It’s Not the AI, It’s Us: Individual Concerns and the Challenge of AI Adoption in Organizations

Artificial intelligence (AI) has gained significant attention both in public discourse (e.g., Special Committee on AIDA, 2022; White House, 2022; OECD, 2019) and in academic research (e.g., Gamma & Magistretti, 2025; Querci et al., 2022). Many organizations continue to face challenges, particularly in the early stages of AI implementation (Atsmon et al., 2021). One potential reason for this slow adoption is the public’s worries and concerns about AI and other emerging technologies. People remain concerned about AI applications in areas such as facial recognition, driverless cars, and detecting false information on social media (Rainie et al., 2022). While AI-driven innovations promise significant progress, widespread anxiety about their societal impact persists (Schiavo et al., 2024); meanwhile, managers often lack a clear scientific grasp of the conditions required for AI to generate organizational value (Gamma & Magistretti, 2025).

Acceptance of AI is closely related to general acceptance of new technology. The literature on new technology acceptance highlights the social, economic, policy, and ethical challenges that arise with emerging technologies (Dwivedi et al., 2021). New technology acceptance is a complex process influenced by various motivators and inhibitors (Blut & Wang, 2020). Researchers have analyzed this process using various models considering both technology-related factors, such as perceived usefulness, ease of use, and risk (Davis et al., 1989; Hubert et al., 2019), and individual-related factors, including anxiety, uncertainty, hedonic motivation, and emotional responses (Tamilmani et al., 2021). Notably, worries and concerns about new technology may act as significant inhibitors to adoption in organizations (Blut & Wang, 2020). While these traits and feelings have been recognized, existing theories often treat worries and concerns as having only a weak or indirect impact on adoption decisions. Beyond AI as a specific new technology, its diffusion is also embedded within the broader digital transformation driven by information and communication technologies (ICT). The adoption of AI thus reflects not only technical change but also the evolution of organizational ICT capabilities and infrastructures that enable intelligent data use and automation (Chugh et al., 2025; Mariani & Dwivedi, 2024). Positioning AI within the ICT continuum underscores that AI is a critical but integral component of modern digital ecosystems in organizations.

Despite growing public and organizational attention to AI, limited empirical research has systematically examined how individuals differ in their concerns about emerging technologies. Understanding such variation is important because individual perceptions and anxieties shape broader public attitudes and, indirectly, the environment in which organizations adopt AI. Such perceptions may also influence how individuals within organizations respond to AI initiatives, affecting their openness, trust, and readiness for change. Therefore, the aim of this paper is to identify and classify patterns of individual worries and concerns about AI and related technologies, and to explore how these attitudinal clusters differ across technology domains.

Using data from over 10,000 U.S. adults collected through the Pew Research Center’s American Trends Panel (ATP) survey (Pew Research Center, 2021), this study examines worries and concerns about new technologies, particularly AI. We ask whether individuals worry uniformly about all technologies, or whether their worries vary depending on the specific technology in question. Identifying such differences, and grouping individuals accordingly, may help firms accelerate AI adoption and provide insights into how organizational members perceive AI in practice.

To analyze individuals’ opinions on new technologies, particularly AI, we used data from the Pew Research Center’s American Trends Panel (ATP) survey (Pew Research Center, 2021). The ATP is a nationally representative online panel of over 10,000 U.S. adults on current issues, with questions available in both English and Spanish (Rainie et al., 2022; Keeter, 2019). This study specifically uses data from Wave 99 of the ATP survey, conducted from November 1 to November 7, 2021, with 10,260 U.S. adults as respondents, including residents of Hawaii and Alaska (Rainie et al., 2022).

