The Four Artificial Intelligence Apprehension Scales: Apprehension Towards Personal AI, General AI, Institutional AI, and Large Language Models

The Four Artificial Intelligence Apprehension Scales: Apprehension Towards Personal AI, General AI, Institutional AI, and Large Language Models | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article The Four Artificial Intelligence Apprehension Scales: Apprehension Towards Personal AI, General AI, Institutional AI, and Large Language Models Ala Yankouskaya, Mohamed Basel Almourad, Magnus Liebherr, Mohammad Mominur Rahman, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9018418/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Recent research has developed measures that conceptualise human factors in relation to AI in relatively broad terms, such as general attitudes and positive, negative or adaptive use. However, more nuanced instruments are still needed to capture specific psychological dimensions, such as apprehension, which is the focus of this paper. Moreover, many existing measures conceptualise AI as a single, undifferentiated object of evaluation, rather than distinguishing between different AI systems, contexts of use, or functional domains. The present study adapted, developed and validated four distinct but structurally parallel instruments measuring apprehensions toward General AI, Large Language Models, Personal AI and Institutional AI. The instruments share identical structure and item content adapted to each type of AI. Apprehension is operationalised across three dimensions (Implied Malice, Undesirability, and Unpredictability). Data from a British sample of 559 adults (age range 18–45, M = 30.64, SD = 6.75, 50.3% males). For each scale, confirmatory factor analyses supported the three-dimensional structure, while internal consistency was strong at both total and subscale levels, and model-based indicators demonstrated well-defined latent constructs. The scales demonstrated good discriminant and convergent validity. These findings establish four distinct and psychometrically robust instruments suitable for research requiring measurement of apprehension toward artificial intelligence both at a general level and across specific classes of AI systems. Psychology Artificial Intelligence and Machine Learning Apprehension large language models personal AI general AI institutional AI Figures Figure 1 Figure 2 1. Introduction Attitudes towards AI may differ depending on the type of AI and the nature of the interaction involved. Direct end-user conversations with Siri or Alexa (an example of Personal AI) about specific events feel fundamentally different from being instructed to reroute to a less busy hospital based on an AI-powered surveillance and resource-management system (an example of Institutional AI). Both involve artificial intelligence, yet they evoke strikingly divergent emotional responses. The former might inspire curiosity or even playfulness, while the latter often generates profound unease about fairness, transparency, and control over consequential life outcomes. This experiential divergence poses a critical challenge for researchers seeking to understand public responses to AI. If individuals experience AI as a collection of distinct systems with different capabilities, purposes, and power dynamics, then treating AI apprehension as a monolithic construct risks obscuring the very psychological processes we seek to understand (Dang and Li, 2026 ; De Freitas et al., 2024 ). This variability in emotional responses to different classes of AI reflects not only individual differences in technological literacy and personal factors, but also the fundamental heterogeneity of AI systems themselves. Different AI technologies vary substantially in their capabilities, purposes, transparency, and proximity to users' personal lives. Yet despite growing recognition that emotional responses to AI represent a critical component of human-AI interaction, existing measurement approaches remain limited. Scientific evaluations increasingly criticise existing measurement instruments for often treating AI as a monolithic category, thereby neglecting the variance in application context, user relationship and risk profile (Montag et al., 2025 ). Additionally, while broad attitudinal measures capture general positivity or negativity toward AI, they often conflate affective, cognitive, and behavioral components, making it difficult to isolate the specific role of apprehension as an anticipatory emotional state characterized by concern, unease, and heightened sensitivity to potential negative outcomes (Schepman and Rodway, 2023 ). Furthermore, existing instruments for measuring technology-related apprehension were developed for general interactive technologies without specific consideration of AI systems' unique characteristics (Woźniak et al., 2021 ) The present study addresses these measurement gaps by examining perceived apprehension across four distinct AI contexts: personal AI, general AI, institutional AI, and Large Language Models. This multi-context approach is grounded in literature suggesting that apprehension toward AI is not uniform but context-dependent, shaped by the specific characteristics, perceived intentions, and power dynamics associated with different AI instantiations (Bialy et al., 2025 ; Grassini et al., 2025 ). By capturing emotional responses across qualitatively different forms of AI, the study aims to provide a more nuanced understanding of how creepiness varies as a function of both system characteristics and individual differences. To operationalize perceived creepiness in these contexts, an adapted version of the Perceived Creepiness of Technology Scale (Woźniak et al., 2021 ) was employed, preserving its theoretically grounded three-factor structure of implied malice, undesirability, and unpredictability while contextualizing items for AI-specific concerns. This work contributes to the literature on human-AI interaction by providing psychometrically validated instruments that are built on an established measure for technology-related apprehension (Woźniak et al., 2021 ), while being specifically tailored to AI systems across theoretically meaningful classes. We first establish the psychometric adequacy of each scale, including evidence for convergent and discriminant validity with theoretically relevant constructs. We then demonstrate that apprehension shows both shared variance across AI types and context-specific variation, supporting the decision to distinguish four separate AI domains. Finally, we examine demographic correlates of AI-related apprehension and discuss their implications for equitable AI design and deployment. We build on the Perceived Creepiness of Technology Scale (PCTS) proposed by Woźniak et al. ( 2021 ) to assess apprehension toward artificial intelligence. The original scale was adapted primarily at the level of wording to reflect a more formal focus on technology-related apprehension rather than creepiness per se. Moreover, whereas the original instrument targets perceptions of specific technological products, the present version is designed to capture attitudes toward broader classes of AI systems. This approach allows researchers and practitioners to assess AI-related unease using a theoretically grounded instrument that is sensitive to both the psychological foundations of technology-related discomfort and the distinctive characteristics of different AI domains. 2. Theoretical Background 2.1. Attitudes Toward AI and the Affective Dimension Attitudes toward technology have long been understood as multifaceted psychological constructs encompassing cognitive beliefs, affective responses, and behavioral intentions (Venkatesh et al., 2012). In the context of AI, attitudes reflect individuals' overall evaluations of AI systems, shaped by perceived benefits, risks, ethical concerns, and personal experiences (Glikson and Woolley, 2020 ). The tripartite model of attitudes has been widely applied to understand technology acceptance, with the affective dimension increasingly recognized as a critical driver of adoption and usage behaviors (Beaudry and Pinsonneault, 2010 ). However, AI systems differ from conventional technologies in ways that necessitate theoretical refinement. Unlike static tools or passive information systems, AI technologies exhibit characteristics traditionally associated with agency. They learn, adapt, make autonomous decisions, and generate novel outputs that exceed their creators' explicit programming (Eloundou et al., 2023 ; Sundar and Lee, 2022 ). These agentic qualities fundamentally alter the psychological dynamics of human-technology interaction, shifting users from a position of control to one of coordination or, in some cases, subordination (Hancock et al., 2011 ; Sun et al., 2025 ). This shift toward machine agency introduces distinct affective dimensions that extend beyond conventional technology acceptance frameworks. When individuals interact with AI, they must navigate not only instrumental concerns about utility and ease of use but also relational concerns about trust, transparency, and alignment of interests (Lee and See, 2004 ; Shin, 2021 ). The opacity of many AI systems, particularly deep learning models whose decision-making processes are neither transparent nor readily interpretable, exacerbates these concerns and creates conditions for apprehension. 2.2. Defining AI Apprehension Apprehension represents a specific affective component within the broader attitude structure toward AI. While general attitudes may range from highly positive to highly negative, apprehension specifically captures the negative affective pole, emphasizing emotional responses rooted in concern, unease, worry, and anticipatory anxiety about AI systems and their implications (Schiavo et al., 2024 ; Uğur and Dursun, 2025 ). This construct is theoretically distinct from several adjacent concepts. AI fear refers to intense emotional distress or anxiety about specific AI-related threats such as job displacement, loss of human autonomy, or existential risks (Schepman and Rodway, 2023 ). While apprehension may encompass elements of fear, it is broader in scope and less tied to discrete threat scenarios. AI skepticism involves cognitive doubt about AI capabilities or benefits, often accompanied by critical evaluation of AI hype or distrust of technology companies (Glikson and Woolley, 2020 ; Wang et al., 2025 ). Unlike apprehension, skepticism is primarily cognitive rather than affective and may not be accompanied by emotional distress. AI resistance encompasses behavioral opposition to AI adoption manifested through avoidance or active rejection (Dietvorst et al., 2015 ; Longoni et al., 2019 ), representing the conative component of attitudes rather than the affective dimension we target. Importantly, apprehension is not simply the inverse of acceptance or trust. An individual may simultaneously recognize the utility of AI systems while feeling uncertain or uneasy about their implications, reflecting the well-documented phenomenon of attitudinal ambivalence toward complex technologies (Cacioppo et al., 1997 ). This ambivalence is particularly pronounced in the AI domain, where many users report concurrent appreciation for AI benefits and concern about AI risks (Mick and Fournier, 1998 ). Our conceptualization of apprehension as a distinct affective construct allows researchers to capture this nuanced emotional landscape without conflating it with cognitive evaluations or behavioral intentions. 2.3. Psychological Dimensions of AI Apprehension Following Woźniak et al. ( 2021 ) conceptual model of creepiness, we apply the three-dimensional structure to the context of AI apprehension. We adopt this framework for two reasons. First, there is a lack of existing scales that explicitly use the term “apprehension” and focus their content on this specific affective state. Second, while “creepiness” serves as an informal designation for technology-related apprehension, the original scale employed colloquial language and phrasing suitable for evaluating novel consumer products e.g. smart glasses or gadget. For the present AI context, we required more formal terminology and item formulations appropriate for a broader range of AI systems. This structure is grounded in psychological theories of threat appraisal, social judgment, and uncertainty (Kleine et al., 2024 ; Metzger and Flanagin, 2015 ) and is further supported by empirical research on public concerns about AI (Zhang and Dafoe, 2019 ). Implied malice captures concerns about the ethical foundations and moral intentions underlying AI development and deployment. This dimension reflects individuals' uncertainty about whether AI creators prioritize human welfare, adhere to ethical principles, or design systems that align with societal values (Taddeo and Floridi, 2018 ). Research consistently demonstrates that public trust in AI is undermined by perceptions that technology companies prioritize profit over safety, lack transparency about AI capabilities and limitations, or fail to address algorithmic bias and discrimination (Yang and Sundar, 2025 ; Zhang and Dafoe, 2019 ). The implied malice dimension is theoretically grounded in research on institutional trust and perceived organizational ethics (Mayer et al., 1995 ), encompassing both benevolence concerns and integrity concerns about AI development processes. Undesirability reflects discomfort with AI presence in social and public contexts, particularly when AI systems are visible, ambient, or integrated into spaces traditionally governed by human norms and interpersonal dynamics (Lee, 2018 ). This dimension captures the sense that AI technologies disrupt social naturalness, create awkwardness or self-consciousness, and introduce elements that feel out of place in human-centered environments. The undesirability dimension is informed by theories of social presence, anthropomorphism (Epley et al., 2007 ), and norm violation (Bicchieri, 2005 ). When AI systems are introduced into public or interpersonal contexts, they may be perceived as infringing on privacy, altering social dynamics, or creating surveillance pressures that constrain authentic behavior (Zuboff, S., 2023 ). Additionally, the presence of AI in traditionally human domains can evoke uncanniness, a sense of unease arising from entities that approximate but do not fully replicate human characteristics (Gray and Wegner, 2012 ). Unpredictability encompasses uncertainty about AI behavior, goals, and outcomes, concerns rooted in the opacity, complexity, and autonomy of contemporary AI systems (Castelo et al., 2019 ; Shin, 2021 ). This dimension reflects individuals' doubts about whether AI systems operate as intended, whether their purposes are transparent and comprehensible, and whether their future behaviors can be reliably anticipated (Burrell and Fourcade, 2021 ; Lipton, 2018 ). The unpredictability dimension is theoretically grounded in research on uncertainty and control (Thompson, 1981 ) and algorithmic opacity (Pasquale, 2016 ). Many AI systems function as black boxes, generating outputs through processes that are opaque (Rudin, 2019 ). This opacity undermines users' ability to form accurate mental models of AI behavior, calibrate their trust appropriately, or predict system responses in novel situations (Bansal et al., 2021 ). Furthermore, the autonomy and adaptability of AI systems introduce behavioral unpredictability, the possibility that AI may behave in unexpected ways or pursue goals that deviate from user intentions. Even when AI systems function correctly from a technical perspective, their goals may be misaligned with human values or their deployment contexts may differ from training conditions, resulting in outcomes that users find surprising or harmful (Amodei et al., 2016 ; Gabriel, 2020 ). 2.4. Differential Apprehension Across AI Types A central theoretical premise of our work is that the degree of apprehension toward AI is not monolithic but varies as a function of the specific type of AI system under consideration. While the underlying construct of apprehension remains consistent across AI types, we argue that meaningfully different categories of AI evoke varying degrees of apprehension across these same dimensions due to differences in their purposes, user relationships, transparency, power dynamics, and potential for harm (Glikson and Woolley, 2020 ; Zhang and Dafoe, 2019 ). This differentiation reflects established findings in human-computer interaction research showing that user responses to technology depend critically on context, perceived agency, and the nature of the human-technology relationship (Sundar and Lee, 2022 ). General AI serves as an umbrella term for intelligent systems as a whole, ranging from simple smart gadgets for healthcare to complex recommendation systems (Triguero et al., 2024 ). Apprehension toward general AI may reflect diffuse concerns about AI's societal trajectory and long-term consequences, including technological unemployment, erosion of human autonomy, and gradual destabilization of critical societal systems (Bucknall and Dori-Hacohen, 2022 ). Large language models are characterized by their ability to generate human-like text and perform complex language tasks (Brown et al., 2020 ). While apprehension toward LLMs is pronounced along the unpredictability dimension due to non-deterministic outputs and hallucinations (Ji et al., 2023 ), it also manifests as undesirability in contexts like education or social discourse, where reliance on machine-generated content may be perceived as socially unacceptable. Furthermore, concerns regarding training data and corporate power concentration evoke the implied malice dimension (Bender et al., 2021 ; Sag, 2023 ). Personal AI systems are designed to be adaptive and integrated with individual users' habits (Luger and Sellen, 2016 ). Here, the undesirability dimension is specifically linked to social discomfort, such as the perceived "awkwardness" of interacting with assistants like Alexa or Siri in public or in front of others. While personalization introduces implied malice through privacy concerns and potential exploitation (Martin, 2018 ), it also triggers unpredictability regarding the opaque ways personal data might be processed (Gagrčin et al., 2026 ). Institutional AI encompasses systems deployed by organizations to monitor or evaluate individuals, such as face recognition at gates, traffic control, or hiring algorithms (Lee et al., 2015 ). These systems involve significant power asymmetries and often operate as "black boxes" (Burrell and Fourcade, 2021 ). We expect elevated apprehension across all dimensions: implied malice regarding organizational motives (Caton and Haas, 2024 ), unpredictability due to systemic opacity (Ananny and Crawford, 2018 ), and undesirability. The latter is particularly relevant as individuals are increasingly subjected to AI interactions in public or semi-public spaces, where the interaction may feel intrusive or socially uncomfortable (Bialy et al., 2025 ). These four AI types capture theoretically meaningful variation in AI characteristics. General AI provides an umbrella; LLMs represent a cognitively focused instantiation; personal AI captures intimacy and data vulnerability concerns; and institutional AI reflects power-asymmetric, high-stakes contexts. By measuring apprehension across these contexts, we can examine both common psychological foundations and context-specific patterns of AI-related concern. 2.5. Measuring Technology-Related Apprehension Research on negative affective responses to technology has produced several validated measurement instruments, each capturing distinct but related constructs. The Perceived Creepiness of Technology Scale (Woźniak et al., 2021 ) provides a validated instrument for assessing initial feelings of unease. However, it was primarily developed for interactive products and wearable gadgets that augment the user's body or daily environment, such as smart glasses and fitness trackers, rather than the autonomous decision-making logic of AI. The General Attitudes towards Artificial Intelligence Scale (Schepman and Rodway, 2023 ) measures broad evaluative stances. As a general attitude scale, it does not differentiate between specific AI applications, which may limit its sensitivity to context-specific concerns. Similarly, the Artificial Intelligence Anxiety Scale (Wang and Wang, 2022 ) focuses on anxiety-related affect, such as job-related threats and the learning curve of using AI, but remains at a monolithic "AI" level. The Fear of Artificial Intelligence Scale (Corradi et al., 2025 ) advances the literature by operationalizing fear via a broad set of content domains, including job displacement and loss of human autonomy. While this scale provides a comprehensive measure of fear as a holistic emotion, our work departs from this by treating apprehension as a multi-dimensional construct. Unlike previous unidimensional measures, we deliver a granular analysis of the components (malice, undesirability, and unpredictability) allowing us to examine how the strength and degree of these dimensions fluctuate across different AI categories. The gap between existing instruments and the need for a nuanced measurement of AI apprehension is substantial. AI technologies exhibit autonomy and learning capabilities that distinguish them from conventional interactive gadgets. These properties suggest that while the fundamental psychological triggers identified by Woźniak et al. ( 2021 ) remain relevant, their strength and expression require contextualization for AI. A conversational assistant, an institutional hiring algorithm, and a personalized health recommender do not just evoke fear, but distinct intensities of the identified apprehension components. 2.6. The Present Study The present study adapted the items from the Perceived Creepiness of Technology Scale (Woźniak et al., 2021 ) to develop and validate four parallel scales measuring apprehension toward artificial intelligence across distinct AI types: general AI, large language models, personal AI, and institutional AI. The items were rephrased to refer specifically to apprehension toward AI, while preserving the underlying meaning of the original scale. For institutional AI in particular, wording was adapted to emphasise exposure to AI systems implemented by organisations. Where possible, we used parallel items across all AI types, making small wording adjustments only when needed for clarity and content. 3. Methods 3.1. Participants Participants were recruited via the Prolific, a worldwide online platform for research which applies quality measures (www.prolific.com) and completed the survey on SurveyMonkey (www.surveymonkey.com). Data were collected between late September and October 2025. Eligibility criteria required participants to be aged 18-45 years, familiar with general, LMMs, personal AI and institutional AI, born in the UK or holding UK citizenship, self-identifying as British in terms of norms and culture, and speaking English as their native language. Participation was anonymous, and all respondents provided digital informed consent before accessing the questionnaires. Participants who completed the survey and passed attention checks received monetary compensation. A total of 589 participants completed the survey. Following exclusion of those who failed attention check or providing monotonous responses, 559 cases contained complete data and were retained for analysis. Simulation-based evidence for small confirmatory factor models shows that when a measurement model comprises three correlated factors defined by a small number of indicators, samples of at least N = 460 are required to achieve acceptable power, low parameter bias, and proper solutions when standardised factor loadings are in the moderate range (= .50) (Wolf et al., 2013). Expressed on an item basis, this corresponds to ≥ 57.5 participants per item for an 8-item instrument, providing a concrete numerical benchmark for sample size adequacy in short multifactor confirmatory factor analysis (CFA) models (Wolf et al., 2013). The achieved participant-to-item ratio in the present study (N = 559) was 69.9:1, exceeding this empirically derived requirement. The final sample had a mean age of 30.64 years (SD = 6.75), 50.3% males, with ages ranging from 19 to 42 years. 3.2. Procedure and materials 3.2.1. Scale development Our proposed apprehension scales were developed by adapting the item content of the Perceived Creepiness of Technology Scale (PCTS) (Woźniak et al., 2021) while retaining its three factors, Implied Malice, Undesirability, and Unpredictability. In the original measure, items are framed as direct evaluations of a certain AI products (for example, that its design is unethical, or would feel uneasy to use in public, or that its purpose is unclear). In the present scale, an introduction sentence was added that requires respondents to indicate how much more would be needed for them to be able to endorse the corresponding positive state (for example, that those behind the products take morality seriously, that interacting with the products in public would feel normal and not uneasy, and that the products work as described and have clear purposes and goals) (Table 1). Adding the introduction sentence has an advantage over direct evaluative wording because it targets a different part of the judgement process: the perception that what is currently in place exists but may not yet be sufficient. This reflects apprehension that is often expressed as caution under uncertainty, not as a fixed negative judgement. People can withhold full confidence or comfort without endorsing strong negative statements. For example, they may not claim that designers are immoral but still feel that more is needed before morality can be taken for granted. An “I think further conditions are necessary for me” format captures this intermediate position directly by measuring the extent of the perceived shortfall on each criterion, without requiring respondents to commit to strongly condemnatory language. This format also allows clearer interpretation of what respondents mean. Low agreement with a positive statement, or high agreement with a negative one, can reflect very different positions, such as mild hesitation, strong concern, lack of information, distrust of those deploying the system, or general scepticism. When responses are framed as direct evaluations, these positions are difficult to distinguish. By asking how much more is needed, the scale requires respondents to indicate a degree. It therefore distinguishes between those who believe only minor adjustments are necessary and those who believe substantial changes are required, even if both might select similar options on a standard agree-disagree item. For each scale, participants received the following instruction: “Think of [the AI class of interest: General AI, LLMs, Personal AI] products in general, the data they collect and use, your interaction with them through text, voice, image, or video, and the actions they take or the recommendations they make.”. For Institutional AI, the instruction was “Think of Institutional AI products in general, the data they collect and use, your exposure to them in public or private contexts, and the actions they take or the recommendations they make”. Participants then responded to eight items introduced by the common stem “I think further conditions are necessary for me…” using an 11-point scale (0 = “Nothing more needed”, 10 = “A great deal more needed”). The instruction text, item stem, and response scale were identical across all four versions of the scale, with only the referenced AI type varied. To ensure conceptual clarity, respondents were provided with a brief description and examples of each AI category before completing the scale. In the case of Personal AI, for example, this included products such as fitness trackers, smartwatches, health-monitoring devices, digital companions, and personal assistant applications. The final version of the scale (see Appendix 1-4) specifies that administration should be preceded by a short clarification of the relevant AI category, adapted as appropriate to the target population. The scale items are provided in Table 1. Table 1 PTCS items adapted to reflect each type of AI (General AI, Personal AI, Institutional AI, and LLM) apprehension Item number Subscale Item abbreviation General AI / LLMs/ Personal AI (X ) Institutional AI I think further conditions are necessary for me… 1 Implied malice MOR …to feel that the people who make X products take morality seriously. …to feel that the people who design and put in place Institutional AI products take morality seriously. 