of such respondents in the collected sample may somehow impact the results of MGA when evaluating the moderating effects of the two moderators.

6. Discussion and Interpretation

6.1. The Holistic Continuous Watching Intention Is Positive but Not Robust

RQ1 attempts to explore users’ sustained watching intention of AI news anchors in the future. While based on the 598 questionnaires returned, the cumulative percentages of respondents who scored “agree”–“strongly agree” (“4”–“6”) for CI items presented in

Table 10. Descriptive statistics for continuance intention (N = 598).

	CI1		CI2		CI3		CI4
Score a	Frequency	Ratio	Frequency	Ratio	Frequency	Ratio	Frequency	Ratio
1	29	4.8%	27	4.5%	40	6.7%	31	5.2%
2	59	9.9%	69	11.5%	92	15.4%	61	10.2%
3	127	21.2%	106	17.7%	148	24.7%	87	14.5%
4	196	32.8%	206	34.4%	165	27.6%	224	37.5%
5	146	24.4%	146	24.4%	118	19.7%	159	26.6%
6	41	6.9%	44	7.4%	35	5.9%	36	6.0%
Mean	3.83		3.85		3.56		3.88
SD ${}^b$	1.239		1.250		1.305		1.231

Note: ${}^a$ Score “1”–“6” represents “Strongly disagree”–“Strongly agree”. ${}^b$ SD denotes standard deviation.

is 64.0%, 66.2%, 53.1% and 70.1%, respectively. The ratio of users who hold a positive attitude and a negative attitude is about 6:4, which suggests an overall positive continuance intention for AI news anchors. Nevertheless, the adoption of innovations follows a social process that requires some time for people to get accustomed to the inclusion of the AI-generated avatars in news programs. Specifically, it is worth noticing that the mean value of CI3 is significantly lower than that of the other three items, which somehow reflects the fact that when confronted with alternative options (namely traditional news programs), users would opt for more conservative choices. Nonetheless, over half of users still exhibit a favorable attitude towards sustained adoption of the emerging communication form during the comparison process. Additionally, the mean value of CI5 is higher than the other three, indicating that over 70% of users are willing to try to access news via AI news anchors in the future after experiencing the new technology.

6.2. The Impact Paths of Direct Factors

To further improve users’ continuous watching intention for AI news anchors, we examined the relevant influencing factors. RQ2 seeks to figure out the direct factors of CI and the associated impact paths. Results showed that SAT, PI, and TRU can directly and positively influence CI. Among which, SAT was proved to be the key driver for CI (H1, $\beta$ = 0.716, $f^2$ = 0.663), which supported Bhattacherjee’s [28] original finding.

Meanwhile, PI and TRU also have direct and positive associations with CI, which is similar to the research of Balakrishnan et al. [58] and Nguyen et al. [48], respectively. For PI → CI (H5b), since the primary purpose of watching news programs is to obtain information, when viewers feel AI news anchors are competent for the news programs, they will be more willing to continue adopting the AI entities in the future. For TRU → CI (H3a), given that AI news anchors are manipulated by computer programs, problems such as inaccurate news and algorithmic discrimination may arise during the communication process. Whereas a sense of trust can reduce users’ perceptions of risk, worry, and uncertainty [86], trust in news media can even impact users’ daily decisions [45]. Therefore, individuals with a high level of trust in AI news anchors will also be inclined to keep adopting them in the future.

6.3. The Influence Mechanisms of the Indirect Factors

RQ3 investigates the indirect influencing factors of CI and their impact mechanisms. The findings revealed that ANT, PA, CE, and IQ indirectly affected CI via different mediators (see Table 8 for details). Therein, satisfaction, as the most salient predictor of CI, plays a positive mediating role between ANT & CI, CE & CI, and PA & CI.