This dataset is particularly well suited for our study because it is one of the largest nationally representative surveys focusing on AI and emerging technologies, and it contains uniquely detailed attitudinal measures across multiple domains. Respondents came from diverse demographic backgrounds (e.g., age, gender, race/ethnicity, education, income, and region). Pew provides survey weights to align the panel with U.S. population benchmarks; however, in this study we analyzed the unweighted dataset. This choice reflects our focus on relative patterns across clusters of respondents rather than on nationally generalizable point estimates. Accordingly, the results should be interpreted as revealing attitudinal structures within the sample, while acknowledging that weighted distributions would more closely reflect the demographic profile of the U.S. adult population. Basic demographics (e.g., age, gender) closely aligned with U.S. Census benchmarks (United States Census Bureau, 2023), confirming sample reliability (Pew Research Center, 2021). Respondents were members of the general U.S. adult population, not sampled by occupational role or management level. While the survey does not provide organizational subgroups, the findings are nonetheless relevant for organizational research, as employees and managers emerge from the broader public and bring these predispositions into organizations. To identify latent attitudinal clusters (H1), we used two composite variables capturing respondents’ excitement or concern about (a) AI applications (POSNEGAI) and (b) potential human enhancements (POSNEGHE).

At the outset of the survey (Pew Research Center, 2021), all participants answered two multi-item questions: one asked how excited or concerned they would be if AI performed six specific types of work, while the other asked about potential new techniques that could change human abilities in six ways. For each item, respondents selected from five options ranging from “Very excited” to “Very concerned,” with nonresponses coded separately and excluded from analysis. These two question blocks offered a uniquely detailed set of attitudinal measures, each with six items and five-point response options, making them well suited for clustering analysis. Notably, while the first block focused directly on AI applications, the second addressed broader technological changes that may be facilitated by AI. Table 2 summarizes the two multi-item questions that served as clustering inputs for H1:

To test whether the clusters differed significantly in their overall attitudes (H2), we examined four additional variables measuring general orientations toward technology, science, and AI. These items extend beyond the specific clustering inputs (POSNEGAI and POSNEGHE) and provide a broader attitudinal context, allowing us to validate whether the clusters reflect meaningful differences in respondents’ general views. Two items assessed attitudes toward AI (CNCEXC and ALGFAIR), while two others measured general attitudes toward technology (TECH1) and science (SC1). Because TECH1 appeared only in Form 1 and SC1 only in Form 2, these variables also serve as a robustness check across subsamples. Table 3 below includes the questions of TECH1, SC1, CNCEXC and ALGFAIR, which are the four variables used to test H2, capturing broader general attitudes.

To assess whether cluster differences followed a consistent directional order across domains (H3), we analyzed six items from the Pew Research Center survey (2021) covering widely discussed AI-related technologies. These included three noninvasive applications (social media, facial recognition, and driverless vehicles) and three more invasive or embodied applications (brain-implanted chips, gene editing, and robotic exoskeletons). The survey was split into two forms, with approximately half of respondents answering the first set (noninvasive, Form 1) and the other half the second set (invasive, Form 2). This design enables us to test not only whether clusters differ in their overall orientations toward AI but also whether their relative positions remain consistent across contrasting domains. Table 4 below shows the six domain-specific variables analyzed to assess H3:

Note: Some variables (e.g., CNCEXC, BCHIP2) originally used the coding in an order similar to 1 = Good idea, 2 = Bad idea, 3 = Not sure. For comparability across items, we recoded them as 1 = Positive, 2 = Neutral, 3 = Negative. This ensured that lower scores consistently reflect more positive attitudes and higher scores more negative attitudes. All ANOVAs and post-hoc tests were conducted on the recoded variables. For all the variables in Table 3 and Table 4, responses that represented nonresponse codes (e.g., “99”) were also excluded.

As mentioned earlier, each of the two variables for clustering purposes (POSNEGAI and POSNEGHE) comprises six sub-questions (a–f). To facilitate clustering, we averaged these responses to create two composite variables, which were then standardized. We acknowledge that averaging the variables may reduce variance and mask item-level heterogeneity (Hair et al., 2019). However, in this study the averaged variables still produced distinct and interpretable clusters, suggesting that substantive group differences were preserved. Using these scaled variables, we applied k-means clustering to determine the optimal number of clusters. For our cluster analysis, we employed k-means clustering because it is computationally efficient for larger samples, produces non-overlapping clusters that are straightforward to interpret, and is widely applied in marketing and management research (Hair et al., 2019). The elbow method, a widely used approach for this purpose (Johnson & Wichern, 1992), indicated that three clusters best fit the data.

