Development of an Instrument to Measure Perceptions of Responsibility for Deepfakes

Abstract

Deepfakes can distort reality and communicate disinformation so convincing that individuals find it difficult to differentiate real from fake, which can have significant real-world effects. Faced with the challenges of Deepfakes, individuals will assign responsibility for Deepfakes to various entities and that responsibility allocation will influence many issues including regulation, distribution, legal responsibility, technological response, and societal impact among other things. To facilitate theory development and testing, our objective is to develop a survey instrument that assesses individual perceptions of responsibility associated with the Deepfake phenomenon. An initial study (N = 535) and replication study (N = 488) were conducted to create and validate this instrument. Results were then tested against a general sample of the U.S. population (N = 340) as a final validation study. Our results demonstrate reliability and discriminant validity of the 39-item survey. By understanding individual perceptions of responsibility, we aim to establish starting points for the creation of tools, techniques, policies, and procedures for improving decision-making and addressing misinformation created by Deepfakes.

Introduction

Deepfakes are hyper-realistic artificial intelligence (AI)-created fake videos that leave little trace of manipulation. They can be created with increasing ease. Human accuracy in identifying high-quality Deepfakes is about 26 percent, but others have found much lower accuracy.^1,2 Deepfakes are relatively recent phenomenon, and academic scholarship is rapidly growing. Many studies have focused on Deepfake detection methods in computer science research and legal implications and regulation in law research. Still, there is a great need for research to focus on the perspective of individual perceptions and understanding of Deepfakes, familiarity with detection options, and their ability to recognize Deepfakes. Furthermore, it is important to understand the extent that Deepfakes undermine users trust in institutions, including media brands, political institutions, and agents such as tech companies or fact-checking organizations.³

Our study addresses these areas by asking the research question: How do individuals assign responsibility for Deepfakes and their effects? This understanding can play a key role in the link between the individual faced with Deepfakes and technological and societal actions to address Deepfakes.

We analyze the results of three independent surveys (identified as initial, replication, and broad population) to develop a 39-item instrument along eight dimensions of Deepfake responsibility. This instrument will be valuable for researchers and practitioners addressing the challenges of Deepfakes. For researchers, this survey has identified reliable constructs that may be beneficial for theory building and theory extension in the context of Deepfakes, humor, trust, or other areas. For practitioners, this survey can be a starting point for assessing how to attack the problems presented by Deepfakes through a variety of means including education of the general population, regulation, platform adjustments, technology solutions, or other means.

Responsible parties

To understand who individuals consider responsible parties in the context of Deepfakes, we conducted semistructured conversations with 20 mid- to executive-level managers from firms including Equifax, Steelcase, Global Payments, Georgia-Pacific, Delta Air Lines, Comcast Business, and Home Depot. We directly tie their responses to the development of our constructs.

The parties identified are in line with research. Terminology from these conversations indicated an externalization of responsibility for control and regulation of Deepfakes. Concerns about dissemination were connected to social networks and media platforms. Ideas related to prevention were tied to government institutions and social networks. Research has found that individual cynicism increases with respect to media overall when individuals question their ability to determine if a source is a Deepfake.⁴ This aligns with the idea that individuals will search for different responsible parties to aid as they try to establish trust in information. Furthermore, prior research has indicated that when confronted with Deepfakes, individuals feel that they have a personal responsibility to recognize artificial media and also expect social media platforms to intervene when Deepfakes are present.^5,6

Following are descriptions of items related to eight areas of interest around Deepfake responsibility. The related items for each of the eight areas are shown in Table 1.

Table 1.

