Initial Validation of the IMPACT Model: Technological Appropriation of ChatGPT by University Faculty

Abstract

This study presents the initial validation of the IMPACT model, a psychometric tool developed to evaluate how university faculty adopt ChatGPT in higher education. It specifically addresses the existing gap in validated instruments designed for educators, as most prior research has focused on student-based adoption models. A total of 206 professors completed a 39-item Likert-scale questionnaire. Exploratory factor analysis using principal axis factoring with oblimin rotation identified the underlying structure of the instrument. Reliability and internal consistency were examined through Cronbach’s alpha and McDonald’s omega. The analysis revealed a five-factor structure comprising functional appropriation, ethical and academic concerns, cost and accessibility, facilitating conditions, and perceived reliability and trustworthiness. Intention to use and performance expectancy merged into a single factor, and social influence did not emerge as a determinant. The model demonstrated strong reliability and internal consistency across all dimensions. The IMPACT model offers a validated framework for understanding faculty adoption of ChatGPT, emphasizing functional, ethical, and infrastructural factors over social influence. These findings provide a foundation for confirmatory analyses and contribute to advancing theoretical and practical insights into AI integration in higher education teaching.

1. Introduction

The incorporation of generative language models based on artificial intelligence, such as ChatGPT, has sparked significant interest in the field of education (Alqahtani et al., 2023) and is redefining the dynamics of teaching and learning in higher education (Mourtajji & Arts-Chiss, 2024). Emerging technologies that use generative AI, including ChatGPT—language models based on deep learning, specifically on a type of neural network known as a transformer—can produce text that closely resembles human writing. This development has driven a profound transformation in educational practice, fueled by its ability to support content structuring, assessment, and instructional design (Javaid et al., 2023; Labadze et al., 2023; Naznin et al., 2025).

In the context of teaching practice, these tools enable the generation of instructional materials, the formulation of questions, grading of texts, the provision of automated feedback, and support for intelligent tutoring processes (Dong et al., 2024; Kasneci et al., 2023; Yan et al., 2023). They also contribute to educational inclusion by enhancing accessibility for students with specific needs and enable the design of more engaging learning environments through gamification strategies (Huber et al., 2024; Zirar, 2023).

However, alongside these benefits, significant challenges emerge that require critical assessment. Several studies warn of the risk of excessive dependence on these tools, which could hinder the development of critical thinking and intellectual autonomy (Kwak & Pardos, 2024; Dong et al., 2024). Moreover, ongoing concerns persist regarding data privacy, algorithmic transparency, and the perpetuation of cultural and linguistic biases, which could exacerbate pre-existing inequalities, particularly in Global South contexts (Bender, 2024; J. Lee et al., 2024; Zirar, 2023).

Considering this scenario, the effective adoption of models like ChatGPT by faculty members requires not only digital competencies and pedagogical judgment, but also a comprehensive framework to understand the factors influencing their acceptance and use. Empirically, most psychometric studies have focused on the student perspective, while evidence concerning faculty remains scarce and, in some cases, methodologically limited.

From the faculty’s perspective, research on the adoption of ChatGPT in higher education reveals divergent methodological approaches. At the international level, Barakat et al. (2025) have proposed a methodological framework based on the Technology Acceptance Model (TAM), employing Principal Components Analysis (PCA) as the extraction method, despite the fact that the goal is not to reduce correlated variables into uncorrelated components but rather to extract latent factors grounded in a preexisting theory (Floyd & Widaman, 1995; Costello & Osborne, 2005; Matsunaga, 2010; Saif et al., 2024).

In turn, Sigüenza Orellana et al. (2024), in Ecuador, developed an instrument to measure teachers’ perceptions of ChatGPT by conducting an exploratory factor analysis to validate their CPDUChatGPT questionnaire. However, their methodological approach does not clearly specify the statistical assumptions verified for the application of the factor analysis, the extraction method used, or the type of rotation employed, elements that are essential to ensure scientific replicability.

Previous research in educational technologies has shown how the lack of clarity in factor analysis procedures can lead to unstable factor structures and inconsistent theoretical models (Thompson & Daniel, 1996; Worthington & Whittaker, 2006). This issue extends to the specific field of teacher technology adoption, where studies such as those by Venkatesh et al. (2003) have demonstrated the importance of following rigorous protocols in factor analysis to establish valid and generalizable technology acceptance models.

The IMPACT model addresses a conceptual gap not covered by theoretical frameworks such as TAM2 or UTAUT2, which explain the adoption of general technologies based on perceived usefulness, ease of use, and social influence. However, with the emergence of generative artificial intelligence, new dimensions have surfaced that were not previously considered and are directly related to the unique characteristics of a tool capable of generating content from user-issued prompts. In this sense, the IMPACT model deepens the understanding of disruptive technology adoption by providing a conceptual framework that integrates constructs such as ethical and academic concerns, perceived reliability, and functional appropriation. Notably, this latter dimension establishes the main difference from the CHASSIS model, as in student contexts, intention to use and perceived usefulness appear as separate dimensions.

