2.1. Generative Ai in the workplace

The popularity of Generative AI has grown with the emergence of ChatGPT(Hussain et al., 2024). By integrating deep learning and language models, ChatGPT has greatly improved the capabilities of traditional chatbots. Its advanced language processing skills allow it to be effectively utilized in various workplace settings. One prominent application is in customer service, where ChatGPT's ability to understand and respond to human language in real time is expected to significantly enhance efficiency. Additionally, it excels in data analysis and report generation, going beyond mere text summarization to analyze data and create detailed reports. In fields like programming, ChatGPT demonstrates impressive problem-solving abilities.

Research on Generative AI has predominantly focused on its application in educational contexts. Studies have explored its potential to aid in scientific writing (Salvagno et al., 2023) and its impact on educational environments, debating whether it would be beneficial or detrimental (Lim et al., 2023). However, most of these studies remain conceptual, with empirical research still in its early stages. Specifically, there is a lack of research examining the role and acceptance of Generative AI in workplace settings.

With the rise of AI driven automation, like ChatGPT, there's a growing fear of job instability among workers (Cao & Song, 2024). Additionally, businesses might be wary of implementing Generative AI tools due to worries about sensitive data leaks. To highlight the importance of further investigation, the upcoming section will focus on the dual factor theory.

2.2. Dual factor theory

Dual factor theory divides the determinants of technology usage intentions into two categories: enablers and inhibitors (Cenfetelli, 2004). The theory was first introduced by Herzberg et al. (1959), who discovered that job satisfaction and dissatisfaction arise from different factors. Contrary to the common belief that satisfaction and dissatisfaction are merely opposite ends of a single spectrum, this theory asserts that they are influenced by distinct variables. Traditional models of technology acceptance, such as the Technology Acceptance Model (Davis, 1989) and the Unified Theory of Acceptance and Use of Technology (Williams et al., 2015), attempt to explain the adoption of new systems based on their perceived usefulness (Park et al., 2022). However, they often overlook inhibitors. Considering the crucial role that barriers to innovation play in technology adoption, both enablers and inhibitors must be examined (Park and Park, 2024; Park et al., 2024). Despite the growing recognition of Generative AI, particularly ChatGPT, its professional application remains limited. This study, therefore, employs Dual Factor Theory to explore and understand the factors that influence the intention to adopt Generative AI in the workplace.

3.1. Enablers

3.2. Inhibitors

4.1. Survey Items

Respondents were asked to address six variables that represent the benefits and risks associated with Generative AI. The benefits include knowledge support, customization, and entertainment, while the risks encompass perceived risks, realistic threat, and identity threat. To ensure the reliability and validity of the measurements, items were adopted from previous studies.

Knowledge support was measured with three items from Wang et al. (2023). Customization was assessed using three items from Cheung et al. (2020), and entertainment was evaluated using three items from Li et al. (2013) and Wolf and Maier (2024). Perceived risk was measured with three items from Lu et al. (2011). Realistic threat and identity threat were each measured with two items from Huang et al. (2021). The intention to use Generative AI was measured with three items from Lin (2011). All items were rated on a seven-point Likert scale (1 = strongly disagree, 7 = strongly agree) to ensure high reliability. The complete list of measurement items is provided in Table 1.

Table 1.

Survey items

4.2. Data Collection

This study surveyed users experienced with Generative AI, collecting data from June 3 to June 5, 2024. A total of 210 office worker participants from the United States were recruited via MTurk, with strict criteria to ensure data reliability (Aguinis et al., 2021). Participants were required to have a 98% or higher approval rating and have completed over 500 tasks on MTurk. Only those meeting these standards were permitted to participate, and each received a $0.50 reward.

To further ensure data integrity, attention-check questions (e.g., "Please select 'Strongly disagree' for this question") were included. Respondents who failed these checks were excluded from the study. At the survey's conclusion, demographic information was gathered from all participants. Table 2 presents an overview of these demographics. After excluding 18 responses that did not pass the attention tests or provided incorrect answers to reverse-coded items, 192 valid responses were retained for analysis.

Table 2.

Demographic statistics of respondents (N=192)

5.1. Measurement validity test

In this study, Smart PLS 4.0 was utilized to conduct data analysis through Partial Least Squares Structural Equation Modelling (PLS-SEM). The choice of PLS-SEM was strategic, given its robustness in handling non-normal data, a prevalent challenge in survey-based research (Hair et al., 2019). Additionally, PLS-SEM is particularly advantageous for modeling intricate structures with smaller sample sizes (Hair et al., 2019)

To ensure the reliability of our analysis, we rigorously tested the validity of the measurement items, focusing on both convergent and discriminant validity. Following the guidelines of Fornell and Larcker (1981), convergent validity was evaluated based on three criteria. Composite Reliability (CR) and Cronbach's alpha values should exceed 0.7, and an Average Variance Extracted (AVE) should be greater than 0.5.

The analysis revealed that all factor loading values ranged from 0.772 to 0.963, meeting the required criteria. The Cronbach's alpha values for all items were robust, with the lowest value being 0.777. Additionally, the lowest AVE recorded was 0.769, well above the minimum threshold of 0.5. Therefore, the convergent validity of the measurement model was confirmed, as detailed in Table 3.

