Understanding the Adoption of Artificial Intelligence in Journalism: An Empirical Study in Vietnam

Artificial Intelligence Technology in Journalism

The application of artificial intelligence (AI) technology in journalism has seen a significant increase in recent years, with various terms used to describe this integration, such as “computational journalism” (S. Cohen et al., 2011), “algorithmic journalism” (Anderson, 2013), and “automated journalism” (Nickolas, 2019). Using algorithms to analyze data from various sources, translate text into audio and video, and determine sentiment are examples of how AI is implemented in newsrooms (Calvo Rubio & Ufarte Ruiz, 2021). The New York Times, The Washington Post, and the Associated Press are examples of traditional news organizations that have effectively integrated AI efforts in their newsrooms (Chan-Olmsted, 2019). Chinese newsrooms, like the state-run Xinhua News Agency, also use AI tools to produce news through platforms like Media Brain, which autogenerates news bulletins (Yu & Huang, 2021).

AI has facilitated the development of news proposal systems (Helberger, 2019; Túñez-López et al., 2021) and has become a crucial technology for exploring data, analyzing trends, identifying topics of interest, and verifying data (Latar, 2015; Noain-Sánchez, 2022). Journalists increasingly use social networks, websites, and online forums to search for news on the internet, allowing them to quickly explore news events (Parratt-Fernández et al., 2021). AI applications in data analysis serve to create content and understand readers (Chou et al., 2021). Charlie (Charlie, 2019) found that machine learning, automation, and data processing are the primary AI technologies used in news production, which comprises gathering, creating, and spreading the news. AI has made mass media more competitive in today’s fragmented media landscape (Chou et al., 2021).

However, AI application in news production faces several limitations. AI models are often designed for specific stories, requiring re-creation and training for new projects, which prevents cost amortization (Stray, 2019). Investigative journalism projects using computer vision necessitate significant investments in technology infrastructure and skilled personnel (de-Lima-Santos & Mesquita, 2021). Furthermore, AI models often use outdated and biased datasets, raising ethical concerns (Guzman & Lewis, 2020). AI implementation in newsrooms also requires substantial expenses (Broussard et al., 2019).

While AI is not a panacea for journalism, it is a novel instrument that necessitates that news agency personnel gain a deeper understanding to promote and advance newsroom AI skills. Power dynamics between stakeholders, AI enforcement tools, and adherence to legal and ethical standards must be explicitly considered (Broussard et al., 2019). AI tools and algorithms may reflect the values and biases of their developers, and technology companies have little incentive for transparency regarding algorithmic bias (Bird et al., 2016). Research has revealed societal gender biases (Bolukbasi et al., 2016), racial biases (Buolamwini & Gebru, 2018; Campolo & Crawford, 2020) or foment violence, spread falsehoods, to diminish self-esteem (Haugen, 2023), and more in AI systems. Such biases can pose challenges when applying AI systems across different countries or cultures, as the encoded values may not automatically adjust to new contexts. Consequently, the study of AI usage in journalism, both in theory and practice, should consider diverse cultural contexts worldwide.

The Unified Theory of Acceptance and Use of Technology (UTAUT) in Media Journalism

UTAUT has been a focal point in media journalism research, with studies examining the influence of factors such as effort expectancy, performance expectancy, and social influence. Additionally, research has explored the evolution track and future research trends based on UTAUT and its application to journalists and technology adaptation (Ahadzadeh et al., 2021; Peng & Miller, 2021; C. T. Pham & Thi Nguyet, 2023). UTAUT has been used to investigate every stage of technology adoption, from pre-adoption through post-adoption use. Performance expectancy, effort expectancy, social influence, and facilitating conditions are among the fundamental ideas of UTAUT (Venkatesh et al., 2003). These concepts have to do with whether someone thinks using a system will improve their job performance, how simple it is to use, whether they believe important others think they should use the new system, and whether they think there is an organizational and technological infrastructure in place to support system use.