Another feature that marks a clear distinction between the TAM2, UTAUT2, and CHASSIS models is that, in the IMPACT model, social influence does not emerge as a relevant dimension. This suggests that emotional maturity and professional experience may cause social influence to lose its decisive role in the adoption of ChatGPT among university faculty.

2. Theoretical Framework

2.1. Technology Acceptance Models

Generative artificial intelligence, particularly ChatGPT, has substantially transformed the contemporary technological and educational landscape, generating growing interest in understanding the factors that determine its adoption and continued use (Almogren et al., 2024; Salih et al., 2025). Regarding acceptance and intention to use, four foundational models have served as primary references in research on the adoption of digital innovations: the Technology Acceptance Model (TAM; Davis, 1989) and its extension TAM2 (Venkatesh & Davis, 2000), as well as the Unified Theory of Acceptance and Use of Technology (UTAUT; Venkatesh et al., 2003) and its later update, UTAUT2 (Venkatesh et al., 2012).

In TAM, Davis (1989) considered only two dimensions: perceived usefulness and ease of use (effort expectancy). Subsequently, TAM2 (Venkatesh & Davis, 2000) suggested that user acceptance is influenced by two types of processes: social influence—including subjective norm, voluntariness, and image—and cognitive instrumental processes—comprising job relevance, output quality, result demonstrability, and perceived ease of use. These factors jointly determine perceived usefulness, while Perceived Ease of Use remains an independent variable that also functions as a predictor of Perceived Usefulness. Building on this foundation, Venkatesh et al. (2003) considered job relevance, output quality, result demonstrability, and perceived ease of use, all of which determine perceived usefulness. In a further extension of this conceptual framework, UTAUT2, Venkatesh et al. (2012) proposed the inclusion of hedonic motivation, habit, and price value to extend its applicability to consumer contexts. These factors, depending on the context, may be moderated by variables such as age, gender, experience, and voluntariness of use (Abdalla et al., 2024; Dwivedi et al., 2019; Soares et al., 2024). UTAUT2 can explain up to 70% of the variance in behavioral intention (Venkatesh & Zhang, 2010), making it a more robust model than those commonly used to understand technology adoption in educational settings.

Research on emerging technology adoption has been primarily grounded in the Technology Acceptance Model (Ashfaq et al., 2020; Dahri et al., 2024; García et al., 2024; M. K. Kim et al., 2025; A. T. Lee et al., 2025; Ma et al., 2024). However, these models were developed before the emergence of generative artificial intelligence and therefore do not incorporate its inherent dimensions, such as ethical and academic concerns or the reliability of generated responses.

Drawing from these theoretical frameworks and current literature, the CHASSIS model (Pereira-González et al., 2025a) proposed seven conceptual dimensions that were subsequently empirically validated. However, its design focuses on learners, overlooking the complexities involved in technological adoption within professional pedagogical practice. University students typically operate in highly socialized environments where peer pressure and social media influence significantly affect their technological choices. Faculty, conversely, acting from a position of professional autonomy, possess a broader perspective that enables them to evaluate digital tools through more independent and academically rigorous criteria.

This distinction is reflected in two key empirical findings: first, social influence did not emerge as a significant factor in the analysis, indicating that pedagogical technology adoption is guided by independent professional judgment rather than social pressures. Second, an integrated dimension of functional appropriation emerged, combining intention to use and performance expectancy into a single construct. This integration reveals a cognitive process characteristic of experienced professionals, in which the assessment of a tool’s utility and its integration into pedagogical practice are not sequential stages but simultaneous and interdependent processes, guided by instrumental rationality oriented toward specific pedagogical objectives. Therefore, a faculty-specific model is needed—one that prioritizes dimensions associated with pedagogical usefulness, curricular integration, and functional appropriation over social or emotional factors, which are highly relevant in student populations but secondary or non-significant when examining technology adoption among faculty.

2.2. ChatGPT and Generative Artificial Intelligence in Educational Contexts

Several studies have adapted these models to analyze ChatGPT in academic settings; however, unlike other technologies (Alotaibi, 2025; Mennella et al., 2024). For this reason, theoretical approaches that incorporate these new determinants are necessary, particularly in university contexts where instructors and students adopt these tools within formal teaching and learning environments.

2.3. Determining Factors in ChatGPT Use Among University Educators

2.3.1. Perceived Usefulness

The possibility of enhancing academic performance is the most influential factor driving university students to adopt generative artificial intelligence. This core belief makes perceived usefulness the strongest and most consistent predictor of their intention to use such technologies (Alshammari & Babu, 2025; Al Murshidi et al., 2024; García-Alonso et al., 2024). Xue et al. (2024) conducted a systematic review of 162 articles—mostly based on samples from Asia and North America—and found that the most commonly used combination in studies of technology adoption is the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT). They also found that performance expectancy is the most influential determinant of usage intention.