Survey Instrument Dimensions and Items

Item		Factorloading^a
Individual concern
IC3	“Deepfakes” will make me less likely to use online videos to understand important issues.	0.68
IC4	“Deepfakes” reduce my willingness to rely on information from online videos.	0.86
Humorous perception
HP1	I am amused by “Deep Fake” videos.	0.86
HP2	“Deepfakes” are funny.	0.84
HP3	“Deepfakes” make me laugh.	0.90
HP4	“Deepfakes” are good entertainment.	0.87
Individual responsibility
IR1	It is my responsibility to be able to determine if an online video is a “Deep Fake.”	0.82
IR2	Viewers of online videos are responsible for figuring out if a video is a “Deep Fake.”	0.70
IR3	We each are responsible for knowing if the online videos we watch are authentic.	0.73
Government responsibility
GR1	It is the government’s responsibility to determine if an online video is a “Deep Fake.”	0.80
GR2	The government is responsible for figuring out if a video is a “Deep Fake.”	0.77
GR4	The government is responsible for policing “Deepfakes.”	0.82
GR5	The government should take action to control “Deepfakes.”	0.83
GR6	Deepfakes are best controlled when the Government takes action.	0.71
GR7	It is the government’s responsibility to ensure verified and honest video content is available to the public.	0.84
GR9	The government should protect us from “Deepfakes.”	0.82
Deepfake software developer
SD1	“Deep Fake” software developers are responsible for preventing misuse of “Deepfakes.”	0.78
SD2	“Deep Fake” software developers should take action to control “Deepfakes.”	0.68
SD4	Technology firms that make the software used to create “Deepfakes” should be held responsible.	0.80
Identification technology developers
IT1	It is the identification technology developers’ responsibility to be able to determine if an online video is a “Deep Fake.”	0.81
IT2	Identification technology developers are responsible for figuring out if a video is a “Deep Fake.”	0.76
IT3	Identification technology developers are responsible for staying ahead of the curve on Deep Fake issues.	0.82
IT5	Identification technology developers are responsible for solving the problem of “Deepfakes.”	0.77
IT6	Identification technology developers should take action to control “Deepfakes.”	0.81
IT7	Deepfakes are best controlled when identification technology developers take action.	0.78
IT8	It is the responsibility of the identification technology developers to ensure verified and honest video content.	0.82
IT10	Identification technology developers should protect us from “Deepfakes.”	0.85
Platform responsibility
PR1	It is the platform’s responsibility to determine if an online video they host is a “Deep Fake.”	0.84
PR2	Platforms that host online videos are responsible for figuring out if a video is a “Deep Fake.”	0.77
PR3	Platforms are responsible for knowing if the online videos that they host are “Deepfakes.”	0.84
PR4	Platforms are responsible for policing “Deepfakes.”	0.77
PR5	Platforms should take action to control “Deepfakes.”	0.78
PR6	“Deepfakes” are best controlled when platforms are held responsible.	0.74
PR7	It is the responsibility of a platform to provide verified and honest content.	0.80
Media outlet responsibility
MO1	It is the media outlet’s responsibility to be able to determine if an online video is a “Deep Fake” before distributing it.	0.81
MO2	Media outlets that post online videos are responsible for figuring out if the video in a post is a “Deep Fake.”	0.79
MO3	Media outlets are responsible for knowing if the online videos they use are “Deepfakes.”	0.82
MO6	It is the responsibility of a media outlet to provide verified and honest content.	0.80
MO8	Media outlets should verify if the video sources they use are authentic.	0.78

Factor loading is from general population study described later.

Individual concern

There was a high level of concern about effects of Deepfakes present throughout our executive conversations. One respondent stated, “I am very concerned. The rate of progress in deepfake technology quality is staggering.” An executive of a remote company focused on how a “fake” video of our CEO could wreak havoc on our operations. Concerns consolidated around identity theft, direct fraud, external consumer-facing reputational risk, and internal communications. See Table 1 (IC).

Individual responsibility

Our executive conversations supported the “responsibility of individuals” to understand the nature of Deepfakes and to take action to identify them and protect themselves. One respondent stated, “Individuals will have to be cautious about where they get their information from,” while another stated, “Individual awareness and skepticism are critical.”

Individuals with a strong internal locus of control are likely to believe in their ability to execute a task effectively.⁷ In the context of Deepfakes, this translates into confidence in detection/identification. See Table 1 (IR).