To assess whether the identified clusters differed significantly in their attitudes toward AI-related technologies, we employed a series of one-way analyses of variance (ANOVAs) (Hair et al., 2019). Each ANOVA tested mean differences across clusters for a given survey item. Where overall F-tests were significant, we conducted Tukey’s Honest Significant Difference (HSD) post-hoc comparisons to identify which cluster means differed from one another while controlling for familywise error. This allowed us to distinguish whether differences followed a linear pattern or whether reversals emerged. In addition to significance testing, we reported effect sizes using eta squared (η²), which quantify the proportion of variance explained by cluster membership. Effect sizes (η2) were calculated for each ANOVA and interpreted following benchmarks recommended by Hair et al. (2019), with values of approximately .01, .06, and .14 indicating small, medium, and large effects, respectively. These effect sizes provide information about the substantive strength of differences beyond statistical significance. This analytic sequence allowed us to (a) identify latent clusters (H1), (b) test their overall attitudinal differences (H2), and (c) assess whether these differences followed a consistent directional structure across domains (H3).

The dataset was collected in November 2021, prior to the widespread emergence of generative AI applications such as ChatGPT. Survey items focused on technologies that were prominent in public debate at the time, including facial recognition, driverless cars, brain chip implants, gene editing, and robotic exoskeletons. Accordingly, the results capture attitudes toward these technologies rather than generative AI specifically. This temporal scope should be borne in mind when interpreting the findings.

We conducted the analysis using R version 4.5.0. As mentioned earlier, we performed a cluster analysis based on POSNEGAI and POSNEGHE (Table 2) to examine whether different clusters exhibit distinct attitudes toward issues related to new technologies. Using the elbow method, the total within-cluster sum of squares declined sharply before stabilizing at three clusters, indicating that three clusters provided the optimal solution. The sample was divided into three clusters based on responses to POSNEGAI (focused on potential AI technologies) and POSNEGHE (focused on potential new technologies in general). Given this classification, differences in responses to these variables are expected.

The k-means analysis yielded three clusters of sizes 2,951, 4,336, and 2,639 respondents, respectively. Table 5 presents descriptive statistics for the two clustering variables (POSNEGAI and POSNEGHE) by cluster. The results show clear differentiation. These distinctions indicate that averaging did not obscure meaningful differences but produced interpretable clusters.

For interpretability, we labeled the clusters according to their dominant attitudinal patterns. Cluster 1 (n = 2951) scored highest on both measures of concern and was labeled the “Skeptics.” Cluster 3 (n = 2639) scored lowest on both sets of concern items and was therefore labeled the “Optimists,” as they appeared more excited about AI and related technologies. Cluster 2 (n = 4336) showed moderate scores and was labeled the “Cautious.” POSNEGAI captures attitudes toward AI-specific applications, while POSNEGHE reflects attitudes toward potential human enhancements, neither of which includes invasive technologies such as brain chips, gene editing, or exoskeletons. These results provide strong evidence for the existence of three distinct attitudinal clusters, consistent with H1.

To further validate the clustering analysis, we examined responses to general attitudes toward technology and science (TECH1 and SC1) and AI (CNCEXC and ALGFAIR) with ANOVA. The results are as follows:

The three clusters identified by responses to POSNEGAI and POSNEGHE also differ significantly in their responses to the four variables listed, even when divided across two questionnaire forms. While POSNEGAI and POSNEGHE do not address invasive technologies, our results show that these clusters maintain distinct response patterns even when invasive technologies are considered.

As expected, the clusters differed significantly on CNCEXC (F(2, 9923) = 2046, p < 2 × 10⁻¹⁶). As in Table Y, clusters also differed significantly on ALGFAIR (F(2, 9855) = 817.2, p < 2 × 10⁻¹⁶), TECH1 (F(2, 4984) = 318.4, p < 2 × 10⁻¹⁶), and SC1 (F(2, 4923) = 209.3, p < 2 × 10⁻¹⁶). On all variables, the Optimists scored lowest, the Skeptics highest, and the Cautious fell in between, mirroring the patterns observed for POSNEGAI and POSNEGHE. These results confirm that the clusters not only exist but also differ systematically in their general orientations toward AI, technology, and science, consistent with H2.

