Digital Skills and Employer Transparency: Two Key Drivers Reinforcing Positive AI Attitudes and Perception Among Europeans

Abstract

Using 2024 Eurobarometer survey data from 26,415 workers in 27 EU countries, this study examines how digital skills and employer transparency shape workers’ attitudes toward and perception of artificial intelligence (AI). Drawing on information systems and behavioral theories, regression analyses reveal that digital skills strongly predict augmentation-dominant attitude. Workers with higher digital skills view AI as complementary rather than threatening, with an augmentation attitude mediating 56% of the skills–perception relationship. Adjacently, employer transparency attenuates the translation of replacement attitude into a negative perception of AI in the workplace. Organizations and policymakers should prioritize digital upskilling and ensure workplace AI transparency requirements to foster a positive attitude and perception, recognizing that skills development and organizational communication are equally vital for the successful integration of AI in the workplace.

1. Introduction

A European survey, conducted in 2024, indicates that Europe remains in an early stage of integrating AI into the workplace. Compared with 2023, the use of AI in European Union (EU) enterprises has only increased by 5.45%, roughly from 8% to 13.5% [1]. This is justified because when compared to other leading nations, including the US and China [2], the EU is committed to cautious and responsible AI integration, guided by a risk-based approach [3]. For instance, although the US has taken several steps in AI regulation, the focus is now shifting to accelerating the use of AI and overall promotion. The recent executive order in the National AI Initiatives Act of 2020 is aimed at removing barriers, deregulating, and accelerating the use. As a consequence, safety acts like Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence are being reevaluated [4]. Moreover, the second latest executive order is aimed at accelerating data center infrastructure [5]. This demonstrates the reverse trend of the US’s earlier focus on stricter control. On the other hand, China’s regulations, “Interim Measures for the Management of Generative Artificial Intelligence Services,” entered into force in 2023. China’s approach is more in line with the EU’s approach, as compared to the US. These interim measures draw from pre-existing legislation, notably the Cybersecurity Law, Data Security Law, and Personal Information Protection Law (PIPL). Overall, the measures focus on AI content labeling, reinforcing adherence to existing AI laws, mandating legal sourcing of training data, and aligning AI services with society’s current value system [6]. However, the EU’s commitment is evident in numerous initiatives instigated by the EU. The EU is among the first to pass an exacting AI legislation, the 2024 EU AI Act. The EU is also reinforcing General Data Protection Regulation (GDPR) and platform–work rules, while the Digital Europe Programme (DEP) 2021–2027 focuses on strengthening Europe’s digital capacities. The European Digital Innovation Hubs, covered in DEP 2021–2023, continue to support SMEs in their efforts to harvest the benefits of AI. At the same time, DEP 2025–2027 invests substantially in digital skills, AI, and cybersecurity (i.e., EUR 1.3 billion) [7]. This is in parallel with European public opinion believing AI integration requires careful and strict management (84%), robust worker-privacy rules and regulation (over 80%), and employee involvement in AI deployment (77%) [8].

In attempts to regulate AI, the Organization of Economic Co-operation and Development’s (OECD) AI principles were the first intergovernmental standard on AI that offered recommendations as well as a definition of AI that has become the de facto standard for regulatory jurisdictions. Today, the EU, the Council of Europe, the US, the United Nations, and other jurisdictions use the OECD’s definition of AI and its lifecycle in their legislative and regulatory frameworks [9]. The EU AI Act, adapted to the revised OECD’s definition [10], enforces a risk-based compliance framework that classifies and places restrictions based on the type and level of risk posed by AI. Practices such as social scoring, real-time remote biometric identification, manipulating people’s decisions or exploiting their vulnerabilities, and inferring emotions in workplaces are categorized as “unacceptable risk” and are prohibited. The majority of the obligations are imposed on AI technologies classified as “high-risk AI systems” (i.e., posing high risk to people’s health, safety, or fundamental rights) and their providers. The Act mandates their strict pre-market conformity assessments, rigorous data governance, human oversight, and continuous risk management systems. Furthermore, general-purpose AI (GPAI) models face tiered restrictions, with all providers required to provide technical documentation and instructions for use to the commission as well as downstream providers, while retaining their copyright. In addition, GPAI with systemic risk must comply with additional restrictions. The goal of the Act is to promote human-centered, trustworthy AI, while ensuring Europeans’ protection [11].

Although the EU has taken numerous steps in appropriately ensuring and regulating informed integration of AI into workplaces, they have yet to directly address the core challenge of job displacement [12] or the fear of replacement posed by AI [8]. This fear shapes workers’ attitudes and perceptions surrounding AI. In AI literature, perceptions and attitudes are seldom studied as distinct constructs; however, research into AI perceptions typically investigates attitudes [13]. Attitudes are individuals’ evaluative opinions with some degree of favor or disfavor [14]; perceptions are sense-making judgments of individuals’ environment and related stimuli, indirectly influenced by attitudes [15,16]. Thus, conflating these constructs obscures how attitudes shape subsequent perceptions. Implying that distinguishing evaluative attitudes from impact perception is necessary to clarify how attitudes form individuals’ perception of AI.

Second, while previous studies identify digital skills as central to technology acceptance in the workplace [17,18,19], the specific mechanisms through which skills influence AI attitudes and perception remain unclear. Specifically, if attitudes dictate the effect of technical competence on AI receptivity requires systematic investigation. This is particularly invaluable as it allows us to further narrow down, isolate, and elucidate the role of the primal lens, i.e., individuals’ attitudes, in the larger scheme of AI’s integration into the workplace. Such novel insights can not only be compatibly utilized in organizational interventions but will also be instrumental in guiding future research to explore attitudes in depth and latitude. Similarly, although transparency is widely advocated in responsible AI frameworks [3,20], empirical evidence for its effects on workers’ attitudes and perceptions is limited.