2 ETH …to feel that the way X products are designed and run is ethical. …to feel that the way Institutional AI products are implemented and operated is ethical. 3 Undesirability GOA …to feel normal and less odd when interacting with X products in public. …to feel normal and less odd in public spaces where Institutional AI products are installed. 4 SUR …to interact with X products in public without feeling uneasy. …to experience Institutional AI products in public without feeling uneasy. 5 EXP …to feel that interacting with X products in public is natural rather than awkward. …to feel that the presence of Institutional AI products in society is natural rather than disturbing. 6 Unpredictability NOR …to feel that [X] products work the way they are described and expected. …to feel that Institutional AI products work the way they are described and expected. 7 UNE …to feel sure about the purpose of use of X products, both for me and for the businesses behind them. …to feel sure about the purpose of use of Institutional AI products, both for me and for the institutions behind them. 8 NAT …to feel that X products have clear goals, both for me and for the businesses behind them. …to feel that Institutional AI products have clear goals, both for me and for the institutions behind them. The original PCTS scale employed a 7-point Likert response format; in the present study, items were administered on an 11-point (0-10) scale to capture greater variation in participants’ judgements. Evidence from rating-scale research indicates that broader response ranges provide increased measurement sensitivity compared with shorter Likert-type scales and are often preferred by respondents, as they allow more fine-grained evaluations without increasing response burden (Jebb et al., 2021; Revilla et al., 2014). In addition, empirical comparisons of scales with different numbers of response categories have shown that respondents tend to prefer scales with a larger number of points, particularly formats around 10 response options (Preston and Colman, 2000). 3.2.2. Survey design and procedure A cross-sectional survey design was employed. The survey comprised a demographic block, four AI apprehension blocks (General AI, LLMs, Personal AI, and Institutional AI), followed by blocks assessing attitudes towards AI and personality traits (Big Five). At the start of each AI apprehension block, respondents were given a brief description of the relevant AI type for that block. Attention-check items were embedded in the survey, each instructing respondents to select a specific response option. Cases with failed attention checks were excluded from the analyses. The study was accessed via a link distributed through the Prolific platform and was completed using SurveyMonkey. Before starting the survey, informed consent was obtained, and respondents were informed that participation was voluntary and that they could withdraw at any time without penalty. 3.2.3. Measures 3.2.2.1. Demographics. We asked participants about their age, gender, and education level. In addition, we assessed their religious beliefs to examine whether apprehension toward different types of artificial intelligence varies as a function of religiosity. Religion is consistently linked to differences in moral attitudes, perceptions of control, and responses to technological change (Alexander, 2020; Hommel, 2023; McPhetres and Zuckerman, 2018). Including this variable therefore allows us to test whether the scales are associated with individual differences that extend beyond basic demographic characteristics. 3.2.2.2. AI Apprehension scales . Each scale (General AI, Personal AI, LLM, and Institutional AI) consisted of eight items and followed the original factor structure. The scale distinguishes three dimensions of apprehension: Implied Malice (2 items assessing concerns about the morality and ethical orientation of those who develop or deploy AI), Undesirability (3 items assessing discomfort with AI in public or social contexts), and Unpredictability (3 items assessing confidence in AI behaviour and clarity of its goals) (Table 1). Items were rated on an 11-point scale ranging from 0 ( Nothing more needed ) to 10 ( A great deal more needed ). 3.2.2.3. Big-Five Inventory (BF-10) . Personality traits were measured using the 10-item Big Five Inventory (Rammstedt and John, 2007), which assesses the five personality dimensions: Openness (associated with intellectual curiosity and aesthetic appreciation), Conscientiousness (associated with dutifulness and adherence to social norms), Extraversion (associated with self-confidence), Agreeableness (associated with harmonious relations with others) and Neuroticism (associated with high levels of negative emotions such as anxiety). Each trait is measured using two items, one positively worded and one negatively worded, rated on a five-point Likert scale from 1 ( strongly disagree ) to 5 ( strongly agree ). The BFI-10 has been validated in large-scale survey research and provides an efficient assessment of personality traits while minimising respondent burden (Costa and McCrae, 1992). 3.2.2.4. Attitudes toward artificial intelligence (ATAI). Attitude Towards Artificial Intelligence (ATAI) scale, a brief five-item measure with a two-factor structure comprising Acceptance (2 items) and Fear (3 items) (Sindermann et al., 2021). Items were rated on the 11-point response format ranging from 0 (“strongly disagree”) to 10 (“strongly agree”). AI Acceptance is operationalised using the items “I trust artificial intelligence” and “Artificial intelligence will benefit humankind”, whereas Fear includes “I fear artificial intelligence”, “Artificial intelligence will destroy humankind”, and “Artificial intelligence will cause many job losses”. Subscale scores are computed by summing scores on the relevant items, with higher scores indicating greater acceptance of AI or greater fear of AI, respectively. In this study, the ATAI scale was administered with respect to four AI types: General artificial intelligence, LLMs, Personal AI, and Institutional AI. For each administration, the AI type was specified in the item stem (e.g. attitudes toward LLMs), while the item content and response format remained unchanged. Acceptance and Fear subscale scores were calculated independently for each AI type. Internal consistency was acceptable across AI types. Cronbach’s α for the Acceptance subscale ranged from α = .81 to .84, and for the Fear subscale from α = .73 to .76, across the four AI types. 3.3 Data analysis . 3.3.1. Monotonicity analysis Monotonicity analysis was performed prior data analysis and was assessed by examining whether conditional item means increased as a function of the rest score. Adjacent violations were defined as decreases in the conditional mean between successive rest-score groups. For each item, the total number of adjacent violations and the maximum observed adjacent drop in conditional means were recorded to quantify the extent and magnitude of local non-monotonicity. Across all scales, items showed positive monotonic associations with their rest scores, with Kendall’s τ ranging from 0.44 to 0.69 and Spearman’s ρ from 0.57 to 0.83. Adjacent monotonicity violations were observed for all scales, reflecting the fine-grained rest-score partitioning, but maximum adjacent drops were limited in magnitude (1-3 points on a 0-10 scale). Institutional and Personal AI scales showed the strongest monotonic structure, followed by the LLM and General AI scales. Overall, departures from monotonicity were not systematic, supporting the ordinal coherence of all four scales (see Supplementary Material 1 for complete report). 3.3.2. Descriptive statistics For each of the four AI apprehension scales item-level descriptive statistics were computed, including means and standard deviations. To examine response distributions, frequency tables were generated for each response option (0-10), and item distributions were inspected using histograms with unit-width bins covering the full-scale range (see Supplementary Material 2). 3.3.3. Confirmatory factor analysis (CFA) Confirmatory factor analyses (CFA) were conducted separately for the four AI apprehension scales to test a hypothesised three-factor structure. For each scale, items were specified to load on three correlated latent factors reflecting Implied Malice, Undesirability, and Unpredictability. Models were estimated using maximum likelihood with robust standard errors (MLR) to account for potential non-normality in item responses. Model fit was evaluated using multiple indices, including the comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation (RMSEA) with 90% confidence intervals, and the standardised root mean square residual (SRMR). In line with current recommendations, model evaluation was based on the joint consideration of comparative and residual-based indices rather than reliance on any single fit statistic (Kenny et al., 2015). 3.3.4. Reliability, construct and discriminant validity Reliability was evaluated separately for each factor within each AI type (General AI, LLMs, Personal AI, and Institutional AI), using a combination of classical internal consistency indices and model-based reliability estimates derived from the confirmatory factor analysis. For all factors, Cronbach’s alpha (α) was computed as a conventional index of internal consistency. Given the short length of the scales, α was calculated from the covariance matrix rather than item frequency tables to avoid distortions associated with discrete response distributions. In addition, the mean inter-item correlation was reported for all factors, as this index is particularly informative for short scales and provides a direct indication of item homogeneity. For the Implied Malice factor, which comprised two items in each AI context, reliability was primarily evaluated using the Spearman-Brown coefficient, which is the recommended reliability estimate for two-item measures. To complement classical reliability indices, model-based reliability was assessed using CFA-derived estimates. Composite reliability (CR) was computed from standardised factor loadings and residual variances using the compRelSEM() function in semTools to estimate the proportion of variance in the observed indicators attributable to the latent construct. In addition, average variance extracted (AVE) was calculated to quantify the proportion of indicator variance captured by the latent factor relative to measurement error; and interpreted as an index of convergent validity. Finally, construct replicability was evaluated using H (MaxR(H)), which reflects the degree to which a latent variable is defined by its indicators. Higher values indicate stronger latent definition and greater stability of the construct. Discriminant validity of the three latent factors was evaluated separately for each AI type. First, the Fornell-Larcker criterion was applied by comparing the square root of AVE for each factor with its correlations with other factors. Discriminant validity was considered supported when the square root of AVE exceeded all corresponding inter-factor correlations. Second, competing CFA models were also estimated to provide a direct test of discriminability. For each AI type, the hypothesised three-factor model was compared against a one-factor model and two-factor models. Model comparisons were conducted using scaled chi-square difference tests appropriate for the robust maximum likelihood estimator (MLR). 3.3.5. Shared variance between apprehension toward four AI types This analysis was conducted to determine the extent to which apprehension across different AI types reflects a common underlying tendency as opposed to AI type variation. Overlap between AI types was quantified using total apprehension scores computed separately for General AI, LLMs, Personal AI, and Institutional AI, with scores treated as continuous indicators of overall apprehension within each context. Pairwise associations between total scores were estimated using Pearson correlations, with Spearman rank correlations calculated as a robustness check for potential departures from normality. Overlap was operationalised as shared variance, defined as the squared Pearson correlation coefficient ( r² ) and expressed as a percentage, representing the proportion of variance in apprehension that is common to each pair of AI types. Uncertainty around these estimates was quantified using 95% confidence intervals derived from Fisher’s z transformation based on pairwise complete observations. In addition, intraclass correlation coefficients (ICC[A,1]) (Liljequist et al., 2019) were computed to assess the degree of absolute agreement in overall apprehension levels across AI types. In addition, we performed the same analysis between the same factors in each scale. 3.3.6. Convergent validity The aim of the convergent validity analyses was to establish whether apprehension toward artificial intelligence shows systematic associations with conceptually related evaluative and dispositional constructs. Convergent validity was assessed using three theoretically grounded indicators: ATAI fear of AI, ATAI acceptance of AI, and the personality trait of openness to experience. Fear of AI represents an affective response characterised by anxiety, perceived threat, and concern regarding the societal, occupational, and personal implications of AI. Empirical research consistently demonstrates that fear of AI is associated with heightened negative affect, risk perception, and avoidance-oriented evaluations of AI technologies (Schepman and Rodway, 2023; Yang and Sundar, 2025). Because AI apprehension similarly reflects a negatively valenced emotional orientation towards AI, grounded in concern and unease rather than neutral uncertainty, a positive association between apprehension and AI-related fear is theoretically expected. Accordingly, we hypothesised that AI apprehension would be positively associated with fear of artificial intelligence across all AI types ( H1 ). Acceptance of AI captures a general evaluative stance reflecting perceived usefulness, desirability, and willingness to engage with AI systems. Within attitude and technology acceptance frameworks, acceptance is consistently positioned in opposition to threat-focused responses, with higher acceptance associated with lower perceived risk, reduced concern, and diminished resistance to AI adoption (Glikson and Woolley, 2020; Schepman and Rodway, 2023). Given that AI apprehension reflects elevated concern and negative evaluation, it is conceptually incompatible with acceptance-oriented attitudes. We therefore hypothesised that AI apprehension would be negatively associated with AI acceptance across all four AI types ( H2 ). Openness to experience was included as a dispositional indicator relevant to convergent validity. Openness reflects cognitive flexibility, curiosity, and tolerance for novelty, and has been shown to predict more favourable evaluations of emerging technologies, including AI (Stein et al., 2024). Opened individuals are more likely to approach technological innovation with interest, suggesting lower apprehension in response to AI (Grassini et al., 2025; Sindermann et al., 2021). Although openness is not an AI-specific construct, its established association with receptivity to novel systems (Svendsen et al., 2013) provides a theoretical basis for a negative association with AI apprehension. We therefore expected higher openness to be associated with lower AI apprehension across AI types ( H3 ). For each AI type, two complementary analyses were employed. First, zero-order associations were estimated using Pearson product-moment correlations, with Spearman rank-order correlations computed in parallel as a robustness check against potential non-normality and the influence of outliers. Pearson correlations were used as the primary metric, and 95% confidence intervals were derived using Fisher’s z transformation. Second, for each AI type, apprehension scores were regressed on ATAI acceptance and ATAI fear using ordinary least squares estimation with HC3. Openness to experience was included alongside AI acceptance and fear to examine whether convergent associations remained stable when accounting for broader dispositional openness to novelty. 3.3.8. Divergent Validity The purpose of the divergent validity analysis was to demonstrate that AI apprehension is empirically distinct from broad personality dispositions and therefore cannot be reduced to general individual differences unrelated to threat-based responses to AI. For this analysis, extraversion and conscientiousness were selected within the Five-Factor Model as they are not theoretically linked to threat appraisal, anxiety, or negative affect toward external systems, including technology. Extraversion reflects sociability, assertiveness, and positive emotionality in interpersonal contexts (Costa and McCrae, 1992). Empirical research on technology attitudes consistently shows that extraversion is, at most, weakly and inconsistently related to technology use and apprehension (McElroy et al., 2007; Svendsen et al., 2013). Conscientiousness reflects self-discipline and goal-directed control (Costa & McCrae, 1992). While conscientiousness predicts rule adherence, task persistence, and reliability, it is not associated with emotional threat responses or evaluations (Kotov et al., 2010). Conscientiousness also shows inconsistent or negligible associations with evaluative attitudes, and when effects are observed they typically relate to compliance or usage behaviour rather than fear or concern (Devaraj et al., 2008). Demonstrating weak or null associations between AI apprehension and extraversion and conscientiousness will support the hypothesis that our apprehension scales are not artefacts of general personality structure ( H4 ). Divergent validity was assessed using the same statistical framework as convergent validity to ensure analytic symmetry. 3.3.9. Network analysis Network models were estimated separately for each AI type (General AI, LLMs, Personal AI, Institutional AI) to examine the structure of associations among the eight apprehension items within each AI type. Networks were specified as regularised partial correlation networks, such that each edge represents the association between two items after conditioning on all remaining items. Networks were estimated using the graphical LASSO with Extended Bayesian Information Criterion (EBIC) model selection, with the tuning parameter fixed at γ = 0.50 to balance sparsity and sensitivity. For each AI type, the resulting weighted adjacency matrix of partial correlations was extracted. Edge accuracy was evaluated using non-parametric bootstrapping (1,000 resamples), and network robustness was assessed via case-dropping bootstrap correlation stability analysis of node strength (1,000 resamples). Additional centrality indices (closeness and betweenness) were also computed. 3.3.10. Additional analysis Demographic correlates of the four AI apprehension scales were examined to provide practitioners and researchers with preliminary estimates of how AI apprehension may vary across gender, age, education, and religious beliefs. Identifying such patterns may offer practical insight into whether certain population groups report systematically higher or lower apprehension, which may inform policy considerations, and future applied research. Separate regression models were estimated for each scale, and detailed results are reported in Supplementary Material 4. 4. Results 4.1. Descriptive statistics of AI Apprehension scales Descriptive statistics for the four AI apprehension scales are summarised in Table 3 . Table 3 Descriptive statistics for AI apprehension items across four scales Factors Item abbreviation General AI LLM AI Personal AI Institutional AI Mean (SD) Mean (SD) Mean (SD) Mean (SD) Implied Malice MOR 6.499 (2.808) 5.737 (2.745) 5.878 (2.809) 6.499 (2.809) ETH 6.862 (2.745) 6.288 (2.559) 6.492 (2.559) 6.352 (2.559) Undesirability GOA 6.200 (2.967) 5.906 (2.945) 6.317 (2.745) 5.907 (2.745) SUR 6.159 (2.816) 5.832 (2.968) 6.278 (2.817) 5.879 (2.817) EXP 5.900 (2.795) 5.598 (2.745) 6.129 (2.745) 5.742 (2.745) Unpredictability NOR 5.055 (2.745) 4.995 (2.983) 5.602 (2.982) 5.004 (2.983) UNE 5.123 (2.817) 5.042 (2.744) 5.745 (2.744) 5.061 (2.744) NAT 5.057 (2.746) 4.979 (2.817) 5.714 (2.817) 4.830 (2.817) 4.2. Confirmatory factor analysis of AI Apprehension scales Confirmatory factor analyses were conducted to test the hypothesised three-factor structure across the four AI apprehension scales. For the General AI scale, the three-factor model showed very good fit to the data (CFI = 0.99; TLI = 0.99; SRMR = 0.03). For the LLMs scale, model fit was also good, with high comparative fit indices (CFI = 0.99; TLI = 0.98) and low residual misfit (SRMR = 0.04). The Personal AI scale demonstrated acceptable to good fit, with strong comparative fit indices (CFI = 0.98; TLI = 0.97) and low residual error (SRMR = 0.04). Similarly, the Institutional AI scale showed good fit, with CFI and TLI values above conventional cut-offs (CFI = 0.98; TLI = 0.97) and a low SRMR (0.03). Across all four scales, incremental and residual-based indices therefore consistently supported the adequacy of the hypothesised three-factor structures. RMSEA was 0.057 (90% CI [0.038, 0.076]) for the General AI scale, 0.085 (90% CI [0.067, 0.103]) for the LLMs scale, 0.109 (90% CI [0.092, 0.127]) for the Personal AI scale, and 0.093 (90% CI [0.076, 0.111]) for the Institutional AI scale. While the latter three values exceed conventional close-fit thresholds, this pattern is consistent with documented sensitivity of RMSEA to model degrees of freedom and parsimony in relatively small-df CFA models (Hayduk and Glaser, 2000 ; Kenny et al., 2015 ; Marsh et al., 2004 ). Importantly, the elevated RMSEA values were not accompanied by deterioration in CFI, TLI, or SRMR. In line with recommendations against rigid reliance on a single fit index (Lai and Green, 2016 ; Montoya and Edwards, 2021 ) model evaluation was therefore based on the overall convergence of fit evidence rather than RMSEA in isolation. Based on this our results indicate that the tested three-factor structure provides an adequate and substantively meaningful representation of apprehension toward General, LLM, Personal, and Institutional AI. Across all four AI types, items loaded strongly on their hypothesised latent factors. Standardised loadings were uniformly high, ranging from 0.78 to 0.98, with the strongest loadings consistently observed for Undesirability items, followed by Implied Malice, and slightly lower, but still substantial, loadings for Unpredictability items (Table 4 ). The similarity of loading patterns across contexts indicates a stable and coherent three-factor structure that generalises across different types of AI. Table 4 Standardised factor loadings (λ) for the three-factor CFA across AI contexts Factor Item abbreviation General AI LLM Personal AI Institutional AI Implied Malice MOR 0.86 0.893 0.904 0.924 ETH 0.885 0.901 0.936 0.962 Undesirability NOR 0.913 0.943 0.963 0.907 UNE 0.956 0.966 0.975 0.966 NAT 0.888 0.957 0.945 0.907 Unpredictability GOA 0.777 0.776 0.831 0.856 SUR 0.909 0.956 0.96 0.958 GOA 0.881 0.931 0.953 0.949 4.3. Shared variance between AI apprehension across AI types Shared-variance analyses indicated partial overlap between total score apprehension across AI types. The greatest overlap was observed between General AI and LLM apprehension, which shared 67.9% of variance, followed by LLM and Personal AI apprehension (47.4%) and General and Personal AI apprehension (40.7%). In contrast, overlap Institutional AI showed consistently lower overlap with other scales, with shared variance ranging from 27.9% to 30.4%, indicating that apprehension toward Institutional AI is less strongly aligned with apprehension for other AI types. Agreement across AI types was further evaluated using a two-way random-effects intraclass correlation coefficient for absolute agreement. The ICC indicated moderate agreement across the four total apprehension scores (ICC[A,1] = .59, 95% CI [.53, .64], F (558, 307) = 7.41, p < .001), suggesting that individuals’ relative levels of apprehension are partly consistent across AI types, while also showing AI type-specific variation. We also examined shared variance for each factor (Implied Malice, Undesirability, and Unpredictability) across the four AI types. This analysis showed that a substantial proportion of variance is not shared across AI types ranging from 6.2% to 62.1% (Fig. 1 ). 4.4. Reliability, construct and discriminant validity of AI Apprehension scales Across all four AI types, the three apprehension factors demonstrated consistently high internal consistency, model-based reliability, and construct quality. For Implied Malice, which comprised two items in each scale, inter-item correlations were strong (r = 0.76–0.89), with Spearman-Brown coefficients ranging from 0.86 to 0.94 and Cronbach’s α values from 0.86 to 0.94, indicating excellent reliability for a short two-item construct. Composite reliability estimates were identical to α values, reflecting the equivalence of these indices in two-indicator models. For the three-item Undesirability and Unpredictability factors, Cronbach’s α values ranged from 0.89 to 0.97, with mean inter-item correlations between 0.73 and 0.92, indicating strong internal coherence without evidence of attenuation due to scale brevity. Composite reliability was uniformly high across AI contexts (CR = 0.90–0.97), closely aligning with α and supporting the stability of the latent constructs. Average variance extracted exceeded 0.73 for all factors, indicating that a substantial proportion of item variance was captured by the intended latent variables. Construct replicability, indexed by H (MaxR(H)), was moderate to high across all factors and contexts (H = 0.56–0.81). Replicability was consistently highest for Undesirability, particularly for the LLM and Personal AI (H = 0.80), indicating strong latent definition. Unpredictability showed moderate to good replicability across contexts (H = 0.60–0.70), while Implied Malice demonstrated moderate replicability (H = 0.56–0.68), consistent with its shorter two-item composition (see Supplementary Material 3 for detail). Discriminant validity of the three latent factors was evaluated using multiple complementary criteria. Inter-factor correlations remained well below conventional thresholds for construct redundancy across all AI types (all r < .82). The Fornell-Larcker criterion was satisfied for all factor pairs, with the square root of AVE for each factor exceeding its correlations with other factors. Crucially, three-factor CFA models fit the data significantly better than one-factor and all two-factor alternative models across AI types (all scaled χ² difference tests p < .001), providing strong evidence that the three factors represent empirically distinct dimensions rather than a single undifferentiated construct. These findings indicate that all three factors show excellent reliability, satisfactory construct validity and excellent discriminant validity across General, LLM, Personal, and Institutional AI apprehension, supporting their use in latent variable modelling and comparative analyses. 