What Drives Effective AI Use in the Newsroom? Communication Barriers, Organizational Support, and Journalist Performance in China

Abstract

As artificial intelligence reshapes professional workflows, understanding what drives effective AI use among employees has become a critical concern for organizations. Moving beyond traditional technology acceptance frameworks, this study develops an integrative multi-level model to examine the behavioral determinants of AI use performance (AUP) among journalists. Drawing on the Technology Acceptance Model (TAM) and the Expectation Confirmation Model (ECM) and incorporating individual and organizational factors, a survey was conducted among 543 journalists in China. Hypotheses are tested via a hybrid PLS-SEM and artificial neural network (ANN) approach to capture both linear and non-linear relationships. The findings reveal that expectation confirmation significantly enhances AUP by driving perceived usefulness and satisfaction. Digital literacy, personal trust in AI, and organizational support positively influence AUP, whereas communication barriers exert the strongest negative effect. Demographic variables (gender, age, education) show no significant impact. Notably, the ANN sensitivity analysis identifies communication barriers as the most influential predictor overall, a finding not apparent from linear analysis alone. This study advances theoretical understanding of employee behavioral responses in AI-integrated professional contexts and offers practical insights into how organizations can foster effective employee–AI collaboration through targeted communication strategies and supportive infrastructure.

1. Introduction

The rise of Artificial Intelligence (AI) in journalism marks a significant shift in the media landscape, empowering journalists with tools that can revolutionize content production, decision-making, and audience engagement. AI technologies, including natural language processing, machine learning, and automated content generation, are increasingly integrated into journalistic practices to streamline workflows, enhance content curation, and improve efficiency (Duan et al., 2019; Zhang et al., 2022). However, despite its transformative potential, the adoption of AI in journalism faces numerous barriers, such as concerns over AI reliability, digital literacy gaps, trust in the technology, and the level of organizational support (Alves et al., 2022; Pham & Nguyet, 2023). These factors collectively influence AI use performance (AUP), which refers to the extent to which journalists are able to effectively integrate AI into their professional activities (Mahony & Chen, 2025). Understanding how these factors interact as a system of interdependent influences is critical to optimizing AI adoption and ensuring its effective implementation within journalistic workflows.

While a growing body of research has examined the adoption of new technologies across various sectors, the specific context of AI in journalism remains underexplored. Much of the existing literature has focused on the technical aspects of AI, such as algorithmic efficiency and machine learning models (Duan et al., 2019), while overlooking the human and organizational factors that jointly shape adoption outcomes. These factors—such as digital literacy, personal trust in AI, organizational support, and communication barriers—have not been sufficiently addressed within an integrated analytical framework applicable to journalistic practice (Dierickx et al., 2024). Furthermore, although some studies have explored the use of AI in media organizations, they have primarily focused on isolated variables such as viewership and user engagement, with little attention paid to the systemic interactions among multiple adoption determinants (de-Lima-Santos & Ceron, 2021). This research aims to bridge these gaps by examining how key individual, organizational, and technological factors jointly shape AI use performance in journalism.

The Technology Acceptance Model (TAM) and the Expectation Confirmation Model (ECM) are foundational frameworks in technology adoption research, focusing on perceived ease of use (PEU) and perceived usefulness (PU) as key drivers of acceptance (Davis, 1989; Bhattacherjee, 2001). However, these models alone fail to fully capture the unique challenges faced in AI adoption in journalism, including the role of trust, organizational support structures, and communication barriers within media organizations (Al-Rahmi et al., 2021; Al-Sharafi et al., 2023). Moreover, most studies on AI adoption in journalism rely on linear models like Structural Equation Modeling (SEM), which may not adequately address the complex, non-linear relationships between the various factors influencing AI use (Hair et al., 2019; V.-H. Lee et al., 2020). To address these limitations, this study employs both PLS-SEM and Artificial Neural Networks (ANNs) to offer a more comprehensive understanding of the factors driving AI adoption in journalism.

From a systems perspective, technology adoption in organizational contexts is not driven by isolated variables but rather by the dynamic interplay among individual capabilities, cognitive evaluations, affective responses, and organizational conditions. Systems thinking emphasizes that the behavior of a complex system—such as a newsroom adopting AI—cannot be fully understood by examining its components in isolation; instead, it requires attention to the interactions, feedback loops, and boundary conditions that connect individual-level factors to organizational-level outcomes (Checkland, 1981; Sterman, 2000). In this study, constructs such as expectation confirmation, perceived usefulness, and perceived satisfaction represent the cognitive–affective subsystem; digital literacy and personal trust constitute the individual capability subsystem; and organizational support and communication barriers form the organizational–environmental subsystem. The hybrid SEM-ANN approach enables the simultaneous examination of both linear structural relationships and non-linear interaction effects across these subsystems, offering a holistic account of AI use performance in journalism.

This research integrates key constructs from the TAM and ECM frameworks with additional variables including digital literacy (DL), personal trust in AI (PT), organizational support (OS), and communication barriers (CBs). By combining PLS-SEM with ANN, this study offers a novel methodological approach that provides deeper insights into how these factors interact and collectively influence AI use performance. Specifically, this research examines how the complex, non-linear dynamics of trust, literacy, and organizational support shape journalists’ use of AI technologies, offering a more holistic perspective on AI adoption in media organizations. This study not only advances the theoretical understanding of AI adoption determinants but also offers empirically grounded insights into how the interplay of these factors shapes journalists’ everyday professional engagement with AI technologies.

2. Review of Literature

2.1. AI Adoption in Journalism: Theoretical Models and Influencing Factors

The adoption of Artificial Intelligence (AI) in journalism has gained significant attention in recent years, with studies exploring how AI technologies can improve news production, content curation, and decision-making processes. Two widely used models to understand technology adoption are the Technology Acceptance Model (TAM) and the Expectation Confirmation Model (ECM). TAM posits that perceived ease of use (PEU) and perceived usefulness (PU) are the primary drivers influencing users’ acceptance of technology (Davis, 1989; Bakı et al., 2018). In the context of journalism, this model suggests that journalists will be more likely to adopt AI tools if they find them easy to use and beneficial for enhancing work efficiency and content personalization. Similarly, ECM highlights the role of expectation confirmation—when users’ expectations align with the actual performance of technology, satisfaction increases, thereby fostering continued use (Bhattacherjee, 2001; M. C. Lee, 2010). In journalism, the confirmation of AI tools’ usefulness and reliability plays a crucial role in driving their adoption and improving performance outcomes (Alves et al., 2022; Venkatesh, 2022).

Beyond these traditional models, several external factors also influence AI adoption in journalism. Digital literacy (DL) is a significant predictor of how well journalists can integrate AI tools into their workflows (Visković et al., 2024). Journalists with higher digital literacy are more confident and efficient in using AI technologies, which enhances their AI use performance (Pham & Nguyet, 2023). Another crucial factor is personal trust in AI (PT). As AI technologies increasingly assist in editorial decisions and content generation, journalists’ trust in the accuracy and reliability of these tools is pivotal to their adoption (Al-Rahmi et al., 2021). Organizational support (OS), which includes access to training, resources, and organizational backing, further influences journalists’ willingness and ability to adopt AI tools (Ngoveni, 2025). Strong organizational support encourages journalists to overcome initial resistance and uncertainties, facilitating smoother integration of AI technologies into newsrooms (Dierickx et al., 2024). Communication barriers (CBs), such as a lack of effective communication between AI developers and journalists or misunderstandings about AI capabilities, can hinder adoption (Munoriyarwa, 2024). Clear communication about AI’s role and potential in the newsroom is essential for overcoming these barriers and ensuring successful AI integration (Zhang & Pérez Tornero, 2023).

A systems perspective suggests that these factors do not operate in isolation but rather interact as components of an interconnected adoption system. Individual cognitive evaluations (PU, PEU, PS), personal capabilities (DL, PT), and organizational conditions (OS, CB) form distinct but interrelated subsystems. The performance of the overall system—reflected in AI use performance—depends on the alignment and mutual reinforcement among these subsystems. Understanding these systemic interactions requires analytical methods that can capture both linear structural relationships and non-linear effects, which motivates the hybrid SEM-ANN approach adopted in this study.