The research agenda presents four primary research opportunities: antecedents/determinants of UTAUT constructs (Morris & Venkatesh, 2010; Morris et al., 2005; Venkatesh et al., 2004), interventions, moderators of UTAUT relationships, new predictors, and consequences. It is crucial to note that UTAUT has been a robust general theoretical model, with its constructs proving predictive in various contexts and technologies (Venkatesh et al., 2012, 2016; Xu et al., 2017).

Venkatesh (2022) discusses the utilization of UTAUT as a theoretical basis for examining antecedents/determinants tailored to specific technologies. This examination framework encompasses individual characteristics such as personality (Thong, 1999); technology characteristics like quality (Venkatesh et al., 2007); environmental characteristics, including the culture of innovation (Venkatesh & Bala, 2008); and interventions such as training (Venkatesh et al., 2016). Individual factors like risk-seeking and desire to learn (Venkatesh, 2000) play a significant role in adopting and using AI tools. Technology characteristics, including perceptions of model errors and AI model transparency, can impact UTAUT predictors (Venkatesh, 1999). Environmental factors like organizational climate promoting innovation can also influence AI tool adoption and usage. Interventions, including training, can be studied to evaluate their effects on adoption and use (Venkatesh & Bala, 2008).

The UTAUT model has undergone many investigations to ensure its validity, and the findings have consistently demonstrated that it is a viable and trustworthy model for forecasting technology adoption. For additional empirical research, Bervell and Umar (2017) suggested comparing the proposed full UTAUT model with non-linear relationships and modifiers to the original UTAUT model. The UTAUT model was a superior technology acceptance and used predictor compared to other competing models. Jacob and Pattusamy (2020) also developed empirical support for the UTAUT model using data from Germany and India. The research discovered that the UTAUT model was a trustworthy indicator of technological adoption and utilization in both nations.

Additionally, Abdullahi et al. (2021) did a study to evaluate language teachers at public institutions in North-Central Nigeria’s awareness of and integration of e-learning amid the COVID-19 pandemic. The UTAUT model was superior to eight theories/models in empirical testing and validation. Empirical testing has shown that the UTAUT model is ideal for other widely used rival models. It is a solid and accurate model for estimating technological adoption and utilization. It will take further research to refine the idea and address its flaws.

Based on these discussions, this study uses UTAUT as a theoretical framework to suggest possible research topics in individual characteristics, technical characteristics, environmental characteristics, and interventions. The recommendations in this study could help organizations positively affect the adoption of AI tools and contribute to the adoption literature and a springboard for additional research (Venkatesh, 2022; Zhang & Venkatesh, 2018).

Questionnaire Design

To gather data, we employed a questionnaire. UTAUT constructs, Trust, Regulatory support, and Technology affinity were selected as the primary constructs of the research model from measurement constructs created in relevant studies (Aldás-Manzano et al., 2009; Kumar Bhardwaj et al., 2021; C. T. Pham & Thi Nguyet, 2023; Venkatesh & Zhang, 2010; Venkatesh et al., 2003; Zhu et al., 2021).

Six journalism and communication specialists were consulted after we had prepared the questionnaire to enhance its semantics and content validity. A pre-test with 25 journalists was conducted after expert revision to ensure the questionnaire’s phrasing, completeness, sequencing, and other potential errors. Following input from respondents, the questionnaire was slightly revised to improve its clarity and thoroughness. The assessment tool utilized in this study had 32 items in the final questionnaire, with 9 components in our research model (see Appendix 1). The first section’s questions were all scored using a 5-point Likert scale, from “strongly disagree” to “strongly agree,” for each response. Basic information about the respondent’s attributes, such as age, gender, and duration of journalism employment, was also provided.

Data Analysis

The Partial Least Squares-Structural Equation Modeling (PLS-SEM) approach was employed to examine the conceptual model, utilizing SmartPLS software version 4 (Ringle et al., 2022). The analysis followed a two-step procedure, beginning with evaluating the measurement model to ensure validity and reliability, then examining the structural model to test the proposed hypotheses (Hair et al., 2017). The measurement model defines the methods for measuring each construct, while the structural model delineates the relationships between these constructs in the overall model.