4.5. Convergent and divergent validity of AI Apprehension scales Convergent validity results are summarised in Table 5 . Across all four AI types, ATAI acceptance and ATAI fear showed statistically significant associations with AI apprehension in the expected directions, with acceptance negatively related and fear positively related to apprehension under HC3-robust estimation. Model-level robust Wald tests indicated that the convergent predictors jointly explained significant variance in apprehension for every AI type. For the acceptance-fear models (M1), Wald F statistics ranged from 36.99 (General AI) to 67.86 (Institutional AI), all p < .001. When openness was added (M2), joint effects remained significant for all AI types (Wald F range = 33.72–53.18, all p < .001), indicating that the combined contribution of acceptance, fear, and openness was robust to heteroskedasticity. The acceptance-fear models accounted for 18.7–26.9% of variance in apprehension across AI types, with the highest values observed for Institutional AI. Adding openness to the acceptance-fear models produced consistent increases in explained variance across all AI types. Specifically, R² increased by 0.024 for General AI (from 0.187 to 0.211), 0.017 for LLMs (from 0.201 to 0.219), 0.006 for Personal AI (from 0.201 to 0.207), and 0.011 for Institutional AI (from 0.269 to 0.279). Table 5 Convergent validity of AI apprehension (robust HC3 regressions) AI type Predictor b 95% CI β p General AI Acceptance -0.91 [-1.33, -0.50] -0.24 < .001 Fear 0.64 [0.39, 0.89] 0.25 < .001 Openness 1.27 [0.66, 1.87] 0.15 < .001 LLMs Acceptance -0.57 [-1.01, -0.14] -0.14 .010 Fear 1.05 [0.79, 1.32] 0.36 < .001 Openness 1.23 [0.54, 1.92] 0.13 < .001 Personal AI Acceptance -0.71 [-1.19, -0.24] -0.16 .003 Fear 1.13 [0.83, 1.43] 0.35 < .001 Openness 0.81 [0.05, 1.57] 0.08 .036 Institutional AI Acceptance -0.98 [-1.42, -0.55] -0.23 < .001 Fear 1.10 [0.80, 1.39] 0.35 < .001 Openness 1.06 [0.32, 1.81] 0.10 .005 Divergent validity was supported by consistently weak and non-significant associations between AI apprehension and extraversion and conscientiousness across all AI types under HC3-robust regression. For extraversion, coefficients were positive but non-significant for all AI types: General AI (b = 0.95, 95% CI [-0.21, 2.12], p = .108), LLM (b = 1.09, 95% CI [-0.26, 2.45], p = .113), Personal AI (b = 0.59, 95% CI [-0.91, 2.10], p = .438), and Institutional AI (b = 1.20, 95% CI [-0.32, 2.72], p = .123). For conscientiousness, coefficients were negative and non-significant across all AI types: General AI (b = -0.39, 95% CI [-1.23, 0.46], p = .369), LLM (b = -0.43, 95% CI [-1.33, 0.48], p = .356), Personal AI (b = -0.51, 95% CI [-1.53, 0.52], p = .333), and Institutional AI (b = -0.69, 95% CI [-1.70, 0.33], p = .186). The uniformity of these null findings under heteroskedasticity-consistent estimation provides evidence of divergent validity. 4.6. Network analysis of AI Apprehension scales Mean absolute edge weights ranged from 0.165 to 0.185 across networks (maximum difference = 0.020), and median absolute edge weights ranged from 0.067 to 0.089 (maximum difference = 0.022) (Table 6 ). These narrow ranges indicate limited variation in global connectivity strength across the four AI networks. The relative difference between the lowest and highest mean absolute edge weights was approximately 12% (0.020/0.165), indicating limited variation in global association strength. Case-dropping bootstrap analyses indicated acceptable centrality stability for all four networks. CS (strength) coefficients exceeded the recommended threshold of 0.50 for General AI (0.594), LLMs (0.594), Personal AI (0.517), and Institutional AI (0.673), indicating that strength centrality estimates were sufficiently stable to permit interpretation across all four networks (Epskamp et al., 2018 ). Table 6 Global connectivity indices and strength centrality stability across AI apprehension networks AI type Edges Density Mean absolute edge Median absolute edge Maximum absolute edge CS (strength) General AI 22 0.786 0.169 0.071 0.607 0.594 LLMs 23 0.821 0.167 0.067 0.709 0.594 Personal AI 21 0.750 0.185 0.089 0.731 0.517 Institutional AI 23 0.821 0.165 0.068 0.718 0.673 Note : Edges = number of non-zero partial correlation coefficients retained following EBICglasso regularisation (γ = 0.50). Density = proportion of all possible edges retained in the estimated network. Mean, median, and maximum values refer to the distribution of absolute edge weights within each network. CS (strength) = case-dropping bootstrap correlation stability coefficient for strength centrality. CS values ≥ .50 indicate sufficient stability for interpretation Across all four networks, the strongest edges were consistently within factors, supporting the intended three-factor structure of the scales. Implied Malice was characterised by a strong association between MOR and ETH in all networks, with large positive edge weights (General AI = 0.607; LLM = 0.628; Personal AI = 0.665; Institutional AI = 0.718). No other Implied Malice edges approached comparable magnitude in any network. Unpredictability showed robust internal coupling across all AI types. The strongest edge consistently linked SUR and GOA (General AI = 0.559; LLM = 0.709; Personal AI = 0.731; Institutional AI = 0.657). Additional positive associations were observed between EXP and SUR (range = 0.191–0.243) and between EXP and GOA (range = 0.110–0.227), indicating a tightly integrated Unpredictability cluster in every network. Undesirability items were also strongly interconnected in every network. The association between NOR and UNE was consistently strong (General AI = 0.556; LLM = 0.445; Personal AI = 0.590; Institutional AI = 0.585). Similarly, UNE and NAT showed substantial positive associations (General AI = 0.468; LLM = 0.548; Personal AI = 0.458; Institutional AI = 0.494). Positive links were also observed between NOR and NAT (range = 0.134–0.374), indicating coherent clustering of public-discomfort items across AI types (Fig. 2 ). Taken together, the consistent dominance of within-factor edges and the replication of the strongest connections across AI types indicate a similar internal organisation of apprehension components across networks, despite minor variation in absolute edge magnitude. 5. Discussion The present study responds to two limitations in existing measures of public responses to artificial intelligence. First, prior research has often relied on global assessments of AI attitudes, acceptance, trust, or fear, without distinguishing between types of AI that differ in autonomy, proximity to the user, and institutional authority. Second, apprehension to AI has rarely been measured directly and it has been inferred from broader attitude or trust scales (Grassini, 2023 ; McGrath et al., 2025 ; Rahman et al., 2025 ). To address these limitations, the present study introduces instruments that assess apprehension explicitly and separately toward distinct AI types (General AI, LLMs, Personal AI, and Institutional AI). Adopted from the Perceived Creepiness of Technology Scale (Woźniak et al., 2021 ) the new scales measure three theoretically grounded dimensions of implied malice, undesirability, and unpredictability for each AI type. This approach allows to assess AI apprehension as a form of concern that does not necessarily imply avoidance or rejection, but that varies according to the context in which AI is deployed. In doing so, these instruments provide a basis for empirical comparisons of AI-related concerns across applications, populations, and policy settings. 5.1 Psychometric properties and theoretical implications The four AI apprehension scales show satisfactory psychometric properties when evaluated against standard criteria used in scale development and validation. Model-based indices of construct quality provide additional support for interpreting the three dimensions as well-defined latent variables. Average variance extracted values indicate that, for each factor, a substantial proportion of item variance is attributable to the intended construct, supporting interpretation each scale at the latent level (Fornell and Larcker, 1981 ). Construct replicability estimates indicate that the dimensions are adequately defined by their indicators and are likely to reproduce under similar measurement conditions, despite the brevity of the subscales. Internal consistency of our scales was evaluated using multiple complementary indices, consistent with recommendations to avoid reliance on Cronbach’s alpha alone (Sijtsma, 2009 ). The strong inter-item correlations and Spearman-Brown coefficients observed for the two-item Implied Malice subscale are consistent with established guidance on the interpretation of short scales (Eisinga et al., 2013 ). The convergence between alpha and composite reliability, together with AVE values above recommended thresholds support interpretation of the subscales as well-defined latent constructs (Raykov, 1997 ). Important, our discriminant validity analyses indicate that Implied Malice, Undesirability, and Unpredictability should not be treated as interchangeable components: although the dimensions share variance, collapsing them resulted in systematic loss of model fit. The external validation results show that AI apprehension is systematically related to conceptually relevant evaluations of AI while remaining distinct from broader dispositional characteristics. Across AI types, apprehension was positively associated with fear of AI and negatively associated with acceptance of AI. Across AI types, apprehension was positively associated with fear of AI and negatively associated with acceptance of AI. This pattern supports the external validity of the scales, as apprehension relates to theoretically relevant constructs in expected directions while remaining empirically distinct from them. The positive association with fear indicates shared affective content, whereas the negative association with acceptance reflects an opposing evaluative orientation. However, apprehension was not reducible to either construct. Its moderate association with fear suggests overlap without equivalence, and its non-equivalence with acceptance indicates that apprehension cannot be interpreted simply as the absence of positive evaluation. Within tripartite models of attitudes (Kaiser and Wilson, 2019 ), our results are consistent with apprehension reflecting a distinct affective component of evaluation. Moreover, the coexistence of apprehension and acceptance aligns with theoretical accounts of attitudinal ambivalence (Cacioppo et al., 1997 ), which propose that complex technologies can evoke simultaneous positive and negative evaluations. Together, our findings support the interpretation of AI apprehension as a specific affective response, broader and less threat-focused than fear, yet not reducible to general negativity or lack of acceptance. As expected, AI apprehension in the present study was largely independent of broad personality traits, supporting the construct validity of our scales. This finding is consistent with recent research showing that personality explains only a limited proportion of variance in affective responses to AI, with evaluations influenced more strongly by perceived characteristics and societal implications of AI systems (Schepman and Rodway, 2023 ; Stein et al., 2024 ). Recent theoretical work consider AI not as a single object of evaluation but as a set of sociotechnical systems whose social meaning depends on deployment context, including distinctions between consumer-facing systems, interpersonal interfaces, and institutionally embedded decision tools (Johnson and Verdicchio, 2025 ; Sartori and Theodorou, 2022 ). By measuring apprehension separately for different AI types, the present scales translate this theoretical position into an affective measurement framework that preserves these contextual distinctions at the level of individual evaluation. This allows theories of AI governance and human-AI interaction to examine how apprehension varies with proximity, responsibility, and accountability across different AI types, using comparable measures that are aligned in structure but differentiated by evaluative target (Singhal et al., 2024 ; Vallor and Vierkant, 2024 ). 5.2 Practical implementation Cross-context consistency and non-interchangeability . The degree of consistency observed across apprehension toward different AI types has psychometric implications for how AI apprehension should be measured and interpreted. Shared-variance analyses showed only partial overlapping between the four scales, with overlap ranging from approximately two thirds of variance between General AI and LLMs to less than one third of variance for comparisons involving Institutional AI. This pattern was mirrored in the level of absolute agreement across scales, which fell in the moderate range, indicating that individuals’ relative standing on apprehension was only partly preserved across AI types. From a measurement perspective, these values indicate that the four scales capture a related evaluative tendency, but not to an extent that would justify treating them as interchangeable indicators of a single construct. Instead, apprehension shows specificity to each type of AI combined with the largest divergence emerging for Institutional AI. Apprehension therefore cannot be interpreted as a context-free individual difference; its expression varies according to the specific AI type being evaluated. Shared variance analyses further showed that, even at the factor level (Implied Malice, Undesirability, and Unpredictability), each scale retained a substantial proportion of variance specific to its respective AI type. This implies that pooling scores across AI types would reduce assessment specificity, because the pooled score would blend shared variance with AI-type-specific variance. Psychometrically, this supports using separate AI-type scales with the same factor structure, so that comparisons are aligned while preserving the context-specific information captured by each scale. The association between apprehension items . The network results indicate that AI apprehension is organised around stable internal linkages within each dimension. Concerns within Implied Malice, Unpredictability, and Undesirability are more strongly connected to each other than to concerns outside their dimension, and this pattern is consistent across AI types. Practically, this suggests that concerns within a dimension tend to reinforce one another. Addressing a single concern within a dimension may therefore be insufficient if closely connected concerns remain unchanged. The replication of the same strongest within-dimension connections across AI types further indicates that the internal configuration of apprehension is not context-specific noise but a stable structural pattern. This stability allows practitioners to anticipate which concerns are likely to co-occur and to design responses that address linked concerns jointly. Because overall connectivity is similar across AI types, differences between AI contexts are unlikely to reflect random variation in how concerns relate to one another. Instead, they reflect the activation of the same underlying structure in different contexts. Practically, this means that assessment tools and intervention strategies can rely on the same structural framework while adapting content to the specific AI domain. 5.3 Scoring Scoring by factors . The validated factor structure for each Apprehension scale comprises three dimensions represented by an unequal number of items (2 + 3 + 3). When subscales differ in item count, computing total or composite scores using simple unit-weighted sums can systematically distort the contribution of individual dimensions, because factors with more items exert greater influence on the total score purely as a function of scale length rather than construct relevance. Moreover, from a reliability perspective, shorter factors are inherently disadvantaged because internal consistency indices are sensitive to the number of indicators (Edelsbrunner et al., 2025 ; McNeish, 2018 ). Two-item factors can show acceptable inter-item correlations and strong factor loadings while nonetheless contributing less variance to a summed score than longer factors. As a result, unweighted scoring may under-represent meaningful individual differences on that dimension despite adequate construct validity. Therefore, when factor scores are used, we suggest applying weighting of 1.5 to the total score of the Implied Malice factor to account for unequal item numbers across factors (Woźniak et al., 2021 ). Total scoring . When users are interested in overall AI apprehension, weighting is unnecessary. In this case, the scale should be scored using a simple unit-weighted sum of all eight items. Total scores range from 0 to 80, with higher scores indicating greater overall apprehension towards AI. Because the total score is intended to capture general apprehension rather than differential contributions of specific domains, unequal item counts across factors do not pose a conceptual or interpretative problem in this use case. 5.4. Limitations and future research Several limitations should be considered when interpreting the present findings. First, the study used a single time-point design. As a result, we cannot say how stable scores on these scales are over time, whether they change as people gain experience with AI systems, or how quickly they respond to changes in the wider AI landscape. Future work should examine test-retest reliability and longer-term stability. Second, the scales were adapted and validated in a UK adult sample, which limits generalisability to non-Western cultures. Although we do not expect large cross-cultural differences in the core content of the three dimensions (Implied Malice, Undesirability, and Unpredictability) because these map onto broadly shared concerns about intent, social acceptability, and uncertainty in complex systems, cultural context may still influence how strongly particular concerns are endorsed. Differences in institutional trust, regulation, media narratives, and everyday exposure to AI could alter item functioning. For that reason, the next step is cross-cultural validation, including formal tests of measurement invariance, before using the scales for direct comparisons across countries or language groups. Third, the scales rely on self-report, and the response format asks respondents to indicate how much more would be required for them to consider the AI system moral, ethical, good, safe, expected, normal, or natural. This framing was deliberately chosen to operationalise apprehension as a perceived shortfall relative to an acceptable standard. However, responses still depend on how individuals internally define these standards. Future work could therefore examine how respondents interpret these evaluative thresholds and whether scores predict behavioural outcomes, such as acceptance, endorsement, or real-world use. Future research can extend the present findings by testing theoretical accounts of AI apprehension that make predictions about how evaluations of different AI types are organised and updated. Because the present study shows that apprehension is structured by AI type within the same individuals, future research can examine whether changes in beliefs, information, or public discourse selectively affect apprehension toward some AI types while leaving others unchanged. This allows competing theoretical accounts to be tested, such as whether apprehension is primarily influenced by perceived agency, institutional authority, or social proximity, depending on the AI type under consideration. In addition, future studies can examine whether relations between AI apprehension and other constructs, such as trust or acceptance, follow the same structure across AI types or diverge systematically, which would further clarify the conceptual boundaries of AI apprehension. These directions build directly on the present findings by using the differentiated structure of the scales to refine theory about how concern toward artificial intelligence is organised and maintained. 6. Conclusion The General AI Apprehension Scale, Large Language Model Apprehension Scale, Personal AI Apprehension Scale, and Institutional AI Apprehension Scale measure apprehension toward four distinct types of artificial intelligence. Each scale using the same underlying dimensions, which allows apprehension to be examined separately across AI contexts and compared when needed. Across all four scales, the measurement models show stable internal structure, good reliability at both total and subscale levels, and separation from related evaluative constructs. The scales are intended for empirical research that requires direct and context-specific measurement of apprehension toward different forms of artificial intelligence. Declarations Corresponding authors : Ala Yankouskaya, Raian Ali Author contribution AY: Conceptualised and designed the study, curated the data, designed and performed and reported the statistical analysis, wrote the first draft. BA: Conceptualised the study, validated the analysis, reviewed and edited the paper. ML: Validated the analysis, wrote parts of the first draft, reviewed and edited the paper. MMR: Designed the study, curated the data, reviewed the paper. SA: Designed the study, validated the analysis, reviewed and edited the paper. RA: Conceptualised and designed the study, curated the data, validated the analysis, reviewed and edited the paper. Funding Open Access funding provided by the Qatar National Library. This publication was supported by NPRP 14 Cluster grant # NPRP 14C-0916–210015 from the Qatar National Research Fund (a member of Qatar Foundation). The findings herein reflect the work and are solely the responsibility of the authors. Author declaration The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Ethical approval and consent to participate : Ethical approval was granted by the Institutional Review Board (IRB) at the lead author’s institution (IRB Protocol Reference Number: HBKU-IRB-2025-185). Informed consent was obtained from all individual participants included in the study. Data availability : The study design, dataset, and analysis files are available on the Open Science Framework (OSF) at the following link view_only=a1d3d422a48643c8b75fd8cf06ae336. The author confirms that all data generated or analysed during this study are included in this published article. References Alexander, J.K., 2020. Introduction: the entanglement of technology and religion. Hist. Technol. 36, 165–186. https://doi.org/10.1080/07341512.2020.1814513 Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Mané, D., 2016. Concrete Problems in AI Safety. https://doi.org/10.48550/ARXIV.1606.06565 Ananny, M., Crawford, K., 2018. Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media Soc. 20, 973–989. https://doi.org/10.1177/1461444816676645 Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M.T., Weld, D., 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Presented at the CHI ’21: CHI Conference on Human Factors in Computing Systems, ACM, Yokohama Japan, pp. 1–16. https://doi.org/10.1145/3411764.3445717 Beaudry, A., Pinsonneault, A., 2010. The Other Side of Acceptance: Studying the Direct and Indirect Effects of Emotions on Information Technology Use1. MIS Q. 34, 689–710. https://doi.org/10.2307/25750701 Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S., 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜, in: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Presented at the FAccT ’21: 2021 ACM Conference on Fairness, Accountability, and Transparency, ACM, Virtual Event Canada, pp. 610–623. https://doi.org/10.1145/3442188.3445922 Bialy, F., Elliot, M., Meckin, R., 2025. Perceptions of AI Across Sectors: A Comparative Review of Public Attitudes. https://doi.org/10.48550/arXiv.2509.18233 Bicchieri, C., 2005. The Grammar of Society: The Nature and Dynamics of Social Norms, 1st ed. Cambridge University Press. https://doi.org/10.1017/CBO9780511616037 Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D., 2020. Language Models are Few-Shot Learners. https://doi.org/10.48550/ARXIV.2005.14165 Bucknall, B.S., Dori-Hacohen, S., 2022. Current and Near-Term AI as a Potential Existential Risk Factor, in: Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society. Presented at the AIES ’22: AAAI/ACM Conference on AI, Ethics, and Society, ACM, Oxford United Kingdom, pp. 119–129. https://doi.org/10.1145/3514094.3534146 Burrell, J., Fourcade, M., 2021. The Society of Algorithms. Annu. Rev. Sociol. 47, 213–237. https://doi.org/10.1146/annurev-soc-090820-020800 Cacioppo, J.T., Gardner, W.L., Berntson, G.G., 1997. Beyond Bipolar Conceptualizations and Measures: The Case of Attitudes and Evaluative Space. Personal. Soc. Psychol. Rev. 1, 3–25. https://doi.org/10.1207/s15327957pspr0101_2 Castelo, N., Bos, M.W., Lehmann, D.R., 2019. Task-Dependent Algorithm Aversion. J. Mark. Res. 56, 809–825. https://doi.org/10.1177/0022243719851788 Caton, S., Haas, C., 2024. Fairness in Machine Learning: A Survey. ACM Comput. Surv. 56, 1–38. https://doi.org/10.1145/3616865 Corradi, G., Theirs, C., Martínez‐Martí, M.L., Isern‐Mas, C., Villar, S., 2025. Who Fears Generative Artificial Intelligence? Scale Development and Predictors of Fears Towards GenAI. Scand. J. Psychol. sjop.70037. https://doi.org/10.1111/sjop.70037 Costa, P.T., McCrae, R.R., 1992. The Five-Factor Model of Personality and Its Relevance to Personality Disorders. J. Personal. Disord. 6, 343–359. https://doi.org/10.1521/pedi.1992.6.4.343 Dang, Q., Li, G., 2026. Unveiling trust in AI: the interplay of antecedents, consequences, and cultural dynamics. AI Soc. 41, 669–692. https://doi.org/10.1007/s00146-025-02477-6 De Freitas, J., Uguralp, A.K., Uguralp, Z.O., Stefano, P., 2024. AI Companions Reduce Loneliness. https://doi.org/10.48550/ARXIV.2407.19096 Devaraj, S., Easley, R.F., Crant, J.M., 2008. Research Note —How Does Personality Matter? Relating the Five-Factor Model to Technology Acceptance and Use. Inf. Syst. Res. 19, 93–105. https://doi.org/10.1287/isre.1070.0153 Dietvorst, B.J., Simmons, J.P., Massey, C., 2015. Algorithm aversion: People erroneously avoid algorithms after seeing them err. J. Exp. Psychol. Gen. 144, 114–126. https://doi.org/10.1037/xge0000033 Edelsbrunner, P.A., Simonsmeier, B.A., Schneider, M., 2025. The Cronbach’s Alpha of Domain-Specific Knowledge Tests Before and After Learning: A Meta-Analysis of Published Studies. Educ. Psychol. Rev. 37, 4. https://doi.org/10.1007/s10648-024-09982-y Eisinga, R., Grotenhuis, M.T., Pelzer, B., 2013. The reliability of a two-item scale: Pearson, Cronbach, or Spearman-Brown? Int. J. Public Health 58, 637–642. https://doi.org/10.1007/s00038-012-0416-3 Eloundou, T., Manning, S., Mishkin, P., Rock, D., 2023. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://doi.org/10.48550/arXiv.2303.10130 Epley, N., Waytz, A., Cacioppo, J.T., 2007. On seeing human: A three-factor theory of anthropomorphism. Psychol. Rev. 114, 864–886. https://doi.org/10.1037/0033-295X.114.4.864 Epskamp, S., Borsboom, D., Fried, E.I., 2018. Estimating psychological networks and their accuracy: A tutorial paper. Behav. Res. Methods 50, 195–212. https://doi.org/10.3758/s13428-017-0862-1 Fornell, C., Larcker, D.F., 1981. Evaluating Structural Equation Models with Unobservable Variables and Measurement Error. J. Mark. Res. 18, 39. https://doi.org/10.2307/3151312 Gabriel, I., 2020. Artificial Intelligence, Values, and Alignment. Minds Mach. 30, 411–437. https://doi.org/10.1007/s11023-020-09539-2 Gagrčin, E., Toth, R., Schaetz, N., Naab, T., Emmer, M., 2026. Perceived personal and societal data harms shape users’ data control preferences. Internet Policy Rev. 15. https://doi.org/10.14763/2026.1.2060 Glikson, E., Woolley, A.W., 2020. Human Trust in Artificial Intelligence: Review of Empirical Research. Acad. Manag. Ann. 14, 627–660. https://doi.org/10.5465/annals.2018.0057 Grassini, S., 2023. Development and validation of the AI attitude scale (AIAS-4): a brief measure of general attitude toward artificial intelligence. Front. Psychol. 