2.2. Challenges and Barriers in AI Adoption for Journalism

The adoption of AI in journalism, while promising numerous benefits such as efficiency, personalization, and data-driven insights, faces several interconnected challenges that hinder its full integration into newsroom practice (Verma, 2024; Kevin-Alerechi et al., 2025). From a systems perspective, these barriers operate at multiple levels—individual, organizational, and institutional—and their interactions create compounding effects that cannot be adequately understood in isolation.

At the individual level, ethical concerns regarding AI’s impact on journalistic integrity and authenticity constitute a prominent barrier (Sonni et al., 2024; Benson et al., 2025). As AI tools become more involved in content creation, fundamental questions about transparency, algorithmic bias, and editorial accountability emerge (Singhal et al., 2024). These concerns are theoretically grounded in the tension between journalistic professional norms—which prioritize human editorial judgment, source verification, and public accountability—and the opaque decision-making processes inherent in machine learning systems (Diakopoulos, 2019). Closely related is the issue of trust. The lack of trust in AI systems poses a significant psychological barrier to adoption (Ivchyk, 2024). Drawing on technology acceptance research, trust serves as a prerequisite for sustained engagement with autonomous systems; without confidence in AI’s reliability and fairness, journalists are unlikely to delegate decision-making authority to algorithmic tools, particularly in domains requiring nuanced human judgment such as investigative reporting or ethical deliberation (Diakopoulos, 2019).

At the organizational level, technological and infrastructural limitations within news organizations present substantial obstacles (Larrondo et al., 2016). Many newsrooms, particularly smaller or resource-constrained outlets, lack the necessary computing infrastructure, data management systems, and technical expertise to integrate AI tools effectively (Munoriyarwa et al., 2023). This resource asymmetry creates what scholars describe as a digital capability divide, where the benefits of AI adoption accrue disproportionately to well-resourced organizations while marginalizing smaller players. Digital literacy further compounds this divide; not all journalists possess the technical competencies required to work productively with AI tools, creating a gap between those who can leverage AI and those who cannot. Cultural resistance within traditional newsrooms also plays a crucial role in shaping adoption outcomes (Bunce, 2019). Rooted in professional identity theory, journalists may resist AI adoption due to perceived threats to occupational autonomy, fears of job displacement, or a sense that algorithmic decision-making undermines the craft-based nature of journalism (Morosoli et al., 2025). Financial constraints further limit adoption, as the initial investment required for AI tools, coupled with ongoing costs of staff training and system maintenance, can be prohibitive for financially struggling outlets (Jamil, 2022). These barriers form a complex system of interacting constraints. Ethical concerns amplify trust deficits, which in turn reduce willingness to invest in digital literacy development; organizational resource limitations constrain infrastructure investment, which reinforces capability gaps. Understanding these systemic interdependencies is essential for designing effective interventions that address not merely individual barriers but the structural conditions that sustain them.

2.3. The Chinese Media Context

China presents a distinctive and highly relevant context for studying AI adoption in journalism. The usage rate of AI tools in the workplace in China has reached 93%, considerably exceeding the global average of 58% (KPMG & University of Melbourne, 2025). China’s internet user base reached 1.123 billion, with an internet penetration rate of 79.7%, and the rapid expansion of AI technologies, particularly in industries like journalism, underscores this growth (CINIC, 2025). Additionally, the proliferation of mobile internet and the adoption of 5G technology, with over 4.55 million 5G base stations, further supports the infrastructure for AI integration (CINIC, 2025).

The Chinese media system operates under a unique institutional framework characterized by state ownership and regulation of major media outlets, a dual structure combining party-affiliated and commercially oriented media organizations, and strong government support for technological innovation in the media sector (Cui & Wu, 2021). The Chinese government has actively promoted AI development through national strategies such as the New Generation AI Development Plan, which has accelerated AI integration across industries including journalism. Major news organizations such as Xinhua News Agency, People’s Daily, and China Media Group have established dedicated AI laboratories and deployed AI-powered tools for automated news writing, content recommendation, and multimedia production. Commercially oriented platforms such as Toutiao (ByteDance) and Tencent News have also pioneered the use of algorithmic content distribution and AI-assisted editorial workflows.

Specific AI tools widely used by Chinese journalists include Xinhua’s “Media Brain” (a media-focused AI platform for content analysis and production), Tencent’s Dreamwriter (an automated news writing system), Baidu’s NLP-based writing assistants, and various AI transcription and translation tools that fully support Chinese language processing. These tools are primarily applied to tasks such as automated financial and sports reporting, real-time transcription and translation, data-driven investigative leads, content recommendation and audience analytics, and video/image recognition for multimedia journalism. While these tools have significantly improved efficiency, journalists have reported limitations in areas requiring nuanced editorial judgment, cultural sensitivity, and investigative depth, as well as occasional accuracy issues in Chinese-specific language contexts such as classical Chinese references and regional dialects.

Beyond the technological infrastructure, the Chinese media context is also characterized by distinctive features of journalism culture and professional practice that are relevant to understanding AI adoption. The Chinese media system encompasses both party-affiliated and commercially oriented organizations, each operating under different editorial priorities and resource conditions (Cui & Wu, 2021). This dual structure means that AI adoption is not a uniform process; rather, it is shaped by the specific institutional expectations and organizational cultures within which journalists work. Notably, the Chinese government’s proactive promotion of AI across industries—including journalism—through national strategies such as the New Generation AI Development Plan has created a top-down facilitative environment for technological integration that distinguishes the Chinese context from many Western media systems, where AI adoption is primarily driven by individual organizational decisions and market competition (Ji et al., 2024). Recent scholarship has further highlighted that Chinese journalism is undergoing significant professional transformation concurrent with the digital expansion, including shifts in professional news standards and evolving debates about journalistic roles in the age of algorithmic media (Wang & Meng, 2023; Lin & Tong, 2022). These professional dynamics are likely to shape how journalists perceive and engage with AI tools, as receptivity to technological innovation in newswork may be mediated by journalists’ underlying conceptions of their professional identity and editorial authority (Ji et al., 2024). It is also important to recognize that the conditions for AI adoption vary considerably across the Chinese media landscape. While major national-level organizations have the resources to establish dedicated AI laboratories and deploy sophisticated tools, smaller regional and municipal outlets often face significant financial and human resource constraints that limit their capacity for technological integration (Tong, 2023). These resource asymmetries are not purely technological; they also reflect organizational management priorities and the extent to which local leadership perceives AI as beneficial to their operational needs. Furthermore, the notably high survey completion rate observed in this study (88.7%) may itself reflect distinctive features of the Chinese professional context, including a strong culture of institutional responsiveness and collective participation in knowledge-producing activities within media organizations. Understanding these cultural and structural dimensions is essential for interpreting the survey findings, as they contextualize both the adoption patterns and the attitudinal responses reported by Chinese journalists.

3. Hypotheses Development

3.1. Expectation Confirmation (EC)

Expectation confirmation, rooted in Oliver’s (1980) expectation-disconfirmation theory, refers to the degree to which users perceive that a technology’s actual performance meets or exceeds their prior expectations. Bhattacherjee (Bhattacherjee, 2001) incorporated this concept into information systems research through the Expectation Confirmation Model (ECM), arguing that confirmation functions as a central post-adoption appraisal mechanism shaping users’ subsequent evaluations of technology.

The effect of expectation confirmation on perceived usefulness can be understood as a process of cognitive updating. When journalists experience AI tools as performing in line with or beyond their expectations, they are more likely to revise their evaluations of the technology’s utility upward, reinforcing the belief that AI can meaningfully support professional tasks (Al-Sharafi et al., 2023; Gupta et al., 2020). In journalism, where AI tools are assessed against demanding standards of speed, accuracy, and reliability, such confirmation is especially important for sustaining judgments of instrumental value. Therefore:

H1. 