In light of ANT, the relationship of ANT → SAT → CI (H4a & H1) can be explained by the homophily principle, which states that individuals tend to interact with people who share similar characteristics with them [87]. To elaborate, when users perceive sufficient anthropomorphic social cues (e.g., similar appearance, natural voice, smooth movements) while watching the AI news anchors’ video, their affective attitudes of the AI entities will be more positive, which in turn trigger increased future watching intention. Moreover, the direct relationship of ANT → CI (H4b) was not supported for the whole investigated population but the multi-group analysis (Table 9) demonstrates that ANT negatively affects CI among female participants. One possible explanation is the uncanny valley effect proposed by Masahiro Mori. The theory posits that when non-human entities such as robots look similar to real humans, but do not perfectly fit real human’s appearance and movement, the non-human features (e.g., the movements of the AI anchor’s lips do not synchronize with the sounds of their words) could lead to eerie sensation and thus attenuate further use and exposure [88]. Besides, existing research demonstrated that women were more sensitive to anthropomorphic features and are more likely to feel anxious and worried than men when faced with robots with the same level of anthropomorphism [89]. Therefore, a higher level of anthropomorphism can indirectly lead to a stronger continuance intention, but for the female population, the conclusion may be reversed.

In terms of CE, it can not impact CI directly (H2b) but exerts an influence via the mediation of SAT (i.e., CE → SAT → CI, H2a & H1), which is consistent with the original findings of Bhattacherjee [28]. Additionally, it is important to note that the construct confirmation of expectation in this research mainly reflects the core traits anthropomorphism and intelligence of AI news anchors (H4c: ANT → CE & H5c: PI → CE, $R_{CE}^2$ = 0.634). The reason behind this can be attributed to the halo effect. Namely, in the process of interpersonal interaction, people tend to remember the most obvious characteristics of each other and ignore the other characteristics, thereby forming a “partial generalization” impression. While during the process of watching the AI news anchors’ video, users are more likely to notice the most evident features of AI anchors (i.e., anthropomorphism and intelligence) and thereby lead to a confirmation that exceeded respondents’ original expectations (Mean${}_{CE}$ = 4.10, SD${}_{CE}$ = 0.98), which then increases users’ satisfaction with AI news anchors and triggers further watching intention.

With regards to PA, the mechanism of PA → SAT → CI (H7 & H1) can be attributed to the trend of beauty value, where a pleasing external image evokes positive emotions in viewers and then increases future intention to continue watching.

Despite SAT, TRU also serves as a mediator in the theoretical model. There exist two antecedents of TRU – PI (H5a) and IQ (H8), which primarily examine TRU from the utilitarian value of AI anchors ($R_{TRU}^2$ = 0.658). Therein, PI is associated with CI directly (H5b). As for IQ, it influences CI via the serial mediation of TRU and SAT (i.e., IQ → TRU → SAT → CI, H8, H3b & H1). The rationale for such a mechanism is that high-quality information will result in a strong sense of trust, and the generation of trust leads to a favorable perception of the AI news anchor, which ultimately brings out the idea of long-term viewing.

Another intriguing finding of this study indicates that PN can not predict SAT (H6), whereas in the study by Wang et al. [20], perceived novelty played a crucial role in explaining the acceptance of AI news anchors. The possible reasons are summarized as follows: (1) We selected two full-length news programs hosted solely by AI anchors as stimulus materials, during which AI anchors played indispensable roles in the programs. Thus, users may value the AI news anchors more as a tool than as a toy for technological curiosity when forming affective attitudes towards the AI anchors. Respondents perceived a high level of novelty (Mean${}_{PN}$ = 4.12, SD${}_{PN}$ = 0.99), but novelty did not lead to increasing satisfaction. (2) The results of the multi-group analysis (Table 9) indicate that the positive relationship between PN & SAT only exists for those who have seen videos of AI news anchors before. As an innovative IT, AI news anchors have not yet been popularized among the general public. Those who have watched videos of AI news anchors before are more likely to be innovators among the public, they generally prefer novel things and have a relatively strong curiosity, so the perception of novelty can evoke pleasant and positive emotions among them.