14, 1191628. https://doi.org/10.3389/fpsyg.2023.1191628 Grassini, S., Thorp, S.O., Cipriani, E., Mannari, V., Sevic, A., 2025. Context matters: Analyzing the factors influencing artificial intelligence acceptance across five different applications. Comput. Hum. Behav. Rep. 20, 100834. https://doi.org/10.1016/j.chbr.2025.100834 Gray, K., Wegner, D.M., 2012. Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition 125, 125–130. https://doi.org/10.1016/j.cognition.2012.06.007 Hancock, P.A., Billings, D.R., Schaefer, K.E., 2011. Can You Trust Your Robot? Ergon. Des. Q. Hum. Factors Appl. 19, 24–29. https://doi.org/10.1177/1064804611415045 Hayduk, L.A., Glaser, D.N., 2000. Jiving the Four-Step, Waltzing Around Factor Analysis, and Other Serious Fun. Struct. Equ. Model. Multidiscip. J. 7, 1–35. https://doi.org/10.1207/S15328007SEM0701_01 Hommel, B., 2023. Religion and cognitive control: An event-coding approach. New Ideas Psychol. 70, 101022. https://doi.org/10.1016/j.newideapsych.2023.101022 Jebb, A.T., Ng, V., Tay, L., 2021. A Review of Key Likert Scale Development Advances: 1995–2019. Front. Psychol. 12, 637547. https://doi.org/10.3389/fpsyg.2021.637547 Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P., 2023. Survey of Hallucination in Natural Language Generation. ACM Comput. Surv. 55, 1–38. https://doi.org/10.1145/3571730 Johnson, D.G., Verdicchio, M., 2025. The sociotechnical entanglement of AI and values. AI Soc. 40, 67–76. https://doi.org/10.1007/s00146-023-01852-5 Kaiser, F.G., Wilson, M., 2019. The Campbell Paradigm as a Behavior-Predictive Reinterpretation of the Classical Tripartite Model of Attitudes. Eur. Psychol. 24, 359–374. https://doi.org/10.1027/1016-9040/a000364 Kenny, D.A., Kaniskan, B., McCoach, D.B., 2015. The Performance of RMSEA in Models With Small Degrees of Freedom. Sociol. Methods Res. 44, 486–507. https://doi.org/10.1177/0049124114543236 Kleine, A.-K., Schmitt, A., Wisse, B.M., 2024. Challenge and threat appraisal of entrepreneurial errors: a latent profile analysis and examination of coping responses. Curr. Psychol. 43, 1206–1220. https://doi.org/10.1007/s12144-023-04370-1 Kotov, R., Gamez, W., Schmidt, F., Watson, D., 2010. Linking “big” personality traits to anxiety, depressive, and substance use disorders: A meta-analysis. Psychol. Bull. 136, 768–821. https://doi.org/10.1037/a0020327 Lai, K., Green, S.B., 2016. The Problem with Having Two Watches: Assessment of Fit When RMSEA and CFI Disagree. Multivar. Behav. Res. 51, 220–239. https://doi.org/10.1080/00273171.2015.1134306 Lee, J.D., See, K.A., 2004. Trust in Automation: Designing for Appropriate Reliance. Hum. Factors J. Hum. Factors Ergon. Soc. 46, 50–80. https://doi.org/10.1518/hfes.46.1.50_30392 Lee, M.K., 2018. Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data Soc. 5, 2053951718756684. https://doi.org/10.1177/2053951718756684 Lee, M.K., Kusbit, D., Metsky, E., Dabbish, L., 2015. Working with Machines: The Impact of Algorithmic and Data-Driven Management on Human Workers, in: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems. Presented at the CHI ’15: CHI Conference on Human Factors in Computing Systems, ACM, Seoul Republic of Korea, pp. 1603–1612. https://doi.org/10.1145/2702123.2702548 Liljequist, D., Elfving, B., Skavberg Roaldsen, K., 2019. Intraclass correlation – A discussion and demonstration of basic features. PLOS ONE 14, e0219854. https://doi.org/10.1371/journal.pone.0219854 Lipton, Z.C., 2018. The Mythos of Model Interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue 16, 31–57. https://doi.org/10.1145/3236386.3241340 Longoni, C., Bonezzi, A., Morewedge, C.K., 2019. Resistance to Medical Artificial Intelligence. J. Consum. Res. 46, 629–650. https://doi.org/10.1093/jcr/ucz013 Luger, E., Sellen, A., 2016. “Like Having a Really Bad PA”: The Gulf between User Expectation and Experience of Conversational Agents, in: Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. Presented at the CHI’16: CHI Conference on Human Factors in Computing Systems, ACM, San Jose California USA, pp. 5286–5297. https://doi.org/10.1145/2858036.2858288 Marsh, H.W., Wen, Z., Hau, K.-T., 2004. Structural Equation Models of Latent Interactions: Evaluation of Alternative Estimation Strategies and Indicator Construction. Psychol. Methods 9, 275–300. https://doi.org/10.1037/1082-989X.9.3.275 Martin, K., 2018. The penalty for privacy violations: How privacy violations impact trust online. J. Bus. Res. 82, 103–116. https://doi.org/10.1016/j.jbusres.2017.08.034 Mayer, R.C., Davis, J.H., Schoorman, F.D., 1995. An Integrative Model of Organizational Trust. Acad. Manage. Rev. 20, 709. https://doi.org/10.2307/258792 McElroy, J.C., Hendrickson, A.R., Townsend, A.M., DeMarie, S.M., 2007. Dispositional Factors in Internet Use: Personality Versus Cognitive Style1. MIS Q. 31, 809–820. https://doi.org/10.2307/25148821 McGrath, M.J., Lack, O., Tisch, J., Duenser, A., 2025. Measuring trust in artificial intelligence: validation of an established scale and its short form. Front. Artif. Intell. 8, 1582880. https://doi.org/10.3389/frai.2025.1582880 McNeish, D., 2018. Thanks coefficient alpha, we’ll take it from here. Psychol. Methods 23, 412–433. https://doi.org/10.1037/met0000144 McPhetres, J., Zuckerman, M., 2018. Religiosity predicts negative attitudes towards science and lower levels of science literacy. PLOS ONE 13, e0207125. https://doi.org/10.1371/journal.pone.0207125 Metzger, M.J., Flanagin, A.J., 2015. Psychological Approaches to Credibility Assessment Online, in: Sundar, S.S. (Ed.), The Handbook of the Psychology of Communication Technology. Wiley, pp. 445–466. https://doi.org/10.1002/9781118426456.ch20 Mick, D.G., Fournier, S., 1998. Paradoxes of Technology: Consumer Cognizance, Emotions, and Coping Strategies. J. Consum. Res. 25, 123–143. https://doi.org/10.1086/209531 Montag, C., Nakov, P., Ali, R., 2025. On the need to develop nuanced measures assessing attitudes towards AI and AI literacy in representative large-scale samples. AI Soc. 40, 1129–1130. https://doi.org/10.1007/s00146-024-01888-1 Montoya, A.K., Edwards, M.C., 2021. The Poor Fit of Model Fit for Selecting Number of Factors in Exploratory Factor Analysis for Scale Evaluation. Educ. Psychol. Meas. 81, 413–440. https://doi.org/10.1177/0013164420942899 Pasquale, F., 2016. The black box society: the secret algorithms that control money and information, First Harvard University Press paperback edition. ed. Harvard University Press, Cambridge, Massachusetts London, England. Preston, C.C., Colman, A.M., 2000. Optimal number of response categories in rating scales: reliability, validity, discriminating power, and respondent preferences. Acta Psychol. (Amst.) 104, 1–15. https://doi.org/10.1016/S0001-6918(99)00050-5 Rahman, M.M., AlShakhsi, S., Babiker, A., Ali, R., Yankouskaya, A., 2025. On the role of personality in attitudes toward AI: Do AI’s freedom of choice and social proof matter? World Wide Web 28, 67. https://doi.org/10.1007/s11280-025-01372-w Rammstedt, B., John, O.P., 2007. Measuring personality in one minute or less: A 10-item short version of the Big Five Inventory in English and German. J. Res. Personal. 41, 203–212. https://doi.org/10.1016/j.jrp.2006.02.001 Raykov, T., 1997. Scale Reliability, Cronbach’s Coefficient Alpha, and Violations of Essential Tau-Equivalence with Fixed Congeneric Components. Multivar. Behav. Res. 32, 329–353. https://doi.org/10.1207/s15327906mbr3204_2 Revilla, M.A., Saris, W.E., Krosnick, J.A., 2014. Choosing the Number of Categories in Agree–Disagree Scales. Sociol. Methods Res. 43, 73–97. https://doi.org/10.1177/0049124113509605 Rudin, C., 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell. 1, 206–215. https://doi.org/10.1038/s42256-019-0048-x Sag, M., 2023. Copyright Safety for Generative AI. SSRN Electron. J. https://doi.org/10.2139/ssrn.4438593 Sartori, L., Theodorou, A., 2022. A sociotechnical perspective for the future of AI: narratives, inequalities, and human control. Ethics Inf. Technol. 24, 4. https://doi.org/10.1007/s10676-022-09624-3 Schepman, A., Rodway, P., 2023. The General Attitudes towards Artificial Intelligence Scale (GAAIS): Confirmatory Validation and Associations with Personality, Corporate Distrust, and General Trust. Int. J. Human–Computer Interact. 39, 2724–2741. https://doi.org/10.1080/10447318.2022.2085400 Schiavo, G., Businaro, S., Zancanaro, M., 2024. Comprehension, apprehension, and acceptance: Understanding the influence of literacy and anxiety on acceptance of artificial Intelligence. Technol. Soc. 77, 102537. https://doi.org/10.1016/j.techsoc.2024.102537 Shin, D., 2021. The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI. Int. J. Hum.-Comput. Stud. 146, 102551. https://doi.org/10.1016/j.ijhcs.2020.102551 Sijtsma, K., 2009. On the Use, the Misuse, and the Very Limited Usefulness of Cronbach’s Alpha. Psychometrika 74, 107–120. https://doi.org/10.1007/s11336-008-9101-0 Sindermann, C., Sha, P., Zhou, M., Wernicke, J., Schmitt, H.S., Li, M., Sariyska, R., Stavrou, M., Becker, B., Montag, C., 2021. Assessing the Attitude Towards Artificial Intelligence: Introduction of a Short Measure in German, Chinese, and English Language. KI - Künstl. Intell. 35, 109–118. https://doi.org/10.1007/s13218-020-00689-0 Singhal, A., Neveditsin, N., Tanveer, H., Mago, V., 2024. Toward Fairness, Accountability, Transparency, and Ethics in AI for Social Media and Health Care: Scoping Review. JMIR Med. Inform. 12, e50048. https://doi.org/10.2196/50048 Stein, J.-P., Messingschlager, T., Gnambs, T., Hutmacher, F., Appel, M., 2024. Attitudes towards AI: measurement and associations with personality. Sci. Rep. 14, 2909. https://doi.org/10.1038/s41598-024-53335-2 Sun, B., Zhang, Y., Meng, X., Zhang, Z., 2025. Social class, public self-awareness, and consumers’ heterogeneous attitude to AI machine anthropomorphism. Curr. Psychol. 44, 10561–10578. https://doi.org/10.1007/s12144-025-07872-2 Sundar, S.S., Lee, E.-J., 2022. Rethinking Communication in the Era of Artificial Intelligence. Hum. Commun. Res. 48, 379–385. https://doi.org/10.1093/hcr/hqac014 Svendsen, G.B., Johnsen, J.-A.K., Almås-Sørensen, L., Vittersø, J., 2013. Personality and technology acceptance: the influence of personality factors on the core constructs of the Technology Acceptance Model. Behav. Inf. Technol. 32, 323–334. https://doi.org/10.1080/0144929X.2011.553740 Taddeo, M., Floridi, L., 2018. How AI can be a force for good. Science 361, 751–752. https://doi.org/10.1126/science.aat5991 Thompson, S.C., 1981. Will it hurt less if I can control it? A complex answer to a simple question. Psychol. Bull. 90, 89–101. https://doi.org/10.1037/0033-2909.90.1.89 Triguero, I., Molina, D., Poyatos, J., Del Ser, J., Herrera, F., 2024. General Purpose Artificial Intelligence Systems (GPAIS): Properties, definition, taxonomy, societal implications and responsible governance. Inf. Fusion 103, 102135. https://doi.org/10.1016/j.inffus.2023.102135 Uğur, N.G., Dursun, F., 2025. The social anatomy of AI anxiety: gender, generations, and technological exposure. Front. Psychiatry 16, 1641546. https://doi.org/10.3389/fpsyt.2025.1641546 Vallor, S., Vierkant, T., 2024. Find the Gap: AI, Responsible Agency and Vulnerability. Minds Mach. 34, 20. https://doi.org/10.1007/s11023-024-09674-0 Venkatesh, Thong, Xu, 2012. Consumer Acceptance and Use of Information Technology: Extending the Unified Theory of Acceptance and Use of Technology. MIS Q. 36, 157. https://doi.org/10.2307/41410412 Wang, Y., Guan, X., Sun, Y., Wang, H., Chen, D., 2025. The cognitive acceptance of generative AI image tools based on TPB-TAM model and multi-theory integration. Adv. Des. Res. 3, 38–54. https://doi.org/10.1016/j.ijadr.2025.08.001 Wang, Y.-Y., Wang, Y.-S., 2022. Development and validation of an artificial intelligence anxiety scale: an initial application in predicting motivated learning behavior. Interact. Learn. Environ. 30, 619–634. https://doi.org/10.1080/10494820.2019.1674887 Wolf, E.J., Harrington, K.M., Clark, S.L., Miller, M.W., 2013. Sample Size Requirements for Structural Equation Models: An Evaluation of Power, Bias, and Solution Propriety. Educ. Psychol. Meas. 73, 913–934. https://doi.org/10.1177/0013164413495237 Woźniak, P.W., Karolus, J., Lang, F., Eckerth, C., Schöning, J., Rogers, Y., Niess, J., 2021. Creepy Technology:What Is It and How Do You Measure It?, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Presented at the CHI ’21: CHI Conference on Human Factors in Computing Systems, ACM, Yokohama Japan, pp. 1–13. https://doi.org/10.1145/3411764.3445299 Yang, H., Sundar, S.S., 2025. AI anxiety: Explication and exploration of effect on state anxiety when interacting with AI doctors. Comput. Hum. Behav. Artif. Hum. 3, 100128. https://doi.org/10.1016/j.chbah.2025.100128 Zhang, B., Dafoe, A., 2019. Artificial Intelligence: American Attitudes and Trends. SSRN Electron. J. https://doi.org/10.2139/ssrn.3312874 Zuboff, S., 2023. The age of surveillance capitalism, in: Social Theory Re-Wired. pp. 203–213. Additional Declarations The authors declare no competing interests. Supplementary Files SupplementaryMaterial.docx Supplementary Material SupplmentrayFiles.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9018418","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":599864723,"identity":"245f9abc-d8e5-4bd8-82ed-5292f3631041","order_by":0,"name":"Ala Yankouskaya","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0003-0794-0989","institution":"Bournemouth University","correspondingAuthor":true,"prefix":"","firstName":"Ala","middleName":"","lastName":"Yankouskaya","suffix":""},{"id":599864724,"identity":"bd8f7ac8-f907-4ce5-8e72-153ccddefd02","order_by":1,"name":"Mohamed Basel Almourad","email":"","orcid":"https://orcid.org/0000-0003-2193-1472","institution":"Zayed University","correspondingAuthor":false,"prefix":"","firstName":"Mohamed","middleName":"Basel","lastName":"Almourad","suffix":""},{"id":599864725,"identity":"790b8053-07e7-4ee9-93ac-87f9074d2bed","order_by":2,"name":"Magnus Liebherr","email":"","orcid":"https://orcid.org/0000-0001-8580-2464","institution":"University Duisburg-Essen","correspondingAuthor":false,"prefix":"","firstName":"Magnus","middleName":"","lastName":"Liebherr","suffix":""},{"id":599864726,"identity":"7a459377-3a6f-43af-87e5-2801ee4ed68f","order_by":3,"name":"Mohammad Mominur Rahman","email":"","orcid":"https://orcid.org/0000-0002-9703-1748","institution":"Hamad Bin Khalifa University","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"Mominur","lastName":"Rahman","suffix":""},{"id":599864727,"identity":"abadd7c2-95a7-43b1-ac2c-9adf562cf0ea","order_by":4,"name":"Sameha AlShakhsi","email":"","orcid":"https://orcid.org/0000-0002-2138-4731","institution":"Hamad Bin Khalifa University","correspondingAuthor":false,"prefix":"","firstName":"Sameha","middleName":"","lastName":"AlShakhsi","suffix":""},{"id":599864728,"identity":"ca6188ec-b37b-47dd-b8cb-935162e6ddb8","order_by":5,"name":"Raian Ali","email":"","orcid":"https://orcid.org/0000-0002-5285-7829","institution":"Hamad Bin Khalifa University","correspondingAuthor":false,"prefix":"","firstName":"Raian","middleName":"","lastName":"Ali","suffix":""}],"badges":[],"createdAt":"2026-03-03 09:13:15","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9018418/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9018418/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103893584,"identity":"8ec310ab-10e0-4c03-9891-60c8b63b86c1","added_by":"auto","created_at":"2026-03-04 08:28:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":226019,"visible":true,"origin":"","legend":"\u003cp\u003eFactor-level shared variance and overall cross-type agreement across AI types\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9018418/v1/aa8ba8b345a2e74ed7b68c51.png"},{"id":103893560,"identity":"7d4765b9-a16f-40ab-bc0c-35ce8c5c97e1","added_by":"auto","created_at":"2026-03-04 08:28:32","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":236350,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork structure of AI apprehension scales\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9018418/v1/7f52c1385696dc61373ef43e.png"},{"id":104401579,"identity":"15da9fec-1f16-453d-98fb-fc6787fb7ced","added_by":"auto","created_at":"2026-03-11 12:13:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1704711,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9018418/v1/159bdaf6-3fc6-44a4-b6fb-a67e446558b2.pdf"},{"id":103893581,"identity":"1e3d2ae5-49ec-4fc3-97a1-691ef2c42c00","added_by":"auto","created_at":"2026-03-04 08:28:36","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":588822,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Material\u003c/p\u003e","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-9018418/v1/ba990f3ca03755e198a5088a.docx"},{"id":103893582,"identity":"e7c1b454-e24f-4a97-81ad-7a07fda5e61c","added_by":"auto","created_at":"2026-03-04 08:28:38","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":49874,"visible":true,"origin":"","legend":"","description":"","filename":"SupplmentrayFiles.docx","url":"https://assets-eu.researchsquare.com/files/rs-9018418/v1/cb0f04dde1efddb952bbc324.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eThe Four Artificial Intelligence Apprehension Scales: Apprehension Towards Personal AI, General AI, Institutional AI, and Large Language Models\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAttitudes towards AI may differ depending on the type of AI and the nature of the interaction involved. Direct end-user conversations with Siri or Alexa (an example of Personal AI) about specific events feel fundamentally different from being instructed to reroute to a less busy hospital based on an AI-powered surveillance and resource-management system (an example of Institutional AI). Both involve artificial intelligence, yet they evoke strikingly divergent emotional responses. The former might inspire curiosity or even playfulness, while the latter often generates profound unease about fairness, transparency, and control over consequential life outcomes. This experiential divergence poses a critical challenge for researchers seeking to understand public responses to AI. If individuals experience AI as a collection of distinct systems with different capabilities, purposes, and power dynamics, then treating AI apprehension as a monolithic construct risks obscuring the very psychological processes we seek to understand (Dang and Li, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2026\u003c/span\u003e; De Freitas et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis variability in emotional responses to different classes of AI reflects not only individual differences in technological literacy and personal factors, but also the fundamental heterogeneity of AI systems themselves. Different AI technologies vary substantially in their capabilities, purposes, transparency, and proximity to users' personal lives. Yet despite growing recognition that emotional responses to AI represent a critical component of human-AI interaction, existing measurement approaches remain limited. Scientific evaluations increasingly criticise existing measurement instruments for often treating AI as a monolithic category, thereby neglecting the variance in application context, user relationship and risk profile (Montag et al., \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Additionally, while broad attitudinal measures capture general positivity or negativity toward AI, they often conflate affective, cognitive, and behavioral components, making it difficult to isolate the specific role of apprehension as an anticipatory emotional state characterized by concern, unease, and heightened sensitivity to potential negative outcomes (Schepman and Rodway, \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Furthermore, existing instruments for measuring technology-related apprehension were developed for general interactive technologies without specific consideration of AI systems' unique characteristics (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eThe present study addresses these measurement gaps by examining perceived apprehension across four distinct AI contexts: personal AI, general AI, institutional AI, and Large Language Models. This multi-context approach is grounded in literature suggesting that apprehension toward AI is not uniform but context-dependent, shaped by the specific characteristics, perceived intentions, and power dynamics associated with different AI instantiations (Bialy et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Grassini et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). By capturing emotional responses across qualitatively different forms of AI, the study aims to provide a more nuanced understanding of how creepiness varies as a function of both system characteristics and individual differences. To operationalize perceived creepiness in these contexts, an adapted version of the Perceived Creepiness of Technology Scale (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) was employed, preserving its theoretically grounded three-factor structure of implied malice, undesirability, and unpredictability while contextualizing items for AI-specific concerns.\u003c/p\u003e \u003cp\u003eThis work contributes to the literature on human-AI interaction by providing psychometrically validated instruments that are built on an established measure for technology-related apprehension (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), while being specifically tailored to AI systems across theoretically meaningful classes. We first establish the psychometric adequacy of each scale, including evidence for convergent and discriminant validity with theoretically relevant constructs. We then demonstrate that apprehension shows both shared variance across AI types and context-specific variation, supporting the decision to distinguish four separate AI domains. Finally, we examine demographic correlates of AI-related apprehension and discuss their implications for equitable AI design and deployment. We build on the Perceived Creepiness of Technology Scale (PCTS) proposed by Woźniak et al. (\u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) to assess apprehension toward artificial intelligence. The original scale was adapted primarily at the level of wording to reflect a more formal focus on technology-related apprehension rather than creepiness per se. Moreover, whereas the original instrument targets perceptions of specific technological products, the present version is designed to capture attitudes toward broader classes of AI systems. This approach allows researchers and practitioners to assess AI-related unease using a theoretically grounded instrument that is sensitive to both the psychological foundations of technology-related discomfort and the distinctive characteristics of different AI domains.\u003c/p\u003e"},{"header":"2. Theoretical Background","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Attitudes Toward AI and the Affective Dimension\u003c/h2\u003e \u003cp\u003eAttitudes toward technology have long been understood as multifaceted psychological constructs encompassing cognitive beliefs, affective responses, and behavioral intentions (Venkatesh et al., 2012). In the context of AI, attitudes reflect individuals' overall evaluations of AI systems, shaped by perceived benefits, risks, ethical concerns, and personal experiences (Glikson and Woolley, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The tripartite model of attitudes has been widely applied to understand technology acceptance, with the affective dimension increasingly recognized as a critical driver of adoption and usage behaviors (Beaudry and Pinsonneault, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). However, AI systems differ from conventional technologies in ways that necessitate theoretical refinement. Unlike static tools or passive information systems, AI technologies exhibit characteristics traditionally associated with agency. They learn, adapt, make autonomous decisions, and generate novel outputs that exceed their creators' explicit programming (Eloundou et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sundar and Lee, \u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These agentic qualities fundamentally alter the psychological dynamics of human-technology interaction, shifting users from a position of control to one of coordination or, in some cases, subordination (Hancock et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Sun et al., \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This shift toward machine agency introduces distinct affective dimensions that extend beyond conventional technology acceptance frameworks. When individuals interact with AI, they must navigate not only instrumental concerns about utility and ease of use but also relational concerns about trust, transparency, and alignment of interests (Lee and See, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; Shin, \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The opacity of many AI systems, particularly deep learning models whose decision-making processes are neither transparent nor readily interpretable, exacerbates these concerns and creates conditions for apprehension.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Defining AI Apprehension\u003c/h2\u003e \u003cp\u003eApprehension represents a specific affective component within the broader attitude structure toward AI. While general attitudes may range from highly positive to highly negative, apprehension specifically captures the negative affective pole, emphasizing emotional responses rooted in concern, unease, worry, and anticipatory anxiety about AI systems and their implications (Schiavo et al., \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Uğur and Dursun, \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This construct is theoretically distinct from several adjacent concepts. AI fear refers to intense emotional distress or anxiety about specific AI-related threats such as job displacement, loss of human autonomy, or existential risks (Schepman and Rodway, \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). While apprehension may encompass elements of fear, it is broader in scope and less tied to discrete threat scenarios. AI skepticism involves cognitive doubt about AI capabilities or benefits, often accompanied by critical evaluation of AI hype or distrust of technology companies (Glikson and Woolley, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Wang et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Unlike apprehension, skepticism is primarily cognitive rather than affective and may not be accompanied by emotional distress. AI resistance encompasses behavioral opposition to AI adoption manifested through avoidance or active rejection (Dietvorst et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Longoni et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), representing the conative component of attitudes rather than the affective dimension we target.