This study addresses these gaps, first, by developing a dual-attitude framework that distinguishes between AI augmentation attitude (viewing AI as complementing human work) and AI replacement attitude (viewing AI as replacing the workforce). Second, it explores the impact of digital skills on AI reception while disentangling attitudes from perception. Third, it examines the role of employer transparency, i.e., whether employees are informed of AI use in their workplace, in the translation of AI replacement attitude into perception. Moreover, we study to what extent digital skills benefit from employer digital support, i.e., training and tools provided by employers to employees to work effectively with AI. Our theoretical approach integrates several key insights. Primarily, from Technology Acceptance Models [21,22], we understand the mechanisms through which digital skills influence attitudes. Drawing on behavioral theories [23,24,25,26], we note attitudes’ influence on individuals’ perception of AI use and shed light on the role of employer transparency. The research models are depicted in Figure 1 and Figure 2.

We analyze data from the 2024 Eurobarometer survey, commissioned by the European Commission, which covers 26,415 workers across 27 EU member states. We conduct regression analyses to estimate relationships among employer digital support, digital skills, employer transparency, attitudes, and perception (Figure 1 and Figure 2). The paper proceeds with a comprehensive literature review establishing our theoretical framework, followed by methodology and results sections synthesizing our key insights and their implications for both theory and practice.

1.1. Digital Skills

The role of digital skills is foreseeably one of the most crucial in regard to the implementation and real-world use of AI in workplaces [12,18,27,28,29,30], and thus, equally critical is their development. The DEP work plan for 2025–2027, focusing on advancing digital skills, seeks to enhance cooperation between EU member states and stakeholders in digital skills and jobs, with a mission to close the skills gap. The program seeks to launch four digital skills academies aiming to bolster the EU’s technological sovereignty and innovation capacity with specialized education and training in key digital areas, including AI [7]. In fact, there is an expected reorganization of technology-driven skills in the AI-powered world, allowing for exploration of relevant and necessary skills [27]. Thus, a consensus in operationalizing or establishing a standard definition of digital skills per se is crucial. In contrast to a globally accepted standard definition, there is a patchwork of frameworks across the globe [31,32,33,34,35,36].

Nevertheless, we define the digital skills used in our study in line with the DigComp 2.2 framework of digital competence [36], and Deursen and Dijk’s types of digital skills [37,38]. Our scale assesses individuals’ skills in using digital technology including AI across four domains: daily life, job performance, future job adaptability, and digital and online learning [8], and thereby encompasses DigComp’s broad definition of digital competence, defined as “confident, critical and responsible use of, and engagement with, digital technologies for learning, work, and participation in society”. Furthermore, digital competence areas composing Dimension 1, which engenders 21 competences composing Dimension 2, further help us ground our scale in the DigComp 2.2 framework [36]. In terms of DigComp’s competence areas, all four of our domains inherently reflect “problem-solving” and “information and data literacy,” with “communication and collaboration” particularly prominent in daily life and work domains, and “digital content creation” most salient in the work domain. The aspect of “safety,” while not directly measured, is implicitly relevant to daily life and work domains.

In parallel, the four domains coherently map onto the types of digital skills distinguished by Deursen and Dijk [37,38]. Specifically, “operational skills” (basic technical use of hardware and software for everyday tasks) are reflected in the daily life item, “formal skills” (mastery of task-specific applications and software) align with performing your job item, “strategic skills” (planning and adapting digital tools for complex, novel tasks) correspond to future job adaptability, and “information skills” (navigating, filtering and evaluating education content online) underpin the digital and online learning item.

1.2. Employer Digital Support

In a late US-based survey, nearly half of the employees agreed that formal training from their organizations would assist generative AI integration and their likelihood of using AI tools daily. Whereas, more than a fifth report that they received minimal to no support [2]. This is in contrast to other leading nations in AI adoption [39,40], where employees received more support [2]. However, more direct statistics on training or tools provided by employers to employees in our context, i.e., the EU, are reflected in the Eurobarometer survey 101.4 (2025). A total of 68% of Europeans working for an employer or manager reported that they were provided with the necessary tools and training to effectively work with AI [8].

Digitalization is not only a transformation requiring change management, but also a demand driving a broad range of workplace transformation toward the future of work. It represents a core work demand that first requires resources to be addressed, and then skills to be met. Drawing upon Job Demands–Resources (JD–R) theory, training delivered in job crafting interventions, allowing individuals to work on job crafting goals such as increasing specific job resources, can be effective. Skill variety is recognized as an important job resource, especially in a high-demand work environment [41]. Thus, digital skills, as a relatively new skill set in skill variety, can benefit from training. Furthermore, self-efficacy theory proposes that mastering an experience significantly shapes one’s self-efficacy, i.e., an individual’s belief in their capacity to execute behaviors necessary to produce specific performance attainments [23]. Building on that foundation, mastering how to work effectively with AI via training and tools can partly assist, if not shape, individuals’ (digital) skills in efficiently working with AI.

Training is a cornerstone of skill development and is proven to amplify AI integration [42] and foster innovation and adaptability by advancing employees’ skills [43]. AI training and development initiatives equip employees with skills to adopt and effectively use AI [44]. A review article by Morandini et al. recommended implementing comprehensive training and development programs, particularly in developing specific skills required for AI adoption and effective use [45]. Taken together, training and tools provided by the employer, hereafter referred to as employer digital support, should positively influence employees’ digital skills. Thus, we hypothesize the following:

Hypothesis 1 (H1).

Employer digital support is positively associated with employees’ digital skills.

1.3. Attitudes Toward AI

Attitudes are mental tendencies to evaluate an entity along an evaluative continuum with some degree of favor or disfavor; attitudes sum up liking or opinions toward an entity [14]. In regard to AI acceptance specifically, two broad and distinct attitudes have been maintained. First, the AI augmentation view, which regards AI as complementing or augmenting the workforce and work processes. In contrast, the AI replacement view reflects the trepidation among people that AI will replace the workforce.