PLS-SEM has gained widespread acceptance as a multivariate method frequently used in social science research (Khan et al., 2018). The choice of PLS-SEM for this study is based on its ability to simultaneously analyze both the measurement and structural models, resulting in more precise estimations (Al-Saedi et al., 2020). Data collected via the questionnaire was entered into an Excel spreadsheet and subsequently imported into SmartPLS for analysis, allowing for the evaluation of the research model and the hypothesized relationships between constructs.

Data Collection and Participants

Most of the study’s participants were journalists from Hanoi, Vietnam, who were chosen using convenience sampling. To gather information, we spoke with them in person. All participants agreed before data was collected, and they were assured that their answers would be kept private and used exclusively for research. From October 2022 to December 2022, data was gathered. We gathered 238 surveys in all. We took into account the Partial Least Squares SEM (PLS-SEM) sample size recommendations; as a result, 188 observations are needed to achieve an 80% statistical power to detect R2 values of at least 0.1 with 1% probability of error (Hair et al., 2017, p. 48) when the measurement and structural models include eight independent variables. Consequently, the study’s sample size was sufficient.

The 238 participants’ demographic and employment-related details are shown in Table 1. The group’s gender composition is primarily male (57.1%), with females accounting for less than half of the total (42.9%). Most of the participants, or 34.4%, are between the ages of 30 and 39 when the age range of the participants is considered. 18.1% of the group is 40 to 49 years old, and a close second, 19.3%, is 50 or older. After that, 28.2% of the participants are between the ages of 22 and 29, indicating a workforce that is still quite young. This demonstrates a good representation of all age groups in the sample, with a bit of bias toward younger ages. The statistics also shed light on how long the individuals worked. Most of the group (38.2%) have been employed for 1 to 2 years, while 35.7% have been employed for 3 to 5 years. The lowest group is six years or more of employment (26.1%). According to this breakdown, most of the workforce has only five years of experience or less.

OPEN IN VIEWER

Table 1.

Participants’ Information (N = 238).

Demographic information	Count	Percent
Gender
Male	136	57.1
Female	102	42.9
Age
22–29	67	28.2
30–39	82	34.4
40–49	43	18.1
≥50	46	19.3
Length of time worked
1–2 years	91	38.2
3–5 years	85	35.7
≥6 years	62	26.1
The most type of AI adoption in working
Text to Speech and Speech to Text	94	39.5
Video, Image, and text processing	88	37.0
Content creation	18	7.6
Content editing	27	11.3
Other	11	4.6

The many types of AI adoption at work are detailed in Table 1. Text-to-Speech and Speech-to-Text technology is the group’s most used type of AI, employed by 39.5% of the participants. Video, picture, and text processing, used by 37.0%, are next. Other kinds of AI technology adoption have fallen significantly. Only 7.6% of participants utilize AI for content production; compared to 11.3% who use content editing, 4.6% of respondents said they used other AI systems. According to the data, this group mainly uses text-to-speech and video, image, and text processing; other AI techniques are employed far less frequently.

Common Method Bias and Multicollinearity

We calculated the Variance Inflation Factor (VIF) values to ensure the acquired data didn’t have common method bias (CMB). If all of the VIFs in the outer model (see Table 1) and all of the VIFs in the inner model (see Table 2) are equal to or lower than 3.3 as a result of a thorough collinearity test, the model is thought to be free of CMB (Kock, 2015; Kock & Lynn, 2012). The results showed that All VIFs had values lower than 3.3; hence, the data gathered had no CMB-related problems. Table 2’s VIF value data reveals that all VIF coefficients are below 3. As a result, according to Hair et al. (2019), there was no multicollinearity in this research model.

OPEN IN VIEWER

Table 2.