Expectation confirmation has a positive influence on the perceived usefulness (PU) of AI technologies among journalists.

Expectation confirmation also influences perceived satisfaction through an affective evaluative process. Satisfaction reflects users’ overall emotional response to the extent to which actual experience aligns with prior expectations (Bhattacherjee, 2001). Unlike perceived usefulness, which concerns instrumental value, satisfaction captures a broader affective assessment of the usage experience (Eren, 2021). When journalists find that AI tools perform as expected, they are more likely to develop favorable feelings toward their use, resulting in higher satisfaction. Prior studies similarly suggest that expectation confirmation positively shapes both cognitive and affective post-adoption responses in technology continuance settings (Al-Sharafi et al., 2023; D. M. Nguyen et al., 2021). Therefore:

H2. 

Expectation confirmation has a positive influence on the perceived satisfaction (PS) of AI technologies among journalists.

3.2. Perceived Usefulness (PU)

Perceived usefulness, defined as the degree to which a user believes that using a technology will enhance work performance, is a core construct in the Technology Acceptance Model and has consistently been identified as a key predictor of technology-related outcomes (Davis, 1989). In journalism, perceived usefulness reflects journalists’ evaluation of whether AI tools can meaningfully support core professional activities such as information gathering, content production, and editorial decision-making.

The influence of perceived usefulness on satisfaction can be understood through an instrumental evaluation process. When journalists perceive that AI tools provide tangible professional benefits—such as faster data processing, more relevant content recommendations, or greater workflow efficiency—they are more likely to evaluate the overall usage experience positively (Bhattacherjee, 2001; Al-Sharafi et al., 2023). In this sense, satisfaction is shaped not only by whether the technology is easy to use, but also by whether it is perceived as genuinely valuable for accomplishing important work goals. Therefore:

H3. 

Perceived usefulness has a positive influence on journalists’ satisfaction with AI tools.

Perceived usefulness also contributes directly to AI use performance through a motivational and behavioral mechanism. When journalists regard AI tools as instrumentally valuable, they are more willing to invest effort in learning, adopting, and integrating these tools into their professional routines, which in turn enhances effective use (Davis, 1989; Gupta et al., 2020). This is consistent with TAM’s central proposition that perceived usefulness extends beyond adoption intention to shape how effectively users incorporate technology into their work practices (Davis et al., 1989). In journalism, where professional performance is judged by concrete criteria such as speed, accuracy, and editorial quality, the perceived relevance of AI tools to these outcomes is likely to directly affect use performance. Therefore:

H4. 

Perceived usefulness has a positive influence on journalists’ AI use performance.

3.3. Perceived Ease of Use (PEU)

Perceived ease of use refers to the degree to which a user believes that interacting with a technology will require minimal effort (Davis, 1989). In journalism, where practitioners often work under tight deadlines and high workload pressure, the usability of AI tools is particularly important in shaping technology-related evaluations and outcomes.

The effect of perceived ease of use on satisfaction can be understood from a cognitive load perspective. When AI tools are perceived as intuitive and easy to operate, journalists experience lower cognitive burden during interaction, which reduces frustration and contributes to a more positive overall usage experience (Gupta et al., 2020). As a result, they are more likely to evaluate the technology favorably and report higher satisfaction. By contrast, tools that require excessive technical effort or complex procedures may undermine satisfaction even when they offer functional benefits (Al-Sharafi et al., 2023). Therefore:

H5. 

Perceived ease of use has a positive influence on journalists’ satisfaction with AI tools.

Perceived ease of use also has a direct effect on AI use performance. When journalists find AI tools easy to use, they are better able to incorporate them into their daily work routines and apply them more efficiently to professional tasks (Davis, 1989; Davis et al., 1989). In this sense, ease of use reduces operational barriers and enables users to focus more fully on the substantive value of AI in journalistic practice. In high-pressure newsroom environments, such usability is likely to directly support more effective use of AI tools. Therefore:

H6. 

Perceived ease of use has a positive influence on journalists’ AI use performance (AUP).

3.4. Perceived Satisfaction (PS)

Perceived satisfaction refers to users’ overall affective evaluation of their technology usage experience, encompassing both the fulfillment of functional expectations and the emotional response to interaction quality (Bhattacherjee, 2001). In the context of AI adoption in journalism, satisfaction reflects whether journalists feel that AI tools have adequately met their professional needs and contributed positively to their work experience.

The effect of perceived satisfaction on AI use performance can be understood through post-adoption engagement. When journalists are satisfied with AI tools, they are more likely to continue using them actively, explore additional functionalities, and invest effort in integrating them more fully into their professional routines (Zhang et al., 2022; Yan et al., 2021). In this way, satisfaction supports more consistent and effective use of AI in journalistic work. Conversely, users who are dissatisfied with a technology are more likely to disengage from it or use it only superficially, thereby limiting performance gains (Roy et al., 2025). This is consistent with ECM’s proposition that satisfaction is a key driver of post-adoption behavior, shaping not only whether users continue using a technology but also how effectively they use it (Bhattacherjee, 2001; Al-Sharafi et al., 2023). Therefore:

H7. 

Perceived satisfaction (PS) has a positive influence on journalists’ AI use performance (AUP).

3.5. Digital Literacy (DL)

Digital literacy extends beyond basic technical skills to encompass the cognitive, evaluative, and social competencies required to navigate and engage productively with digital environments (Martin & Madigan, 2006). In journalism, digital literacy reflects journalists’ ability not only to operate AI tools but also to understand their underlying logic, critically interpret their outputs, and make informed editorial decisions based on AI-generated information (Sarji & Aziz, 2025; Fischer, 2025).

The effect of digital literacy on AI use performance can be understood through the logic of knowledge absorption and application. Journalists with higher levels of digital literacy are better equipped to understand how AI tools function, adapt them to specific professional tasks, evaluate AI outputs more critically, and address technical issues more independently (Pham & Nguyet, 2023; Nikou et al., 2022). These capabilities make it more likely that AI tools will be used effectively and confidently in journalistic practice. By contrast, journalists with limited digital competencies are more likely to engage with AI only at a superficial level, leading to underutilization and lower performance outcomes. Digital literacy may also support more active experimentation and self-directed learning, further strengthening effective AI use over time (Aldás-Manzano et al., 2009). Therefore:

H8. 

Digital literacy (DL) has a positive influence on journalists’ AI use performance (AUP).

3.6. Personal Trust (PT)

Personal trust in AI technologies refers to users’ confidence in the reliability, fairness, and accuracy of a technology’s outputs and processes (Choung et al., 2023; Yang & Wibowo, 2022). In journalism, where editorial decisions involve both professional standards and public accountability, trust in AI is especially important because journalists must be willing to rely on AI-generated information in content production, editorial judgment, and audience-oriented decision-making (Verma, 2024; von Sikorski & Hameleers, 2025).

The effect of personal trust on AI use performance can be understood in terms of uncertainty reduction. When journalists trust AI systems, they perceive lower risk in acting upon AI outputs or delegating certain tasks to AI tools, which reduces hesitation and supports more extensive use in professional practice (Pham & Nguyet, 2023; Stojanović et al., 2023). In this sense, trust increases journalists’ willingness to rely on AI in consequential editorial contexts and facilitates deeper functional integration of AI into journalistic routines. By contrast, low trust is likely to generate skepticism, repeated verification, and a tendency to avoid AI tools in favor of conventional methods, thereby constraining use performance regardless of the technology’s actual capability (Ivchyk, 2024). Therefore:

H9. 

Personal trust (PT) in AI has a positive influence on journalists’ AI use performance (AUP).

3.7. Organizational Support (OS)

Organizational support refers to the resources, training programs, technical assistance, and managerial encouragement provided by an organization to facilitate the adoption and effective use of new technologies (Melitski et al., 2010). In journalism, organizational support shapes the conditions under which journalists encounter, learn, and integrate AI tools into their daily workflows (Jamil, 2021).