Humorous perception

While not directly identified in our executive conversations, many Deepfakes are presented as humor. Humor increases trust across multiple settings including broadcast media where it increases positive affect and reduces negative affect.⁸ Humor has been found to be a moderating factor in perceptions of Deepfakes.⁹ See Table 1 (HP).

Government responsibility

In our executive conversations, a typical response was “The government needs to get involved. There is a need for regulations to be put in place. There should be regulations around the use of someone’s image.” While people have varying expectations of governmental roles generally, one might expect government to play a role in setting forth and enforcing regulations related to the potential consequences of Deepfakes. See Table 1 (GR).

Deepfake software developer responsibility

The role of Deepfake software developers was prominent in our executive conversations. One executive said, “Developers and those related to the creation of this technology should be responsible for addressing Deepfakes.” There is a growing consensus to hold firms accountable for the consequences of their products and actions.¹⁰ See Table 1 (SD).

Platform responsibility

The role of social media platforms was highlighted in our conversations with executives. One manager stated, “Social media has already created paths of distrust that have been evident in the past several elections, large news stories, scientific discoveries.” Social media platforms host and rapidly transmit Deepfakes through their social networks. Their algorithms quickly amplify content, pushing it “viral.” However, these algorithms are internal secrets unavailable for external scrutiny, and the “codes of conduct” these firms espouse are haphazardly enforced.¹¹ Among some, there is a growing call to hold platforms responsible for the material transmitted through their networks.¹² See Table 1 (PR).

Media outlet responsibility

Media outlets play an important role in video that could be used to create Deepfakes. Media outlets are recognized media organizations that include their brand when distributing a video, such as CNN, Fox News, and the New York Times. Media outlets play two roles. First, they may inadvertently spread false information by unknowingly using a Deepfake. Second, authentic videos they release may be used in the generation of Deepfakes but contain the media outlet’s brand.

Brand is an important asset for businesses and a critical factor in consumer intentions.¹³ Use of branded videos for Deepfakes may influence user perception of truthfulness when the content is artificial. As such, we expect that individuals will hold media outlets accountable for preventing the deceptive use of their branded video. See Table 1 (MO).

Identification technology developers

During our executive conversations, some participants presumed that an identification technology would emerge to blunt the risk. One executive stated, “I’m not numb to it, but I won’t go into work tomorrow asking others to make changes in our department.” Another was “… confident that there will be technology to determine if a video is fake or not.” An Identification Technology is any potential technology solution to identify Deepfake material. See Table 1 (IT).

Data analysis

A survey questionnaire featuring items based on the eight dimensions elaborated upon in the prior section was completed by study participants. Undergraduate students enrolled at a state university made up our initial study and replication study samples. The replication study involved a second round of data collection and analysis featuring the items that were retained in the measurement model. We performed a final validation study with a general U.S. population sample to bolster our standing on the reliability and validity goals mentioned above since the initial and replication samples featured undergraduate students. Qualtrics survey software was used to collect data via an anonymous survey instrument throughout. The characteristics of the study participants including the sample sizes for each data collection effort are summarized in Table 2.

Table 2.

Participant Characteristics

Variable	Initial (N = 535)	Replication (N = 488)	General population (N = 340)
Gender	27 percent males/73 percent females	30 percent males/70 percent females	50.8 percent males/49.1 percent females
Average age	22.72 years (7.15 years)	22.72 (7.47 years)	50.7 (17.8 years)
Range of age	17–72 years	17–64 years	18–91 years
Race	White 58 percent, Black 23 percent, Asian 7 percent, Hispanic 5 percent, Multirace 5 percent, Other 2 percent	White 57 percent, Black 23 percent, Asian 7 percent/Hispanic 3 percent, Multirace 5 percent, Other 5 percent	White 77.9 percent, Black 12.4 percent, Asian 3.2 percent, Hispanic 3.8 percent, Other 2.6 percent
Education	Undergraduate 100 percent	Undergraduate 100 percent	Less than high school 3.8 percent, high school graduate 25 percent, college/no degree 22.9 percent, associate degree 14.7 percent, bachelor degree 20.3 percent, master degree 10.3 percent, PhD degree 1.2 percent, professional JD, MD 1.8 percent

Value in brackets—standard deviation for continuous variables.