Table 3.

Loadings and Convergent validity

For assessing discriminant validity, we employed the Heterotrait-Monotrait Ratio (HTMT) method, as recommended by Hair et al. (2019). Discriminant validity is established when HTMT values are below 0.9. As illustrated in Table 4, all variables satisfied this stringent criterion, thereby confirming the discriminant validity of the measurement model.

Table 4.

Discriminant validity results (HTMT)

5.2. Hypotheses test

Next, we tested our hypotheses. Initially, we examined the R-square score, which stands at 0.429, indicating that our research model explains 42.9% of the variance in the intention to adopt Generative AI. We then applied the bootstrap method with 5000 samples for robust hypothesis testing. The results indicate that among the three enablers, knowledge support (β=0.411,p=0.019) and entertainment(β =0.400, p=0.026) have a significant impact on adopt intention. However, customization does not have a significant effect on adoption intention (p=0.513). Therefore, hypotheses H1 and H3 are supported while H2 is rejected.

Next, among the three inhibitors, only realistic threat has a significant impact on adoption intention (β=-0.187, p=0.019). Thus, H5 is supported. Contrary to our expectations, perceived risk (p=0.778), and identity threat (p=0.878), do not have a significant impact on adoption intention. Therefore, H4 and H6 are rejected. According to Cohen (1988), all adopted relationships showed a weak effect (f2 ≥ 0.02). Table 5 presents a summary of the results.

Table 5.

Hypotheses test results

6.1. Theoretical implication

First, it is significant as an initial exploration of how Generative AI is being used in workplaces. While the use of Generative AI in real world scenarios is on the rise, its application at work is mostly limited to a few areas like programming or entertainment. Previous studies have mainly focused on the potential downsides of Generative AI in educational settings, giving less attention to how it interacts with and is embraced by users in professional settings. This study fills this research gap by offering valuable insights for future studies.

Second, this study addresses a key issue in technology adoption research, which often concentrates solely on the advantages of new technologies. Models such as the TAM (Davis, 1989) and UTAUT (Williams et al., 2015) provide valuable perspectives on how the benefits of new technologies affect adoption decisions. However, these models consider only the facilitators of technology adoption, undermining the importance of inhibitors. To overcome this limitation, this study adopted the dual factor theory as a framework, examining both facilitators and inhibitors. The results indicate that the perception of a tangible threat plays a crucial role in discouraging users from embracing new technologies. This thorough investigation enhances our overall comprehension and lays a solid groundwork for future studies.

6.2. Practical implication

The study's findings suggest several practical implications for effectively encouraging the adoption of Generative AI in the workplace. Organizations should emphasize features related to knowledge support and entertainment. Specifically, promoting how Generative AI can aid in tasks like automating repetitive processes or providing instant information can enhance user acceptance. Additionally, highlighting the AI's capability to engage in interactive, entertaining communication can further drive its adoption. Companies should invest in showcasing these aspects through training and communication strategies.

The significant negative impact of realistic threat on adoption intentions indicates that employees might fear AI as a potential job disruptor. Organizations need to address these concerns proactively by clarifying how Generative AI can complement rather than replace human roles. Providing clear examples of AI's role in augmenting tasks rather than replacing them, and involving employees in the implementation process, can help mitigate these fears. Moreover, transparent communication about how AI will impact job roles and create new opportunities can alleviate concerns and foster a more positive reaction.

6.3. Limitations and Future Research

The study has pointed out some interesting discoveries, but there are certain limitations that suggest areas for future investigation. First, the sample was restricted to participants from the U.S. recruited through MTurk, which may not capture the full range of perspectives among global white-collar workers. To enhance the applicability of the findings, future studies should involve a more diverse and international sample. Moreover, this study considered white-collar workers from various industries and with differing levels of experience as a single, homogeneous group, which is a limitation. In addition, the skewed nature of the data regarding Generative AI tools and usage frequency is a limitation. Most participants used chat-based AI like ChatGPT, highlighting its current dominance in Generative AI. This suggests limited adoption of other AI types, such as image and audio generation in workplaces. Over half of respondents reported no AI usage, indicating many have yet to integrate these tools into their work routines. This imbalance may affect the generalizability of the findings across AI tools and usage patterns. Future research should use control or independent variables to explore differences across industries and sectors. A more balanced sample with a wider range of AI tools and use cases would provide clearer insights into Generative AI adoption.

Furthermore, while this research delved into factors like knowledge support, customization, entertainment value, perceived risk, realistic threat, and identity threat affecting Generative AI adoption, it overlooked other potential factors. To gain a more comprehensive understanding of adoption patterns, upcoming studies should explore additional drivers and barriers such as organizational culture, user experience aspects, and specific features of AI tools.

Lastly, this study focused on Generative AI adoption in a broad workplace setting. Future investigations could narrow down to specific industries or job roles to uncover the distinct challenges and opportunities presented by AI in diverse professional contexts. Examining industry-specific applications and their varied impacts on different job functions could yield more targeted insights and practical recommendations.