The effect of organizational support on AI use performance can be understood in terms of practical enablement. When news organizations provide structured AI training, dedicated technical assistance, and sufficient resources, they reduce the learning costs and operational uncertainties that journalists face during AI adoption (F. Li & Wang, 2025; Badghish & Soomro, 2024). These supportive conditions reduce adoption barriers and help journalists develop the competence and confidence needed for effective AI use. Organizational support may also signal institutional legitimacy for AI use, thereby alleviating concerns about the professional risks of relying on algorithmic tools in editorial work (Mohaimen et al., 2025). In the absence of such support, journalists are more likely to experience AI adoption as an individual burden rather than an organizational priority, resulting in fragmented or inconsistent use that limits performance outcomes. Therefore:

H10. 

Organizational support (OS) has a positive influence on journalists’ AI use performance (AUP).

3.8. Communication Barriers (CBs)

Communication barriers refer to obstacles that impede the effective exchange of information within organizations, particularly between journalists and AI developers or across editorial and technical departments (Rink, 2024). In journalism, such barriers may take the form of unclear guidance about AI tool capabilities, insufficient feedback channels between users and system designers, or a lack of shared understanding regarding the role of AI in editorial processes (Jones et al., 2022; Georgakakis, 2004).

The negative effect of communication barriers on AI use performance can be understood in terms of information asymmetry. When organizational communication regarding AI tools is inadequate—whether between journalists and dedicated technical teams where such teams exist, or through other institutional channels such as training programs, editorial guidelines, and peer knowledge-sharing—journalists may lack the contextual information needed to evaluate AI capabilities and limitations accurately (F. Li & Wang, 2025; Ateeq et al., 2024). It should be noted that not all news organizations, particularly smaller regional outlets, maintain dedicated AI technical teams; in such settings, communication barriers may manifest as the absence of any structured guidance on AI use rather than as breakdowns in cross-departmental dialogue. Moreover, while inadequate communication may lead to misaligned expectations, it is equally important to acknowledge that journalists may also develop well-informed understandings of AI through independent professional experience and self-directed learning. The concern here is not that journalists are inherently unable to assess AI tools, but rather that systemic communication deficiencies within organizations may constrain access to the timely and accurate information that supports effective use. Poor communication also restricts the flow of user feedback to developers, limiting the iterative refinement of AI tools for journalistic needs. In addition, communication barriers may create role ambiguity regarding the division of responsibility between human editors and AI systems, thereby increasing uncertainty and weakening journalists’ willingness to engage with AI tools in daily practice. As a systemic constraint, communication barriers can reduce the effectiveness of otherwise supportive conditions by limiting access to the clear, accurate, and timely information required for effective AI use. Furthermore, communication barriers may obscure the complementary relationship between AI-generated efficiency and the irreplaceable value of human editorial judgment, creativity, and authenticity—qualities that journalists and their audiences continue to regard as central to credible newswork, Therefore:

H11. 

Communication barriers (CBs) have a negative influence on journalists’ AI use performance (AUP).

3.9. Demographic Variables

Demographic characteristics such as gender, age, and educational level have long been examined as potential predictors of technology adoption and use in information systems research (Syed et al., 2025; Venkatesh et al., 2003). Digital divide scholarship suggests that variation in technology exposure, self-efficacy, risk perception, and educational background may shape individuals’ engagement with emerging technologies (Méndez-Suárez et al., 2023; Nouraldeen, 2023). However, the relevance of demographic factors to technology use performance is increasingly contested in professionalized work settings, where standardized training, structured workflows, and technical support may attenuate such differences (Venkatesh et al., 2003; J. Li et al., 2022). In journalism, the widespread normalization of digital tools may further reduce the explanatory power of demographic characteristics relative to cognitive, affective, and organizational factors. Given these competing expectations, this study examines whether demographic variables retain predictive relevance for AI use performance in this context. Thus, the following hypothesis is proposed:

H12. 

Demographic variables (age, gender, and educational level) significantly affect journalists’ AI use performance (AUP).

3.10. Conceptual Model

The conceptual model (Figure 1) presents the proposed relationships among the study constructs. The framework is organized into three interrelated subsystems. First, the cognitive–affective subsystem captures the pathway through which expectation confirmation influences perceived usefulness and perceived satisfaction, while perceived usefulness, perceived ease of use, and perceived satisfaction jointly shape AI use performance. Second, the individual capability subsystem includes digital literacy and personal trust in AI, both of which exert direct effects on AUP. Third, the organizational–environmental subsystem comprises organizational support as a facilitating condition and communication barriers as a constraining condition for effective AI use. Demographic variables (gender, age, and educational level) are included to assess whether individual characteristics retain explanatory relevance beyond these systemic factors. Overall, the framework reflects the study’s central argument that AI use performance in journalism arises from the combined influence of cognitive, individual, and organizational conditions rather than from any single factor alone.

4. Materials and Method

4.1. Research Context and Sampling

Given that China represents a highly conducive environment for the study of AI adoption in journalism, this study targeted journalists within China who have direct experience using AI technologies in their professional practice. The target population comprised journalists actively employed in Chinese media organizations who reported using AI tools as part of their regular journalistic practice. The inclusion criteria required that respondents: (a) were currently employed as journalists, editors, or content producers in a recognized media organization in China; (b) had at least three months of experience using one or more AI tools in their daily work; and (c) were at least 18 years of age. Respondents who did not meet these criteria were excluded from the final sample. Data were collected through a self-administered online questionnaire distributed via professional and social media platforms, including WeChat and QQ, using a purposive snowball sampling approach. The survey was open for a period of six weeks. A total of 612 responses were received, of which 69 were excluded due to incomplete responses, failed attention checks, or ineligibility. This yielded 543 valid and eligible responses, representing an effective completion rate of 88.7%. The sample included journalists from a mix of organizational types: approximately 38% were affiliated with state-owned or party-affiliated media organizations, approximately 42% worked for commercially oriented digital media platforms and online news outlets, and approximately 20% were employed by regional or municipal media organizations.

4.2. Profile of Respondents

The sample comprised 543 Chinese journalists. Of these, 36.1% were female and 63.9% were male. In terms of age distribution, the largest group was aged 26–30 (35.7%), followed by 31–40 (24.1%) and 41–50 (16.8%). A smaller proportion were in the 18–25 age group (14.4%), 51–60 years (7.0%), and 2.0% were above 60. Regarding educational qualifications, the majority (53.4%) held a Bachelor’s degree, 44.8% had a Master’s degree, and 1.8% held a Doctoral degree. This composition is consistent with the profile of journalists actively engaged with AI tools in Chinese newsrooms (Sun et al., 2024).

4.3. Research Instrument

The research instrument was a structured questionnaire divided into three sections. Part A collected demographic and behavioral usage information. Part B evaluated key constructs related to AI adoption (

Table 1. Constructs and indicators.