Indicator reliability and internal consistency

Cronbach’s alpha was used for reliability estimation.¹⁴ Item loading was evaluated on each dimension and the square of the total of factor loadings for a dimension. Confirmatory factor analysis was used for this analysis and was the first step in the validation and confirmation of our measurement model (Table 1). We used IBM, SPSS AMOS (version 27) software to build the models.

We noted the following observations of our initial study: (a) 8 of 71 factor loadings were below the minimum limit of 0.50; (b) Cronbach’s alpha of the initial model found 6 dimensions above the minimum limit of 0.70 and 2 dimensions below; (c) The chi-square (C_min/df) statistic was 2.61, suggesting a poor fit. The Kaiser–Meyer–Olkin (KMO) measure was 0.94, Bartlett’s test of sphericity (BTS) was associated with a p value of 0.000, all communalities were above 0.30, and the initial eigen values indicated that about 64.37 percent of the total variance can be explained by the factors—all providing support that the data are adequate for factor analysis.

By deleting nine measures with low factor loadings individually and gauging their impact in tandem, we obtained the measurement model. Individual concern was the only construct with two remaining items. We ensured that there were no identification issues for individual concern.^15,16 All other constructs had at least three items, and their errors are uncorrelated.

Analysis of the measurement model in the initial study found each variable associated with a factor loading above 0.50 demonstrating reliability. The square of the total of factor loadings for each dimension is well above 0.70 (Table 3), indicating internal consistency. Cronbach’s alpha surpassed 0.70 (Table 4). KMO and BTS measures, communalities, and eigen values were consistent with the initial model. Fit indices (Table 5) are within acceptable limits.¹⁴ Our measurement model exhibits excellent fit and reliability. Root mean square error of approximation (RMSEA) was 0.046 (with a p value associated with PCLOSE equal to 0.98), and the chi-square statistic (C_min/df) was 2.13 (Table 5).

Table 3.

Construct Reliability Measurement

	Individual concern	Humorous perception	Individual responsibility	Government responsibility	Deepfake software developer	Identification technology	Platform responsibility	Media outlet
Initial	74.89 percent	91.65 percent	84.18 percent	93.10 percent	78.98 percent	90.42 percent	91.13 percent	85.71 percent
Replication	78.72 percent	93.05 percent	84.76 percent	93.57 percent	76.74 percent	89.84 percent	91.77 percent	88.16 percent
General population	74.96 percent	92.95 percent	79.51 percent	92.81 percent	79.62 percent	93.44 percent	92.06 percent	90.36 percent

Table 4.

Reliability (Measurement Model)

Dimension	Number ofquestions	Cronbach’salpha
Individual concern	2	0.74
Humorous perception	4	0.92
Individual responsibility	3	0.84
Government responsibility	7	0.93
Deepfake software developer	3	0.79
Identification technology	8	0.91
Platform responsibility	7	0.91
Media outlet	5	0.86

Table 5.

Fit Metrics (Measurement Model)

Goodness of fit index	Initial	Replication	General population
χ²/df	2.13	2.15	2.052
RMSEA	0.046	0.049	0.056
Comparative Fit Index (CFI)	0.94	0.94	0.93
Goodness of Fit Index (GFI)	0.90	0.89	0.85
Normed Fit Index (NFI)	0.89	0.89	0.87
Standardized Root Mean Square Residual (SRMR)	0.043	0.050	0.051

We performed a replication study to confirm the construct validity of the measures and to test the relationships of the overall instrument. Fit indices (Table 5) associated with the replication study are within acceptable limits. The RMSEA value of 0.049 (with a p value associated with PCLOSE equal to 0.76) and a chi-square statistic (C_min/df) equal to 2.15 serve as good examples of this effect and compare favorably to the initial study. The factor loading for each variable surpassed 0.50 (Table 1), demonstrating reliability. The square of the total of factor loadings for a dimension surpassed 0.70 (Table 3), indicating internal consistency. Cronbach’s alpha for the measurement items surpassed 0.70, the KMO was 0.94, the BTS was associated with a p value of 0.000, all communalities were above 0.50, and the initial eigen values indicated that 70.24 percent of the total variance is explained by the factors.