7. Conclusions, Implications, and Limitations

Based on the expectation confirmation model, and combining constructs of perceived anthropomorphism (ANT), perceived intelligence (PI), perceived attractiveness (PA), information quality (IQ), perceived novelty (PN), and trust (TRU), this study establishes a theoretical model to explain users’ continuance intention (CI) for AI news anchor. The results suggest that the direct influencing factors on CI include SAT, TRU, and PI. CE cannot directly impact CI but exerts a influence through the mediation of SAT. Similarly, SAT positively mediates the relationship between PA & CI, ANT & CI, while IQ impacts CI through a serial mediation of trust and satisfaction. Although there is no direct correlation between ANT and CI, PN and SAT, the relationships are moderated by user gender and related experiences, respectively. Theoretical implications and practical implications are summarized as follows.

7.1. Theoretical Implications

First, this study takes the first step to examine human-like agents acceptance research under the context of post-adoption, which further enriches the line of virtual avatar acceptance research. On the one hand, CI, SAT, and CE which are derived from ECM are all significant determinants to reflect users’ long-term attention. The investigated results suggest that the actual performances of AI news anchors commonly exceed respondents’ original expectations (Mean${}_{CE}$ = 4.10). And the positive cognitive discrepancy associates with SAT, which further trigger continuous intention (CE → SAT → CI). Namely, the constructs and hypotheses generated from ECM have been validated in the context of AI news anchors empirically. On the other hand, most of the new hypotheses different from those in ECM (H2b-H8) were supported in the present study. Particularly, PI and TRU have been verified as direct predictors of CI, which means they may also be variables that can reflect users’ psychological intention in the long run under the context of AI news anchors besides SAT and CE. The original and newly developed constructs constitute a possible model that reflects the formation process of users’ long-term attention to the human-like agents. Second, responding to the call by Wang et al. [20], this study further developed a measurement scale for AI news anchors. As a novel phenomenon in recent years, scant attention is given to the empirical investigation of AI news anchors. The findings and measurement instruments of this study may provide some references for subsequent related research. Third, our findings suggest that gender and prior relevant experience moderate the relationships of ANT & CI, and PN & SAT, respectively. This indicates that research on human-AI interaction could take the moderation effect of consumer gender into account when investigating the construct anthropomorphism and the uncanny effect. Likewise, studies of new technology acceptance can consider relevant experiences as a moderating variable as well.

7.2. Practical Implications

Practical insights are also garnered to provide references for further commercialization and promotion of AI news anchors. Firstly, we advise media outlets to pay more attention to the “tool value” of AI news anchors and weaken their “toy attributes”. Results of the study show that users focus more on the ability of AI anchors and the quality of news information to achieve long-term development for AI news anchors, whereas stronger perceived novelty will not lead to higher satisfaction. Therefore, despite taking full advantage of AI anchors’ intelligence and efficiency, providing informative and reliable information, relevant practitioners are also suggested to focus on enhancing AI news anchors’ professionalism, training them to be “experts” in a certain type of news program, and obtaining users’ long-term attention with differentiated and unique information. Excessive changes in AI anchors’ styling, clothing, and program formats are recommended to be avoided.

Secondly, improve the transparency of algorithms and construct emotional bonds. In the present study, trust, which was evaluated from both cognitive aspect and emotional aspect, is verified as a direct predictor of CI. In light of user cognitive perception, the logic of intelligent news production of AI news anchors may raise concerns about fairness, accountability among users. When the gatekeeper of news content is gradually replaced by automated algorithms, we may need to pay more attention to improving the transparency of algorithms to constitute cognitive trust in the long run. To be more specific, explanations to non-professional users of the “black box” existing in the news production process, mutual supervision by different subjects in terms of internal management, and clarity of responsibility subjects and restriction of legal provision after the event are all needed to achieve long-term success for AI news anchors [90]. On the other hand, to cultivate emotional trust and enhance stickiness, media practitioners may consider making more efforts to construct emotional bonds with users. For example, media organizations may consider setting up social media accounts for AI news anchors, crafting distinctive personas to attract users’ attention, meanwhile utilizing the virtual image to imitate social interaction, and offering customized services for each user to strengthen emotional connection with users.