\u003c/p\u003e \u003cp\u003eImportantly, apprehension is not simply the inverse of acceptance or trust. An individual may simultaneously recognize the utility of AI systems while feeling uncertain or uneasy about their implications, reflecting the well-documented phenomenon of attitudinal ambivalence toward complex technologies (Cacioppo et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e1997\u003c/span\u003e). This ambivalence is particularly pronounced in the AI domain, where many users report concurrent appreciation for AI benefits and concern about AI risks (Mick and Fournier, \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e1998\u003c/span\u003e). Our conceptualization of apprehension as a distinct affective construct allows researchers to capture this nuanced emotional landscape without conflating it with cognitive evaluations or behavioral intentions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Psychological Dimensions of AI Apprehension\u003c/h2\u003e \u003cp\u003eFollowing Woźniak et al. (\u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) conceptual model of creepiness, we apply the three-dimensional structure to the context of AI apprehension. We adopt this framework for two reasons. First, there is a lack of existing scales that explicitly use the term \u0026ldquo;apprehension\u0026rdquo; and focus their content on this specific affective state. Second, while \u0026ldquo;creepiness\u0026rdquo; serves as an informal designation for technology-related apprehension, the original scale employed colloquial language and phrasing suitable for evaluating novel consumer products e.g. smart glasses or gadget. For the present AI context, we required more formal terminology and item formulations appropriate for a broader range of AI systems. This structure is grounded in psychological theories of threat appraisal, social judgment, and uncertainty (Kleine et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Metzger and Flanagin, \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and is further supported by empirical research on public concerns about AI (Zhang and Dafoe, \u003cspan citationid=\"CR92\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eImplied malice captures concerns about the ethical foundations and moral intentions underlying AI development and deployment. This dimension reflects individuals' uncertainty about whether AI creators prioritize human welfare, adhere to ethical principles, or design systems that align with societal values (Taddeo and Floridi, \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Research consistently demonstrates that public trust in AI is undermined by perceptions that technology companies prioritize profit over safety, lack transparency about AI capabilities and limitations, or fail to address algorithmic bias and discrimination (Yang and Sundar, \u003cspan citationid=\"CR91\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Zhang and Dafoe, \u003cspan citationid=\"CR92\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The implied malice dimension is theoretically grounded in research on institutional trust and perceived organizational ethics (Mayer et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e1995\u003c/span\u003e), encompassing both benevolence concerns and integrity concerns about AI development processes.\u003c/p\u003e \u003cp\u003eUndesirability reflects discomfort with AI presence in social and public contexts, particularly when AI systems are visible, ambient, or integrated into spaces traditionally governed by human norms and interpersonal dynamics (Lee, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). This dimension captures the sense that AI technologies disrupt social naturalness, create awkwardness or self-consciousness, and introduce elements that feel out of place in human-centered environments. The undesirability dimension is informed by theories of social presence, anthropomorphism (Epley et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2007\u003c/span\u003e), and norm violation (Bicchieri, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). When AI systems are introduced into public or interpersonal contexts, they may be perceived as infringing on privacy, altering social dynamics, or creating surveillance pressures that constrain authentic behavior (Zuboff, S., \u003cspan citationid=\"CR93\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Additionally, the presence of AI in traditionally human domains can evoke uncanniness, a sense of unease arising from entities that approximate but do not fully replicate human characteristics (Gray and Wegner, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2012\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eUnpredictability encompasses uncertainty about AI behavior, goals, and outcomes, concerns rooted in the opacity, complexity, and autonomy of contemporary AI systems (Castelo et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Shin, \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This dimension reflects individuals' doubts about whether AI systems operate as intended, whether their purposes are transparent and comprehensible, and whether their future behaviors can be reliably anticipated (Burrell and Fourcade, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Lipton, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The unpredictability dimension is theoretically grounded in research on uncertainty and control (Thompson, \u003cspan citationid=\"CR82\" class=\"CitationRef\"\u003e1981\u003c/span\u003e) and algorithmic opacity (Pasquale, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Many AI systems function as black boxes, generating outputs through processes that are opaque (Rudin, \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This opacity undermines users' ability to form accurate mental models of AI behavior, calibrate their trust appropriately, or predict system responses in novel situations (Bansal et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Furthermore, the autonomy and adaptability of AI systems introduce behavioral unpredictability, the possibility that AI may behave in unexpected ways or pursue goals that deviate from user intentions. Even when AI systems function correctly from a technical perspective, their goals may be misaligned with human values or their deployment contexts may differ from training conditions, resulting in outcomes that users find surprising or harmful (Amodei et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Gabriel, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Differential Apprehension Across AI Types\u003c/h2\u003e \u003cp\u003eA central theoretical premise of our work is that the degree of apprehension toward AI is not monolithic but varies as a function of the specific type of AI system under consideration. While the underlying construct of apprehension remains consistent across AI types, we argue that meaningfully different categories of AI evoke varying degrees of apprehension across these same dimensions due to differences in their purposes, user relationships, transparency, power dynamics, and potential for harm (Glikson and Woolley, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Zhang and Dafoe, \u003cspan citationid=\"CR92\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This differentiation reflects established findings in human-computer interaction research showing that user responses to technology depend critically on context, perceived agency, and the nature of the human-technology relationship (Sundar and Lee, \u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cem\u003eGeneral AI\u003c/em\u003e serves as an umbrella term for intelligent systems as a whole, ranging from simple smart gadgets for healthcare to complex recommendation systems (Triguero et al., \u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Apprehension toward general AI may reflect diffuse concerns about AI's societal trajectory and long-term consequences, including technological unemployment, erosion of human autonomy, and gradual destabilization of critical societal systems (Bucknall and Dori-Hacohen, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cem\u003eLarge language models\u003c/em\u003e are characterized by their ability to generate human-like text and perform complex language tasks (Brown et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). While apprehension toward LLMs is pronounced along the unpredictability dimension due to non-deterministic outputs and hallucinations (Ji et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), it also manifests as undesirability in contexts like education or social discourse, where reliance on machine-generated content may be perceived as socially unacceptable. Furthermore, concerns regarding training data and corporate power concentration evoke the implied malice dimension (Bender et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Sag, \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cem\u003ePersonal AI systems\u003c/em\u003e are designed to be adaptive and integrated with individual users' habits (Luger and Sellen, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Here, the undesirability dimension is specifically linked to social discomfort, such as the perceived \"awkwardness\" of interacting with assistants like Alexa or Siri in public or in front of others. While personalization introduces implied malice through privacy concerns and potential exploitation (Martin, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), it also triggers unpredictability regarding the opaque ways personal data might be processed (Gagrčin et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2026\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cem\u003eInstitutional AI\u003c/em\u003e encompasses systems deployed by organizations to monitor or evaluate individuals, such as face recognition at gates, traffic control, or hiring algorithms (Lee et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). These systems involve significant power asymmetries and often operate as \"black boxes\" (Burrell and Fourcade, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). We expect elevated apprehension across all dimensions: implied malice regarding organizational motives (Caton and Haas, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), unpredictability due to systemic opacity (Ananny and Crawford, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), and undesirability. The latter is particularly relevant as individuals are increasingly subjected to AI interactions in public or semi-public spaces, where the interaction may feel intrusive or socially uncomfortable (Bialy et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese four AI types capture theoretically meaningful variation in AI characteristics. General AI provides an umbrella; LLMs represent a cognitively focused instantiation; personal AI captures intimacy and data vulnerability concerns; and institutional AI reflects power-asymmetric, high-stakes contexts. By measuring apprehension across these contexts, we can examine both common psychological foundations and context-specific patterns of AI-related concern.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Measuring Technology-Related Apprehension\u003c/h2\u003e \u003cp\u003eResearch on negative affective responses to technology has produced several validated measurement instruments, each capturing distinct but related constructs. The Perceived Creepiness of Technology Scale (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) provides a validated instrument for assessing initial feelings of unease. However, it was primarily developed for interactive products and wearable gadgets that augment the user's body or daily environment, such as smart glasses and fitness trackers, rather than the autonomous decision-making logic of AI. The General Attitudes towards Artificial Intelligence Scale (Schepman and Rodway, \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) measures broad evaluative stances. As a general attitude scale, it does not differentiate between specific AI applications, which may limit its sensitivity to context-specific concerns. Similarly, the Artificial Intelligence Anxiety Scale (Wang and Wang, \u003cspan citationid=\"CR88\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) focuses on anxiety-related affect, such as job-related threats and the learning curve of using AI, but remains at a monolithic \"AI\" level. The Fear of Artificial Intelligence Scale (Corradi et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) advances the literature by operationalizing fear via a broad set of content domains, including job displacement and loss of human autonomy. While this scale provides a comprehensive measure of fear as a holistic emotion, our work departs from this by treating apprehension as a multi-dimensional construct. Unlike previous unidimensional measures, we deliver a granular analysis of the components (malice, undesirability, and unpredictability) allowing us to examine how the strength and degree of these dimensions fluctuate across different AI categories. The gap between existing instruments and the need for a nuanced measurement of AI apprehension is substantial. AI technologies exhibit autonomy and learning capabilities that distinguish them from conventional interactive gadgets. These properties suggest that while the fundamental psychological triggers identified by Woźniak et al. (\u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) remain relevant, their strength and expression require contextualization for AI. A conversational assistant, an institutional hiring algorithm, and a personalized health recommender do not just evoke fear, but distinct intensities of the identified apprehension components.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6. The Present Study\u003c/h2\u003e \u003cp\u003eThe present study adapted the items from the Perceived Creepiness of Technology Scale (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) to develop and validate four parallel scales measuring apprehension toward artificial intelligence across distinct AI types: general AI, large language models, personal AI, and institutional AI. The items were rephrased to refer specifically to apprehension toward AI, while preserving the underlying meaning of the original scale. For institutional AI in particular, wording was adapted to emphasise exposure to AI systems implemented by organisations. Where possible, we used parallel items across all AI types, making small wording adjustments only when needed for clarity and content.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methods","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.1. Participants\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eParticipants were recruited via the Prolific, a worldwide online platform for research which applies quality measures (www.prolific.com) and completed the survey on SurveyMonkey (www.surveymonkey.com). Data were collected between late September and October 2025. Eligibility criteria required participants to be aged 18-45 years, familiar with general, LMMs, personal AI and institutional AI, born in the UK or holding UK citizenship, self-identifying as British in terms of norms and culture, and speaking English as their native language. Participation was anonymous, and all respondents provided digital informed consent before accessing the questionnaires. Participants who completed the survey and passed attention checks received monetary compensation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eA total of 589 participants completed the survey. Following exclusion of those who failed attention check or providing monotonous responses, 559 cases contained complete data and were retained for analysis. Simulation-based evidence for small confirmatory factor models shows that when a measurement model comprises three correlated factors defined by a small number of indicators, samples of at least N = 460 are required to achieve acceptable power, low parameter bias, and proper solutions when standardised factor loadings are in the moderate range (= .50) (Wolf et al., 2013). Expressed on an item basis, this corresponds to \u0026ge; 57.5 participants per item for an 8-item instrument, providing a concrete numerical benchmark for sample size adequacy in short multifactor confirmatory factor analysis (CFA) models (Wolf et al., 2013). The achieved participant-to-item ratio in the present study (N = 559) was 69.9:1, exceeding this empirically derived requirement.\u003c/p\u003e\n\u003cp\u003eThe final sample had a mean age of 30.64 years (SD = 6.75), 50.3% males, with ages ranging from 19 to 42 years.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.2. Procedure and materials\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.1. Scale development\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eOur proposed apprehension scales were developed by adapting the item content of the Perceived Creepiness of Technology Scale (PCTS)\u0026nbsp;(Woźniak et al., 2021)\u0026nbsp;while retaining its three factors, Implied Malice, Undesirability, and Unpredictability. In the original measure, items are framed as direct evaluations of a certain AI products (for example, that its design is unethical, or would feel uneasy to use in public, or that its purpose is unclear). In the present scale, an introduction sentence was added \u0026nbsp;that requires respondents to indicate how much more would be needed for them to be able to endorse the corresponding positive state (for example, that those behind the products take morality seriously, that interacting with the products in public would feel normal and not uneasy, and that the products work as described and have clear purposes and goals) (Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAdding the introduction sentence has an advantage over direct evaluative wording because it targets a different part of the judgement process: the perception that what is currently in place exists but may not yet be sufficient. This reflects apprehension that is often expressed as caution under uncertainty, not as a fixed negative judgement. People can withhold full confidence or comfort without endorsing strong negative statements. For example, they may not claim that designers are immoral but still feel that more is needed before morality can be taken for granted. An \u0026ldquo;I think further conditions are necessary for me\u0026rdquo; format captures this intermediate position directly by measuring the extent of the perceived shortfall on each criterion, without requiring respondents to commit to strongly condemnatory language. This format also allows clearer interpretation of what respondents mean. Low agreement with a positive statement, or high agreement with a negative one, can reflect very different positions, such as mild hesitation, strong concern, lack of information, distrust of those deploying the system, or general scepticism. When responses are framed as direct evaluations, these positions are difficult to distinguish. By asking how much more is needed, the scale requires respondents to indicate a degree. It therefore distinguishes between those who believe only minor adjustments are necessary and those who believe substantial changes are required, even if both might select similar options on a standard agree-disagree item.\u003c/p\u003e\n\u003cp\u003eFor each scale, participants received the following instruction: \u0026ldquo;Think of [the AI class of interest: General AI, LLMs, Personal AI] products in general, the data they collect and use, your interaction with them through text, voice, image, or video, and the actions they take or the recommendations they make.\u0026rdquo;. For Institutional AI, the instruction was \u0026ldquo;Think of Institutional AI products in general, the data they collect and use, your exposure to them in public or private contexts, and the actions they take or the recommendations they make\u0026rdquo;.\u0026nbsp;Participants then responded to eight items introduced by the common stem \u0026ldquo;I think further conditions are necessary for me\u0026hellip;\u0026rdquo; using an 11-point scale (0 = \u0026ldquo;Nothing more needed\u0026rdquo;, 10 = \u0026ldquo;A great deal more needed\u0026rdquo;). The instruction text, item stem, and response scale were identical across all four versions of the scale, with only the referenced AI type varied. To ensure conceptual clarity, respondents were provided with a brief description and examples of each AI category before completing the scale. In the case of Personal AI, for example, this included products such as fitness trackers, smartwatches, health-monitoring devices, digital companions, and personal assistant applications. The final version of the scale (see Appendix 1-4) specifies that administration should be preceded by a short clarification of the relevant AI category, adapted as appropriate to the target population. The scale items are provided in Table 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003ePTCS items adapted to reflect each type of AI (General AI, Personal AI, Institutional AI, and LLM) apprehension\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eItem number\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSubscale\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eItem abbreviation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGeneral AI / LLMs/ Personal AI \u0026nbsp;(X )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInstitutional AI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 600px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003eI think further conditions are necessary for me\u0026hellip;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003eImplied malice\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eMOR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that the people who make X products take morality seriously.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that the people who design and put in place Institutional AI products take morality seriously.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eETH\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that the way X products are designed and run is ethical.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that the way Institutional AI products are implemented and operated is ethical.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003eUndesirability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eGOA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel normal and less odd when interacting with X products in public.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel normal and less odd in public spaces where Institutional AI products are installed.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eSUR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to interact with X products in public without feeling uneasy.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to experience Institutional AI products in public without feeling uneasy.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eEXP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that interacting with X products in public is natural rather than awkward.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that the presence of Institutional AI products in society is natural rather than disturbing.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 119px;\"\u003e\n \u003cp\u003eUnpredictability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eNOR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that [X] products work the way they are described and expected.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that Institutional AI products work the way they are described and expected.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eUNE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel sure about the purpose of use of X products, both for me and for the businesses behind them.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel sure about the purpose of use of Institutional AI products, both for me and for the institutions behind them.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003eNAT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 150px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that X products have clear goals, both for me and for the businesses behind them.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 174px;\"\u003e\n \u003cp\u003e\u0026hellip;to feel that Institutional AI products have clear goals, both for me and for the institutions behind them.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe original PCTS scale employed a 7-point Likert response format; in the present study, items were administered on an 11-point (0-10) scale to capture greater variation in participants\u0026rsquo; judgements. Evidence from rating-scale research indicates that broader response ranges provide increased measurement sensitivity compared with shorter Likert-type scales and are often preferred by respondents, as they allow more fine-grained evaluations without increasing response burden (Jebb et al., 2021; Revilla et al., 2014). In addition, empirical comparisons of scales with different numbers of response categories have shown that respondents tend to prefer scales with a larger number of points, particularly formats around 10 response options (Preston and Colman, 2000).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.2. Survey design and procedure\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA cross-sectional survey design was employed. The survey comprised a demographic block, four AI apprehension blocks (General AI, LLMs, Personal AI, and Institutional AI), followed by blocks assessing attitudes towards AI and personality traits (Big Five). At the start of each AI apprehension block, respondents were given a brief description of the relevant AI type for that block. Attention-check items were embedded in the survey, each instructing respondents to select a specific response option. Cases with failed attention checks were excluded from the analyses.\u003c/p\u003e\n\u003cp\u003eThe study was accessed via a link distributed through the Prolific platform and was completed using SurveyMonkey. Before starting the survey, informed consent was obtained, and respondents were informed that participation was voluntary and that they could withdraw at any time without penalty.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.3. Measures\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.2.1. Demographics.\u003c/em\u003e We asked participants about their age, gender, and education level. In addition, we assessed their religious beliefs to examine whether apprehension toward different types of artificial intelligence varies as a function of religiosity. Religion is consistently linked to differences in moral attitudes, perceptions of control, and responses to technological change (Alexander, 2020; Hommel, 2023; McPhetres and Zuckerman, 2018). Including this variable therefore allows us to test whether the scales are associated with individual differences that extend beyond basic demographic characteristics.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.2.2. AI Apprehension scales\u003c/em\u003e.\u0026nbsp;Each scale (General AI, Personal AI, LLM, and Institutional AI) consisted of eight items and followed the original factor structure. The scale distinguishes three dimensions of apprehension:\u0026nbsp;Implied Malice\u0026nbsp;(2 items assessing concerns about the morality and ethical orientation of those who develop or deploy AI),\u0026nbsp;Undesirability\u0026nbsp;(3 items assessing discomfort with AI in public or social contexts), and\u0026nbsp;Unpredictability\u0026nbsp;(3 items assessing confidence in AI behaviour and clarity of its goals) (Table 1). Items were rated on an 11-point scale ranging from 0 (\u003cem\u003eNothing more needed\u003c/em\u003e) to 10 (\u003cem\u003eA great deal more needed\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.2.3. Big-Five Inventory (BF-10)\u003c/em\u003e. Personality traits were measured using the 10-item Big Five Inventory (Rammstedt and John, 2007), which assesses the five personality dimensions: Openness (associated with intellectual curiosity and aesthetic appreciation), Conscientiousness (associated with dutifulness and adherence to social norms), Extraversion (associated with self-confidence), Agreeableness (associated with harmonious relations with others) and Neuroticism (associated with high levels of negative emotions such as anxiety).\u0026nbsp;Each trait is measured using two items, one positively worded and one negatively worded, rated on a five-point Likert scale from 1 (\u003cem\u003estrongly disagree\u003c/em\u003e) to 5 (\u003cem\u003estrongly agree\u003c/em\u003e). The BFI-10 has been validated in large-scale survey research and provides an efficient assessment of personality traits while minimising respondent burden\u0026nbsp;(Costa and McCrae, 1992).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.2.2.4. Attitudes toward artificial intelligence (ATAI).