A final validation study with a general U.S. population sample was conducted. The results in Tables 1, 3, and 5 find the reliability and validity of the instrument stable across the initial sample, the replication study, and the general U.S. population study. Description of demographic characteristics is shown in Table 2.

Convergent, discriminant, and nomological validity

We examined convergent and discriminant validity for each of these dimensions to assess whether the dimensions are different enough to be useful for research purposes. Convergent validity (variance extracted) results for the initial study, the replication study, and the general U.S. population study, each surpassed the desired minimum of 50 percent (Table 6).

Table 6.

Convergent Validity: Variance Extracted

	Individualconcern	Humorousperception	Individualresponsibility	Gov’tresponsibility	Deepfakesoftwaredeveloper	ID Tech	Platformresponsibility	Mediaoutlet
Initial study	59.98 percent	73.34 percent	63.96 percent	65.89 percent	55.61 percent	54.17 percent	59.51 percent	54.56 percent
Replication study	64.93 percent	77.06 percent	65.00 percent	67.53 percent	52.37 percent	52.56 percent	61.50 percent	59.88 percent
General population	60.29 percent	76.72 percent	56.52 percent	64.91 percent	56.69 percent	64.07 percent	62.41 percent	65.23 percent

Validation of discriminant validity can be accomplished by a comparison of variance extracted (Table 6; these values also make up the diagonal of the matrix in Table 7, also referred to as AVE) with the square of the interitem correlation (Table 7 for results for the initial study and the replication study). All the average variance extracted estimates (AVE) exceeded corresponding squared interitem correlation estimates. The indicators have more in common with the dimension they are associated with than they do with other dimensions. This demonstrates discriminant validity. This analysis holds for the general U.S. population sample.

Table 7.

Discriminant Validity (Reduced Model—Initial Study) Using CFA: Squared Interconstruct Correlation Estimates vs. Variance Extracted

	Individualconcern	Humorousperception	Individualresponsibility	Governmentresponsibility	Deepfakesoftwaredeveloper	Identificationtechnology	Platformresponsibility	Mediaoutlet
Individual concern	0.60
Humorous perception	0.028	0.73
Individual responsibility	0.011	0.038	0.64
Government responsibility	0.11	0.036	0.031	0.66
Software developer	0.20	0.12	0.006	0.39	0.56
Identification technology	0.19	0.035	0.002	0.301	0.40	0.542
Platform responsibility	0.13	0.077	0.001	0.39	0.51	0.40	0.60
Media outlet	0.037	0.029	0.022	0.087	0.15	0.14	0.267	0.55
Discriminant Validity (Reduced Model—Replication Study) Using CFA: Squared Interconstruct Correlation Estimates vs. Variance Extracted
Individual concern	0.65
Humorous perception	0.050	0.77
Individual responsibility	0.010	0.020	0.65
Government responsibility	0.086	0.040	0.068	0.68
Software developer	0.25	0.076	0.057	0.29	0.52
Identification technology	0.17	0.016	0.026	0.31	0.30	0.53
Platform responsibility	0.21	0.051	0.049	0.27	0.52	0.30	0.62
Media outlet	0.066	0.005	0.003	0.033	0.14	0.088	0.32	0.60

CFA, confirmatory factor analysis.

