Institutional archetypes of resistance to artificial intelligence

1 Introduction

The integration of generative artificial intelligence within media and marketing organisations does not encounter a uniform type of resistance, but rather qualitatively different epistemic attitudes. Experts who approach this integration with hesitation raise fundamentally different objections; specifically, some point to the absence of the organisational infrastructure necessary for the responsible deployment of artificial intelligence tools, whilst others express deep epistemological distrust of the final outputs produced by these tools. The two types of resistance mentioned above do not differ in degree but in their fundamental nature, and as the study's results empirically demonstrate, common methodological approaches—whether in the form of surveys on the frequency of obstacles or composite resistance scales—are unable to adequately capture this distinction. Research focused on the challenges of implementing artificial intelligence in a media context is divided into two methodological strands. The first catalogs the occurrence of individual barriers using descriptive statistics (Newman et al., 2024; Thurman et al., 2019), while the second uses regression models to verify whether the level of support differs across organizational groups (Beckett and Deuze, 2016). These two approaches ultimately define a limitation: within them, barriers are either treated as mutually uncorrelated items or aggregated into a single composite index of resistance, with both implicitly assuming the one-dimensional nature of the phenomenon under study. The assumption is not sufficiently theoretically grounded and, as research suggests, empirically verifiable. The theory of resistance to innovation (Huang et al., 2021) explicitly distinguishes between functional barriers, practical obstacles that can be addressed through targeted training and organizational policies, and psychological barriers that reflect value conflicts and epistemic dispositions that require systematic trust building and normative reassessment. The categories refer to empirically distinguishable professional profiles, which has direct and significant implications for the design of intervention strategies, where treating passive and active resisters equally will yield less than ideal outcomes for both groups. Research defines the importance of the above question. Focus group surveys of news organizations in the countries surveyed identify persistent skepticism toward AI in journalism, with only a minority comfortable with news stories generated solely by AI and expressing significantly higher trust when the process is led by a human (Fletcher and Nielsen, 2025). Parallel focus group research links skepticism to already low trust in news media and calls for greater transparency about where and how AI is being used (Minnesota Journalism Center, 2025). In terms of production, research on British journalists and newsrooms documents uneven adoption of AI, influenced not only by attitudinal and ethical reservations but also by skills gaps (Thurman et al., 2025), while qualitative studies show that media professionals are showing deliberate restraint and aligning the use of AI with established professional values, rather than rejecting or fully embracing the technology (Cools et al., 2024). It is important to add that in the field of marketing communications, there has been extraordinary progress in the adoption of AI, although legal, quality and ethical concerns remain (Buder et al., 2024). The aforementioned studies identify trust, transparency, and professional identity as recurring problem areas, but continue to address the barriers either individually or as a whole, without examining the underlying structure of resistance, a gap that the study addresses.

The research focus of the study is defined by three research questions:

• RQ1: What is the structure of the conditional dependencies of barriers to the implementation of artificial intelligence; do they form an interconnected syndrome or a sparse, partially orthogonal network?
• RQ2: Do media and marketing experts form latent classes defined by qualitatively distinct barrier profiles?
• RQ3: Do the identified archetypes differ in their management requirements, perceptions of artificial intelligence, and epistemological orientation based on IRT?