\u003c/em\u003e Attitude Towards Artificial Intelligence (ATAI) scale, a brief five-item measure with a two-factor structure comprising Acceptance (2 items) and Fear (3 items) (Sindermann et al., 2021). Items were rated on the 11-point response format ranging from 0 (\u0026ldquo;strongly disagree\u0026rdquo;) to 10 (\u0026ldquo;strongly agree\u0026rdquo;). AI Acceptance is operationalised using the items \u0026ldquo;I trust artificial intelligence\u0026rdquo; and \u0026ldquo;Artificial intelligence will benefit humankind\u0026rdquo;, whereas Fear includes \u0026ldquo;I fear artificial intelligence\u0026rdquo;, \u0026ldquo;Artificial intelligence will destroy humankind\u0026rdquo;, and \u0026ldquo;Artificial intelligence will cause many job losses\u0026rdquo;. Subscale scores are computed by summing scores on the relevant items, with higher scores indicating greater acceptance of AI or greater fear of AI, respectively.\u003c/p\u003e\n\u003cp\u003eIn this study, the ATAI scale was administered with respect to four AI types: General artificial intelligence, LLMs, Personal AI, and Institutional AI. For each administration, the AI type was specified in the item stem (e.g. attitudes toward LLMs), while the item content and response format remained unchanged. Acceptance and Fear subscale scores were calculated independently for each AI type.\u0026nbsp;Internal consistency was acceptable across AI types. Cronbach\u0026rsquo;s \u0026alpha; for the Acceptance subscale ranged from \u0026alpha; = .81 to .84, and for the Fear subscale from \u0026alpha; = .73 to .76, across the four AI types.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3 Data analysis\u003c/em\u003e\u003c/strong\u003e\u003cem\u003e.\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.1. Monotonicity analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Monotonicity analysis was performed prior data analysis and was assessed by examining whether conditional item means increased as a function of the rest score. Adjacent violations were defined as decreases in the conditional mean between successive rest-score groups. For each item, the total number of adjacent violations and the maximum observed adjacent drop in conditional means were recorded to quantify the extent and magnitude of local non-monotonicity. Across all scales, items showed positive monotonic associations with their rest scores, with Kendall\u0026rsquo;s \u0026tau; ranging from 0.44 to 0.69 and Spearman\u0026rsquo;s \u0026rho; from 0.57 to 0.83. Adjacent monotonicity violations were observed for all scales, reflecting the fine-grained rest-score partitioning, but maximum adjacent drops were limited in magnitude (1-3 points on a 0-10 scale). Institutional and Personal AI scales showed the strongest monotonic structure, followed by the LLM and General AI scales. Overall, departures from monotonicity were not systematic, supporting the ordinal coherence of all four scales (see Supplementary Material 1 for complete report).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.2. Descriptive statistics\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFor each of the four AI apprehension scales item-level descriptive statistics were computed, including means and standard deviations. To examine response distributions, frequency tables were generated for each response option (0-10), and item distributions were inspected using histograms with unit-width bins covering the full-scale range (see Supplementary Material 2).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.3. Confirmatory factor analysis (CFA)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eConfirmatory factor analyses (CFA) were conducted separately for the four AI apprehension scales to test a hypothesised three-factor structure. For each scale, items were specified to load on three correlated latent factors reflecting Implied Malice, Undesirability, and Unpredictability. Models were estimated using maximum likelihood with robust standard errors (MLR) to account for potential non-normality in item responses. Model fit was evaluated using multiple indices, including the comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation (RMSEA) with 90% confidence intervals, and the standardised root mean square residual (SRMR). In line with current recommendations, model evaluation was based on the joint consideration of comparative and residual-based indices rather than reliance on any single fit statistic (Kenny et al., 2015).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.4. Reliability, construct and discriminant validity\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eReliability was evaluated separately for each factor within each AI type (General AI, LLMs, Personal AI, and Institutional AI), using a combination of classical internal consistency indices and model-based reliability estimates derived from the confirmatory factor analysis. For all factors, Cronbach\u0026rsquo;s alpha (\u0026alpha;) was computed as a conventional index of internal consistency. Given the short length of the scales, \u0026alpha; was calculated from the covariance matrix rather than item frequency tables to avoid distortions associated with discrete response distributions. In addition, the mean inter-item correlation was reported for all factors, as this index is particularly informative for short scales and provides a direct indication of item homogeneity. For the Implied Malice factor, which comprised two items in each AI context, reliability was primarily evaluated using the Spearman-Brown coefficient, which is the recommended reliability estimate for two-item measures.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo complement classical reliability indices, model-based reliability was assessed using CFA-derived estimates. Composite reliability (CR) was computed from standardised factor loadings and residual variances using the compRelSEM() function in \u003cem\u003esemTools\u003c/em\u003e to estimate the proportion of variance in the observed indicators attributable to the latent construct. In addition, average variance extracted (AVE) was calculated to quantify the proportion of indicator variance captured by the latent factor relative to measurement error; and interpreted as an index of convergent validity. Finally, construct replicability was evaluated using H (MaxR(H)), which reflects the degree to which a latent variable is defined by its indicators. Higher values indicate stronger latent definition and greater stability of the construct.\u003c/p\u003e\n\u003cp\u003eDiscriminant validity of the three latent factors was evaluated separately for each AI type. First, the Fornell-Larcker criterion was applied by comparing the square root of AVE for each factor with its correlations with other factors. Discriminant validity was considered supported when the square root of AVE exceeded all corresponding inter-factor correlations. Second, competing CFA models were also estimated to provide a direct test of discriminability. For each AI type, the hypothesised three-factor model was compared against a one-factor model and two-factor models. Model comparisons were conducted using scaled chi-square difference tests appropriate for the robust maximum likelihood estimator (MLR).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.5. Shared variance between apprehension toward four AI types\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis analysis was conducted to determine the extent to which apprehension across different AI types reflects a common underlying tendency as opposed to AI type variation. Overlap between AI types was quantified using total apprehension scores computed separately for General AI, LLMs, Personal AI, and Institutional AI, with scores treated as continuous indicators of overall apprehension within each context. Pairwise associations between total scores were estimated using Pearson correlations, with Spearman rank correlations calculated as a robustness check for potential departures from normality. Overlap was operationalised as shared variance, defined as the squared Pearson correlation coefficient (\u003cem\u003er\u0026sup2;\u003c/em\u003e) and expressed as a percentage, representing the proportion of variance in apprehension that is common to each pair of AI types. Uncertainty around these estimates was quantified using 95% confidence intervals derived from Fisher\u0026rsquo;s \u003cem\u003ez\u003c/em\u003e transformation based on pairwise complete observations. In addition, intraclass correlation coefficients (ICC[A,1]) (Liljequist et al., 2019) were computed to assess the degree of absolute agreement in overall apprehension levels across AI types. In addition, we performed the same analysis between the same factors in each scale.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.6. Convergent validity\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe aim of the convergent validity analyses was to establish whether apprehension toward artificial intelligence shows systematic associations with conceptually related evaluative and dispositional constructs. Convergent validity was assessed using three theoretically grounded indicators: ATAI fear of AI, ATAI acceptance of AI, and the personality trait of openness to experience.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFear of AI\u003c/em\u003e represents an affective response characterised by anxiety, perceived threat, and concern regarding the societal, occupational, and personal implications of AI. Empirical research consistently demonstrates that fear of AI is associated with heightened negative affect, risk perception, and avoidance-oriented evaluations of AI technologies (Schepman and Rodway, 2023; Yang and Sundar, 2025). Because AI apprehension similarly reflects a negatively valenced emotional orientation towards AI, grounded in concern and unease rather than neutral uncertainty, a positive association between apprehension and AI-related fear is theoretically expected. Accordingly, we hypothesised that AI apprehension would be positively associated with fear of artificial intelligence across all AI types (\u003cstrong\u003eH1\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAcceptance of AI\u003c/em\u003e captures a general evaluative stance reflecting perceived usefulness, desirability, and willingness to engage with AI systems. Within attitude and technology acceptance frameworks, acceptance is consistently positioned in opposition to threat-focused responses, with higher acceptance associated with lower perceived risk, reduced concern, and diminished resistance to AI adoption (Glikson and Woolley, 2020; Schepman and Rodway, 2023). Given that AI apprehension reflects elevated concern and negative evaluation, it is conceptually incompatible with acceptance-oriented attitudes. We therefore hypothesised that AI apprehension would be negatively associated with AI acceptance across all four AI types (\u003cstrong\u003eH2\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eOpenness to experience\u003c/em\u003e was included as a dispositional indicator relevant to convergent validity. Openness reflects cognitive flexibility, curiosity, and tolerance for novelty, and has been shown to predict more favourable evaluations of emerging technologies, including AI (Stein et al., 2024). Opened individuals are more likely to approach technological innovation with interest, suggesting lower apprehension in response to AI (Grassini et al., 2025; Sindermann et al., 2021). Although openness is not an AI-specific construct, its established association with receptivity to novel systems (Svendsen et al., 2013) provides a theoretical basis for a negative association with AI apprehension. We therefore expected higher openness to be associated with lower AI apprehension across AI types (\u003cstrong\u003eH3\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eFor each AI type, two complementary analyses were employed. First, zero-order associations were estimated using Pearson product-moment correlations, with Spearman rank-order correlations computed in parallel as a robustness check against potential non-normality and the influence of outliers. Pearson correlations were used as the primary metric, and 95% confidence intervals were derived using Fisher\u0026rsquo;s \u003cem\u003ez\u003c/em\u003e transformation. Second, for each AI type, apprehension scores were regressed on ATAI acceptance and ATAI fear using ordinary least squares estimation with HC3. Openness to experience was included alongside AI acceptance and fear to examine whether convergent associations remained stable when accounting for broader dispositional openness to novelty.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.8. Divergent Validity\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe purpose of the divergent validity analysis was to demonstrate that AI apprehension is empirically distinct from broad personality\u0026nbsp;dispositions and therefore cannot be reduced to general individual differences unrelated to threat-based responses to AI. For this analysis, extraversion\u0026nbsp;and\u0026nbsp;conscientiousness\u0026nbsp;were selected within the Five-Factor Model as they are\u0026nbsp;not theoretically linked to threat appraisal, anxiety, or negative affect toward external systems, including technology.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eExtraversion\u003c/em\u003e reflects sociability, assertiveness, and positive emotionality in interpersonal contexts (Costa and McCrae, 1992). Empirical research on technology attitudes consistently shows that extraversion is, at most, weakly and inconsistently related to technology use and apprehension (McElroy et al., 2007; Svendsen et al., 2013). \u003cem\u003eConscientiousness\u003c/em\u003e reflects self-discipline and goal-directed control (Costa \u0026amp; McCrae, 1992). While conscientiousness predicts rule adherence, task persistence, and reliability, it is not associated with emotional threat responses or evaluations (Kotov et al., 2010). Conscientiousness also shows inconsistent or negligible associations with evaluative attitudes, and when effects are observed they typically relate to compliance or usage behaviour rather than fear or concern (Devaraj et al., 2008). Demonstrating weak or null associations between AI apprehension and extraversion and conscientiousness will support the hypothesis that our apprehension scales are not artefacts of general personality structure (\u003cstrong\u003eH4\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003eDivergent validity was assessed using the same statistical framework as convergent validity to ensure analytic symmetry.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp;3.3.9. Network analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eNetwork models were estimated separately for each AI type (General AI, LLMs, Personal AI, Institutional AI) to examine the structure of associations among the eight apprehension items within each AI type. Networks were specified as regularised partial correlation networks, such that each edge represents the association between two items after conditioning on all remaining items. Networks were estimated using the graphical LASSO with Extended Bayesian Information Criterion (EBIC) model selection, with the tuning parameter fixed at \u0026gamma; = 0.50 to balance sparsity and sensitivity. For each AI type, the resulting weighted adjacency matrix of partial correlations was extracted. Edge accuracy was evaluated using non-parametric bootstrapping (1,000 resamples), and network robustness was assessed via case-dropping bootstrap correlation stability analysis of node strength (1,000 resamples). Additional centrality indices (closeness and betweenness) were also computed.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3.3.10. Additional analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eDemographic correlates of the four AI apprehension scales were examined to provide practitioners and researchers with preliminary estimates of how AI apprehension may vary across gender, age, education, and religious beliefs. Identifying such patterns may offer practical insight into whether certain population groups report systematically higher or lower apprehension, which may inform policy considerations, and future applied research. Separate regression models were estimated for each scale, and detailed results are reported in Supplementary Material 4.\u003c/p\u003e"},{"header":"4. Results","content":"\u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Descriptive statistics of AI Apprehension scales\u003c/h2\u003e \u003cp\u003eDescriptive statistics for the four AI apprehension scales are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDescriptive statistics for AI apprehension items across four scales\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFactors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eItem abbreviation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGeneral AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLLM AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePersonal AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eInstitutional AI\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eMean (SD)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eMean (SD)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eMean (SD)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cem\u003eMean (SD)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eImplied Malice\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMOR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.499 (2.808)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.737 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.878 (2.809)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6.499 (2.809)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eETH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.862 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.288 (2.559)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.492 (2.559)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6.352 (2.559)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eUndesirability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGOA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.200 (2.967)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.906 (2.945)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.317 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.907 (2.745)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSUR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6.159 (2.816)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.832 (2.968)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.278 (2.817)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.879 (2.817)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEXP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.900 (2.795)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.598 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6.129 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.742 (2.745)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eUnpredictability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNOR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.055 (2.745)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.995 (2.983)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.602 (2.982)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.004 (2.983)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUNE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.123 (2.817)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.042 (2.744)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.745 (2.744)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5.061 (2.744)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNAT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.057 (2.746)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.979 (2.817)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5.714 (2.817)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.830 (2.817)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Confirmatory factor analysis of AI Apprehension scales\u003c/h2\u003e \u003cp\u003eConfirmatory factor analyses were conducted to test the hypothesised three-factor structure across the four AI apprehension scales. For the General AI scale, the three-factor model showed very good fit to the data (CFI\u0026thinsp;=\u0026thinsp;0.99; TLI\u0026thinsp;=\u0026thinsp;0.99; SRMR\u0026thinsp;=\u0026thinsp;0.03). For the LLMs scale, model fit was also good, with high comparative fit indices (CFI\u0026thinsp;=\u0026thinsp;0.99; TLI\u0026thinsp;=\u0026thinsp;0.98) and low residual misfit (SRMR\u0026thinsp;=\u0026thinsp;0.04). The Personal AI scale demonstrated acceptable to good fit, with strong comparative fit indices (CFI\u0026thinsp;=\u0026thinsp;0.98; TLI\u0026thinsp;=\u0026thinsp;0.97) and low residual error (SRMR\u0026thinsp;=\u0026thinsp;0.04). Similarly, the Institutional AI scale showed good fit, with CFI and TLI values above conventional cut-offs (CFI\u0026thinsp;=\u0026thinsp;0.98; TLI\u0026thinsp;=\u0026thinsp;0.97) and a low SRMR (0.03). Across all four scales, incremental and residual-based indices therefore consistently supported the adequacy of the hypothesised three-factor structures.\u003c/p\u003e \u003cp\u003eRMSEA was 0.057 (90% CI [0.038, 0.076]) for the General AI scale, 0.085 (90% CI [0.067, 0.103]) for the LLMs scale, 0.109 (90% CI [0.092, 0.127]) for the Personal AI scale, and 0.093 (90% CI [0.076, 0.111]) for the Institutional AI scale. While the latter three values exceed conventional close-fit thresholds, this pattern is consistent with documented sensitivity of RMSEA to model degrees of freedom and parsimony in relatively small-df CFA models (Hayduk and Glaser, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2000\u003c/span\u003e; Kenny et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Marsh et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). Importantly, the elevated RMSEA values were not accompanied by deterioration in CFI, TLI, or SRMR. In line with recommendations against rigid reliance on a single fit index (Lai and Green, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Montoya and Edwards, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) model evaluation was therefore based on the overall convergence of fit evidence rather than RMSEA in isolation. Based on this our results indicate that the tested three-factor structure provides an adequate and substantively meaningful representation of apprehension toward General, LLM, Personal, and Institutional AI.\u003c/p\u003e \u003cp\u003eAcross all four AI types, items loaded strongly on their hypothesised latent factors. Standardised loadings were uniformly high, ranging from 0.78 to 0.98, with the strongest loadings consistently observed for Undesirability items, followed by Implied Malice, and slightly lower, but still substantial, loadings for Unpredictability items (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The similarity of loading patterns across contexts indicates a stable and coherent three-factor structure that generalises across different types of AI.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStandardised factor loadings (λ) for the three-factor CFA across AI contexts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFactor\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eItem abbreviation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGeneral AI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLLM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePersonal AI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eInstitutional AI\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eImplied Malice\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMOR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.893\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.904\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.924\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eETH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.885\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.936\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.962\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eUndesirability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNOR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.963\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.907\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUNE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.975\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.966\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNAT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.957\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.907\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eUnpredictability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGOA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.777\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.776\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.831\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.856\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSUR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.958\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGOA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.881\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.953\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.949\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e4.3. Shared variance between AI apprehension across AI types\u003c/h2\u003e \u003cp\u003eShared-variance analyses indicated partial overlap between total score apprehension across AI types. The greatest overlap was observed between General AI and LLM apprehension, which shared 67.9% of variance, followed by LLM and Personal AI apprehension (47.4%) and General and Personal AI apprehension (40.7%). In contrast, overlap Institutional AI showed consistently lower overlap with other scales, with shared variance ranging from 27.9% to 30.4%, indicating that apprehension toward Institutional AI is less strongly aligned with apprehension for other AI types. Agreement across AI types was further evaluated using a two-way random-effects intraclass correlation coefficient for absolute agreement. The ICC indicated moderate agreement across the four total apprehension scores (ICC[A,1] = .59, 95% CI [.53, .64], \u003cem\u003eF\u003c/em\u003e(558, 307)\u0026thinsp;=\u0026thinsp;7.41, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001), suggesting that individuals\u0026rsquo; relative levels of apprehension are partly consistent across AI types, while also showing AI type-specific variation.\u003c/p\u003e \u003cp\u003eWe also examined shared variance for each factor (Implied Malice, Undesirability, and Unpredictability) across the four AI types. This analysis showed that a substantial proportion of variance is not shared across AI types ranging from 6.2% to 62.1% (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e4.4. Reliability, construct and discriminant validity of AI Apprehension scales\u003c/h2\u003e \u003cp\u003eAcross all four AI types, the three apprehension factors demonstrated consistently high internal consistency, model-based reliability, and construct quality.\u003c/p\u003e \u003cp\u003eFor Implied Malice, which comprised two items in each scale, inter-item correlations were strong (r\u0026thinsp;=\u0026thinsp;0.76\u0026ndash;0.89), with Spearman-Brown coefficients ranging from 0.86 to 0.94 and Cronbach\u0026rsquo;s α values from 0.86 to 0.94, indicating excellent reliability for a short two-item construct. Composite reliability estimates were identical to α values, reflecting the equivalence of these indices in two-indicator models. For the three-item Undesirability and Unpredictability factors, Cronbach\u0026rsquo;s α values ranged from 0.89 to 0.97, with mean inter-item correlations between 0.73 and 0.92, indicating strong internal coherence without evidence of attenuation due to scale brevity. Composite reliability was uniformly high across AI contexts (CR\u0026thinsp;=\u0026thinsp;0.90\u0026ndash;0.97), closely aligning with α and supporting the stability of the latent constructs. Average variance extracted exceeded 0.73 for all factors, indicating that a substantial proportion of item variance was captured by the intended latent variables.