One of the first systematic literature reviews of human–AI augmentation in the workplace revealed that AI can yield both benefits and drawbacks, and that the human–AI relationship is complex and multifaceted [46]. For instance, AI can relieve humans of basic and repetitive tasks, allowing workers more resources to focus on more complex, multiple responsibilities [47,48,49,50], ultimately improving decision-making, efficiency, and productivity [46]. On the other hand, in the presence of AI, employees across industries report substantial fear of being replaced by AI [51,52,53,54].

Most importantly, there is a lack of a clearer, direct assessment of people’s attitudes toward AI acceptance. While some studies measure people’s attitude toward AI [55,56,57], they are not operationalized in a way to measure whether people are accepting or rejecting AI in the workplace. It is also unclear whether their accepting belief leads their rejecting belief, or vice versa. Here, we investigate people’s attitudes toward AI in such a dichotomy. Moreover, by calculating their difference, we analyze if an individual’s augmentation attitude leads their replacement attitude and vice versa, and how differently these attitudes are predicted by the level of digital skills.

A growing literature indicates that individuals with stronger digital skills tend to hold more positive attitudes toward AI [58,59,60], whereas those with weaker skills exhibit more negative expectations [61]. Studies report that employees with higher occupational and digital self-efficacy experience lower AI anxiety and job insecurity [57,62]. In light of Management Information System (MIS) attitude research and the Technology Acceptance Model (TAM) by Davis [21], we note that attitude toward technology is a function of perceived ease of use. We further note that the effort to develop skills or the ease of adopting skills in the technology use is one of the six sub-constructs of perceived ease of use, specifically designed to “directly tap ease of learning”. Similarly, United Theory of Acceptance and Use of Technology (UTAUT) notes that attitude toward using technology stems from effort expectancy, which envelops “learning to operate the system” as well as ‘ease to become skillful at using the system’ [22]. Thus, individuals with higher digital skills should require less effort to be skillful in using or operating AI, which in turn dictates their attitude toward AI accordingly.

On these premises, we propose that individuals’ digital skills systematically predict their AI attitudes. Individuals with higher digital skills are more likely to endorse an augmentation (favoring) attitude over a replacement (disfavoring) attitude. Thus, we hypothesize the following:

Hypothesis 2 (H2).

Digital skills are positively associated with the augmentation–replacement attitude balance. Specifically, individuals with higher digital skills will report a more favorable attitude (i.e., a net augmentation-dominant attitude), whereas those with lower digital skills will report a more unfavorable attitude (i.e., a net replacement-dominant attitude).

1.4. AI Individual Perception

Individuals’ perception involves making sense of the environment and related stimuli; this process is influenced by past experiences, expectations, motives, and attitudes [16]. Attitudes encapsulate positive and negative feelings and beliefs [63]; affective emotions and cognitive beliefs are two integral components of attitudes [25]. By virtue of automatic activation, the attitude indirectly biases perceptions of the qualities of the object [15]. Furthermore, Lord et al.’s (1979) study on confirmation bias [26] highlighted that people with strong opinions process empirical information in a biased manner, and the biased perception and judgment formation depend on the state of beliefs [64].

However inverted, perception-attitude association toward technology has been studied and noted to be profound [65]. Perception of AI’s capabilities was positively related to affective and cognitive attitudes toward AI among employees [66]. Another study attempting to explain attitudes toward AI and how beliefs influence perceptions noted positive perceptions and negative attitudes, but no association between attitudes and perception was studied [67].

Furthermore, individuals’ perceptions should also be guided by their level of digital skills. Digitally skilled individuals should positively perceive AI in the workplace compared to individuals with lower levels of digital skills. Studies note that individuals’ perceptions can provide insight into their level of AI literacy [13]. Moreover, familiarity with AI is seen to influence perceptions [68], and evidence suggests a positive association between the perceived impact of AI use and perception of digital literacy skills [69].

On these premises, we expect that digital skills influence AI perception, and individuals’ attitude toward AI guides how positively or negatively they perceive the use of AI in the workplace. Consequently, attitude balance should also mediate the relationship between digital skills and individuals’ AI perception. Thus, we hypothesize the following:

Hypothesis 3 (H3).

Digital skills are positively associated with AI perception. Specifically, digitally skilled individuals should report a positive perception of AI’s use in the workplace.

Hypothesis 4 (H4).

Augmentation–replacement attitude balance is positively associated with AI perception. Specifically, a favoring, augmentation-dominant attitude should predict a positive, high AI perception.

Hypothesis 5 (H5).

Augmentation–replacement attitude balance will mediate the positive effect of digital skills on AI perception. Specifically, higher digital skills will exhibit a stronger net favor of augmentation (rather than replacement), which in turn leads to a more positive perception of AI use.

1.5. Employer Transparency

AI concerned information and decisions remain a private, top-down conversation among executives. There is a general ambivalence toward AI, where perception of the need for top-down formal approaches to AI-related change can obscure important new information, and the vast “unknown” makes it prudent to await directives from above [70]. Uncertainty Reduction theory, one of the first theories on interpersonal communication, defines uncertainty as a function of the number and likelihood of alternatives that may occur. This uncertainty requires coping to reduce anxiety and unpredictability caused [24]. In this light, a new AI tool or lack of information on AI deployment status from the employer can indeed exacerbate (but also curb, in cases of informed employees) the negative affective state of employees, an integral part of attitude [25]. Van De Poel’s ethical framework for evaluating the acceptability of experimental technologies [71,72] highlights the requirement for human subjects to be informed about new technologies along with their risks and benefits, and to be involved in the approval regarding their deployment.

Moreover, leaders must use effective communication strategies to prepare workers for AI, explaining AI’s implementation and effects on their jobs [73]. Glikson and Woolley [20] suggested that workers’ cognitive trust, a component of attitude [25], in workplace AI is influenced by its transparency. The employees are keen to learn more about AI, its possibilities, and its impact on work [74]. Workers report that the communication quality of AI guidelines is important for effective adoption of AI [75]. Furthermore, drawing on the self-efficacy theory, Bandura [23] notes that the outcome expectancy’s (i.e., particular behavior leading to certain outcomes) effect on behavior change is of small magnitude and significance, but it does reduce phobic behavior. Further, “informing/telling people what to expect” does raise outcome expectations. Outcome expectancy is prominent in shaping cognitive views of behavior [23], an important component of attitude [25].

