Evaluating public sector employee perceptions towards artificial intelligence and generative artificial intelligence integration

1. Introduction

In spite of the increased prevalence of digital technology in today’s corporate landscape, there still needs to be a comprehensive understanding of the constituents of digitisation capabilities. The extant body of literature on digital transformation predominantly examines its impact on external stakeholders, particularly the dynamics between organisations and their customers [1]. However, there appears to be a relative paucity of research addressing the role of employees in the adoption and implementation of digital transformation within organisations [2]. A significant gap in the academic literature exists concerning the role of employees’ perceptions and attitudes towards new technologies, particularly artificial intelligence (AI), and the subsequent impact these factors have on the adoption and endorsement processes of such technologies. This oversight represents a crucial problem in understanding the comprehensive dynamics of technological integration in the workplace. Further scholarly exploration in this domain is warranted to decipher and appreciate the underlying factors that shape the incorporation of these transformative technologies in businesses.

The study aims to create a dialogue between theoretical studies and empirical studies, filling in this research gap by providing an overview of how public organisations’ employees perceive AI technology and the role of that perception in adopting and endorsing AI technologies in public administration practices. This research aims to bridge two different points of view about digitalisation in the public sector: the first concentrates on external stakeholders. In contrast, the second focuses on the role of employees’ perception of innovative technologies. Authors embed the discussion in the theory of human knowledge assets, treating employees as key organisational assets determining the process of AI implementation. Authors contribute to the existing research by analysing the influence of different demographics in the process, to explore how various demographic factors – such as age, gender, educational background and job experience – influence these perceptions and attitudes. Understanding these demographic nuances is vital, as it can lead to more effective and inclusive approaches to technology adoption, ensuring that the benefits of new technologies are maximised across the entire workforce. By addressing these gaps, future research can provide a more holistic view of the factors that drive successful technology integration, ultimately fostering a more adaptive and forward-thinking organisational culture.

This study explores how civil servants perceive the use of AI technologies in the public sector, aiming to fill gaps in the existing literature. The authors added the effect of gender and age on the technology acceptance variables, addressing the following research questions:

The remainder of the article is structured as follows. The second section consists of the theoretical background, where the authors present the concept of AI, focusing specifically on public sector specificity. The third section outlines the methodology, precisely a questionnaire-based approach. The fourth section presents the main results, and the fifth section presents the discussion. The article concludes with theoretical and managerial implications and study limitations, paving the way for future research.

2. Theoretical background

3. Methodology

The exploratory research used a questionnaire as the primary method of data collection. The authors took the questions from the survey by Ahn and Chen [62] as a reference, as it was considered comprehensive. Participants used a 5-point Likert-type scale (ranging from no knowledge/impact to high knowledge/impact) to indicate the response to each demand. The 17 questions investigated several aspects of the phenomenon:

Participants were chosen from various hierarchical levels, including department heads, directors, managers and employees of multiple ranks, to guarantee a broad and representative sample. The questionnaire was constructed using the Google platform and disseminated online via email.

Following the construction of the questionnaire, a pretesting process was conducted. The questionnaire was assessed by 30 individuals who provided feedback regarding its complexity and question formulation. The pretest aimed to identify any apparent defects, and fortunately, this phase did not reveal any issues.

Data collection took place in Italy between June and August 2023. Out of the 470 participants, 439 returned complete questionnaires. These respondents comprised 196 females and 243 males domiciled in Italy, with an average age of 47 years (standard deviation, σ = 11.5).

It is important to note that incomplete questionnaires, representing only 6.59% of the total, were not included in our analysis. This approach allowed us to disregard the handling of missing data. Authors consider the obtained sample significant, and its representativeness has been determined using Cochran’s method [63].

The authors declare their objectivity in the face of competing commercial interests or personal relationships that may influence the outcomes presented in this article.

4. Analysis methodology

The collected data were analysed using R software, and various statistical tests were performed, including measures of dispersion, adjusted dispersion, the Student’s t-test, tests for heteroskedasticity and criteria such as Akaike, Schwarz and Hannan–Quinn. These tests were carried out to evaluate and confirm the model’s reliability.