Common method bias is a concern.¹⁷ First, a Harmon one-factor test showed that the most covariance explained by one factor is 26.62 percent, indicating that our results are not contaminated by common method biases. Second, a common method factor whose indicators included all the principal dimension indicators was included and followed by a computation of each indicator’s variances substantively explained by the method.¹⁷ The average method variance was found to be 0.027, and most method variances were insignificant. We also compared the standardized regression weights from this model with the standardized regression weights of a model without the common method factor and did not note large differences.¹⁸ Third, a marker variable test was performed through the addition of another latent factor—health environment sensitivity, with six measures, uncorrelated with the other latent factors in the model. A subsequent zero constraints test indicated that the null hypothesis cannot be rejected (i.e., the constrained and unconstrained models are the same or “invariant,” p value ≈1).¹⁹ We did not detect any response bias affecting our model and contend that common method bias does not pose a concern. Our measurement model showed acceptable results throughout, indicating that it is an acceptable measurement model for individual perceptions of Deepfakes responsibility.

Discussion

Our empirical analysis affirms that Deepfake responsibility is a distributed concept with multiple parties held responsible with individual traits likely impacting the allocation of responsibility. Individuals do not assess any one party primarily responsible for the impact of Deepfakes, reinforcing the importance of taking an encompassing approach. Understanding responsibility around Deepfakes is important to the development of interventions and the eventual development of accountability including legislation, legal action, social activism, and social stigma. This instrument and understanding that Deepfake responsibility is a distributed concept may assist those individuals and organizations that are attempting to deal with the challenges presented by Deepfakes in a practical manner.

The relationship among the responsible parties deserves deeper conceptualization and further empirical exploration, which this scale development facilitates and other research has begun to examine.^4,6,9 Researchers could leverage the validated items and concepts presented in this study for theory building and theory extension. Importantly, this instrument includes responsibility constructs that can be used as dependent variables in the assessment of intervention designs as well as other constructs that may be important moderators. By providing validated dependent variables and moderators, scholars can begin to examine theory-based interventions, policies, and societal expectations around Deepfakes.

As AI technology advances, the sophistication of Deepfakes will increase, eliminating the subtle contextual triggers that raise “red” flags. Individuals with a strong belief in their own ability to identify Deepfakes may find themselves fooled and seek to externalize responsibility. Aside from use in cross-sectional studies, this survey may help scholars conducting longitudinal studies examining the shifting landscape of responsibility allocation as the technology develops and individuals become more aware and experienced with Deepfakes.

Conclusions

Often absent with new phenomena are validated measures of how those experiencing the phenomena are perceiving the experience. A better understanding of individual perceptions of Deepfake responsibility needs to develop for Deepfakes and their consequences to be effectively studied. Our survey instrument is a validated way to assess individual perceptions of responsibility for multiple internal and external aspects of the phenomenon using 39 items across 8 interest areas related to Deepfake responsibility. This survey instrument can be a starting point for the creation of tools, techniques, policies, and procedures needed to address the context.

As with any research project, there are limitations. We took the time and effort to conduct a replication study as well as an extension to validate our results against a broad demographic representative of the U.S. population. The study was conducted relatively early in the evolution of Deepfakes. Continuing efforts should be made to validate and extend this instrument.

Footnotes

Authors’ Contributions

S.N.: Conceptualization (equal), methodology (equal), original draft (lead), investigation (lead), and project administration (lead). J.A.P.: Conceptualization (equal), methodology (lead), data curation (lead), formal analysis (lead), validation (lead), and review and editing (equal). J.C.: Conceptualization (equal), methodology (equal), and review and editing (equal).

Author Disclosure Statement

The authors have no employment relationship with an organization that would gain or lose financially from publication of this research. The authors have no other competing interest that would impact the integrity of this research. The authors have no personal financial interest in the impact of this research or its publication.

Funding Information

No funding source was used that would gain or lose financially from publication of this research.