Thus, replacement attitude enveloping negative affect and cognitive views of employees should vary depending on the status, whether employees are informed of AI use or not, i.e., employer transparency. Furthermore, due to the curbing power of transparency over the negative attitude, employer transparency should attenuate the translation of replacement attitude into a negative perception of AI. Based on this, we hypothesize the following:

Hypothesis 6 (H6).

The negative relationship between AI replacement attitude and AI perception is positively moderated by employer transparency. Specifically, among uninformed employees, the negative effect of AI replacement attitude on AI perception will be stronger than among informed employees.

2. Materials and Methods

2.1. Data and Participants

Data for this study come from the 2024 Standard Eurobarometer 104.1, commissioned by the European Commission, a cross-national survey of the populations of the 27 EU member states. The survey employed a multi-stage, random probability sampling design stratified by region and density (metropolitan, urban, and rural). Following this, households within these Primary Sampling Units (PSUs) were chosen via systematic random route methods. From each household, one respondent aged 15 and over was selected using a Kish grid to achieve nationally representative samples of EU residents. Up to two recalls were made to minimize non-response, and no more than one interview per household was conducted [8]. Trained interviewers conducted face-to-face, computer-assisted personal interviews (CAPI) between April 2024 and May 2024 with a sample size of 1000 respondents per country, resulting in 26,415 total respondents. Interviews followed a standardized questionnaire translated and back-translated in each national language to ensure conceptual equivalence. Out of all the member states, the Netherlands had the highest response rate (81.6%), and Italy had the lowest (28.5). Thus, post-fieldwork, country-level weights were calculated to correct for unequal selection probabilities, non-response, and to align the sample with official population benchmarks on age, gender, region, and education. Microdata and documentation were obtained from the GESIS Eurobarometer Data Service under an academic-use license. All respondents provided informed consent in accordance with the ethical standards of the Eurobarometer consortium and the European Commission’s Directorate General for Communication. This rigorous, probability-based design and weighting protocol ensures that our analyses rest on a robust, high-quality dataset, fully suitable for examining pan-European attitudes toward AI in the workplace.

2.2. Measures

Out of 349 variables in the dataset, 5 variables, along with some demographics such as age, gender, and education level, were selected in our study. The majority of the variables are retained in their original form, albeit with varied levels of recoding. Whereas, one measure (AI attitudes) is derived from a variable and is (re)coded into two distinct variables. Furthermore, we note that although multi-item scales are preferred for psychometric robustness, we use three single-item variables. Single-item global measures have been shown to perform on par with multi-item scales when the construct is concrete and familiar to respondents [76,77]. For instance, AI perception directly captures the unambiguous, domain-general perception we study, has very high face validity, and correlated r = 0.57 (p < 0.001) with an 8-item domain-specific AI-for-management perception scale, demonstrating adequate concurrent validity for a broad, global AI perception scale. Similarly, the other two single-item scales used contain global, face-valid items, but more importantly, they capture the unidimensional nature we study, i.e., to what extent employees received training/tools, and whether they were informed about AI’s use in their workplace. They are distinct from digital skills and attitudes per se, where the constructs are inherently multidimensional. Lastly, although participants’ extent of familiarity with AI was not discretely measured, both AI perception and attitude explicitly assess individuals’ reflections on AI. Leveraging this, “do not know” responses (ranging from only 4% to 9%) were omitted from the analyses in our best attempt to control familiarity with AI influencing our study.

Digital skills are measured with a 4-item scale, on a 4-point Likert scale, i.e., 1 (totally disagree) to 4 (totally agree). It assesses individuals’ skills in using digital technologies, including AI, across four domains: daily life, job performance, future job adaptability, and digital and online learning. For example, “to what extent do you agree … regarding your skills in the use of the most recent digital technologies, including AI… in your daily life”. The internal consistency measured by Cronbach’s alpha (α) is 0.88.

Participants’ perceptions of employer digital support were measured with a single-item, 4-point Likert scale (1 = strongly disagree; 4 = strongly agree): “To what extent do you agree or disagree that your employer provides you with the necessary tools or training to work effectively with the most recent digital technologies, including AI?”

Individuals’ perception of AI was assessed by a single-item, face-valid measure: “How positively or negatively do you perceive the use of robots and AI in the workplace?” Responses were recorded on a 4-point Likert scale (1 = very negatively; 4 = very positively). Due to the possibility of a convergent validity test, we correlated AI perception with an 8-item scale: individuals’ perception of AI managing workplace activities, such as hiring and monitoring. Aligning with our expectations [78], the test resulted in a moderate correlation, r = 0.57 (p < 0.001).

For the AI augmentation–replacement attitude balance index, we combine a four-item AI augmentation attitude scale (e.g., “Robots and AI help people do their jobs”) with a two-item AI replacement attitude scale (“More jobs will disappear…”, “Robots and AI steal people’s jobs”). Both scales are recorded on a 4-point Likert scale (1 = strongly disagree; 4 = strongly agree). The Cronbach’s α for the augmentation attitude scale is 0.80, and the replacement attitude scale is 0.74. Each scale is mean-centered, and their difference (augmentation–replacement) yields a continuous balance score reflecting augmentation versus replacement views about AI in the workplace. Specifically, a score > 0 represents an augmentation-dominant attitude, a score < 0 represents a replacement-dominant attitude, and a score = 0 represents a perfectly balanced attitude.

To assess employer transparency, a total of 7 responses (e.g., “employer informed … without further details” or “yes, and you have been given access to personal data”) were recoded into dichotomous responses (0 = uninformed; 1 = informed). The face-valid item is as follows: “Did your employer inform you about the use of digital technologies, including Artificial Intelligence, to manage activities in your workplace?”