To test the RQ1, authors implemented an econometric model based on a multiple linear regression (MLR) function. This method, widely used in various studies [64–66], enables understanding the relationship between a dependent variable and multiple independent variables. It aids in predicting outcomes and clarifying the contribution of various factors to these outcomes, allowing simultaneous management of multiple predictors and providing a clearer picture of how these factors collectively impact the dependent variable. In addition, it offers precise estimates of the strength and direction of each predictor’s effect [67–69].

The MLR distribution utilises the ordinary least squares (OLS) method

y = α 0 + α 1 * X 1 + α 2 * X 2 … + α n * X n + ε

where:

The MLR (authors’ values) best fit the data and minimise the sum of squared residuals (the differences between the predicted values and the actual values).

RQ2 was verified using Spearman’s ρ, which allows authors to measure the strength and direction of monotonic association between variables [70–72]. This coefficient is used in many studies to verify the convergent validity of assumptions considered [73–75]. Differences in mean scores were assessed through the Wilcoxon–Mann–Whitney test [76], while the relationships between nominal variables were tested using chi-square statistics [77]. The level of statistical significance for tests was set at p ≤ 0.05.

5. Results

Table 1 shows the characteristics of the respondents.

OPEN IN VIEWER

Table 1.

Sample demographics features.

Variable	Description	Number	Frequency (%)
Gender	Male	196	44.6
	Female	243	55.4
Age	Minus or equal to 30	47	10.7
	31–40	74	16.9
	41–50	102	23.2
	51–60	172	39.2
	Over 60	44	10.0
Role	Department heads	34	7.7
	Directors	92	21.0
	Managers	112	25.5
	Employees of multiple ranks	201	45.8
Work sectors	National Health Service	85
	Regions and Local Autonomies	69
	Ministries	53
	Education	93
	Police and Armed Forces	103
	Prefectural and Penitentiary Career	36

The data distribution table enabled us to understand the background of Italian respondents, dominated by female employees with an average age of 47 years (σ = 11.5). The obtained values are consistent with the data distributed in Italy by ISTAT [78]; therefore, the sample can be considered as representative of the population.

To examine RQ1, the authors employed an MLR model. The data analysis was carried out utilising R software. The initial stage involved assessing the multicollinearity of the variables under consideration. This step was essential, as correlations between multiple independent variables in a multiple regression model could lead to regression coefficient estimates that lack statistical significance. With this aim, authors computed the variance inflation factor (VIF). The obtained value (VIF = 1.59), following Debbie et al. [79], indicates that no multicollinearity was observed.

The second phase aggregated the questions related to each aspect to obtain a representative index. To do this, authors calculated the average of the responses to individual questions and then assessed the robustness of this aggregation using the standard error. By combining the participants’ responses, researchers aimed to obtain a general indicator or score that reflected the overall measure of each aspect considered. The average calculation was used as the aggregation method; at the same time, the standard error was considered to assess how consistent and reliable the responses were as the basis for creating the representative index. This process is intended to simplify the interpretation of the data and provide a clear view of the overall trends in the data related to the aspects considered (Table 2).

OPEN IN VIEWER

Table 2.

Layout of the aggregative method.

Acronym	Description	N items	Standard errors
FAMBLEC	Overall familiarity with big data, machine learning, cloud computing, blockchain and practical experience in their usage	5	0.0361
TUPID	Perception of technology’s impact on public sector decision-making and privacy concerns regarding personal data	3	0.0281
P-GAI	People’s perception of Generative AI’s future applications	1	0.000
WTUWT	Willingness to embrace and support the adoption of new workplace technologies	4	0.0416

To determine how various dimensions of perception influence public servants’ inclination to adopt and endorse AI technologies in public administration practices, the authors established a multiple linear equation based on the OLS method

P − GAI = 0 . 701 * TUPID + 0 . 549 * WTUWT + 0 . 261 * FAMBLEC

where P − GAI is the dependent variable of the model with TUPID , WTUWT and FAMBLEC are the independent variables. The coefficient of determination (R²) is equal to 0.891.

The formulation of the equation indicates that the model accounts for 89.1% of the studied phenomenon, namely the perception of the utility and impact of new technologies for advancing decision-making in the public sector, combined with the willingness and desire to support their implementation and use in work practices, as well as the general level of knowledge of digital technologies, generates a greater inclination towards the use of generative AI in public administration practices.