Construct	Item	Indicator	Source
Expectation Confirmation (EC)	EC1	My experience with AI tools in journalism was better than I expected	(Bhattacherjee, 2001)
	EC2	The functions provided by AI tools were better than I expected
	EC3	Overall, most of my expectations from using AI tools were confirmed
	EC4	The service level provided by AI tools was better than I expected
Perceived Usefulness (PU)	PU1	Using AI tools improves my work performance in journalism	(Saadé, 2007)
	PU2	Using AI tools increases my productivity in journalistic tasks
	PU3	Using AI tools enhances the effectiveness of my work
	PU4	Using AI tools makes it easier to do my job
	PU5	Overall, I find AI tools useful in my journalistic practice
Perceived Ease of Use (PEU)	PEU1	I find AI tools easy to use in my journalistic work	(Segars & Grover, 1993)
	PEU2	Learning to operate AI tools is easy for me
	PEU3	My interaction with AI tools is clear and understandable
Perceived Satisfaction (PS)	PS1	I am satisfied with my overall experience of using AI tools in journalism	(Udo et al., 2010)
	PS2	I am pleased with the experience of using AI tools for my work
	PS3	My decision to use AI tools in journalism was a wise one
	PS4	Overall, I am satisfied with AI tools for my journalistic activities
Digital Literacy (DL)	DL1	I am confident in my ability to use digital technologies effectively	(Covello & Lei, 2010; L. A. T. Nguyen & Habók, 2024)
	DL2	I can evaluate and critically assess digital information and tools
	DL3	I am able to learn new digital tools and platforms independently
	DL4	I understand how digital systems and algorithms work in general
Personal Trust (PT)	PT1	I believe AI tools produce accurate and reliable outputs	(Johnson-George & Swap, 1982; Couch et al., 1996)
	PT2	I feel confident that AI tools will perform consistently
	PT3	I believe AI tools handle information fairly and without bias
	PT4	I can rely on AI tools to function as intended
Organizational Support (OS)	OS1	My organization provides adequate training for using AI tools	(Shore & Tetrick, 1991; Kurtessis et al., 2017)
	OS2	My organization provides the resources needed to use AI tools effectively
	OS3	My organization encourages the use of AI tools in journalistic work
	OS4	My organization values my efforts to integrate AI into my work
	OS5	Technical assistance is available when I encounter problems with AI tools
	OS6	My organization’s leadership supports the adoption of AI technologies
Communication Barriers (CBs)	CB1	There is a lack of clear information about the capabilities of AI tools in my organization	(Back et al., 1972; Paluck et al., 2003)
	CB2	Communication between journalists and AI developers is insufficient
	CB3	I often receive unclear or contradictory guidance about how to use AI tools
	CB4	There are limited channels to provide feedback about AI tools to technical teams
	CB5	Misunderstandings about AI’s role in the newsroom are common
	CB6	Information about updates or changes to AI tools is not communicated effectively
	CB7	There is a gap in understanding between editorial and technical staff regarding AI
AI Use Performance (AUP)	AUP1	I can effectively use AI tools to accomplish my journalistic tasks	Adapted from (Venkatesh et al., 2003; Goodhue & Thompson, 1995)
	AUP2	AI tools help me produce higher quality journalistic outputs
	AUP3	I am able to integrate AI tools into my daily workflow efficiently
	AUP4	Using AI tools has improved my overall professional performance
	AUP5	I can use AI tools to make better editorial decisions
	AUP6	AI tools have enhanced my ability to meet deadlines and work demands

). Part C assessed respondents’ perceptions of the constructs under study. A pre-test was conducted to ensure face and content validity, with feedback from practitioners and academicians, leading to slight modifications. The measurement items were adapted from established scales in technology adoption research.

4.4. Statistical Analysis

The collected data were analyzed using IBM SPSS Statistics (Version 24; IBM Corp., Armonk, NY, USA) and SmartPLS (Version 4.1.0.8; SmartPLS GmbH, Bönningstedt, Germany). The proposed framework was tested using a combination of ANN (Haykin, 2001) and PLS-SEM (Hair et al., 2019). PLS-SEM was utilized in the initial phase for two primary reasons. First, PLS-SEM is preferable for predicting complex models and facilitating theory development (Ooi et al., 2018). Second, PLS-SEM does not impose numerous constraints on non-normal distributions and sample sizes (Lew et al., 2020). As Mardia’s multivariate skewness (β = 503.05) and kurtosis (β = 2893.46) both had p-values less than 0.001, PLS-SEM was considered more appropriate than CB-SEM.

At a power level of 0.95, an alpha value of 0.05, and an effect size of 0.15, the 543-person sample exceeded the minimal 262-person sample required by G*Power (Version 3.1; Heinrich-Heine-Universität Düsseldorf, Düsseldorf, Germany) (Faul et al., 2009), demonstrating sufficient statistical power. Since PLS-SEM exclusively captures linear correlations, the ANN methodology was employed in the subsequent phase to elucidate non-linear relationships and ascertain the relative importance of variables (Al-Sharafi et al., 2021; Mohaimen et al., 2025). The combination of PLS-SEM and ANN provides a richer understanding of the systemic drivers of AI use performance.

4.5. Common Method Bias (CMB)

Given the cross-sectional design, both procedural and statistical remedies were applied to assess the potential influence of common method bias (CMB) (Tan & Ooi, 2018). Procedurally, respondents were assured of confidentiality and anonymity. Statistically, Harman’s single-factor test showed that the largest factor explained 46.77% of the total variance, which is below the 50% threshold (MacKenzie & Podsakoff, 2012). In addition, the full collinearity assessment indicated that most VIF values were below the conservative threshold of 3.3, with only one value (CB = 3.362) marginally exceeding this cutoff (

Table 2. Full collinearity assessment (variance inflation factors).

	Full Collinearity Test			Full Collinearity Test with a Random Variable
	PU	PS	AUP	Random Variable
AUP				2.455
CB			3.362	3.071
DL			2.119	2.068
EC	1.000	2.010		1.717
OS			2.185	1.162
PEU		1.941	2.302	2.279
PS			2.334	2.267
PT			1.577	1.586
PU		2.012	2.477	1.578

). As this deviation was slight and the remaining indicators were within acceptable limits, common method bias is unlikely to pose a serious threat to the validity of the findings.

5. Results

5.1. Assessing the Outer Measurement Model

The measurement model was evaluated in terms of internal consistency reliability, convergent validity, and discriminant validity. Internal consistency reliability was assessed using Cronbach’s alpha (α) and composite reliability (CR). As shown in

Table 3. Loadings, Cronbach’s alpha, composite reliability and average variance extracted.

Constructs	Items	λ	α	CR	AVE
AI Use Performance (AUP)	AUP1	0.824	0.903	0.925	0.673
	AUP2	0.800
	AUP3	0.826
	AUP4	0.826
	AUP5	0.811
	AUP6	0.835
Communication Barriers (CBs)	CB1	0.806	0.902	0.923	0.631
	CB2	0.853
	CB3	0.824
	CB4	0.806
	CB5	0.792
	CB6	0.736
	CB7	0.738
Digital Literacy (DL)	DL1	0.860	0.899	0.930	0.768
	DL2	0.891
	DL3	0.892
	DL4	0.861
Expectation Confirmation (EC)	EC1	0.884	0.900	0.930	0.769
	EC2	0.898
	EC3	0.876
	EC4	0.847
Organizational Support (OS)	OS1	0.866	0.915	0.935	0.706
	OS2	0.869
	OS3	0.862
	OS4	0.864
	OS5	0.857
	OS6	0.710
Perceived Ease of Use (PEU)	PEU1	0.905	0.886	0.930	0.815
	PEU2	0.900
	PEU3	0.903
Perceived Satisfaction (PS)	PS1	0.883	0.898	0.929	0.765
	PS2	0.883
	PS3	0.862
	PS4	0.871
Personal Trust (PT)	PT1	0.842	0.846	0.896	0.684
	PT2	0.831
	PT3	0.802
	PT4	0.832
Perceived Usefulness (PU)	PU1	0.859	0.907	0.930	0.728
	PU2	0.862
	PU3	0.856
	PU4	0.854
	PU5	0.835

, all α values ranged from 0.846 to 0.915 and all CR values ranged from 0.896 to 0.935, exceeding the recommended threshold of 0.70 (Hair et al., 2019).

Convergent validity was assessed using outer loadings (λ) and average variance extracted (AVE). All outer loadings exceeded the recommended threshold of 0.70, and all AVE values were above the minimum criterion of 0.50 (Hair et al., 2019), indicating adequate convergent validity.

Discriminant validity was examined using the heterotrait–monotrait ratio (HTMT). As reported in

Table 4. Heterotrait–Monotrait ratio.