In demographics, gender and the recoded variable of age were utilized. To measure the level of education, 9 variables were recoded as one, where value 1 represents no education and value 9 represents doctoral or equivalent.

2.3. Data Analysis

The weights were applied in all analyses. Furthermore, after listwise deletion of missing values, the final analytical sample for H1 to H5 ranges from N = 3638 to N = 11,424, and H6 contains N = 10,529.

First, an exploratory factor analysis (EFA) was conducted to examine the underlying three-factor structure. This guided the subsequent confirmatory factor analysis (CFA) to validate the model. To test convergent validity and reliability, Cronbach’s α, composite reliability (CR), and average variance extracted (AVE) were calculated. Prior to regression analyses, necessary correlations and descriptive statistics were extracted. During preliminary analysis, a noteworthy insight was noted, to address which, attitude balance’s tertiles were extracted, and replacement attitude was recoded dichotomously for a cross-tabulation informing an additional inquiry of two mediations. All the hypotheses were tested using linear regression, and PROCESS macro’s Model 1 for moderation analysis and Model 4 for mediation analyses. A variety of assumption checks were performed. Residual normality and homoscedasticity were examined; when significant, reported HC3 (heteroscedasticity-consistent) standard error. Multicollinearity was calculated using VIF and tolerance for predictors and interaction terms. Lastly, previous research has noted that age, generational gap, level of education, and gender (though seldom) affect public receptivity to AI and automation [79,80,81,82]. Accordingly, to mitigate potential confounding effects, all regression analyses were controlled for age, gender, and education level. All the analyses were performed using SPSS V. 25 and PROCESS macro v4.2.

3. Results

3.1. Reliability and Validity

Digital skills, augmentation attitude, and replacement attitude were tested with EFA, CFA, and AVE and CR. The EFA using principal axis factoring with oblimin rotation resulted in an excellent KMO (0.833) and significant Bartlett’s test (χ² = 59,215.90, p < 0.001), indicating suitability for factor analysis. The pattern matrix revealed strong loadings for all items on their intended factors, with no substantial cross-loadings. The subsequent CFA confirmed the model. The model demonstrated good fit: χ²(33) = 6309.76, p < 0.001; CFI = 0.982; TLI = 0.975; RMSEA = 0.085 (90% CI [0.083, 0.087]); SRMR = 0.038. All standardized loadings were strong and statistically significant (range: 0.67–0.88; 0.82 for the constrained two-item factor). Factor correlations were theoretically consistent in the expected directions and ranged from low to moderate (−0.01 to 0.57), supporting discriminant validity. CR and AVE calculated were above recommended thresholds (AVE ≥ 0.50; CR ≥ 0.70) for all constructs (AVE = 0.55–0.73; CR = 0.80–0.92), supporting convergent validity and internal consistency.

Note: For full technical details, please refer to the Appendix A.

3.2. Descriptives and Preliminary Analysis

Required correlations and descriptives are reported in

Table 1. Descriptives and correlations.

Variable	M	SD	1	2	3	4	5	6
1. Digital Skills	2.81	0.86	—	—	—	—	—	—
2. AI Augmentation Attitude (AI Aug.)	2.75	0.73	0.50 **	—	—	—	—	—
3. AI Replacement Attitude (AI Rep.)	3.65	0.60	−0.03 *	−0.02	—	—	—	—
4. AI Perception (AI Per.)	2.70	0.80	0.44 **	0.65 **	−0.16 **	—	—	—
5. AI Perception Management	2.05	0.82	0.42 **	0.62 **	−0.03	0.57 **	—	—
6. Employer Transparency	—	—	0.23 **	0.20 **	−0.08 **	0.25 **	0.39 **	—
7. Employer Digital Support	2.85	0.87	0.52 **	0.40 **	−0.02	0.40 **	0.36 **	0.32 **

Note: ** p < 0.001; * p < 0.01; M: mean; SD: standard deviation.

. Regarding correlations underlying H1 to H5, digital skills were positively and strongly correlated with augmentation attitude (r = 0.50, p < 0.001) and employer support (r = 0.52, p < 0.001). Although of a negligible magnitude, the direction of the relationship was confirmed by correlation between digital skills and replacement attitude (r = −0.03, p < 0.01). AI perception was positively correlated with digital skills (r = 0.44, p < 0.001), augmentation attitude (r = 0.65, p < 0.001), and negatively related to replacement attitude (r = −0.16, p < 0.001). Regarding correlations concerning H6, employer transparency was positively correlated with AI perception (r = 25, p < 0.001) and, although marginally, negatively related with replacement attitude (r = −0.08, p < 0.001).

3.3. Assumption Checks

In linear regression for H1, residual diagnostic plots confirmed approximate normality and homoscedasticity. In mediation analysis, a-path regression exhibited normally distributed, homoscedastic residuals. In the outcome regression (Y on X + M), residual plots and White’s tests indicated heteroscedasticity (F = 33.62, p < 0.001), so we used HC3 standard errors and bootstrapped (5000-sample) confidence intervals (see Appendix A for assumption diagnostics). Multicollinearity diagnostics yielded VIFs of 1.002–1.288 and tolerances of 0.776–0.998, all well within acceptable limits. Concerning linear moderation, residual plots supported normality and homoscedasticity; VIFs for the predictor, moderator, and interaction term ranged from 2.38 to 4.03 (tolerances 0.25–0.42), all below conventional cut-offs.