The conducted statistical tests reveal a substantial homogeneity among the regression variables, as indicated by the p-value coefficient ( p < 0 . 0001 ) falling below the significance threshold of 0.005. Concerning RQ1, the test for normality of residuals (Hi square (2) = 267.861 with p = 2.89963e-041) demonstrated that, under the null hypothesis, the error follows a normal distribution. This leads to rejecting the null hypothesis, confirming the alternative hypothesis and validating the model that addresses the specific hypothesis. The Breusch–Pagan test for heteroskedasticity (LM = 4.67941 with p = P (Hi square (2) > 4.67941) = 0.096786) indicates endogeneity within the econometric model. Consequently, it suggests that the errors or residuals in the observed model exhibit varying variances between one observation and another, leading to rejection of the null hypothesis and confirmation of the model. In Figure 1, the histogram distribution illustrates the variation in predicted data for the dependent variable, which differs across the four segments. This variation is also reflected in the differences in adherence to the profession based on respondents’ professional category. The normality test indicates a significant relationship between the variables, with data concentrated in the quartile area. This causality enables us to quantify the studied impact accurately, validating RQ1.

OPEN IN VIEWER

Figure 1.

Histogram distribution.

RQ2 aims to verify whether the desire and willingness to support the implementation and use of innovative technologies in workplace practices are influenced by factors such as age, gender, job position and type of occupation.

First, the authors analysed the relationship between age and willingness to embrace and support adopting new workplace technologies (WTUWT). For this purpose, the authors computed Spearman’s rho correlation ( ρ ) (Table 3). The coefficient value detects a positive correlation among young workers while highlighting a negative correlation with increasing age.

OPEN IN VIEWER

Table 3.

Spearman’s correlations between age and WTUWT.

Age	WTUWT
	ρ	p
Minus or equal to 30	0.781	0.563
31–40	0.571	0.399
41–50	0.319	0.265
51–60	−0.289	0.302
Over 60	−0.488	0.578

Differences between men and women regarding the frequency of willingness to embrace and support adopting new workplace technologies were tested using chi-square statistics ( χ 2 ) . The results ( χ 2 ( 1 ) = 0 . 07 , p = 0 . 689 ) highlight no significant differences.

To compare the job position (role), the Wilcoxon–Mann–Whitney test was used, a non-parametric test that allows determining if there are statistically significant differences between the means of two independent groups (skewness and kurtosis exceeding 1). The results are shown in Table 4.

OPEN IN VIEWER

Table 4.

Job position decisions.

	Department heads	Directors	Managers	Employees of multiple ranks
	pZ	pZ	pZ	pZ
Department heads		0.781 − 0.391	0.583 − 0.514	0.107 − 1.275
Directors	0.781 − 0.391		0.854 − 0.175	0.061 − 2.151
Managers	0.583 − 0.514	0.854 − 0.175		0.037 − 2.351
	0.107 − 1.275	0.061 − 2.151	0.037 − 2.351

The obtained values highlight a different desire and willingness among various job positions. In particular, the results show considerable differences between employees and the other clusters. On the contrary, significant differences must be evident among the other groups.

The Wilcoxon–Mann–Whitney test was employed to compare the job sectors. The results are shown in Table 5.

OPEN IN VIEWER

Table 5.

Work sector decisions.

	National Health Service	Regions and Local Autonomies	Ministries	Education	Police and Armed Forces	Prefectural and Penitentiary Career
	pZ	pZ	pZ	pZ	pZ	pZ
National Health Service		0.828 − 0.302	0.723 − 0.498	0.239 − 1.198	0.089 − 2.175	0.069 − 2.189
Regions and Local Autonomies	0.828 − 0.302		0.574 − 0.872	0.281 − 2.018	0.078 − 2.307	0.059 − 2.336
Ministries	0.723 − 0.498	0.574 − 0.872		0.347 − 2.009	0.101 − 2.168	0.075 − 2.203
Education	0.239 -1.198	0.281 − 2.018	0.347 − 2.009		0.264 − 1.985	0.268 − 1.986
Police and Armed Forces	0.089 -2.175	0.078 − 2.307	0.101 − 2.168	0.264 − 1.985		0.683 − 0.686
Prefectural and Penitentiary Career	0.069 − 2.189	0.059 − 2.336	0.075 − 2.203	0.268 − 1.986	0.683 − 0.686

The obtained values highlight a different desire and willingness among various work positions. In particular, the test indicates significant differences between clusters of National Health Service, Police and Armed Forces, or Prefectural and Penitentiary Careers. These differences could be attributed to the distinct job roles, work environments, stress levels and the nature of the tasks associated with each position.