\u003c/p\u003e \u003cp\u003eConstruct replicability, indexed by H (MaxR(H)), was moderate to high across all factors and contexts (H\u0026thinsp;=\u0026thinsp;0.56\u0026ndash;0.81). Replicability was consistently highest for Undesirability, particularly for the LLM and Personal AI (H\u0026thinsp;=\u0026thinsp;0.80), indicating strong latent definition. Unpredictability showed moderate to good replicability across contexts (H\u0026thinsp;=\u0026thinsp;0.60\u0026ndash;0.70), while Implied Malice demonstrated moderate replicability (H\u0026thinsp;=\u0026thinsp;0.56\u0026ndash;0.68), consistent with its shorter two-item composition (see Supplementary Material 3 for detail).\u003c/p\u003e \u003cp\u003eDiscriminant validity of the three latent factors was evaluated using multiple complementary criteria. Inter-factor correlations remained well below conventional thresholds for construct redundancy across all AI types (all r \u0026lt; .82). The Fornell-Larcker criterion was satisfied for all factor pairs, with the square root of AVE for each factor exceeding its correlations with other factors. Crucially, three-factor CFA models fit the data significantly better than one-factor and all two-factor alternative models across AI types (all scaled χ\u0026sup2; difference tests p \u0026lt; .001), providing strong evidence that the three factors represent empirically distinct dimensions rather than a single undifferentiated construct.\u003c/p\u003e \u003cp\u003eThese findings indicate that all three factors show excellent reliability, satisfactory construct validity and excellent discriminant validity across General, LLM, Personal, and Institutional AI apprehension, supporting their use in latent variable modelling and comparative analyses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e4.5. Convergent and divergent validity of AI Apprehension scales\u003c/h2\u003e \u003cp\u003eConvergent validity results are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e5\u003c/span\u003e. Across all four AI types, ATAI acceptance and ATAI fear showed statistically significant associations with AI apprehension in the expected directions, with acceptance negatively related and fear positively related to apprehension under HC3-robust estimation.\u003c/p\u003e \u003cp\u003eModel-level robust Wald tests indicated that the convergent predictors jointly explained significant variance in apprehension for every AI type. For the acceptance-fear models (M1), Wald F statistics ranged from 36.99 (General AI) to 67.86 (Institutional AI), all \u003cem\u003ep\u003c/em\u003e \u0026lt; .001. When openness was added (M2), joint effects remained significant for all AI types (Wald F range\u0026thinsp;=\u0026thinsp;33.72\u0026ndash;53.18, all \u003cem\u003ep\u003c/em\u003e \u0026lt; .001), indicating that the combined contribution of acceptance, fear, and openness was robust to heteroskedasticity.\u003c/p\u003e \u003cp\u003eThe acceptance-fear models accounted for 18.7\u0026ndash;26.9% of variance in apprehension across AI types, with the highest values observed for Institutional AI. Adding openness to the acceptance-fear models produced consistent increases in explained variance across all AI types. Specifically, R\u0026sup2; increased by 0.024 for General AI (from 0.187 to 0.211), 0.017 for LLMs (from 0.201 to 0.219), 0.006 for Personal AI (from 0.201 to 0.207), and 0.011 for Institutional AI (from 0.269 to 0.279).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConvergent validity of AI apprehension (robust HC3 regressions)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePredictor\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eb\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eβ\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ep\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGeneral AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAcceptance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[-1.33, -0.50]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFear\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.39, 0.89]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOpenness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.66, 1.87]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLLMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAcceptance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[-1.01, -0.14]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFear\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.79, 1.32]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOpenness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.54, 1.92]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePersonal AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAcceptance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[-1.19, -0.24]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.003\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFear\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.83, 1.43]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOpenness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.05, 1.57]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.036\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInstitutional AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAcceptance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[-1.42, -0.55]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFear\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.80, 1.39]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOpenness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[0.32, 1.81]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eDivergent validity was supported by consistently weak and non-significant associations between AI apprehension and extraversion and conscientiousness across all AI types under HC3-robust regression. For extraversion, coefficients were positive but non-significant for all AI types: General AI (b\u0026thinsp;=\u0026thinsp;0.95, 95% CI [-0.21, 2.12], p = .108), LLM (b\u0026thinsp;=\u0026thinsp;1.09, 95% CI [-0.26, 2.45], p = .113), Personal AI (b\u0026thinsp;=\u0026thinsp;0.59, 95% CI [-0.91, 2.10], p = .438), and Institutional AI (b\u0026thinsp;=\u0026thinsp;1.20, 95% CI [-0.32, 2.72], p = .123). For conscientiousness, coefficients were negative and non-significant across all AI types: General AI (b = -0.39, 95% CI [-1.23, 0.46], p = .369), LLM (b = -0.43, 95% CI [-1.33, 0.48], p = .356), Personal AI (b = -0.51, 95% CI [-1.53, 0.52], p = .333), and Institutional AI (b = -0.69, 95% CI [-1.70, 0.33], p = .186). The uniformity of these null findings under heteroskedasticity-consistent estimation provides evidence of divergent validity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e4.6. Network analysis of AI Apprehension scales\u003c/h2\u003e \u003cp\u003eMean absolute edge weights ranged from 0.165 to 0.185 across networks (maximum difference\u0026thinsp;=\u0026thinsp;0.020), and median absolute edge weights ranged from 0.067 to 0.089 (maximum difference\u0026thinsp;=\u0026thinsp;0.022) (Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e6\u003c/span\u003e). These narrow ranges indicate limited variation in global connectivity strength across the four AI networks. The relative difference between the lowest and highest mean absolute edge weights was approximately 12% (0.020/0.165), indicating limited variation in global association strength.\u003c/p\u003e \u003cp\u003eCase-dropping bootstrap analyses indicated acceptable centrality stability for all four networks. CS (strength) coefficients exceeded the recommended threshold of 0.50 for General AI (0.594), LLMs (0.594), Personal AI (0.517), and Institutional AI (0.673), indicating that strength centrality estimates were sufficiently stable to permit interpretation across all four networks (Epskamp et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eGlobal connectivity indices and strength centrality stability across AI apprehension networks\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEdges\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDensity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean absolute edge\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMedian absolute edge\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMaximum absolute edge\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCS (strength)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGeneral AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.786\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.169\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.071\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.607\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.594\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLLMs\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.167\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.067\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.709\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.594\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePersonal AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.750\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.185\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.089\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.731\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.517\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInstitutional AI\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.165\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.068\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.718\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.673\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cem\u003eNote\u003c/em\u003e: Edges\u0026thinsp;=\u0026thinsp;number of non-zero partial correlation coefficients retained following EBICglasso regularisation (γ\u0026thinsp;=\u0026thinsp;0.50). Density\u0026thinsp;=\u0026thinsp;proportion of all possible edges retained in the estimated network. Mean, median, and maximum values refer to the distribution of absolute edge weights within each network. CS (strength) = case-dropping bootstrap correlation stability coefficient for strength centrality. CS values \u0026ge; .50 indicate sufficient stability for interpretation\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAcross all four networks, the strongest edges were consistently within factors, supporting the intended three-factor structure of the scales. Implied Malice was characterised by a strong association between MOR and ETH in all networks, with large positive edge weights (General AI\u0026thinsp;=\u0026thinsp;0.607; LLM\u0026thinsp;=\u0026thinsp;0.628; Personal AI\u0026thinsp;=\u0026thinsp;0.665; Institutional AI\u0026thinsp;=\u0026thinsp;0.718). No other Implied Malice edges approached comparable magnitude in any network. Unpredictability showed robust internal coupling across all AI types. The strongest edge consistently linked SUR and GOA (General AI\u0026thinsp;=\u0026thinsp;0.559; LLM\u0026thinsp;=\u0026thinsp;0.709; Personal AI\u0026thinsp;=\u0026thinsp;0.731; Institutional AI\u0026thinsp;=\u0026thinsp;0.657). Additional positive associations were observed between EXP and SUR (range\u0026thinsp;=\u0026thinsp;0.191\u0026ndash;0.243) and between EXP and GOA (range\u0026thinsp;=\u0026thinsp;0.110\u0026ndash;0.227), indicating a tightly integrated Unpredictability cluster in every network.\u003c/p\u003e \u003cp\u003eUndesirability items were also strongly interconnected in every network. The association between NOR and UNE was consistently strong (General AI\u0026thinsp;=\u0026thinsp;0.556; LLM\u0026thinsp;=\u0026thinsp;0.445; Personal AI\u0026thinsp;=\u0026thinsp;0.590; Institutional AI\u0026thinsp;=\u0026thinsp;0.585). Similarly, UNE and NAT showed substantial positive associations (General AI\u0026thinsp;=\u0026thinsp;0.468; LLM\u0026thinsp;=\u0026thinsp;0.548; Personal AI\u0026thinsp;=\u0026thinsp;0.458; Institutional AI\u0026thinsp;=\u0026thinsp;0.494). Positive links were also observed between NOR and NAT (range\u0026thinsp;=\u0026thinsp;0.134\u0026ndash;0.374), indicating coherent clustering of public-discomfort items across AI types (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTaken together, the consistent dominance of within-factor edges and the replication of the strongest connections across AI types indicate a similar internal organisation of apprehension components across networks, despite minor variation in absolute edge magnitude.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eThe present study responds to two limitations in existing measures of public responses to artificial intelligence. First, prior research has often relied on global assessments of AI attitudes, acceptance, trust, or fear, without distinguishing between types of AI that differ in autonomy, proximity to the user, and institutional authority. Second, apprehension to AI has rarely been measured directly and it has been inferred from broader attitude or trust scales (Grassini, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; McGrath et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Rahman et al., \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). To address these limitations, the present study introduces instruments that assess apprehension explicitly and separately toward distinct AI types (General AI, LLMs, Personal AI, and Institutional AI). Adopted from the Perceived Creepiness of Technology Scale (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) the new scales measure three theoretically grounded dimensions of implied malice, undesirability, and unpredictability for each AI type. This approach allows to assess AI apprehension as a form of concern that does not necessarily imply avoidance or rejection, but that varies according to the context in which AI is deployed. In doing so, these instruments provide a basis for empirical comparisons of AI-related concerns across applications, populations, and policy settings.\u003c/p\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Psychometric properties and theoretical implications\u003c/h2\u003e \u003cp\u003eThe four AI apprehension scales show satisfactory psychometric properties when evaluated against standard criteria used in scale development and validation.\u003c/p\u003e \u003cp\u003eModel-based indices of construct quality provide additional support for interpreting the three dimensions as well-defined latent variables. Average variance extracted values indicate that, for each factor, a substantial proportion of item variance is attributable to the intended construct, supporting interpretation each scale at the latent level (Fornell and Larcker, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1981\u003c/span\u003e). Construct replicability estimates indicate that the dimensions are adequately defined by their indicators and are likely to reproduce under similar measurement conditions, despite the brevity of the subscales.\u003c/p\u003e \u003cp\u003eInternal consistency of our scales was evaluated using multiple complementary indices, consistent with recommendations to avoid reliance on Cronbach\u0026rsquo;s alpha alone (Sijtsma, \u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). The strong inter-item correlations and Spearman-Brown coefficients observed for the two-item Implied Malice subscale are consistent with established guidance on the interpretation of short scales (Eisinga et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). The convergence between alpha and composite reliability, together with AVE values above recommended thresholds support interpretation of the subscales as well-defined latent constructs (Raykov, \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e1997\u003c/span\u003e). Important, our discriminant validity analyses indicate that Implied Malice, Undesirability, and Unpredictability should not be treated as interchangeable components: although the dimensions share variance, collapsing them resulted in systematic loss of model fit.\u003c/p\u003e \u003cp\u003eThe external validation results show that AI apprehension is systematically related to conceptually relevant evaluations of AI while remaining distinct from broader dispositional characteristics. Across AI types, apprehension was positively associated with fear of AI and negatively associated with acceptance of AI. Across AI types, apprehension was positively associated with fear of AI and negatively associated with acceptance of AI. This pattern supports the external validity of the scales, as apprehension relates to theoretically relevant constructs in expected directions while remaining empirically distinct from them. The positive association with fear indicates shared affective content, whereas the negative association with acceptance reflects an opposing evaluative orientation. However, apprehension was not reducible to either construct. Its moderate association with fear suggests overlap without equivalence, and its non-equivalence with acceptance indicates that apprehension cannot be interpreted simply as the absence of positive evaluation. Within tripartite models of attitudes (Kaiser and Wilson, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), our results are consistent with apprehension reflecting a distinct affective component of evaluation. Moreover, the coexistence of apprehension and acceptance aligns with theoretical accounts of attitudinal ambivalence (Cacioppo et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e1997\u003c/span\u003e), which propose that complex technologies can evoke simultaneous positive and negative evaluations. Together, our findings support the interpretation of AI apprehension as a specific affective response, broader and less threat-focused than fear, yet not reducible to general negativity or lack of acceptance.\u003c/p\u003e \u003cp\u003eAs expected, AI apprehension in the present study was largely independent of broad personality traits, supporting the construct validity of our scales. This finding is consistent with recent research showing that personality explains only a limited proportion of variance in affective responses to AI, with evaluations influenced more strongly by perceived characteristics and societal implications of AI systems (Schepman and Rodway, \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Stein et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRecent theoretical work consider AI not as a single object of evaluation but as a set of sociotechnical systems whose social meaning depends on deployment context, including distinctions between consumer-facing systems, interpersonal interfaces, and institutionally embedded decision tools (Johnson and Verdicchio, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Sartori and Theodorou, \u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). By measuring apprehension separately for different AI types, the present scales translate this theoretical position into an affective measurement framework that preserves these contextual distinctions at the level of individual evaluation. This allows theories of AI governance and human-AI interaction to examine how apprehension varies with proximity, responsibility, and accountability across different AI types, using comparable measures that are aligned in structure but differentiated by evaluative target (Singhal et al., \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Vallor and Vierkant, \u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Practical implementation\u003c/h2\u003e \u003cp\u003e \u003cem\u003eCross-context consistency and non-interchangeability\u003c/em\u003e. The degree of consistency observed across apprehension toward different AI types has psychometric implications for how AI apprehension should be measured and interpreted. Shared-variance analyses showed only partial overlapping between the four scales, with overlap ranging from approximately two thirds of variance between General AI and LLMs to less than one third of variance for comparisons involving Institutional AI. This pattern was mirrored in the level of absolute agreement across scales, which fell in the moderate range, indicating that individuals\u0026rsquo; relative standing on apprehension was only partly preserved across AI types. From a measurement perspective, these values indicate that the four scales capture a related evaluative tendency, but not to an extent that would justify treating them as interchangeable indicators of a single construct. Instead, apprehension shows specificity to each type of AI combined with the largest divergence emerging for Institutional AI. Apprehension therefore cannot be interpreted as a context-free individual difference; its expression varies according to the specific AI type being evaluated. Shared variance analyses further showed that, even at the factor level (Implied Malice, Undesirability, and Unpredictability), each scale retained a substantial proportion of variance specific to its respective AI type. This implies that pooling scores across AI types would reduce assessment specificity, because the pooled score would blend shared variance with AI-type-specific variance. Psychometrically, this supports using separate AI-type scales with the same factor structure, so that comparisons are aligned while preserving the context-specific information captured by each scale.\u003c/p\u003e \u003cp\u003e \u003cem\u003eThe association between apprehension items\u003c/em\u003e. The network results indicate that AI apprehension is organised around stable internal linkages within each dimension. Concerns within Implied Malice, Unpredictability, and Undesirability are more strongly connected to each other than to concerns outside their dimension, and this pattern is consistent across AI types. Practically, this suggests that concerns within a dimension tend to reinforce one another. Addressing a single concern within a dimension may therefore be insufficient if closely connected concerns remain unchanged.\u003c/p\u003e \u003cp\u003eThe replication of the same strongest within-dimension connections across AI types further indicates that the internal configuration of apprehension is not context-specific noise but a stable structural pattern. This stability allows practitioners to anticipate which concerns are likely to co-occur and to design responses that address linked concerns jointly. Because overall connectivity is similar across AI types, differences between AI contexts are unlikely to reflect random variation in how concerns relate to one another. Instead, they reflect the activation of the same underlying structure in different contexts. Practically, this means that assessment tools and intervention strategies can rely on the same structural framework while adapting content to the specific AI domain.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Scoring\u003c/h2\u003e \u003cp\u003e \u003cem\u003eScoring by factors\u003c/em\u003e. The validated factor structure for each Apprehension scale comprises three dimensions represented by an unequal number of items (2\u0026thinsp;+\u0026thinsp;3 + 3). When subscales differ in item count, computing total or composite scores using simple unit-weighted sums can systematically distort the contribution of individual dimensions, because factors with more items exert greater influence on the total score purely as a function of scale length rather than construct relevance. Moreover, from a reliability perspective, shorter factors are inherently disadvantaged because internal consistency indices are sensitive to the number of indicators (Edelsbrunner et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; McNeish, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Two-item factors can show acceptable inter-item correlations and strong factor loadings while nonetheless contributing less variance to a summed score than longer factors. As a result, unweighted scoring may under-represent meaningful individual differences on that dimension despite adequate construct validity. Therefore, when factor scores are used, we suggest applying weighting of 1.5 to the total score of the Implied Malice factor to account for unequal item numbers across factors (Woźniak et al., \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cem\u003eTotal scoring\u003c/em\u003e. When users are interested in overall AI apprehension, weighting is unnecessary. In this case, the scale should be scored using a simple unit-weighted sum of all eight items. Total scores range from 0 to 80, with higher scores indicating greater overall apprehension towards AI. Because the total score is intended to capture general apprehension rather than differential contributions of specific domains, unequal item counts across factors do not pose a conceptual or interpretative problem in this use case.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec36\" class=\"Section2\"\u003e \u003ch2\u003e5.4. Limitations and future research\u003c/h2\u003e \u003cp\u003eSeveral limitations should be considered when interpreting the present findings.\u003c/p\u003e \u003cp\u003eFirst, the study used a single time-point design. As a result, we cannot say how stable scores on these scales are over time, whether they change as people gain experience with AI systems, or how quickly they respond to changes in the wider AI landscape. Future work should examine test-retest reliability and longer-term stability.\u003c/p\u003e \u003cp\u003eSecond, the scales were adapted and validated in a UK adult sample, which limits generalisability to non-Western cultures. Although we do not expect large cross-cultural differences in the core content of the three dimensions (Implied Malice, Undesirability, and Unpredictability) because these map onto broadly shared concerns about intent, social acceptability, and uncertainty in complex systems, cultural context may still influence how strongly particular concerns are endorsed. Differences in institutional trust, regulation, media narratives, and everyday exposure to AI could alter item functioning. For that reason, the next step is cross-cultural validation, including formal tests of measurement invariance, before using the scales for direct comparisons across countries or language groups.\u003c/p\u003e \u003cp\u003eThird, the scales rely on self-report, and the response format asks respondents to indicate how much more would be required for them to consider the AI system moral, ethical, good, safe, expected, normal, or natural. This framing was deliberately chosen to operationalise apprehension as a perceived shortfall relative to an acceptable standard. However, responses still depend on how individuals internally define these standards. Future work could therefore examine how respondents interpret these evaluative thresholds and whether scores predict behavioural outcomes, such as acceptance, endorsement, or real-world use.\u003c/p\u003e \u003cp\u003eFuture research can extend the present findings by testing theoretical accounts of AI apprehension that make predictions about how evaluations of different AI types are organised and updated. Because the present study shows that apprehension is structured by AI type within the same individuals, future research can examine whether changes in beliefs, information, or public discourse selectively affect apprehension toward some AI types while leaving others unchanged. This allows competing theoretical accounts to be tested, such as whether apprehension is primarily influenced by perceived agency, institutional authority, or social proximity, depending on the AI type under consideration. In addition, future studies can examine whether relations between AI apprehension and other constructs, such as trust or acceptance, follow the same structure across AI types or diverge systematically, which would further clarify the conceptual boundaries of AI apprehension. These directions build directly on the present findings by using the differentiated structure of the scales to refine theory about how concern toward artificial intelligence is organised and maintained.