3.4. Regression Model Testing

3.4.1. Research Model 1

This model involves a linear (OLS) regression and a mediation analysis covering hypotheses 1 through 5. All regression results were adjusted for age, gender, and education. For H1, we regressed digital skills to the extent of employer digital support. The regression coefficient was β = 0.42 (p < 0.001), supporting the hypothesis and implying that individuals’ skills benefit from training and tools provided by the employer. The mediation analysis assessing H2, H3, H4, and H5 confirmed all the hypotheses. Digital skills had a positive total effect on AI perception (β = 0.42, p < 0.001), indicating that higher skills relate to a more positive formed perception of AI among workers. Digital skills positively predicted augmentation–replacement attitude balance (β = 0.43, p < 0.001), supporting H2, and in turn, augmentation–replacement attitude balance positively predicted AI perception (β = 0.47, p < 0.001), supporting H5. The indirect effect of digital skills on perception through attitude balance was significant; effect = 0.20, BootSE = 0.0, 95% BootCI [0.186, 0.224]. Because the direct effect remained significant (c’ = 0.22, p < 0.001), mediation is partial rather than full and supports H5. Attitude balance accounts for 48% of the total effect of digital skills on perception (indirect effect/total effect = 0.2045/0.4243). This implies that digitals skill help build a favorable attitude among individuals, which dictates a positive perception of AI’s use in the workplace.

Table 2. Research Model 1.

Effect	Path	Predictor	Outcome	β	SE	t (df)	p	95% CI
H1	—	Employer Digital Support	Digital Skills	0.42	0.00	60.58 (11410)	<0.001	[0.40, 0.43]
H2–H5: Mediation	a	Digital Skills	Attitude Balance	0.43	0.02	24.30 (3630)	<0.001	[0.40, 0.47]
	b	Attitude Balance	AI Per.	0.47	0.01	38.18 (3629)	<0.001	[0.45, 0.50]
	c′	Digital Skills	AI Per.	0.22	0.01	15.35 (3629)	<0.001	[0.20, 0.25]
	c	Digital Skills	AI Per.	0.42	0.01	26.99 (3630)	<0.001	[0.39, 0.45]
	indirect	Digital Skills→ Attitude Balance→ AI Per.	—	0.20	0.009	— (bootstrapped)		[0.19, 0.22]

Note: Employer Digital Support: training and tools provided by the employer; AI Per.: individuals’ perception of AI in the workplace.

summarizes these results.

Here, in an additional inquiry based on the crosstabulation reported in

Table 3. Additional inquiry crosstabulation.

Augmentation–Replacement Balance	AI Replacement Attitude
	Disagree	Agree
Low-Augmentation-Dominant (0.88)	3.7%	96.3%
Medium-Augmentation-Dominant (1.88)	34.4%	65.6%

, we tested two additional mediation models. The rationale for this strict additional inquiry is duly addressed and reported in the discussion, respectively. In a model testing the mediation of replacement attitude on the effect between digital skills and AI perception, the indirect effect (b = 0.007, BootSE = 0.001, 95% CI [0.004, 0.010]) as well as the total effect (b = 0.40, SE = 0.008, 95% CI [0.383, 0.415]) were significant. The indirect effect accounted for 1.8% of the total effect (indirect effect/total effect = 0.0073/0.4062), indicating a minimal mediating role and justifying negligible practical significance. This signifies that digital skills’ effect on AI perception is only minutely carried out by replacement attitude. In a model testing the mediation of augmentation attitude on the effect between digital skills and AI perception, the indirect effect (b = 0.23, BootSE = 0.007, 95% CI [0.219, 0.249]) as well as the total effect (b = 0.42, SE = 0.009, 95% CI [0.400, 0.436]) were significant. The indirect effect accounted for 56% of the total effect (indirect effect/total effect = 0.2346/0.4182), indicating a substantial mediation role of AI augmentation attitude as compared to practically negligible mediation by replacement attitude. This demonstrates that digital skills’ effect on AI perception is largely carried out by an augmentation attitude.

Table 4. Additional inquiry mediations.

Effect	Path	Predictor	Outcome	β	SE	t (df)	p	95% CI
AI Aug. Mediation	c	Digital Skills	AI Per.	0.42	0.00	45.25 (7940)	<0.001	[0.40, 0.44]
	ind	Digital→AI Aug.→AI Per.	—	0.23	0.007	— (bootstrapped)		[0.22, 0.25]
AI Rep. Mediation	c	Digital Skills	AI Per.	0.40	0.00	49.69 (11342)	<0.001	[0.39, 0.42]
	ind	Digital→AI Rep.→AI Per.	—	0.007	0.001	— (bootstrapped)		[0.00, 0.01]

Note: AI Aug.: favoring attitude that AI will augment the workplace; AI Rep.: disfavoring attitude that AI will replace the workforce.

summarizes these results.

3.4.2. Research Model 2

Hypothesis 6 examined whether employer transparency moderated the impact of replacement attitude on individual perception, while controlling for age, gender, and education. In the linear moderation analysis, replacement attitude had a negative effect on AI perception (β = −0.26, p < 0.001), whereas employer transparency had a strong positive effect (β = 0.43, p < 0.001). This signifies that replacement attitude predicts a negative perception of AI among individuals, and, in contrast, being informed about AI use in the workplace predicts a positive perception. The interaction was significant (β = 0.14, p < 0.001); employer transparency positively moderated the negative relationship between replacement attitude and AI perception. This indicated that the negative impact of replacement attitude was attenuated by transparency and confirmed H6. Specifically, for uninformed respondents, the attitude–perception slope was β = −0.26 (p < 0.001), whereas for informed respondents it was only β = −0.12 (p < 0.001). This implies that being informed about AI use lowers the levels of replacement attitude among individuals and its translation into a negative perception. This interaction is illustrated in Figure 3.

4. Discussion

This study utilized a 2024 cross-national survey of populations of the 27 EU member states, exploring the public’s views on digitalization and AI. We primarily aimed to understand Europeans’ most recent attitudes toward, and perception of AI via two novel, unexplored pathways. First, we employed a dichotomous operationalization of attitudes toward AI, distinguishing between AI augmentation attitude and AI replacement attitude. Second, we established a justified distinction of individuals’ perception of AI from attitude, and the direction of their relationship, i.e., attitudes form individuals’ perception of AI use in the workplace [15,16]. In particular, we studied how employer transparency and workers’ digital skills influenced their attitudes and perceptions, while controlling for age, gender, and education. In total, we tested six hypotheses.