HTMTs	AUP	CB	DL	EC	OS	PEU	PS	PT	PU
AUP
CB	0.774
DL	0.677	0.713
EC	0.712	0.794	0.673
OS	0.688	0.770	0.578	0.634
PEU	0.711	0.788	0.654	0.707	0.632
PS	0.715	0.739	0.705	0.673	0.663	0.663
PT	0.592	0.604	0.581	0.621	0.456	0.579	0.587
PU	0.718	0.790	0.663	0.718	0.695	0.703	0.698	0.510

, all HTMT values were below the conservative threshold of 0.85 (Kline, 2023), confirming satisfactory discriminant validity. These results indicate that the measurement model exhibited adequate reliability and validity to support the structural model assessment.

5.2. Inspecting the Inner Structural Model

The structural model was assessed using bootstrapping with 5000 subsamples. As shown in Figure 2 and

Table 5. Structural Model Results.

PLS Path	Original Sample (O)	T Statistics	p Values	BC-CIs		Remarks
H1: EC -> PU ***	0.648	15.431	0.000	0.559	0.720	Significant
H2: EC -> PS ***	0.250	4.705	0.000	0.147	0.354	Significant
H3: PU -> PS ***	0.326	5.950	0.000	0.217	0.431	Significant
H4: PU -> AUP **	0.138	2.597	0.009	0.033	0.240	Significant
H5: PEU -> PS ***	0.229	5.336	0.000	0.146	0.313	Significant
H6: PEU -> AUP **	0.135	2.711	0.007	0.039	0.235	Significant
H7: PS -> AUP **	0.142	2.640	0.008	0.043	0.255	Significant
H8: DL -> AUP *	0.113	2.552	0.011	0.029	0.202	Significant
H9: PT -> AUP *	0.106	2.205	0.027	0.014	0.200	Significant
H10: OS -> AUP **	0.158	2.841	0.005	0.060	0.275	Significant
H11: CB -> AUP **	−0.180	2.874	0.004	−0.301	−0.058	Significant
H12: Gender -> AUP n.s.	−0.046	0.823	0.410	−0.155	0.066	Not significant
Age -> AUP n.s.	−0.027	0.938	0.348	−0.083	0.030	Not significant
Educational level -> AUP n.s.	−0.016	0.623	0.533	−0.068	0.034	Not significant

Notes: * p < 0.05, ** p < 0.01, *** p < 0.001; n.s. = Not significant; BC-CIs = Bias Corrected Confidence Intervals.

, all hypothesized relationships were supported except H12.

Within the cognitive–affective subsystem, expectation confirmation had a significant positive effect on perceived usefulness (β = 0.648, p < 0.001), supporting H1. Perceived satisfaction was, in turn, positively influenced by expectation confirmation (β = 0.250, p < 0.001), perceived usefulness (β = 0.326, p < 0.001), and perceived ease of use (β = 0.229, p < 0.001), supporting H2, H3, and H5, respectively.

Regarding AI use performance, perceived usefulness (β = 0.138, p < 0.05), perceived ease of use (β = 0.135, p < 0.01), and perceived satisfaction (β = 0.142, p < 0.01) all exerted significant positive effects on AUP, supporting H4, H6, and H7. Within the individual capability subsystem, digital literacy (β = 0.113, p < 0.05) and personal trust in AI (β = 0.106, p < 0.05) also had significant positive effects on AUP, supporting H8 and H9. Within the organizational–environmental subsystem, organizational support had a significant positive effect on AUP (β = 0.158, p < 0.01), supporting H10, whereas communication barriers had a significant negative effect (β = −0.180, p < 0.01), supporting H11.

By contrast, demographic variables—gender (β = −0.046, p > 0.05), age (β = −0.027, p > 0.05), and educational level (β = −0.016, p > 0.05)—did not show significant effects on AUP. Therefore, H12 was not supported.

5.3. Predictive Relevance and Effect Size

The predictive relevance and explanatory power of the structural model were assessed using effect size (f²), predictive relevance (Q²), coefficient of determination (R²), and PLSpredict. Following Cohen (Cohen, 1992), f² values of 0.02, 0.15, and 0.35 indicate small, medium, and large effects, respectively, whereas values below 0.02 suggest negligible effects. As shown in

Table 6. Effect Size, Explanatory Power, and PLSpredict.

Effect Sizes		Explanatory Power		PLSpredict
Relations	f2	Endogenous Construct	R2	Indicator	PLS SEM RMSE	LM RMSE	Difference	Predictive Power
EC -> PU	0.725	AUP	0.612	AUP1	0.846	0.841	0.005	Medium
EC -> PS	0.061	PS	0.494	AUP2	0.884	0.900	−0.016
PEU -> PS	0.053	PU	0.420	AUP3	0.846	0.847	−0.001
PEU -> AUP	0.020			AUP4	0.837	0.848	−0.011
PU -> PS	0.104			AUP5	0.889	0.908	−0.019
PU -> AUP	0.020	Predictive relevance		AUP6	0.835	0.843	−0.008
PS -> AUP	0.022		Q2
DL -> AUP	0.015	AUP	0.404
PT -> AUP	0.018	PS	0.369
OS -> AUP	0.028	PU	0.302
CB -> AUP	0.025

, the f² values ranged from 0.015 to 0.725. Notably, the path from expectation confirmation to perceived usefulness (EC → PU) exhibited the largest effect size (f² = 0.725), indicating a substantial effect, whereas most other relationships showed small effects. The effects of digital literacy (f² = 0.015) and personal trust (f² = 0.018) on AUP were comparatively weak.

All Q² values for the endogenous constructs were greater than zero (AUP = 0.404, PS = 0.369, PU = 0.302), indicating adequate predictive relevance. In addition, the model explained 61.2% of the variance in AUP (R² = 0.612), 49.4% of the variance in PS (R² = 0.494), and 42.0% of the variance in PU (R² = 0.420), suggesting relatively strong explanatory power overall.

To further assess out-of-sample predictive performance, PLSpredict was employed (Shmueli et al., 2019). For the primary target construct AUP, five of the six indicators showed lower RMSE values for the PLS-SEM model than for the linear model (LM) benchmark, while only one indicator (AUP1) performed slightly worse. This pattern indicates medium predictive power. Taken together, these findings support the model’s ability to both explain and predict AI use performance in journalism.

5.4. Artificial Neural Network (ANN) Analysis

To complement the linear relationships identified by the PLS-SEM analysis, artificial neural network (ANN) analysis was conducted to capture potential non-linear patterns among the study constructs. Only variables found to be statistically significant in the PLS-SEM analysis were included as input neurons, and a ten-fold cross-validation procedure was employed to reduce the risk of overfitting (Al-Sharafi et al., 2021; Mohaimen et al., 2025). Three ANN models were estimated: Model A predicted AUP using PU, PEU, PS, DL, PT, OS, and CB as inputs; Model B predicted PS using EC, PU, and PEU; and Model C predicted PU using EC.

As reported in

Table 7. RMSE values for AI use performance (AUP), Perceived Satisfaction (PS) and Perceived Usefulness (PU).

Neural Network	Model 1		Model 2		Model 3
	Input: PU, PEU, PS, DL, PT, OS, CB		Input: EC, PU, PEU		Input: EC
	Output: AUP		Output: PS		Output: PU
	Training	Testing	Training	Testing	Training	Testing
ANN1	0.068	0.067	0.071	0.064	0.084	0.076
ANN2	0.066	0.079	0.069	0.078	0.077	0.090
ANN3	0.069	0.063	0.070	0.066	0.079	0.073
ANN4	0.064	0.069	0.069	0.069	0.090	0.064
ANN5	0.064	0.067	0.070	0.082	0.083	0.074
ANN6	0.072	0.070	0.072	0.075	0.079	0.079
ANN7	0.065	0.058	0.072	0.069	0.079	0.085
ANN8	0.067	0.049	0.078	0.075	0.087	0.080
ANN9	0.065	0.063	0.071	0.064	0.079	0.087
ANN10	0.073	0.085	0.070	0.058	0.081	0.062
Mean	0.067	0.067	0.071	0.070	0.082	0.077
SD	0.003	0.010	0.003	0.007	0.004	0.009

, the RMSE values for both the training and testing datasets were low across all ten runs. The mean RMSE values were 0.067 (training) and 0.067 (testing) for Model A, 0.071 and 0.070 for Model B, and 0.082 and 0.077 for Model C. The small standard deviations further indicate that the models were stable across repeated estimations. Because the training and testing RMSE values were closely aligned, no substantial overfitting was observed. These results suggest that the ANN models achieved satisfactory predictive accuracy and provide a sound basis for the subsequent sensitivity analysis (Lau et al., 2021).