Also as mentioned earlier, we next ran ANOVAs on six domain-specific items, covering both noninvasive and invasive technologies (SMALG2, FACEREC2, DCARS2, BCHIP2, GENEV2, EXOV2): Social media, facial recognition, driverless vehicles, brain-implanted computer chips, gene editing, and robotic exoskeletons (Pew Research Center, 2021). The results are below:

Clusters differed significantly across all six measures: SMALG2 (F(2, 4979) = 152.9, p < 2 × 10⁻¹⁶), FACEREC2 (F(2, 4977) = 40.83, p < 2 × 10⁻¹⁶), DCARS2 (F(2, 4977) = 592.4, p < 2 × 10⁻¹⁶), BCHIP2 (F(2, 4910) = 452.8, p < 2 × 10⁻¹⁶), GENEV2 (F(2, 4917) = 374, p < 2 × 10⁻¹⁶), and EXOV2 (F(2, 4921) = 498.8, p < 2 × 10⁻¹⁶). Across all the variables above, Skeptics reported the highest concern levels, Cautious respondents fell in between, and Optimists consistently scored lowest, similar to the pattern observed in POSNEGAI and POSNEGHE. These findings indicate that the clusters’ relative positions remain stable across diverse and contrasting domains, consistent with H3.

Our research suggests that individuals can be meaningfully clustered by their attitudes toward new technology, and these attitudinal differences remain relatively stable across diverse AI-related domains. This consistency indicates that technology attitudes are not merely context-specific reactions but reflect deeper, enduring dispositions. The robustness of these patterns aligns with prior research highlighting the influence of individual characteristics on technology acceptance. For instance, studies have linked personality traits such as openness, conscientiousness, and agreeableness to technology adoption (Fuglsang, 2024; Stein et al., 2024; Park & Woo, 2022; Barnett et al., 2015). However, while earlier work often reported only weak correlations (Fuglsang, 2024; Park & Woo, 2022), our results suggest that individuals’ broader orientations toward new technology manifest as distinct and stable clusters. This implies that dispositional differences – potentially shaped by personality but not reducible to it – underlie how people respond to emerging technologies.

Consistent with H1, the cluster analysis revealed three distinct groups with differing attitudes toward AI. ANOVAs further confirmed significant between-group differences (H2). Moreover, in line with H3, these differences followed a consistent directional order across domains: Skeptics expressed the greatest concerns, Optimists the least, and Cautious respondents fell in between. This suggests that attitudinal clusters not only exist but also exhibit stability across diverse AI-related technologies, reinforcing the view that individuals’ underlying dispositions shape their responses more strongly than contextual variation.

This study provides population-level evidence that attitudinal heterogeneity toward AI is structured, stable, and cross-domain. Three clusters (Skeptics, Cautious, Optimists) differ consistently in concern across both applied/embodied AI domains and general orientations toward AI, technology, and science. In line with our hypotheses, the analysis identified three distinct clusters with differing attitudes toward AI (H1) and confirmed significant between-group differences across multiple variables (H2). Importantly, H3 was also supported: cluster differences followed a consistent directional order across domains, with Skeptics expressing the greatest concerns, Optimists the least, and Cautious respondents falling in between. This consistency underscores that individuals’ orientations toward AI remain stable across diverse contexts, reinforcing the role of underlying dispositions in shaping technology attitudes. The findings make theoretical advances and yield actionable managerial implications. The results highlight why organizational adoption hinges not only on technical performance but also on aligning governance, communication, and rollout strategies with personal dispositions. Practically, segmentation and tailored engagement may reduce resistance and accelerate value realization.

Limitations include the focus on concerns rather than motivators, the use of composite measures that may mask item-level nuances, and the pre-GenAI timing of the dataset. These limitations nonetheless point to meaningful avenues for future research using post-2023 data, cross-cultural samples, and designs that connect attitudinal segments with actual adoption behaviors. Overall, the results underscore a central insight: in organizational contexts, the decisive constraint is often not the capability of AI itself but the diversity of human dispositions toward it. Addressing this challenge requires not only strategic and governance-aligned interventions but also a deeper sense of empathy and understanding toward the individuals whose experiences ultimately shape the success of AI adoption.