As hypothesized (H1), employer digital support (i.e., training/tools received) had a very strong, positive effect on employees’ digital skills. This buttresses our theoretical pivot to JD-R and self-efficacy theory, suggesting that training can increase skill variety, particularly where digital skills constitute a relatively new skill set, and mastery through training can assist individuals’ belief in their capacity [23,41]. It also validates academicians’ recommendation of implementing training for effective AI integration via skill building [45]. The findings supporting H2 strongly suggested and confirmed that individuals with stronger digital skills tend to hold a more favorable, augmentation attitude toward AI over a disfavoring, replacement attitude. Digitally skilled people believe AI will augment the workforce rather than replace it. This is consistent with MIS attitude research and theory of TAM [21], along with UTAUT [22], suggesting that attitude is a function of perceived ease of use and effort expectancy, i.e., effort required to be skillful in using or operating the system. Further implying that individuals with higher digital skills will require less of this effort, which in turn, systematically dictates their attitude toward technology. This finding is in support of multiple studies examining digital competence’s positive influence on favoring attitudes toward AI use [59,60]. It also extends support to previous studies confirming the protective influence of digital competence over job replacement insecurity [28,61,62].

Furthermore, digital skills were strongly associated with AI perception (H4), implying that digital skills also influence how positively or negatively workers perceive AI’s use in the workplace. This is in support of numerous contemporary studies noting the positive influence of digital skills on workers’ perception of AI [30,65,69]. In fact, our study builds on their limitations with a larger, improved representative sample, along with influential control variables [83,84]. Moreover, as hypothesized, individuals’ attitudes guided their perception. Augmentation–replacement balance was strongly and positively associated with AI perception (H4). More importantly, confirming one of the core inquiries of this study (H5), the mediation analysis revealed that the positive effect of digital skills on perception was substantially mediated by attitudes (48%). This directly addresses the research gap in AI literature and corroborates the theoretical framework developed in our study. Specifically, individuals’ perception of an entity and its qualities is systematically influenced by their attitudes [15,16], and due to confirmation bias, the biased perception depends on the state of beliefs (an integral component of attitude) [25,64].

However, it is necessary to note that a crosstabulation (Table 3) informs us that we see moderate to high replacement attitude levels even in individuals with low (96.3%) and medium augmentation-dominant attitude (65.6%), reflecting cautious optimism in support of the findings of Dong et al. [52]. Thereby, in an additional inquiry, we investigated whether augmentation attitude alone mediates the relationship between digital skills and perception more profoundly than replacement attitude. Evidently, the augmentation attitude explains 56% of the relationship, compared to only 1.8% explained by the replacement attitude (Table 4). Additionally, there is a difference in r = 0.47 in the strength of correlation between the augmentation attitude and digital skills, compared to the replacement attitude. This is crucially important and might be one of the other novel insights of this study, in contrast to numerous studies alarmingly cautioning against the replacement fear. As it may be suggesting, the replacement attitude is a separate and disparate, yet crucially important parameter that perhaps plays a key role in driving the cautious, responsible adoption of AI [40], rather than AI use in the workplace. Further implying that in the discrete landscape of AI use, the development of digital skills and, as a result, a positive attitude about AI augmenting (and not replacing), assists the individuals in perceiving AI use more positively, and may be more important, as prioritized by the EU in its policies and rigorous public programs [7].

AI integration decisions remaining private among employers, giving rise to the vast “unknown” [70], is another source that can cause unrest among employees via uncertainty [24], which transparency can alleviate [71,72]. Included in our study, evidently, employer transparency moderated a key relationship and indicated that simply being informed about planned AI use lowers the anxious, disfavoring attitude and its translation into a negative perception. This validates our theoretical rationale that a lack of communication on AI engenders uncertainty that exacerbates the negative affective state of employees [24]. Moreover, findings also support that transparency influences workers’ cognitive trust [20], and informing employees helps raise their outcome expectations, shaping their cognitive views [23]. Affective and cognitive components are two of the three components of attitude [25]. Precisely, transparency attenuated the impact of replacement attitude on perception. Among uninformed workers, replacement attitude strongly depressed perception, whereas among informed employees, this effect was much weaker and the levels of replacement attitude were overall low (Figure 3). Implying that being informed about AI plans moderates the extent to which replacement attitude is translated into pessimistic perception, the findings support calls for transparent AI communication [73,74].

4.1. Implications and Recommendations

Digital skills emerged as a pivotal driver of AI acceptance, i.e., higher digital skills engender a favorable, augmentation attitude toward AI and a positive perception of AI’s integration into the workplace. Aligned with technology-acceptance theories (TAM, UTAUT), it suggests that skill-building is one of the most viable paths to improving AI receptivity, and digital skills can allow workers to accept AI as augmenting and hold a positive attitude rather than an attitude that AI will replace the workforce, further implying that digital skills serve a role of protection as well as preparation for AI integration in the workplace. Furthermore, the findings also noted the extreme importance of providing AI-related tools and training to employees in elevating digital skills. Thus, in practice, governments and firms should invest significantly in AI-related training and digital literacy programs. By equipping the workforce with tools and skills to use AI, organizations can create a workforce that views AI as a collaborator rather than a threat.

Second, the replacement attitude persists even in augmentation-dominant attitude respondents, representing cautious optimism. Moreover, our study shows that the positive impact of skills on perception operates mainly through instilling an augmentation mindset, as the augmentation attitude mediated over half of the skills→perception effect. The mediation via replacement attitude was statistically significant but practically negligible. This sheds light on two distinct roles played by two distinct attitudes. The role of replacement attitude is imperative in AI adoption, whereas an augmentation attitude is to be developed by digital skills to perceive AI use in the workplace positively. The replacement attitude essentially establishes the very ground upon which pathways of AI adoption rely, and drives the steps to ensure a responsible, harmless integration of AI in the workplace. Thereupon, the steps to follow include policies and programs (such as those initiated by the EU) that focus on digital skills development and thus, fostering a positive, augmentation attitude. Therefore, future research should systematically and more thoroughly examine workers’ replacement attitude, with an aim to uncover key as well as implicit concerns of workers regarding AI deployment that are methodically nurturing the replacement attitude. Such insights can profoundly inform adoption frameworks that both mitigate job-loss risk and emphasize AI’s capacity to augment human capabilities. In addition, AI promotional communications should emphasize framing AI as a supportive tool under responsible AI policies.