6. Discussion

The results of this survey can provide valuable information for policymakers and practitioners seeking to improve the effectiveness and efficiency of the public sector through technological progress.

Concerning the first research question, as proved by the early studies by Davis et al. [80], the perceived importance of innovative technologies has a more significant impact on the intention to use rather than the ease of use. However, the latter also has a considerable influence. There is a positive and significant correlation between the perceived importance of technology and the intention to disseminate it; however, the opposite cannot be confirmed. Even those most sceptical about innovative technologies did not oppose their dissemination. The research gave clear evidence about the impact of dimensions of perception on civil servants’ inclination to adopt and endorse AI technologies in public administration practices. Online questionnaires showed positive attitudes towards implementing AI technologies in the public sector. Contrary to Craig et al. [36], most surveyed people assumed that AI would not replace their roles.

Technological background acquired skills or previous experience positively influence acceptance regarding attitude, intentions and behaviour towards AI. In contrast, there is no inverse correlation between the level of knowledge of the technologies and the willingness to use them. Our results indicate that even those with little knowledge are still willing to use innovative technologies to support their work. An explanation of this phenomenon can be found in the concept of trust: user trust in technology, exemplified by studies like Chang et al. [81], is essential. It empowers users to rely on devices like Google Maps for restaurant directions. Successful outcomes reinforce this trust, creating a positive cycle of reliance on technology for achieving the desired goals.

Concerning RQ2, the results of this study showed significant differences in respondents’ answers not only by age, as initially assumed by the authors, but also related to role and worker sector. The study also reveals that predictions about AI development in the public sector can be done, considering bias about differences of perceptions among employees.

The perceived usefulness of innovative technologies is more relevant for younger people. In contrast, mature clusters, who believe they have low technological skills, find technologies challenging. Many studies have proved that older adults remain at the opposite end of the digital divide [82–84]. Studies also proved that older adults who adopt ICT feel its value even if it is for a limited number of functions, such as email [85].

The study gives an important contribution to academic research that focuses on knowledge asset frameworks and bias. An example can be found in recent studies that deepen the concept of trust as not necessarily affected by age [86]. Results also contribute to the present field of study concerning how cognitive biases can affect our judgement and ability to make rational decisions in personal and professional environments [87]. Finally, the article integrates important academic research focusing on knowledge contributions of the stakeholders of AI auditing, such as the recent study conducted by Laine et al. [88].

Furthermore, a correlation exists between role and intention to use. The Mann–Whitney test indicates significant differences between cluster employees and directors or managers, as shown in Table 4. Consistent with Bin Taher et al.’s [57] findings, middle and senior managers (executives and department heads, male) were less willing to support introducing enabling technologies in their work practices, while most employees accept and are committed to using and supporting AI.

According to Abdullah and Fakieh [89], this difference in perceptions regarding job tenure may be motivated by the fact that employees are more probable to meet the potential of innovative technologies by working ‘on the front line’ directly with the patient, unlike top managers [90]. Recent studies underscore the significance of empowering frontline workers to provide specific and critical feedback on designing and deploying AI systems [91].

According to the findings of this study, organisations could benefit from the result findings to enhance their digital competence levels to disseminate a credible digitisation culture. Only with an awareness of top management can the latter approve successful technology initiatives promoted by the lower hierarchy ranks [92].

Another significant discovery pertains to variations in willingness among distinct occupational sectors. The data underscore that healthcare professionals exhibit a more pronounced proclivity towards technology adoption when compared with their counterparts in the military sector. This inclination can be attributed to the healthcare sector’s pressing demand for swift access to critical medical information, an environment characterised by ongoing technological advancements, an imperative for real-time communication and information exchange, rigorous record-keeping and monitoring prerequisites, evolving demographics and a prevailing culture that champions innovation.