2 Theoretical framework

The theory of resistance to innovation (Huang et al., 2021) identifies two mutually independent sources of resistance to technology. The identified barriers stem from the perceived incompatibility of the innovation with current work processes, skills, and organizational norms. The obvious potential lies in the fact that these barriers can be overcome, particularly through appropriate training, clear guidelines, and communication with stakeholders. Psychological barriers, on the other hand, reflect value conflicts and epistemic dispositions, as well as doubts about whether the innovation is trustworthy, ethically appropriate, or consistent with one's professional identity. These barriers cannot be removed simply by providing information, as they require deeper processes of trust-building and a reevaluation of norms. The distinction is epistemological rather than purely attitudinal. In particular, the difference in the degree of trust reflects the extent to which an individual views the technology positively, the epistemological orientation relates to a prior dispositional stance from which evidence about the technology is interpreted—whether AI outputs are considered provisionally reliable until proven otherwise, or are inherently suspect regardless of case-specific evidence. Research examining journalists’ attitudes toward generative artificial intelligence finds that experts neither simply accept nor reject this technology, but rather negotiate their stance through attitudes rooted in their professional values and understanding of trustworthy knowledge (Cools et al., 2024). From this perspective, the archetype is not a transient preference but a relatively stable interpretive disposition that determines which obstacles an individual emphasizes, as it is precisely this dispositional, knowledge-and-value-based character—and not the credibility score itself—that matters. Talke and Heidenreich (2014) expand on this framework by distinguishing between passive resistance—characterized by inaction and a lack of organizational capacity—and active resistance, which is intentional in nature and rooted in the expert's value beliefs. Passive resisters are primarily prevented from adopting innovation by functional barriers, while active resisters face barriers of a psychological nature. These two profiles therefore require fundamentally different intervention approaches. For the first type, capacity building is relevant; for the second type, the systematic building of epistemic trust is essential. In the context of introducing artificial intelligence into the media environment, functional barriers include, in particular, a lack of professional skills, the absence of clear organizational guidelines, and resistance from clients. Psychological barriers primarily include distrust of AI-generated outputs and ethical concerns regarding the potential impact of AI on the professional integrity of experts. Network analysis of psychological constructs (Borsboom and Cramer, 2013; Epskamp and Fried, 2018) offers an alternative to factor analysis when examining the structure of agreement regarding barriers. While factor models assume a common latent cause, network models map direct conditional associations between observed variables. The Ising model (Haslbeck and Waldorp, 2020; Van Borkulo et al., 2014) is designed for binary node data, making it particularly suitable for dichotomous assessments of barriers. If the barriers form a dense network of conditional relationships with a high number of strong partial correlations, they can be interpreted as a resistance syndrome; in such cases, a targeted solution focused on key nodes can trigger a cascade effect and also contribute to reducing peripheral barriers. Conversely, when the network is sparse, individual barriers are conditionally independent, and each requires a specific intervention approach. Previous work applying the EBICglasso approach (Epskamp et al., 2012) to binary data may be subject to a tendency towards oversimplification, manifested in the generation of empty networks. The Ising model represents a methodologically more appropriate alternative for binary data and therefore serves as the primary network analytical method in this study. The theory of institutional logic (Lounsbury et al., 2021; Thornton et al., 2012) is based on the assumption that professional behaviour reflects a set of competing logics: journalistic logic, centred on truth and verification; market logic, where efficiency and profit are key; and professional logic, defined as autonomy and expertise. These coexist within organisations rather than being organisationally separated from one another. It follows from the above that archetypes of resistance to artificial intelligence will not cluster along lines of company type, sector or degree of AI usage, but will ultimately reflect the individual epistemological orientations of individual professionals. Insofar as affiliation with a given archetype cannot be predicted on the basis of structural variables, this provides empirical confirmation of the interpretation in the spirit of institutional logic theory and, at the same time, suggests that organisations must operate at the individual level when deploying diagnostic tools, rather than applying blanket strategies applicable to the entire sector.