Third, and perhaps the most noteworthy insight is the power of transparency unraveled in our findings. Being informed about AI plans greatly weakens the negative impact of replacement attitude on AI perception. The finding reaffirms that transparent AI communication effectively raises outcome expectancy, an influential belief, and is prominent in shaping attitudes and their translation. Thus, organizations, along with regulators, should prioritize clear and proactive AI communication in order to reduce the negative outlook toward AI among employees. This could even be formalized in corporate AI governance guidelines or labor policies, given the large mitigating effect observed.

In summary, the study suggests two avenues for practice, i.e., building skills and exercising transparency. First, boost digital skills across the workforce, since skills strongly drive positive AI attitude and perception. Second, increase institutional transparency concerning AI adoption, as being informed consistently lowers the negative reception of AI. In regard to research, scholars should further investigate the cautious optimism phenomenon observed in our study, examining why a favorable attitude co-exists with a disfavoring view.

4.2. Limitations and Future Research

While this research offers novel insights into the drivers of AI attitudes and perception, several limitations merit acknowledgement. First, all data derive from a single Eurobarometer wave, preventing any causal inferences. Furthermore, our study’s research design is not strictly experimental, even though it sufficiently practices randomization and rigorous post-collection protocols. This potentially paves the way for several biases. For instance, all variables were self-reported in the same survey, allowing for common-method bias. On the other hand, response bias may especially be relevant as our study contains measures of digital skills and employer support, where an objective measurement may represent a more accurate assessment. Furthermore, for behavior measurements such as attitudes and perception, self-reported measures may lack accuracy, and extensively validated measures may be better suited. Thus, future longitudinal and experimental studies are recommended to adequately address these concerns.

Generalizability beyond Europe is another restraint of our study. Although Eurobarometer offers a large, weighted, representative sample of the EU-27, cultural, regulatory, and labor-market differences limit direct extensions to non-EU contexts (e.g., the U.S., Asia). Another noteworthy limitation is the lack of cross-industry effects of AI integration. This is extremely crucial as some job roles are more at risk for automation and replacement than others. Although capturing these nuances was deemed seemingly impossible by Dong et al. [52], future research should gain insights into cross-industry effects of AI, examining how differently and uniquely AI affects workplaces.

Last, the most crucial limitation of the study lies in its utilization of single-item measures. Nevertheless, all three single-item measures were face-valid, global measures capturing the unambiguous, unidimensional nature that we study. Furthermore, dichotomization of employer transparency’s seven responses (where four responses are dedicated to ‘yes’, one to ‘no’, and two comprise ‘missing/absent’ values) is intentional and pivoted on increasing robustness. The motivation lies behind the non-linearity, randomness, and specificity of four ‘yes’ responses, as the scale does not increase in intensity, unlike a normal scale, implying that it does not measure ‘the degree of transparency’. For instance, one response reads, “yes, and you have been given access to your personal data”, while the other response reads, “yes, and you have been given access to the results of the automated analysis carried out”. Thus, we retain the unambiguous, robust, and binary knowledge that we required, i.e., whether employees were informed about AI use or not. Notably, the scale is originally dichotomized in the Eurobarometer survey as well. Moreover, due to the possibility, the domain-general AI perception measure was validated against the domain-specific longer scale. However, these measures may still under-capture constructs’ conceptual complexity, do not allow estimation of internal consistency (e.g., Cronbach’s alpha), and are more susceptible to random measurement error. Furthermore, estimated relationships may be attenuated, and results may not support price construct-level calibration. Thus, findings should be interpreted as reflecting upon general, global perception, not fine-grained dimensions of perception. As a result, future research may benefit from multi-item batteries of these variables and should also explore the attitudinal distinction made in our study for further validation.

5. Conclusions

This study explored Europeans’ digital skills, attitudes, and perceptions surrounding AI in the workplace. First, it investigated employer digital support for employees’ digital skills. Second, it uncovered AI attitudes’ (augmentation vs. replacement) role in influencing individuals’ perception of AI. Third, it examined the employer transparency’s integral function in the translation of replacement attitude into negative perception. This research utilized the Eurobarometer survey of over 26,000 EU workers, carried out by the European Commission in 2024. The results revealed that employer-provided training and tools significantly boost digital skills, which in turn foster an augmentation-dominant attitude and reduce replacement attitude. Augmentation attitude strongly mediated the effect of digital skills on positive AI perception, whereas replacement attitude had only a minimal mediating impact. Moreover, employer transparency buffered the negative effect of replacement attitude on perception. This study serves as the foundation stone of formal research inquiries into two broad, distinct attitudes individuals hold toward AI’s deployment in the workplace. In addition, this study’s contribution is particularly invaluable in addressing the research gap and guiding future researchers in being mindful about the distinction between attitude toward and perception of AI in the workplace, as they are conceptually disparate. Furthermore, by highlighting the importance of upskilling, it substantiates the motivator of the Digital Europe Programme (DEP), i.e., to tackle the digital skills gap. It also substantiates the EU’s goals in digital transformation of establishing four academies to close the digital skills gap. Moreover, the EU AI Act demands a great degree of transparency about AI deployment from deployers to the European AI office. However, beyond governmental transparency, our study highlights the value of employee-level transparency, i.e., transparency from employers to employees regarding AI’s integration in the European workplaces. Lastly, this study calls for future research into how replacement attitude shapes cautious AI adoption.