Collinearity Statistics—Inner Model.

Inner paths	BI→AA	EE→BI	FC→BI	PE→BI	RS→AA	SI→BI	TA→AA	TR→BI
Variance Inflation Factor (VIF)	1.743	1.872	1.282	2.003	1.465	2.065	1.829	1.658

Reliability and Validity

At first, the reliability and validity were checked for all constructs. Internal consistency usually indicates reliability, measured by Cronbach’s alpha, composite reliability rho_c and rho_a. According to the findings (see Table 3), Cronbach’s alpha, Composite reliability rho_c, and rho_a were greater than .7, indicating that all scales fulfilled reliability standards (Hair et al., 2012, 2017).

OPEN IN VIEWER

Table 3.

Construct Reliability and Validity.

Constructs	Items	Factor loadings	Cronbach’s alpha	Composite reliability (rho_a)	Composite reliability (rho_c)	Average variance extracted (AVE)	VIF—outer model
AA	AA1	0.842	0.720	0.731	0.843	0.642	1.564
	AA2	0.827					1.486
	AA3	0.730					1.299
BI	BI1	0.836	0.774	0.774	0.869	0.689	1.619
	BI2	0.836					1.678
	BI3	0.817					1.505
EE	EE1	0.757	0.755	0.758	0.844	0.576	1.450
	EE2	0.778					1.485
	EE3	0.781					1.487
	EE4	0.719					1.380
FC	FC1	0.626	0.745	0.823	0.827	0.548	1.366
	FC2	0.661					1.456
	FC3	0.841					1.412
	FC4	0.809					1.465
PE	PE1	0.802	0.812	0.812	0.876	0.639	1.675
	PE2	0.823					1.868
	PE3	0.778					1.545
	PE4	0.794					1.666
RS	RS1	0.794	0.764	0.776	0.863	0.677	1.573
	RS2	0.818					1.476
	RS3	0.856					1.624
SI	SI1	0.826	0.773	0.780	0.868	0.687	1.572
	SI2	0.802					1.538
	SI3	0.858					1.669
TA	TA1	0.771	0.821	0.822	0.882	0.650	1.538
	TA2	0.819					1.726
	TA3	0.823					1.866
	TA4	0.811					1.810
TR	TR1	0.809	0.826	0.827	0.885	0.657	1.843
	TR2	0.807					1.686
	TR3	0.835					1.972
	TR4	0.792					1.608

The concept of convergent validity is the measurement of conceptual similarity across items. To investigate the concurrent validity, this study investigated factor loadings and average variance extracted (AVE). According to the findings in Table 3, the AVE values varied from 0.548 to 0.689, exceeding the recommended value of 0.5 (Hair et al., 2017). Only FC1 and FC2 factor loadings were lower than the recommended value of 0.7, ranging from 0.6 to 0.7, but in social research studies, and this value can be accepted (Hair et al., 2012; Hulland, 1999), so most items of factor loadings have satisfied the requirements, according to Table 3. These findings enable convergent validity to be established.

This study uses the Heterotrait-Monotrait ratio (HTMT) criterion to evaluate the discriminant validity. The HTMT has been proven to perform better than the prior criteria for the measurement of discriminant validity (Henseler et al., 2015). Table 4 indicates that the HTMT criterion is met because all values were below the suggested threshold of 0.9 (Hair et al., 2017).

OPEN IN VIEWER

Table 4.

Heterotrait-Monotrait Ratio (HTMT) Matrix.