3 Methodology

Data were collected in March 2026 via the Prolific platform, which has been validated for professional research (Palan and Schitter, 2018; Peer et al., 2022). Prolific was selected over alternative professional sampling strategies (e.g., panel providers or LinkedIn-based recruitment) because it combines pre-screened, verifiable employment information with documented data-quality standards and transparent participant compensation, while allowing targeted recruitment of currently employed media, journalism and marketing professionals in the UK and the US. Participants were required to be currently employed in the media, journalism or marketing communications in the UK or the US. The final sample comprised N = 151 respondents. In terms of gender, 53.0% were women (n = 80), 42.4% were men (n = 64), 2.6% were non-binary (n = 4), and 2.0% did not wish to specify (n = 3). The largest age group was 26–35 years (38.4%, n = 58). The type of organisation was freelancer 35.1% (n = 53), agency 21.2% (n = 32), other/public sector 16.6% (n = 25), media 15.9% (n = 24), corporate sector 11.3% (n = 17). 60.9% of respondents used generative AI several times a week or daily. Participants selected up to seven barriers from a list of seven options, with responses being binary (1 = selected, 0 = not selected): (1) Lack of trust in the reliability and accuracy of the technology; (2) Ethical concerns and moral dilemmas; (3) Legal risks (copyright, GDPR); (4) Insufficient quality, consistency or superficiality of outputs; (5) Resistance from clients or the public; (6) Lack of digital skills and insufficient staff training; (7) Lack of internal guidelines and unclear organisational rules. The average number of selected barriers was M = 3.31 (SD = 1.04). Respondents rated three statements on a scale from 1 (strongly disagree) to 5 (strongly agree): (1) AI in the workplace requires stricter internal guidelines; (2) audiences and clients must always be informed about AI-generated content; (3) human oversight and fact-checking of AI content are essential. The average overall management score was M = 4.23 (SD = 0.69). The analysis was conducted in four stages. In the first stage, a network analysis was performed using the Ising model via IsingFit (Van Borkulo et al., 2014) and the comparative EBICglasso method (Epskamp et al., 2012), with three gamma values (0.25, 0.50, 0.75) tested and stability verified by bootstrapping (200 repetitions; Epskamp et al., 2018). In the second phase, latent class analysis was performed using poLCA (Linzer and Lewis, 2011) with K = 2–5 classes, with model selection based on BIC and entropy, and stability confirmed by 200 resamplings and 500 repeated LCA selections. In the third phase, the archetypes were compared using Kruskal–Wallis tests with rank-biserial r and Cohen's h effect sizes. In the fourth phase, binary logistic regression tested organisational and demographic predictors of archetype membership, and the results were supplemented with IRT composites of barriers (Huang et al., 2021). All analyses were conducted in R 4.5 using the IsingFit, bootnet, poLCA, ggplot2 and patchwork packages.

4 Results

The most frequently reported barriers were legal risks (84.8%, n = 128) and quality concerns (75.5%, n = 114) (Table 1), which were present in almost the entire sample. Mistrust (57.6%, n = 87) and ethical concerns (57.6%, n = 87) had a moderate prevalence. Less common were client resistance (27.2%), lack of skills (16.6%) and lack of guidelines (11.9%). As legal and qualitative barriers were reported by almost everyone, they have limited ability to distinguish between archetypes. The main distinguishing feature is therefore mistrust, which occurred unevenly.

Table 1. Prevalence of barriers to AI implementation and effect size across archetypes (N = 151).

Cohen's h calculated from archetype-specific conditional probabilities. Mean number of obstacles reported per respondent: M = 3.31, SD = 1.04.

The Ising model identified two significant correlations out of 21 possible ones (Table 2), specifically between mistrust and legal aspects (weight 2.832) and between mistrust and quality (weight 3.431). The remaining five barriers showed no significant correlations following the analysis. EBICglasso confirmed the same results. When tested with gamma values of 0.25, 0.50 and 0.75, 4, 2 and 2 edges emerged, with both key connections persisting in all cases. The network stability (CS coefficient = 0.675) was acceptable. The structure described is theoretically comprehensible. Distrust functions as the sole central node of the network (strength centrality = 6.263, betweenness = 1) and links perceptions of legal and qualitative risks. The network is neither a dense cluster—five of the seven barriers remain mutually independent—nor entirely unrelated. It is rather a specific cluster, as experts who distrust AI have both legal and qualitative concerns, suggesting a common source in the form of epistemic distrust in the reliability of AI. The remaining four barriers (client resistance, lack of skills, missing guidelines and ethical concerns) operate independently.

Table 2. Ising network, edge weights and node centrality statistics.

Ising model (IsingFit, family = binomial). Total number of edges: 2/21 (9.5%). EBICglasso confirms the same structure with 2 edges. CS coefficient (case-drop bootstrap, 200 resamples) = 0.675.

Bold values indicate the central node of the network (highest strength centrality).