A sensitivity analysis was conducted to assess the relative importance of the input variables in the ANN models.

Table 8. Sensitivity analysis of the ANN models.

Neural Network	Model A (Output: AUP)							Model B (Output: PS)			Model C (Output: PU)
	PU	PEU	PS	DL	PT	OS	CB	EC	PU	PEU	EC
ANN1	0.160	0.142	0.149	0.113	0.158	0.130	0.148	0.282	0.394	0.324	1.000
ANN2	0.141	0.104	0.140	0.109	0.213	0.132	0.161	0.318	0.401	0.281	1.000
ANN3	0.132	0.085	0.141	0.096	0.133	0.160	0.253	0.32	0.423	0.257	1.000
ANN4	0.144	0.165	0.174	0.111	0.107	0.127	0.171	0.209	0.413	0.378	1.000
ANN5	0.132	0.151	0.125	0.132	0.135	0.130	0.195	0.309	0.347	0.344	1.000
ANN6	0.137	0.119	0.167	0.156	0.039	0.180	0.202	0.363	0.435	0.203	1.000
ANN7	0.202	0.130	0.139	0.094	0.130	0.110	0.195	0.28	0.423	0.297	1.000
ANN8	0.166	0.130	0.140	0.107	0.148	0.106	0.203	0.306	0.406	0.288	1.000
ANN9	0.151	0.131	0.137	0.145	0.129	0.146	0.161	0.327	0.35	0.323	1.000
ANN10	0.045	0.074	0.127	0.128	0.185	0.206	0.235	0.248	0.45	0.302	1.000
Average relative importance	0.141	0.123	0.144	0.119	0.138	0.143	0.192	0.296	0.404	0.299	1.000
Normalized relative importance (%)	75.89	66.82	76.65	63.59	72.42	74.76	100.0	93.30	100.0	92.30	100.0

Note: Normalized relative importance (%) is calculated by dividing each variable’s average relative importance by the highest average relative importance within the same model and multiplying by 100.

presents the average relative importance and normalized relative importance of each predictor across the ten ANN runs.

For Model A (output: AUP), communication barriers emerged as the most influential predictor, with the highest average relative importance (0.192) and a normalized importance of 100.0%. This was followed by perceived satisfaction (0.144; 76.65%), perceived usefulness (0.141; 75.89%), organizational support (0.143; 74.76%), personal trust (0.138; 72.42%), perceived ease of use (0.123; 66.82%), and digital literacy (0.119; 63.59%). This result is noteworthy because, in the PLS-SEM analysis, organizational support had the largest positive path coefficient among the direct predictors of AUP, whereas communication barriers showed the strongest effect in absolute terms. The ANN results further indicate that communication barriers exert the greatest overall predictive influence on AUP, suggesting the presence of non-linear effects that are not fully captured by linear modeling.

For Model B (output: PS), perceived usefulness was the most influential predictor, with an average relative importance of 0.404 and a normalized importance of 100.0%, followed by expectation confirmation (0.296; 93.30%) and perceived ease of use (0.299; 92.30%). The relatively balanced contributions of these three predictors suggest that perceived satisfaction is jointly shaped by multiple cognitive evaluations rather than by a single dominant factor.

For Model C (output: PU), expectation confirmation was the sole input variable and therefore accounted for 100.0% of the normalized importance, consistent with its strong effect in the PLS-SEM analysis.

To further examine whether the relative importance of predictors remained consistent across linear and non-linear modeling approaches, the rankings derived from PLS-SEM and ANN were compared. Whereas PLS-SEM reflects the relative strength of linear effects, ANN captures the normalized predictive importance of input variables under potentially non-linear conditions.

Table 9. Comparison of PLS-SEM and ANN rankings.

PLS Path	Path Coefficient	ANN Results: Normalized Relative Importance (%)	Ranking (PLS-SEM) [Based on Path Coefficient]	Ranking (ANN) [Based on Normalized Relative Importance (%)]	Remark
Model A: (Output AUP)
PU -> AUP	0.138	75.89	4	3	Not match
PEU -> AUP	0.135	66.82	5	6	Not match
PS -> AUP	0.142	76.65	3	2	Not match
DL -> AUP	0.113	63.59	6	7	Not match
PT -> AUP	0.106	72.42	7	5	Not match
OS -> AUP	0.158	74.76	2	4	Not match
CB -> AUP	−0.180	100.00	1	1	Match
Model B: (Output PS)
EC -> PS	0.250	93.30	2	2	Match
PU -> PS	0.326	100.00	1	1	Match
PEU -> PS	0.229	92.30	3	3	Match
Model C: (Output PU)
EC -> PU	0.648	100.00	1	1	Match

presents the comparative rankings across the three models.

For Model A (output: AUP), substantial differences were observed between the two approaches. Communication barriers ranked first in both PLS-SEM and ANN, indicating strong consistency regarding its central role in shaping AI use performance. However, the rankings of the remaining predictors did not match. For instance, organizational support ranked second in PLS-SEM but fourth in ANN, whereas perceived satisfaction rose from third in PLS-SEM to second in ANN, and personal trust increased from seventh to fifth. These discrepancies suggest that the determinants of AUP may involve non-linear patterns that are not fully captured by linear path coefficients alone.

In contrast, the results for Model B (output: PS) were fully consistent across the two methods. Perceived usefulness, expectation confirmation, and perceived ease of use ranked first, second, and third, respectively, in both PLS-SEM and ANN. Similarly, for Model C (output: PU), expectation confirmation ranked first in both approaches.

The comparison indicates that while PLS-SEM and ANN produce convergent results for the simpler submodels predicting PS and PU, ANN provides additional insight into the more complex predictive structure of AUP by revealing differences in the relative importance of its predictors.

6. Discussion, Conclusions, Implications, and Limitations

6.1. Discussion

This study examined the key drivers of AI use performance among journalists by integrating the ECM and TAM with additional individual and organizational factors. Adopting a systems perspective, the findings reveal that journalists’ AI use performance emerges from the combined and interacting effects of cognitive evaluations, affective responses, individual capabilities, and organizational conditions.

First, Expectation confirmation as a cognitive gateway. Consistent with ECM, expectation confirmation significantly and positively influenced both PU and PS (Bhattacherjee, 2001; Eren, 2021). This suggests that journalists’ assessments of AI tools are strongly anchored in the degree to which actual performance aligns with pre-use expectations. The positive effect on satisfaction supports the notion that satisfaction is generated through confirmation of anticipated outcomes (C. Y. Li & Fang, 2019). In journalism, where professional credibility and workflow reliability are critical, unmet expectations may weaken confidence in AI-assisted practices. In practical terms, this finding highlights the critical importance of expectation management in newsroom AI implementation. Journalists’ pre-use expectations of AI tools are shaped by multiple sources, including organizational promotion, media discourse about AI capabilities, and peer experiences (Beckett, 2019). When these expectations are inflated—for example, when AI tools are introduced with promises of transforming editorial workflows but fail to deliver meaningful improvements in day-to-day reporting tasks—the resulting disconfirmation may not only reduce satisfaction but also generate lasting skepticism that discourages future engagement with AI-assisted practices. This suggests that news organizations should prioritize realistic onboarding processes that align journalists’ expectations with the actual capabilities and limitations of AI tools, rather than relying on aspirational narratives about technological transformation. From a systems perspective, EC functions as a cognitive gateway initiating positive feedback loops: confirmed expectations enhance PU and PS, which drive sustained AI use (D. M. Nguyen et al., 2021; Al-Sharafi et al., 2023).