References

Tahir

, Batool

, Jamshed

, et al. Seeing is believing: Exploring perceptual differences in deepfake videos. In Proceedings of the 2021 CHI conference on human factors in computing systems 2021 May 6 (pp. 1–16); doi: 10.1145/3411764.3445699

Shahid

, Kamath

, Sidotam

, et al. It Matches My Worldview”: Examining perceptions and attitudes around fake videos. In proceedings of the 2022 CHI conference on human factors in computing systems 2022 Apr 29 (pp. 1–15); doi: 10.1145/3491102.3517646

Vasist

, Krishnan

, Indian Institute of Management Kozhikode. Deepfakes: An integrative review of the literature and an agenda for future research. CAIS, 2022; 51(1):590–636; doi: 10.17705/1cais.05126

Liu

, Wang

, Luo

. When seeing is not believing: Self-efficacy and cynicism in the era of intelligent media. Humanit Soc Sci Commun, 2025; 12(1):1–3; doi: 10.1057/s41599-025-04594-5

Lima

, Han

, Cha

. Others are to blame: Whom people consider responsible for online misinformation. Proceedings of the ACM on Human-Computer Interaction. 2022; 6(CSCW1):1–25; doi: 10.1145/3512953

Cochran

, Napshin

. Deepfakes: Awareness, concerns, and platform accountability. Cyberpsychol Behav Soc Netw, 2021; 24(3):164–172; doi: 10.1089/cyber.2020.0100

Lefcourt

. Locus of Control: Current Trends in Theory & Research. Psychology Press; 2014.

Strick

, Holland

, van Baaren

, et al. Those who laugh are defenseless: How humor breaks resistance to influence. J Exp Psychol Appl, 2012; 18(2):213–223; doi: 10.1037/a0028534

Napshin

, Paul

, Cochran

. Individual responsibility around deepfakes: It’s no laughing matter. Cyberpsychol Behav Soc Netw, 2024; 27(2):105–110; doi: 10.1089/cyber.2023.0274

10.

Shane

, Wakabayashi

. ‘The business of war’: Google employees protest work for the Pentagon. The New York Times. 2018. Available from: https://www.nytimes.com/2018/04/04/technology/google-letter-ceo-pentagon-project.html [Last accessed: August 20, 2021].

11.

Horwitz

. Facebook has made lots of new rules this year. It doesn’t always enforce them. The Wall Street Journal. 2020. Available from: https://www.wsj.com/articles/facebook-has-made-lots-of-new-rules-this-year-it-doesnt-always-enforce-them-11602775676?mod=article_inline [Last accessed: August 21, 2021].

12.

Eisenstat

. How to hold social media accountable for undermining democracy. Harvard Business Review. 2021. Available from: https://hbr.org/2021/01/how-to-hold-social-media-accountable-for-undermining-democracy [Last accessed: August 21, 2021].

13.

Delgado‐Ballester

, Luis Munuera‐Alemán

. Does brand trust matter to brand equity? Journal of Product & Brand Management, 2005; 14(3):187–196; doi: 10.1108/10610420510601058

14.

Hair

, Anderson

, Tatham

, et al. Multivariate Data Analysis (5th ed.). Prentice Hall: New Jersey; 2005.

15.

Rigdon

. A necessary and sufficient identification rule for structural models estimated in practice. Multivariate Behav Res, 1995; 30(3):359–383; doi: 10.1207/s15327906mbr3003_4

16.

Kenny

, Kashy

, Bolger

. Data analysis in social psychology. In Handbook of social psychology, 4th ed. ( Gilbert

, Fiske

, Lindzey

. eds) McGraw-Hill: New York; 1998; pp. 233–265.

17.

Podsakoff

, MacKenzie

, Lee

, et al. Common method biases in behavioral research: A critical review of the literature and recommended remedies. J Appl Psychol, 2003; 88(5):879–903; doi: 10.1037/0021-9010.88.5.879

18.

Serrano Archimi

, Reynaud

, Yasin

, et al. How perceived corporate social responsibility affects employee cynicism: The mediating role of organizational trust. J Bus Ethics, 2018; 151(4):907–921; doi: 10.1007/s10551-018-3882-6

19.

Simmering

, Fuller

, Richardson

, et al. Marker variable choice, reporting, and interpretation in the detection of common method variance: A review and demonstration. Organizational Research Methods, 2015; 18(3):473–511; doi: 10.1177/1094428114560023