BIC clearly favoured the two-class model (BIC = 1,063.6) (Table 3) over the three-class models (BIC = 1,091.6; ΔBIC = + 28.0), four (BIC = 1,122.2; ΔBIC = + 30.6) and five classes (BIC = 1,153.2; ΔBIC = + 31.1). Entropy = 1.000 indicates perfect class separation, with every respondent reliably assigned to a single class. The bootstrap BIC confirmed this choice in 99.5% of cases (199 out of 200 repetitions favoured K = 2).

Table 3. LCA model fit statistics (K = 2–5).

* = selected model. nrep = 50, maxiter = 3,000, set.seed = 42. Bootstrapped BIC: K = 2 selected in 99.5% of 200 repeated samples. ΔBIC (K = 2 vs. K = 3) = + 28.0.

Bold values indicate the selected model (K = 2, lowest BIC).

T1 are practical opponents (n = 64; Table 4); these respondents express no distrust of the technology (0.000), but have relatively high concerns regarding ethics (0.609) and legal risks (0.641). They are less concerned about the quality of outputs (0.422). Client resistance and lack of skills are moderate, whilst a lack of guidelines is almost non-existent. T2 are distrustful sceptics (n = 87); these respondents have the highest level of distrust in the technology (1.000) and, at the same time, the highest concerns regarding legal risks and the quality of outputs (both 1.000). Ethical concerns are moderate (0.552). Client resistance, lack of skills and lack of guidance are low. T1 is primarily concerned about practical obstacles (ethics, law, quality), but does not resist the technology itself. T2 combines complete distrust of the technology with maximum legal and quality concerns, which corresponds to the epistemic cluster from the previous network analysis. An important test was whether the division into two classes arose solely due to the ‘distrust’ variable. An LCA analysis without this variable confirmed the same two-class solution (BIC = 972.2; ΔBIC compared to K = 3 is +15.8; Entropy = 1.000; MinClass = 24.5%). The remaining six barriers were also able to perfectly separate the two classes on their own, confirming that these are genuine differences in the configuration of the barriers, not merely the presence or absence of distrust. Kruskal–Wallis tests (Table 5) showed that the overall difference in governance between the classes (T1 where M = 4.08, T2 where M = 4.33; H = 4.033, p = 0.045, r = 0.163) conceals an important pattern. The only item that differed significantly between the archetypes was transparency; sceptics in T2 demanded significantly higher transparency of AI content (M = 4.44 vs. 3.97; H = 9.401, p = 0.002, r = 0.250, small to medium effect). Regulation (H = 0.253, p = 0.615) and human oversight (H = 1.098, p = 0.295) did not differ between groups. The pattern corresponds to the T2 profile: experts who fundamentally distrust AI outputs, specifically requiring that the audience be aware of the origin of AI content, and who perceive transparency as a systemic safeguard. Practical opponents (T1), whose barriers are more capacity-related than epistemic, showed slightly higher, though statistically insignificant, support for human oversight (M = 4.36 vs. 4.55), which corresponds to their preference for practical verification protocols.

Table 4. Conditional LCA probabilities P(barrier=1|class) with 95% bootstrap confidence intervals.

Bold is defining, increased conditional probability according to archetype. CI = 95% bootstrap confidence intervals [500 LCA resamples with class assignment via P(Distrust)]. Entropy = 1.000; all classifications are deterministic.

Table 5. Comparison of long-term outcomes by LCA archetype (kruskal–wallis tests).

Bold values indicate statistically significant differences between archetypes (p < 0.05).

r = effect size according to rank-biserial [r = sqrt(H)/sqrt(N); N = 151]. Interpretation:.1 = small,.3 = medium,.5 = large.

Transparency is a sub-item of the Governance component.

*** p < .001.

** p < .01.

* p < .05.