Thirdly, relevant media organizations may consider further reducing the anthropomorphic degree of AI news anchors. Our findings confirmed that while a higher degree of anthropomorphism, in general, can indirectly evoke a stronger willingness to use, perceived anthropomorphism and sustained watching intention are negatively associated among female users. Given the fact that women make up the primary audience for AI anchors (women account for 61.4% of the user profiles of virtual anchors according to iiMedia Research [91]), it is thereby recommended to deliberately animate the images of AI anchors further to make them distinguishable from human presenters to alleviate the discomfort caused by the imperfect anthropomorphic features. As suggested by Masahiro Mori (proposer of the uncanny effect) himself [88], “I predict that it is possible to create a safe level of affinity by deliberately pursuing a nonhuman design.” AI news anchors with a lower degree of anthropomorphism that undertake fundamental but intense broadcasting work could be a future trajectory of AI news anchors, which is expected to obtain a safe level of affinity. But to achieve a higher level of affinity in the long run, more exploration is needed to figure out the best way of media combination that can reach a consistency of AI news anchors’ appearance, body movement, voice, mind, personification, character setting, etc.

Finally, media outlets could pay more attention to users who have sufficient exposure to AI news anchors and give play to their strengths. Results indicate that the relationship between perceived novelty and satisfaction only shows significance for people who had previously watched AI news anchors’ videos. Such users who have sufficient exposure to AI news anchors are likely to be innovators among the general public. To take full advantage of their strengths, three suggestions are given as follows: First, invite some of the keen innovators to participate in improving AI news anchors. By doing so, they could obtain an incredible sense of loyalty and ownership of AI news anchors and the product could get continuous improvement. Second, utilize the power of peer networks. Interpersonal communication can be more influential than mass media when persuading others to adopt innovations [92]. It is thereby recommended to encourage early adopters (especially those who are key opinion leaders among users) to spread the innovation and reward them with media coverage. Third, understand the needs of different user segments and create tailored service strategies. For example, users with different levels of exposure to AI news anchors will have different “personalities”. Innovators may favor constant iterations but laggards may need convenient and low-cost designs.

7.3. Limitations and Future Research

The study also has some limitations. Firstly, limited to the current popularization state of AI anchors and time and resource cost, we chose a cross-sectional approach to investigate continuance intention. In the future, after AI news anchors are popularized among the public, a longitudinal approach can be conducted to shed a different light on the dynamics of AI news anchors’ post-adoption. Relevant constructs such as para-social interaction, habit, and continuance behavior can be tested in a longitudinal survey. Despite its limitations, a cross-sectional method was employed in many continuance studies (e.g., Bhattacherjee [28], Kim et al. [93], Moussawi et al. [54]). Secondly, as we have mentioned before, there only exist a few respondents who have sufficient exposure to the AI entities in the present sample. Whereas different extents of exposure to the non-human anchors may lead to different choices and outcomes. Therefore, future research could make more efforts in balancing the distribution of respondents who have sufficient exposure to AI news anchors and ones who have not, meanwhile specifically examine the moderating role of individual previous consumption. Thirdly, to better capture users’ long-term attention about AI news anchors, future research could consider using observational data to record users’ online continuance behavior unobtrusively by conducting online experiment (for detailed operation, see measurement for information seeking behavior in the study of So et al. [94]). Finally, as the present study validated the complex relationship between perceived anthropomorphism, satisfaction and continuance intention, future research may consider using experimental approach with more objective indicators such as electroencephalogram (EEG) and functional MRI (fMRI) to examine the nuanced relationship between different anthropomorphic degrees of AI news anchors and human-computer emotion, so as to systematically understand the influence of anthropomorphic traits of AI news anchors on human cognition and emotion, and figure out the best way of media combination.