In summary, our research underscores several critical insights. Regarding the impact of various dimensions of perception on civil servants’ inclination to adopt and endorse AI technologies in public administration practices, public sector employees hold positive attitudes towards AI implementation [93], believing it will enhance, rather than replace, their roles. The perceived importance of innovative technologies significantly influences their adoption, although ease of use also plays a role. A technological background positively influences attitudes, intentions and behaviours towards AI, while the level of knowledge does not deter willingness. Finally, healthcare professionals are more ready to embrace technology, reflecting the sector’s unique demands and innovation-oriented environment. Recent studies confirm that ChatGPT, as a language model, is a groundbreaking healthcare innovation [94].

An important theoretical contribution of this research is related to the field of knowledge assets and the connection with demographic data: results show that gender, age, work position or work sectors differed in the responses. These variables emerged as critical determinants of attitudes and intentions, emphasising the need for tailored strategies aimed at managing human knowledge assets. According to the results of this study, open innovation and knowledge-sharing are pivotal within public sector organisations. The article suggests an interesting direction to manage employees as key organisation’s knowledge assets. The methodology of the study could be developed as a theoretical content for corporate organisations. Top management could benefit from this analysis to acquire digital competence to drive a culture of digitisation. Digital leadership, as discussed by El Sawy et al. [95], pertains to leaders undertaking effective actions to navigate the digitalisation of organisations. This concept encompasses qualities that positively impact the attitudes and behaviours of organisational members influenced by digital technologies [96].

7. Conclusion

According to Gartner, by 2026, enterprises that apply Trust, Risk and Security Management (TRiSM) controls to AI applications will increase the accuracy of their decision-making by eliminating 80% of faulty and illegitimate information. By 2028, 75% of enterprise software engineers will use AI coding assistants, up from less than 10% in early 2023.

According to the results discussed, public sector organisations could operate effectively by managing their knowledge assets with some suggested strategies: developing an open innovation environment, encouraging knowledge access and sharing it among employees at all levels. Although hampered by rigid hierarchies, disseminating knowledge in the public organisation seems to encourage enabling technological progress at all levels [97].

The methodological framework adopted incorporates a meticulously designed questionnaire, referencing antecedent research for its construction. The findings articulate a prevailing optimism among respondents regarding integrating AI technologies within the public sector, aligning with contemporaneous studies [98]. This study accentuates the pivotal role of perceived technological importance in shaping adoption intentions. Furthermore, the analysis reveals that a technological background positively influences attitudes and intentions towards AI adoption, underscoring the pivotal role of trust in technology. The study advocates for nuanced strategies in the public sector to accommodate employees’ diverse perceptions and preferences. Insights about management underscore the evolving nature of leadership in the face of increasing digitisation [99].

As public administrations undergo transformative shifts to meet emerging societal needs, civil servants necessitate digital competencies for more efficient and effective public service delivery [100]. In a broader context, this research augments the knowledge base available to policymakers and practitioners endeavouring to enhance the efficacy of the public sector through technological advancements.

Distinguished by its methodological rigour, exhaustive exploration of diverse technologies, nuanced analysis of demographic variables and theoretical contributions, this research emerges as a novel and valuable addition to the academic literature in different fields, such as technology adoption within organisational contexts, analysing knowledge assets frameworks and bias, and stakeholders mapping of AI auditing.

Nevertheless, limitations of the study can be found in the territorial restriction on the stakeholder involved. More comprehensive research could confront other European contexts apart from the Italian one. Nationality is an essential limit to this field of research, as European laws tend to find a better definition and regulation of AI adoption. The General Data Protection Regulation (GDPR) within the European Union (EU) stands out as a premier framework, adeptly addressing AI decision-making within governmental contexts.

Another limitation of the article can be found in the specificity of the legislations considered about corporate and public contexts and adoption of AI. An interesting comparison could also be made with a country like the United Kingdom, close enough to Italy and the European culture but outside the EU legislation. A critical case study is the Open Government Data of Gov.uk. The United States and China, among other nations, have acknowledged the substantial value of AI for public benefit, initiating costly AI initiatives that underscore the diverse potential applications in various fields.

The article wants to open an original field of research about bias affecting the perception of AI adoption in the public sector, but still lacks to consider a wider range of AI technologies and flows. Future research agenda can be developed for the entire AI innovation lifecycle of adoption, implementation and diffusion in the public sector. Developing research fields that help predict directions in AI development is crucial.