The greatest difference was in the number of barriers (H = 88.406, p < 0.001, r = 0.765), where T2 sceptics reported almost twice as many barriers (M = 3.95) as T1 (M = 2.44), which is a direct consequence of the fact that distrust, legal risks and quality occur simultaneously in T2. The benefits, risks and frequency of AI use did not differ between the groups. Belonging to an archetype therefore does not depend on how much someone uses AI or how they rate it overall. The results confirmed differences between the groups (Table 6), where functional barriers (legal, quality, client, skills, guidelines) were higher in T2 (M = 0.480, SD = 0.099) than in T1 (M = 0.366, SD = 0.153; H = 23.264, p < 0.001, r = 0.392, which is a moderate effect), reflecting strong legal and quality concerns in T2. Psychological barriers (distrust, ethics) were significantly higher in T2 (M = 0.776, SD = 0.250) than in T1 (M = 0.305, SD = 0.246; H = 70.324, p < 0.001, r = 0.682, large effect). The ratio of functional to psychological barriers confirms the typology: T1 has a ratio of 1.16 (functional barriers predominate, passive resistance) and T2 has a ratio of 0.61 (psychological barriers predominate, active resistance).

Table 6. IRT validation: composite functional vs. psychological barriers by archetype.

Functional barriers is average (legal, quality, client, skills, guidelines). Psychological barriers = average (distrust, ethics). F/P ratio > 1 is functional dominance (passive resistance to IRT); <1 = psychological dominance (active resistance to IRT). r (functional) = 0.392; r (psychological) = 0.682.

Bold values indicate the dominant barrier category within each archetype.

Binary logistic regression (dependent variable where T2 = 1, T1 = 0; predictors such as frequency of AI use, organisation type, perceived benefits, perceived risks for N = 151) revealed no significant predictors (all p > 0.24; Nagelkerke's pseudo-R² = 0.040). Organisation type was not significant in any contrast (all p > 0.43). Frequency of AI use also showed no association (B = −0.026, SE = 0.180, p = 0.887). Neither perceived benefits (B = −0.254, p = 0.244) nor perceived risks (B = −0.004, p = 0.987) were predictive. Pseudo-R² = 0.040 means that the factors examined explain only 4% of the variance in archetype affiliation. This supports the interpretation of institutional logic: if resistance were primarily structural, the type of organisation or frequency of AI use would predict it. Their absence confirms that T1 and T2 reflect epistemological dispositions that occur across organisations regardless of sector, size or degree of exposure to technologies.