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThe General AI Apprehension Scale, Large Language Model Apprehension Scale, Personal AI Apprehension Scale, and Institutional AI Apprehension Scale measure apprehension toward four distinct types of artificial intelligence. Each scale using the same underlying dimensions, which allows apprehension to be examined separately across AI contexts and compared when needed. Across all four scales, the measurement models show stable internal structure, good reliability at both total and subscale levels, and separation from related evaluative constructs. The scales are intended for empirical research that requires direct and context-specific measurement of apprehension toward different forms of artificial intelligence.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCorresponding authors\u003c/strong\u003e: Ala Yankouskaya, Raian Ali\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAY: Conceptualised and designed the study, curated the data, designed and performed and reported the statistical analysis, wrote the first draft.\u003c/p\u003e\n\u003cp\u003eBA: Conceptualised the study, validated the analysis, reviewed and edited the paper.\u003c/p\u003e\n\u003cp\u003eML: Validated the analysis, wrote parts of the first draft, reviewed and edited the paper.\u003c/p\u003e\n\u003cp\u003eMMR: Designed the study, curated the data, reviewed the paper.\u003c/p\u003e\n\u003cp\u003eSA: Designed the study, validated the analysis, reviewed and edited the paper.\u003c/p\u003e\n\u003cp\u003eRA: Conceptualised and designed the study, curated the data, validated the analysis, reviewed and edited the paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOpen Access funding provided by the Qatar National Library. This publication was supported by NPRP 14 Cluster grant # NPRP 14C-0916–210015 from the Qatar National Research Fund (a member of Qatar Foundation). The findings herein reflect the work and are solely the responsibility of the authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval and consent to participate\u003c/strong\u003e: Ethical approval was granted by the Institutional Review Board (IRB) at the lead author’s institution (IRB Protocol Reference Number: HBKU-IRB-2025-185). Informed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e: The study design, dataset, and analysis files are available on the Open Science Framework (OSF) at the following link view_only=a1d3d422a48643c8b75fd8cf06ae336. The author confirms that all data generated or analysed during this study are included in this published article.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAlexander, J.K., 2020. Introduction: the entanglement of technology and religion. Hist. Technol. 36, 165\u0026ndash;186. https://doi.org/10.1080/07341512.2020.1814513\u003c/li\u003e\n\u003cli\u003eAmodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., Man\u0026eacute;, D., 2016. Concrete Problems in AI Safety. https://doi.org/10.48550/ARXIV.1606.06565\u003c/li\u003e\n\u003cli\u003eAnanny, M., Crawford, K., 2018. Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media Soc. 20, 973\u0026ndash;989. https://doi.org/10.1177/1461444816676645\u003c/li\u003e\n\u003cli\u003eBansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M.T., Weld, D., 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Presented at the CHI \u0026rsquo;21: CHI Conference on Human Factors in Computing Systems, ACM, Yokohama Japan, pp. 1\u0026ndash;16. https://doi.org/10.1145/3411764.3445717\u003c/li\u003e\n\u003cli\u003eBeaudry, A., Pinsonneault, A., 2010. The Other Side of Acceptance: Studying the Direct and Indirect Effects of Emotions on Information Technology Use1. MIS Q. 34, 689\u0026ndash;710. https://doi.org/10.2307/25750701\u003c/li\u003e\n\u003cli\u003eBender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S., 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜, in: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Presented at the FAccT \u0026rsquo;21: 2021 ACM Conference on Fairness, Accountability, and Transparency, ACM, Virtual Event Canada, pp. 610\u0026ndash;623. https://doi.org/10.1145/3442188.3445922\u003c/li\u003e\n\u003cli\u003eBialy, F., Elliot, M., Meckin, R., 2025. Perceptions of AI Across Sectors: A Comparative Review of Public Attitudes. https://doi.org/10.48550/arXiv.2509.18233\u003c/li\u003e\n\u003cli\u003eBicchieri, C., 2005. The Grammar of Society: The Nature and Dynamics of Social Norms, 1st ed. Cambridge University Press. https://doi.org/10.1017/CBO9780511616037\u003c/li\u003e\n\u003cli\u003eBrown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D., 2020. Language Models are Few-Shot Learners. https://doi.org/10.48550/ARXIV.2005.14165\u003c/li\u003e\n\u003cli\u003eBucknall, B.S., Dori-Hacohen, S., 2022. Current and Near-Term AI as a Potential Existential Risk Factor, in: Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society. Presented at the AIES \u0026rsquo;22: AAAI/ACM Conference on AI, Ethics, and Society, ACM, Oxford United Kingdom, pp. 119\u0026ndash;129. https://doi.org/10.1145/3514094.3534146\u003c/li\u003e\n\u003cli\u003eBurrell, J., Fourcade, M., 2021. The Society of Algorithms. Annu. Rev. Sociol. 47, 213\u0026ndash;237. https://doi.org/10.1146/annurev-soc-090820-020800\u003c/li\u003e\n\u003cli\u003eCacioppo, J.T., Gardner, W.L., Berntson, G.G., 1997. Beyond Bipolar Conceptualizations and Measures: The Case of Attitudes and Evaluative Space. Personal. Soc. Psychol. Rev. 1, 3\u0026ndash;25. https://doi.org/10.1207/s15327957pspr0101_2\u003c/li\u003e\n\u003cli\u003eCastelo, N., Bos, M.W., Lehmann, D.R., 2019. Task-Dependent Algorithm Aversion. J. Mark. Res. 56, 809\u0026ndash;825. https://doi.org/10.1177/0022243719851788\u003c/li\u003e\n\u003cli\u003eCaton, S., Haas, C., 2024. Fairness in Machine Learning: A Survey. ACM Comput. Surv. 56, 1\u0026ndash;38. https://doi.org/10.1145/3616865\u003c/li\u003e\n\u003cli\u003eCorradi, G., Theirs, C., Mart\u0026iacute;nez‐Mart\u0026iacute;, M.L., Isern‐Mas, C., Villar, S., 2025. Who Fears Generative Artificial Intelligence? Scale Development and Predictors of Fears Towards GenAI. Scand. J. Psychol. sjop.70037. https://doi.org/10.1111/sjop.70037\u003c/li\u003e\n\u003cli\u003eCosta, P.T., McCrae, R.R., 1992. The Five-Factor Model of Personality and Its Relevance to Personality Disorders. J. Personal. Disord. 6, 343\u0026ndash;359. https://doi.org/10.1521/pedi.1992.6.4.343\u003c/li\u003e\n\u003cli\u003eDang, Q., Li, G., 2026. Unveiling trust in AI: the interplay of antecedents, consequences, and cultural dynamics. AI Soc. 41, 669\u0026ndash;692. https://doi.org/10.1007/s00146-025-02477-6\u003c/li\u003e\n\u003cli\u003eDe Freitas, J., Uguralp, A.K., Uguralp, Z.O., Stefano, P., 2024. AI Companions Reduce Loneliness. https://doi.org/10.48550/ARXIV.2407.19096\u003c/li\u003e\n\u003cli\u003eDevaraj, S., Easley, R.F., Crant, J.M., 2008. \u003cstrong\u003eResearch Note\u003c/strong\u003e \u0026mdash;How Does Personality Matter? Relating the Five-Factor Model to Technology Acceptance and Use. Inf. Syst. Res. 19, 93\u0026ndash;105. https://doi.org/10.1287/isre.1070.0153\u003c/li\u003e\n\u003cli\u003eDietvorst, B.J., Simmons, J.P., Massey, C., 2015. Algorithm aversion: People erroneously avoid algorithms after seeing them err. J. Exp. Psychol. Gen. 144, 114\u0026ndash;126. https://doi.org/10.1037/xge0000033\u003c/li\u003e\n\u003cli\u003eEdelsbrunner, P.A., Simonsmeier, B.A., Schneider, M., 2025. The Cronbach\u0026rsquo;s Alpha of Domain-Specific Knowledge Tests Before and After Learning: A Meta-Analysis of Published Studies. Educ. Psychol. Rev. 37, 4. https://doi.org/10.1007/s10648-024-09982-y\u003c/li\u003e\n\u003cli\u003eEisinga, R., Grotenhuis, M.T., Pelzer, B., 2013. The reliability of a two-item scale: Pearson, Cronbach, or Spearman-Brown? Int. J. Public Health 58, 637\u0026ndash;642. https://doi.org/10.1007/s00038-012-0416-3\u003c/li\u003e\n\u003cli\u003eEloundou, T., Manning, S., Mishkin, P., Rock, D., 2023. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://doi.org/10.48550/arXiv.2303.10130\u003c/li\u003e\n\u003cli\u003eEpley, N., Waytz, A., Cacioppo, J.T., 2007. On seeing human: A three-factor theory of anthropomorphism. Psychol. Rev. 114, 864\u0026ndash;886. https://doi.org/10.1037/0033-295X.114.4.864\u003c/li\u003e\n\u003cli\u003eEpskamp, S., Borsboom, D., Fried, E.I., 2018. Estimating psychological networks and their accuracy: A tutorial paper. Behav. Res. Methods 50, 195\u0026ndash;212. https://doi.org/10.3758/s13428-017-0862-1\u003c/li\u003e\n\u003cli\u003eFornell, C., Larcker, D.F., 1981. Evaluating Structural Equation Models with Unobservable Variables and Measurement Error. J. Mark. Res. 18, 39. https://doi.org/10.2307/3151312\u003c/li\u003e\n\u003cli\u003eGabriel, I., 2020. Artificial Intelligence, Values, and Alignment. Minds Mach. 30, 411\u0026ndash;437. https://doi.org/10.1007/s11023-020-09539-2\u003c/li\u003e\n\u003cli\u003eGagrčin, E., Toth, R., Schaetz, N., Naab, T., Emmer, M., 2026. Perceived personal and societal data harms shape users\u0026rsquo; data control preferences. Internet Policy Rev. 15. https://doi.org/10.14763/2026.1.2060\u003c/li\u003e\n\u003cli\u003eGlikson, E., Woolley, A.W., 2020. Human Trust in Artificial Intelligence: Review of Empirical Research. Acad. Manag. Ann. 14, 627\u0026ndash;660. https://doi.org/10.5465/annals.2018.0057\u003c/li\u003e\n\u003cli\u003eGrassini, S., 2023. Development and validation of the AI attitude scale (AIAS-4): a brief measure of general attitude toward artificial intelligence. Front. Psychol. 14, 1191628. https://doi.org/10.3389/fpsyg.2023.1191628\u003c/li\u003e\n\u003cli\u003eGrassini, S., Thorp, S.O., Cipriani, E., Mannari, V., Sevic, A., 2025. Context matters: Analyzing the factors influencing artificial intelligence acceptance across five different applications. Comput. Hum. Behav. Rep. 20, 100834. https://doi.org/10.1016/j.chbr.2025.100834\u003c/li\u003e\n\u003cli\u003eGray, K., Wegner, D.M., 2012. Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition 125, 125\u0026ndash;130. https://doi.org/10.1016/j.cognition.2012.06.007\u003c/li\u003e\n\u003cli\u003eHancock, P.A., Billings, D.R., Schaefer, K.E., 2011. Can You Trust Your Robot? Ergon. Des. Q. Hum. Factors Appl. 19, 24\u0026ndash;29. https://doi.org/10.1177/1064804611415045\u003c/li\u003e\n\u003cli\u003eHayduk, L.A., Glaser, D.N., 2000. Jiving the Four-Step, Waltzing Around Factor Analysis, and Other Serious Fun. Struct. Equ. Model. Multidiscip. J. 7, 1\u0026ndash;35. https://doi.org/10.1207/S15328007SEM0701_01\u003c/li\u003e\n\u003cli\u003eHommel, B., 2023. Religion and cognitive control: An event-coding approach. New Ideas Psychol. 70, 101022. https://doi.org/10.1016/j.newideapsych.2023.101022\u003c/li\u003e\n\u003cli\u003eJebb, A.T., Ng, V., Tay, L., 2021. A Review of Key Likert Scale Development Advances: 1995\u0026ndash;2019. Front. Psychol. 12, 637547. https://doi.org/10.3389/fpsyg.2021.637547\u003c/li\u003e\n\u003cli\u003eJi, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P., 2023. Survey of Hallucination in Natural Language Generation. ACM Comput. Surv. 55, 1\u0026ndash;38. https://doi.org/10.1145/3571730\u003c/li\u003e\n\u003cli\u003eJohnson, D.G., Verdicchio, M., 2025. The sociotechnical entanglement of AI and values. AI Soc. 40, 67\u0026ndash;76. https://doi.org/10.1007/s00146-023-01852-5\u003c/li\u003e\n\u003cli\u003eKaiser, F.G., Wilson, M., 2019. The Campbell Paradigm as a Behavior-Predictive Reinterpretation of the Classical Tripartite Model of Attitudes. Eur. Psychol. 24, 359\u0026ndash;374. https://doi.org/10.1027/1016-9040/a000364\u003c/li\u003e\n\u003cli\u003eKenny, D.A., Kaniskan, B., McCoach, D.B., 2015. The Performance of RMSEA in Models With Small Degrees of Freedom. Sociol. Methods Res. 44, 486\u0026ndash;507. https://doi.org/10.1177/0049124114543236\u003c/li\u003e\n\u003cli\u003eKleine, A.-K., Schmitt, A., Wisse, B.M., 2024. Challenge and threat appraisal of entrepreneurial errors: a latent profile analysis and examination of coping responses. Curr. Psychol. 43, 1206\u0026ndash;1220. https://doi.org/10.1007/s12144-023-04370-1\u003c/li\u003e\n\u003cli\u003eKotov, R., Gamez, W., Schmidt, F., Watson, D., 2010. Linking \u0026ldquo;big\u0026rdquo; personality traits to anxiety, depressive, and substance use disorders: A meta-analysis. Psychol. Bull. 136, 768\u0026ndash;821. https://doi.org/10.1037/a0020327\u003c/li\u003e\n\u003cli\u003eLai, K., Green, S.B., 2016. The Problem with Having Two Watches: Assessment of Fit When RMSEA and CFI Disagree. Multivar. Behav. Res. 51, 220\u0026ndash;239. https://doi.org/10.1080/00273171.2015.1134306\u003c/li\u003e\n\u003cli\u003eLee, J.D., See, K.A., 2004. Trust in Automation: Designing for Appropriate Reliance. Hum. Factors J. Hum. Factors Ergon. Soc. 46, 50\u0026ndash;80. https://doi.org/10.1518/hfes.46.1.50_30392\u003c/li\u003e\n\u003cli\u003eLee, M.K., 2018. Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data Soc. 5, 2053951718756684. https://doi.org/10.1177/2053951718756684\u003c/li\u003e\n\u003cli\u003eLee, M.K., Kusbit, D., Metsky, E., Dabbish, L., 2015. Working with Machines: The Impact of Algorithmic and Data-Driven Management on Human Workers, in: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems. Presented at the CHI \u0026rsquo;15: CHI Conference on Human Factors in Computing Systems, ACM, Seoul Republic of Korea, pp. 1603\u0026ndash;1612. https://doi.org/10.1145/2702123.2702548\u003c/li\u003e\n\u003cli\u003eLiljequist, D., Elfving, B., Skavberg Roaldsen, K., 2019. Intraclass correlation \u0026ndash; A discussion and demonstration of basic features. PLOS ONE 14, e0219854. https://doi.org/10.1371/journal.pone.0219854\u003c/li\u003e\n\u003cli\u003eLipton, Z.C., 2018. The Mythos of Model Interpretability: In machine learning, the concept of interpretability is both important and slippery. Queue 16, 31\u0026ndash;57. https://doi.org/10.1145/3236386.3241340\u003c/li\u003e\n\u003cli\u003eLongoni, C., Bonezzi, A., Morewedge, C.K., 2019. Resistance to Medical Artificial Intelligence. J. Consum. Res. 46, 629\u0026ndash;650. https://doi.org/10.1093/jcr/ucz013\u003c/li\u003e\n\u003cli\u003eLuger, E., Sellen, A., 2016. \u0026ldquo;Like Having a Really Bad PA\u0026rdquo;: The Gulf between User Expectation and Experience of Conversational Agents, in: Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. Presented at the CHI\u0026rsquo;16: CHI Conference on Human Factors in Computing Systems, ACM, San Jose California USA, pp. 5286\u0026ndash;5297. https://doi.org/10.1145/2858036.2858288\u003c/li\u003e\n\u003cli\u003eMarsh, H.W., Wen, Z., Hau, K.-T., 2004. Structural Equation Models of Latent Interactions: Evaluation of Alternative Estimation Strategies and Indicator Construction. Psychol. Methods 9, 275\u0026ndash;300. https://doi.org/10.1037/1082-989X.9.3.275\u003c/li\u003e\n\u003cli\u003eMartin, K., 2018. The penalty for privacy violations: How privacy violations impact trust online. J. Bus. Res. 82, 103\u0026ndash;116. https://doi.org/10.1016/j.jbusres.2017.08.034\u003c/li\u003e\n\u003cli\u003eMayer, R.C., Davis, J.H., Schoorman, F.D., 1995. An Integrative Model of Organizational Trust. Acad. Manage. Rev. 20, 709. https://doi.org/10.2307/258792\u003c/li\u003e\n\u003cli\u003eMcElroy, J.C., Hendrickson, A.R., Townsend, A.M., DeMarie, S.M., 2007. Dispositional Factors in Internet Use: Personality Versus Cognitive Style1. MIS Q. 31, 809\u0026ndash;820. https://doi.org/10.2307/25148821\u003c/li\u003e\n\u003cli\u003eMcGrath, M.J., Lack, O., Tisch, J., Duenser, A., 2025. Measuring trust in artificial intelligence: validation of an established scale and its short form. Front. Artif. Intell. 8, 1582880. https://doi.org/10.3389/frai.2025.1582880\u003c/li\u003e\n\u003cli\u003eMcNeish, D., 2018. Thanks coefficient alpha, we\u0026rsquo;ll take it from here. Psychol. Methods 23, 412\u0026ndash;433. https://doi.org/10.1037/met0000144\u003c/li\u003e\n\u003cli\u003eMcPhetres, J., Zuckerman, M., 2018. Religiosity predicts negative attitudes towards science and lower levels of science literacy. PLOS ONE 13, e0207125. https://doi.org/10.1371/journal.pone.0207125\u003c/li\u003e\n\u003cli\u003eMetzger, M.J., Flanagin, A.J., 2015. Psychological Approaches to Credibility Assessment Online, in: Sundar, S.S. (Ed.), The Handbook of the Psychology of Communication Technology. Wiley, pp. 445\u0026ndash;466. https://doi.org/10.1002/9781118426456.ch20\u003c/li\u003e\n\u003cli\u003eMick, D.G., Fournier, S., 1998. Paradoxes of Technology: Consumer Cognizance, Emotions, and Coping Strategies. J. Consum. Res. 25, 123\u0026ndash;143. https://doi.org/10.1086/209531\u003c/li\u003e\n\u003cli\u003eMontag, C., Nakov, P., Ali, R., 2025. On the need to develop nuanced measures assessing attitudes towards AI and AI literacy in representative large-scale samples. AI Soc. 40, 1129\u0026ndash;1130. https://doi.org/10.1007/s00146-024-01888-1\u003c/li\u003e\n\u003cli\u003eMontoya, A.K., Edwards, M.C., 2021. The Poor Fit of Model Fit for Selecting Number of Factors in Exploratory Factor Analysis for Scale Evaluation. Educ. Psychol. Meas. 81, 413\u0026ndash;440. https://doi.org/10.1177/0013164420942899\u003c/li\u003e\n\u003cli\u003ePasquale, F., 2016. The black box society: the secret algorithms that control money and information, First Harvard University Press paperback edition. ed. Harvard University Press, Cambridge, Massachusetts London, England.\u003c/li\u003e\n\u003cli\u003ePreston, C.C., Colman, A.M., 2000. Optimal number of response categories in rating scales: reliability, validity, discriminating power, and respondent preferences. Acta Psychol. (Amst.) 104, 1\u0026ndash;15. https://doi.org/10.1016/S0001-6918(99)00050-5\u003c/li\u003e\n\u003cli\u003eRahman, M.M., AlShakhsi, S., Babiker, A., Ali, R., Yankouskaya, A., 2025. On the role of personality in attitudes toward AI: Do AI\u0026rsquo;s freedom of choice and social proof matter? World Wide Web 28, 67. https://doi.org/10.1007/s11280-025-01372-w\u003c/li\u003e\n\u003cli\u003eRammstedt, B., John, O.P., 2007. Measuring personality in one minute or less: A 10-item short version of the Big Five Inventory in English and German. J. Res. Personal. 41, 203\u0026ndash;212. https://doi.org/10.1016/j.jrp.2006.02.001\u003c/li\u003e\n\u003cli\u003eRaykov, T., 1997. Scale Reliability, Cronbach\u0026rsquo;s Coefficient Alpha, and Violations of Essential Tau-Equivalence with Fixed Congeneric Components. Multivar. Behav. Res. 32, 329\u0026ndash;353. https://doi.org/10.1207/s15327906mbr3204_2\u003c/li\u003e\n\u003cli\u003eRevilla, M.A., Saris, W.E., Krosnick, J.A., 2014. Choosing the Number of Categories in Agree\u0026ndash;Disagree Scales. Sociol. Methods Res. 43, 73\u0026ndash;97. https://doi.org/10.1177/0049124113509605\u003c/li\u003e\n\u003cli\u003eRudin, C., 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell. 1, 206\u0026ndash;215. https://doi.org/10.1038/s42256-019-0048-x\u003c/li\u003e\n\u003cli\u003eSag, M., 2023. Copyright Safety for Generative AI. SSRN Electron. J. https://doi.org/10.2139/ssrn.4438593\u003c/li\u003e\n\u003cli\u003eSartori, L., Theodorou, A., 2022. A sociotechnical perspective for the future of AI: narratives, inequalities, and human control. Ethics Inf. Technol. 24, 4. https://doi.org/10.1007/s10676-022-09624-3\u003c/li\u003e\n\u003cli\u003eSchepman, A., Rodway, P., 2023. The General Attitudes towards Artificial Intelligence Scale (GAAIS): Confirmatory Validation and Associations with Personality, Corporate Distrust, and General Trust. Int. J. Human\u0026ndash;Computer Interact. 39, 2724\u0026ndash;2741. https://doi.org/10.1080/10447318.2022.2085400\u003c/li\u003e\n\u003cli\u003eSchiavo, G., Businaro, S., Zancanaro, M., 2024. Comprehension, apprehension, and acceptance: Understanding the influence of literacy and anxiety on acceptance of artificial Intelligence. Technol. Soc. 77, 102537. https://doi.org/10.1016/j.techsoc.2024.102537\u003c/li\u003e\n\u003cli\u003eShin, D., 2021. The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI. Int. J. Hum.-Comput. Stud. 146, 102551. https://doi.org/10.1016/j.ijhcs.2020.102551\u003c/li\u003e\n\u003cli\u003eSijtsma, K., 2009. On the Use, the Misuse, and the Very Limited Usefulness of Cronbach\u0026rsquo;s Alpha. Psychometrika 74, 107\u0026ndash;120. https://doi.org/10.1007/s11336-008-9101-0\u003c/li\u003e\n\u003cli\u003eSindermann, C., Sha, P., Zhou, M., Wernicke, J., Schmitt, H.S., Li, M., Sariyska, R., Stavrou, M., Becker, B., Montag, C., 2021. Assessing the Attitude Towards Artificial Intelligence: Introduction of a Short Measure in German, Chinese, and English Language. KI - K\u0026uuml;nstl. Intell. 35, 109\u0026ndash;118. https://doi.org/10.1007/s13218-020-00689-0\u003c/li\u003e\n\u003cli\u003eSinghal, A., Neveditsin, N., Tanveer, H., Mago, V., 2024. Toward Fairness, Accountability, Transparency, and Ethics in AI for Social Media and Health Care: Scoping Review. JMIR Med. Inform. 12, e50048. https://doi.org/10.2196/50048\u003c/li\u003e\n\u003cli\u003eStein, J.-P., Messingschlager, T., Gnambs, T., Hutmacher, F., Appel, M., 2024. Attitudes towards AI: measurement and associations with personality. Sci. Rep. 14, 2909. https://doi.org/10.1038/s41598-024-53335-2\u003c/li\u003e\n\u003cli\u003eSun, B., Zhang, Y., Meng, X., Zhang, Z., 2025. Social class, public self-awareness, and consumers\u0026rsquo; heterogeneous attitude to AI machine anthropomorphism. Curr. Psychol. 44, 10561\u0026ndash;10578. https://doi.org/10.1007/s12144-025-07872-2\u003c/li\u003e\n\u003cli\u003eSundar, S.S., Lee, E.-J., 2022. Rethinking Communication in the Era of Artificial Intelligence. Hum. Commun. Res. 48, 379\u0026ndash;385. https://doi.org/10.1093/hcr/hqac014\u003c/li\u003e\n\u003cli\u003eSvendsen, G.B., Johnsen, J.-A.K., Alm\u0026aring;s-S\u0026oslash;rensen, L., Vitters\u0026oslash;, J., 2013. Personality and technology acceptance: the influence of personality factors on the core constructs of the Technology Acceptance Model. Behav. Inf. Technol. 32, 323\u0026ndash;334. https://doi.org/10.1080/0144929X.2011.553740\u003c/li\u003e\n\u003cli\u003eTaddeo, M., Floridi, L., 2018. How AI can be a force for good. Science 361, 751\u0026ndash;752. https://doi.org/10.1126/science.aat5991\u003c/li\u003e\n\u003cli\u003eThompson, S.C., 1981. Will it hurt less if I can control it? A complex answer to a simple question. Psychol. Bull. 90, 89\u0026ndash;101. https://doi.org/10.1037/0033-2909.90.1.89\u003c/li\u003e\n\u003cli\u003eTriguero, I., Molina, D., Poyatos, J., Del Ser, J., Herrera, F., 2024. General Purpose Artificial Intelligence Systems (GPAIS): Properties, definition, taxonomy, societal implications and responsible governance. Inf. Fusion 103, 102135. https://doi.org/10.1016/j.inffus.2023.102135\u003c/li\u003e\n\u003cli\u003eUğur, N.G., Dursun, F., 2025. The social anatomy of AI anxiety: gender, generations, and technological exposure. Front. Psychiatry 16, 1641546. https://doi.org/10.3389/fpsyt.2025.1641546\u003c/li\u003e\n\u003cli\u003eVallor, S., Vierkant, T., 2024. Find the Gap: AI, Responsible Agency and Vulnerability. Minds Mach. 34, 20. https://doi.org/10.1007/s11023-024-09674-0\u003c/li\u003e\n\u003cli\u003eVenkatesh, Thong, Xu, 2012. Consumer Acceptance and Use of Information Technology: Extending the Unified Theory of Acceptance and Use of Technology. MIS Q. 36, 157. https://doi.org/10.2307/41410412\u003c/li\u003e\n\u003cli\u003eWang, Y., Guan, X., Sun, Y., Wang, H., Chen, D., 2025. The cognitive acceptance of generative AI image tools based on TPB-TAM model and multi-theory integration. Adv. Des. Res. 3, 38\u0026ndash;54. https://doi.org/10.1016/j.ijadr.2025.08.001\u003c/li\u003e\n\u003cli\u003eWang, Y.-Y., Wang, Y.-S., 2022. Development and validation of an artificial intelligence anxiety scale: an initial application in predicting motivated learning behavior. Interact. Learn. Environ. 30, 619\u0026ndash;634. https://doi.org/10.1080/10494820.2019.1674887\u003c/li\u003e\n\u003cli\u003eWolf, E.J., Harrington, K.M., Clark, S.L., Miller, M.W., 2013. Sample Size Requirements for Structural Equation Models: An Evaluation of Power, Bias, and Solution Propriety. Educ. Psychol. Meas. 73, 913\u0026ndash;934. https://doi.org/10.1177/0013164413495237\u003c/li\u003e\n\u003cli\u003eWoźniak, P.W., Karolus, J., Lang, F., Eckerth, C., Sch\u0026ouml;ning, J., Rogers, Y., Niess, J., 2021. Creepy Technology:What Is It and How Do You Measure It?, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Presented at the CHI \u0026rsquo;21: CHI Conference on Human Factors in Computing Systems, ACM, Yokohama Japan, pp. 1\u0026ndash;13. https://doi.org/10.1145/3411764.3445299\u003c/li\u003e\n\u003cli\u003eYang, H., Sundar, S.S., 2025. AI anxiety: Explication and exploration of effect on state anxiety when interacting with AI doctors. Comput. Hum. Behav. Artif. Hum. 3, 100128. https://doi.org/10.1016/j.chbah.2025.100128\u003c/li\u003e\n\u003cli\u003eZhang, B., Dafoe, A., 2019. Artificial Intelligence: American Attitudes and Trends. SSRN Electron. J. https://doi.org/10.2139/ssrn.3312874\u003c/li\u003e\n\u003cli\u003eZuboff, S., 2023. The age of surveillance capitalism, in: Social Theory Re-Wired. pp. 203\u0026ndash;213.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Apprehension, large language models, personal AI, general AI, institutional AI","lastPublishedDoi":"10.21203/rs.3.rs-9018418/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9018418/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eRecent research has developed measures that conceptualise human factors in relation to AI in relatively broad terms, such as general attitudes and positive, negative or adaptive use. However, more nuanced instruments are still needed to capture specific psychological dimensions, such as apprehension, which is the focus of this paper. Moreover, many existing measures conceptualise AI as a single, undifferentiated object of evaluation, rather than distinguishing between different AI systems, contexts of use, or functional domains. The present study adapted, developed and validated four distinct but structurally parallel instruments measuring apprehensions toward General AI, Large Language Models, Personal AI and Institutional AI. The instruments share identical structure and item content adapted to each type of AI. Apprehension is operationalised across three dimensions (Implied Malice, Undesirability, and Unpredictability). Data from a British sample of 559 adults (age range 18\u0026ndash;45, M\u0026thinsp;=\u0026thinsp;30.64, SD\u0026thinsp;=\u0026thinsp;6.75, 50.3% males). For each scale, confirmatory factor analyses supported the three-dimensional structure, while internal consistency was strong at both total and subscale levels, and model-based indicators demonstrated well-defined latent constructs. The scales demonstrated good discriminant and convergent validity. These findings establish four distinct and psychometrically robust instruments suitable for research requiring measurement of apprehension toward artificial intelligence both at a general level and across specific classes of AI systems.\u003c/p\u003e","manuscriptTitle":"The Four Artificial Intelligence Apprehension Scales: Apprehension Towards Personal AI, General AI, Institutional AI, and Large Language Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-04 08:27:44","doi":"10.21203/rs.3.rs-9018418/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c31d8531-eb54-4a29-ae08-243dc61263b2","owner":[],"postedDate":"March 4th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":63833328,"name":"Psychology"},{"id":63833329,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-03-04T08:27:44+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-04 08:27:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9018418","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9018418","identity":"rs-9018418","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}