Second, Joint influence of cognitive evaluations and affective responses. Perceived usefulness exerted a direct positive influence on AUP (Davis, 1989; Pérez et al., 2020). Similarly, PEU positively affected both PS and AUP (Davis, 1989; Bhagat et al., 2023). Moreover, PS itself emerged as a significant predictor of AUP (Yan et al., 2021; Al-Sharafi et al., 2023). These results demonstrate that the cognitive–affective subsystem operates as an integrated unit: PU and PEU feed into PS, which amplifies their combined effect on AUP. For journalistic practice, this interconnection carries specific implications. The perceived usefulness of AI tools is likely to vary across different types of journalistic tasks: AI may be evaluated favorably for routine, data-intensive operations such as automated transcription, real-time translation, and audience analytics, while being perceived as less useful—or even counterproductive—for tasks demanding editorial judgment, narrative creativity, and source relationship management (Diakopoulos, 2019). Similarly, the significance of perceived ease of use reflects the high-pressure, deadline-driven nature of newsroom work, where journalists cannot afford extended learning curves or complex technical procedures. AI tools that require substantial technical expertise to operate effectively may be abandoned in favor of established workflows, regardless of their potential benefits. The mediating role of satisfaction further suggests that journalists’ continued engagement with AI is not solely a rational calculation of utility but also an affective response shaped by the quality of everyday interactions with these tools. Interventions targeting any single element may yield limited gains unless satisfaction and perceived value are simultaneously addressed.

Third, Individual capabilities and organizational conditions as system enablers and constraints. Digital literacy was identified as an essential determinant of AUP (Pham & Nguyet, 2023; Nikou et al., 2022). Personal trust in AI also demonstrated a positive influence (Stojanović et al., 2023). Organizational support emerged as a significant enabler (F. Li & Wang, 2025; J. Li et al., 2022), consistent with recent evidence that organizational conditions facilitating knowledge sharing and resource allocation are critical for translating AI capabilities into performance outcomes (Olan et al., 2022). In contrast, communication barriers negatively affected AUP (Ateeq et al., 2024). Notably, the ANN analysis identified communication barriers as the single most influential predictor of AUP (normalized importance: 100%), a finding not apparent from PLS-SEM path coefficients alone. This discrepancy suggests that communication barriers may exert their influence through threshold effects or interaction patterns that linear models cannot capture—underscoring the value of the hybrid SEM-ANN methodology for understanding complex adoption systems. The prominence of communication barriers as the strongest predictor of AUP carries important implications for how newsrooms manage the human side of technological change. In practice, the introduction of AI into editorial workflows is not merely a technical upgrade but a sociotechnical transition that reconfigures professional roles, task boundaries, and knowledge hierarchies within news organizations (Lewis & Westlund, 2015). When journalists lack clear, accessible information about what AI tools can and cannot do—and about how their professional roles relate to AI-assisted processes—they face uncertainty that undermines effective use regardless of the technology’s actual capability. This finding challenges the assumption that providing AI tools and training alone is sufficient; rather, it suggests that the communicative infrastructure surrounding AI implementation—including feedback channels, cross-functional dialogue, and transparent editorial policies on AI use—is at least as important as the technology itself. These findings gain additional explanatory depth when situated within the specific characteristics of the Chinese media context. The strong negative effect of communication barriers may be partly attributable to the structural diversity of the Chinese media system, which encompasses large national-level organizations with dedicated AI infrastructure alongside smaller regional and municipal outlets that often lack specialized technical teams and structured guidance on AI use (Tong, 2023). In well-resourced organizations such as Xinhua or People’s Daily, communication barriers may take the form of cross-departmental coordination failures between editorial and technical units, whereas in smaller outlets, they are more likely to manifest as a fundamental absence of institutional channels for AI-related knowledge dissemination. The state’s proactive top-down promotion of AI in journalism may further amplify the salience of communication barriers: when AI adoption is driven by policy directives rather than organic organizational demand, the gap between institutional expectations and journalists’ actual preparedness may widen unless accompanied by effective internal communication (Ji et al., 2024). Similarly, the significant positive effect of organizational support aligns with the observation that Chinese media organizations vary considerably in their capacity to provide training, resources, and managerial encouragement for AI integration, suggesting that organizational-level investment remains a decisive factor in translating national-level policy ambitions into effective frontline practice.

Finally, Demographic factors and the attenuation of digital divides. Contrary to expectations, age, gender, and educational level were not significantly associated with AUP. This may reflect the increasing normalization of digital technologies in journalistic work, where institutional training and standardized workflows mitigate demographic differences (Venkatesh et al., 2003; J. Li et al., 2022). In systems terms, the organizational–environmental subsystem may serve as an equalizer that buffers the influence of individual demographic variation. This interpretation is further supported by the Chinese professional context, in which institutional training programs and standardized editorial workflows are widespread across media organizations of varying types and sizes, potentially attenuating the influence of individual demographic characteristics on technology use outcomes. From a practical standpoint, this finding suggests that efforts to improve AI use performance in newsrooms should prioritize organizational and communicative interventions rather than demographic-targeted programs, as the professional environment appears to be a more decisive factor than individual characteristics in shaping effective AI engagement.

6.2. Conclusions

Drawing on the Technology Acceptance Model and the Expectation Confirmation Model, and adopting a systems perspective, this study examined how cognitive evaluations, individual capabilities, and organizational conditions jointly shape AI use performance among journalists in China. Using a hybrid PLS-SEM and ANN methodology with data from 543 journalists, the analysis tested twelve hypotheses spanning three interrelated subsystems. The PLS-SEM results confirmed that expectation confirmation, perceived usefulness, perceived ease of use, perceived satisfaction, digital literacy, personal trust, and organizational support all significantly contribute to AI use performance, while communication barriers exert a significant negative effect. Demographic variables showed no significant influence. The ANN sensitivity analysis further revealed that the rank ordering of predictor importance differs from that suggested by linear path coefficients, with communication barriers emerging as the most influential determinant overall. Taken together, these findings suggest that effective AI integration in journalism is not primarily a technological challenge but an organizational and communicative one, requiring news organizations to attend to expectation management, communicative infrastructure, and the alignment of AI tools with the distinctive demands of professional journalistic practice.

6.3. Implications

6.3.1. Theoretical Implications

First, this research advances the field by examining AI use performance as a concrete behavioral outcome rather than adoption intention. Second, the findings demonstrate that AUP is jointly shaped by cognitive evaluations and affective responses, with perceived satisfaction playing a pivotal mediating role. Third, by adopting a systems perspective, the study embeds individual-level responses within organizational and communicative conditions, providing a process-oriented account of AI adoption. Fourth, the hybrid SEM-ANN methodology reveals non-linear relationships that linear models alone cannot detect. Finally, the insignificance of demographic variables suggests that traditional digital divides may be attenuated in highly institutionalized journalistic environments.

6.3.2. Practical Implications

First, news organizations should prioritize expectation management when introducing AI technologies. Second, usability and satisfaction should be treated as central design priorities. Third, continuous skills training can enhance digital literacy, while transparency and explainability can strengthen trust. Fourth, communication barriers emerged as the strongest predictor of reduced AUP; establishing clear communication channels between editorial and technical teams is therefore a priority. Finally, organizations should focus on creating standardized workflows and supportive infrastructures rather than targeting individual demographics.

6.4. Limitations and Future Directions

First, the data were collected within a single national context (China), which may limit generalizability. The Chinese media system’s unique characteristics may amplify certain adoption dynamics that would operate differently elsewhere. Second, the cross-sectional design constrains causal inference. Third, other situational variables were not examined (e.g., task complexity, news genre, algorithmic transparency). Fourth, the purposive snowball sampling approach introduces potential self-selection bias. Fifth, the analysis does not differentiate among specific AI tools or platforms. Future research may benefit from cross-national designs, longitudinal approaches, probability-based sampling, tool-specific analyses, and additional theoretical perspectives such as human–AI collaboration and sociotechnical systems theory.