5 Discussion

Ising's network revealed a structure that refutes both the ’syndrome’ model, in which barriers form a tight cluster, and the model of complete independence, in which all barriers are independent of one another. Instead, a ‘hub-and-spoke’ structure emerges, where epistemic distrust links legal and quality concerns, whilst the five other barriers—ethics, client resistance, skills, guidelines and their interconnections—operate independently. This structure has direct implications for intervention. Focusing on epistemic mistrust may lead to a reduction in legal and quality concerns, but only among T2 professionals, whose profile is encompassed by this cluster. For practical T1 opponents, building trust would not help; they need the development of organisational infrastructure. The central role of distrust builds on theoretical insights into epistemic trust in AI systems (Gillath et al., 2021; Glikson and Woolley, 2020). In a media context, legal and quality concerns are not independent assessments; they are partly rooted in a deeper dispositional distrust of the reliability of AI. This suggests that the concerns reported in T2 are not primarily the result of specific negative experiences or knowledge of legal details, but rather a more fundamental epistemic attitude towards AI outputs. The T1 and T2 typology provides the first empirical validation of the IRT functional-psychological distinction in the context of AI acceptance, using a person-centred LCA method instead of composite scales. Previous applications of IRT (Kleijnen et al., 2009; Laukkanen et al., 2008) used variable-centred approaches; the LCA approach verifies whether the categories mentioned actually correspond to latent groups with distinct configurations of barriers. It should be noted that the elevated psychological-barrier composite in T2 is driven almost entirely by distrust rather than by ethical concerns, which were comparable across archetypes (Cohen's h = 0.117); the functional–psychological contrast therefore rests primarily on the distrust dimension. The finding regarding transparency (T2 > T1; H = 9.401, p = 0.002, r = 0.250) provides important specificity. T2 sceptics do not call for more regulation in general; specifically, they demand the systematic disclosure of the origin of AI-generated content. This is consistent with Diakopoulos (2020) and Newman et al. (2024); experts who distrust AI want the audience to have sufficient information for their own evaluation, rather than automatically trusting AI content. Practical opponents of T1, conversely, favour human oversight as a practical tool for quality control, not as a statement of AI's untrustworthiness. The absence of significant structural predictors (pseudo-R² = 0.040) is a notable finding; rather than confirming a purely epistemological interpretation, it provides no evidence that T2 sceptics cluster in journalistic organisations, small firms without guidelines, or among experienced experts. This null result is, however, also compatible with the modest sample size and the limited statistical power available for detecting structural effects, and should therefore be interpreted with caution. Both archetypes occur in relatively even proportions across all structural categories. This is consistent with the theory of institutional logic: competing logics coexist within organisations, creating different epistemological orientations that transcend structural boundaries. We note, however, that the present data were collected at the individual level, and the term ‘institutional’ is used here to denote the logics that individuals internalise, not a measured organisational-level property. Inferences about organisational resistance should therefore be treated as theoretical rather than empirically established. The practical implication is straightforward: organisations cannot simply identify which departments contain more sceptics. The typology of resistance to AI is distributed across organisations and requires diagnostic tools at the individual level, such as a short five-item screening tool assessing distrust, client concerns and self-efficacy of skills. The following are the limitations of the study. Firstly, the sample (N = 151) was recruited exclusively in the United Kingdom and the United States via Prolific. This limits the generalisability of the findings to other geographical and cultural contexts, and future research should extend the sample to continental European, Latin American and Asian settings. Secondly, although Prolific is a validated platform for professional research, participants who self-select into a study explicitly concerned with AI may hold attitudes that differ systematically from the broader population of media and marketing professionals; the choice of Prolific over alternative professional sampling frames is a pragmatic one, justified by access to a verified, employed professional pool, but it is not free of selection bias. Thirdly, the relatively modest sample size, in relation to the complexity of the combined Ising–LCA analytical strategy, constrains statistical power, particularly for the logistic regression predicting archetype membership; null structural effects should be read in this light. Fourthly, the K = 2 solution is significantly influenced by the distrust variable. Although a sensitive LCA without the distrust subitem confirms a two-class structure (BIC = 972.2, entropy = 1.000), the main classification signal originates from this single binary variable. Future research should employ graded measures of barrier intensity to enable more nuanced results. Fifthly, binary measurements of barriers, where ’selected’ and ‘not selected’ lose information about intensity; resistance may vary within each archetype in ways that binary LCA cannot capture. Sixthly, the cross-sectional design does not allow us to determine whether experts change their archetype affiliation over time or respond to organisational interventions. Finally, a two-archetype typology, while parsimonious and empirically well-separated in the present data, may not exhaust the diversity of epistemological positions towards AI; richer, graded measures may reveal additional or intermediate profiles, and the two-class structure should be regarded as a first approximation rather than an exhaustive taxonomy.

6 Conclusion

This study demonstrates that resistance to AI in media and marketing organisations shifts from a single-construct barrier problem to a typology of two archetypes based on the theory of resistance to innovation and institutional logic. The Ising network reveals that barriers to AI adoption form a ‘hub-and-spoke’ structure rather than a syndrome structure, with mistrust being the central epistemic node linking legal and qualitative concerns. LCA with perfect entropy (E = 1.000) and 99.5% bootstrap stability identifies two qualitatively distinct archetypes: practical opponents (42.4%) and distrustful sceptics (57.6%), whose inter-organisational distribution confirms their epistemological, rather than structural, nature. An analysis of governance requirements shows that the archetypes are not distinguished by the overall level of requirements, but by their type. T2 sceptics specifically demand transparency regarding AI content, which corresponds to their focus on systematic disclosure. For T1 practical opponents, interventions aimed at capacity building, training, the creation of guidelines and communication with clients are appropriate. For distrustful T2 sceptics, it is necessary to build epistemic trust through transparent governance, verifiable quality protocols and clear disclosure of the origin of AI. Treating both groups equally is not only ineffective; it represents a fundamental misunderstanding of the nature of their resistance.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Narcisa Medranda - Morales, Universidad Politecnica Salesiana - Sede Quito, Ecuador

Reviewed by: Abel Suing, Universidad Técnica Particular de Loja, Ecuador

Büşra Ökten, Istanbul Medeniyet University, Türkiye

Ionut Virgil Serban, University of Craiova, Romania