Construct	AA	BI	EE	FC	PE	RS	SI	TA
BI	0.836
EE	0.862	0.783
FC	0.412	0.386	0.528
PE	0.841	0.818	0.691	0.370
RS	0.748	0.637	0.626	0.239	0.603
SI	0.786	0.774	0.820	0.505	0.756	0.552
TA	0.816	0.781	0.721	0.375	0.782	0.663	0.733
TR	0.684	0.812	0.584	0.220	0.721	0.620	0.63	0.686

Structural Model Assessment

The structural model is the next stage after the validated measurement model. To find the coefficient of determination and the path coefficients, a bootstrapping technique with 5,000 additional samples is required (Hair et al., 2017). According to the results, Q² values of AA = 0.334 and BI = 0.400 from the construct cross-validated redundancy and PLS Predict showed that the model had a stronger predictive potential (see Table 5), and the PLS-path model had a moderate to sizeable predictive relevance (Hair et al., 2019). The robustness of the prediction model was evaluated by the holdout test’s Q² values (AA = 0.534, BI = 0.572) from the PLS predict (Hair et al., 2017). These findings demonstrated that the PLS-SEM model demonstrates acceptable predictive relevance and accuracy for AA and BI. This shows that the model could be able to predict the values of AA and BI if it makes use of the relationships between the independent variables.

OPEN IN VIEWER

Table 5.

PLS Predict and Construct Cross-Validated Redundancy.

Construct	Q2 (=1 − SSE/SSO)	Q2 predict
AA	0.334	0.534
BI	0.400	0.572

As indicated in Table 6 and Figure 2, the structural model of the assessment indicates hypotheses testing. The findings indicated that PE, EE, and SI significantly impact AI adoptions’ behavioral intention to use in journalism. Thus, H1 (β = .245, t = 3.554), H2 (β = .223, t = 3.717), H3 (β = .148, t = 2.345) are supported. Furthermore, the results pointed out that Trust significantly positively affects behavioral intention to use AI in journalism. Hence, H6 is supported (β = .327, t = 5.312). Moreover, the results revealed that perceived FC, BI, RS, and TA significantly positively affect AI adoption in journalism. Therefore, H4 (β = .109, t = 2.211), H5 (β = .291, t = 4.202), H7 (β = .257, t = 3.944), and H8 (β = .276, t = 3.621) are supported.

OPEN IN VIEWER

Table 6.

Hypothesis Testing.

Hypothesis	PLS Paths	Beta-values (β)	t-Value	p-Values	Decision	f-square
H1	PE → BI	.245	3.554	.000	Supported	0.075
H2	EE → BI	.223	3.717	.000	Supported	0.068
H3	SI → BI	.148	2.345	.019	Supported	0.027
H4	FC → AA	.109	2.211	.027	Supported	0.022
H5	BI → AA	.291	4.202	.000	Supported	0.103
H6	TR → BI	.327	5.312	.000	Supported	0.160
H7	RS → AA	.257	3.944	.000	Supported	0.098
H8	TA → AA	.276	3.621	.000	Supported	0.089

The effect size (f-square) quantifies each predictor’s relative contribution to the dependent variable’s variance. It is employed to ascertain the relationships’ practical importance in the model (J. Cohen, 1988). The results showed that FC, BI, RS, and TA have a medium effect on AA, whereas PE, EE, and SI have a medium effect on BI, and TR has a larger effect on BI, according to the f-square values in Table 6.

OPEN IN VIEWER

Figure 2.

Structural model.

Importance-Performance Map Analysis (IPMA)

As an advanced method of PLS-SEM, the importance-performance map analysis (IPMA) is used in this study, with BI and AA as the goal variables. The IPMA technique, which considers the average value of latent variables and associated indicators (i.e., performance measure), can be used to understand better PLS-SEM results (Hair et al., 2017; Ringle & Sarstedt, 2016). The findings of the IPMA are shown in Figures 3 and 4. All of the independent variables’ significance and effectiveness were evaluated. The results of IPMA in Figure 3 indicate that TR has the highest importance for predicting the behavioral intention to use, with the highest effect and performance scores. Concerning the importance of measures in AI adoption, Figure 4 shows that BI is the most important factor impacting AI adoption, followed by TA and RS.

OPEN IN VIEWER

Figure 3.

IPMA for behavioral intention.

OPEN IN VIEWER

Figure 4.

IPMA for AI adoption.