This impact has been corroborated by findings from studies employing various statistical techniques, including exploratory (Sallam et al., 2023) and confirmatory factor analyses (Abdaljaleel et al., 2024), multivariate regression (Sallam et al., 2024), and structural equation modeling (Ngo et al., 2024; Saxena & Doleck, 2023). Similarly, Makkonen et al. (2023) demonstrated that perceived usefulness has a direct, positive, and highly significant effect on the intention to continue using generative AI chatbots in the workplace, identifying it as the key component of their model. These results reinforce the notion that this perception is essential in any model aiming to explain technology adoption.

2.3.2. Ethical and Cognitive Concerns

Ethical concerns can limit the intention to use ChatGPT. Theseissues can be categorized into four main domains: legal, humanistic, algorithmic, and informational. In the realm of algorithmic ethics, Bender et al. (2021) highlight the dangers associated with large language models like GPT-3, pointing to issues such as algorithmic bias, high energy consumption, and lack of transparency in their operation. Zhou et al. (2024) emphasize the cultural, ideological, and gender biases that these technologies may perpetuate. Stahl (2025) also addresses algorithmic ethics but broadens the focus to a more humanistic perspective, expressing concern over human dignity, autonomy, and moral responsibility. He advocates for a contextual and relational ethics framework, grounded in humanistic philosophy and applied to the development and use of emerging technologies such as ChatGPT.

Sebastian (2023) reviews the dual challenge of safeguarding users’ sensitive information while maintaining the efficiency of machine learning models and discusses critical concerns regarding data protection and privacy in the context of chatbot systems based on large language models (LLMs), with particular focus on ChatGPT. Meanwhile, S. J. Kim (2024a) highlights key legal implications such as intellectual property and copyright, plagiarism, attribution, and the confidentiality and privacy required during peer review processes when using AI tools that may store or process sensitive data. The absence of clear legal frameworks to define accountability among users, service providers, and AI developers generates uncertainty that can inhibit technology adoption.

Regarding concerns about the impact of generative artificial intelligence on memory, critical thinking, and cognitive skills, recent literature in educational contexts presents both positive and negative findings related to the use of tools like ChatGPT. Some studies have reported that the immediate access to answers provided by generative AI may limit the development of working memory and consequently reduce students’ ability to retain information, as it diminishes the cognitive effort required to process, solve problems, and consolidate knowledge (Abbas et al., 2024; León-Domínguez, 2024; Zhai et al., 2024). This phenomenon also appears to be associated with increased procrastination and a decline in academic performance when AI use is intensive and lacks reflective engagement (Bai et al., 2023; Gerlich, 2025).

Nevertheless, research has shown that when AI is appropriately integrated into the educational process, it can foster creativity, the development of critical thinking skills, and problem-solving abilities (Huamani-Anco & Maraza-Quispe, 2025; S. J. Kim, 2024b; Toma & Yánez-Pérez, 2024). AI supports the presentation of complex problems and personalized feedback, which stimulates analysis, self-reflection, and decision-making. In fact, the use of AI has been found to significantly enhance critical thinking, particularly when its implementation is oriented toward personalized learning and the critical evaluation of information (Ashraf et al., 2025; Lawasi et al., 2024; Fuchs & Aguilos, 2023; Mastrogiacomi, 2024). As a result, educational approaches are increasingly advocating for a sociocognitive architecture that integrates human and artificial intelligence in a complementary way to maximize benefits and minimize the risks associated with these technologies (Dubey et al., 2024; Mogavi et al., 2024). It is especially important to properly align technological innovation with the preservation of academic integrity (Fajt & Schiller, 2025).

2.3.3. Economic and Technological Barriers

The extension of the Unified Theory of Acceptance and Use of Technology—UTAUT2—proposed by Venkatesh et al. (2012), was designed to shift the application of UTAUT from an organizational context to one focused on voluntary adoption and consumer behavior. To implement this adaptation, the authors introduced three new determinants: hedonic motivation, habit, and price value—the latter defined as the balance between the cost of using the technology and the benefits it provides to the user (Y. Wang & Zhang, 2023; Suryavanshi et al., 2025). In contexts where the cost is borne individually by the user, price value becomes the most significant factor in the UTAUT2 model (Strzelecki et al., 2024).

In this regard, Parveen et al. (2024), using structural equation modeling, found that price value constitutes a significant latent variable when users are university students. Similarly, Saranza et al. (2024) identified a direct correlation between frequency of use and perceived value in the use of ChatGPT, and further noted that this correlation is moderated by users’ disposable income. These studies provide a valid foundation to support proposals aimed at adapting traditional technology acceptance models to the distinctive characteristics associated with the use of generative artificial intelligence.

2.3.4. Facilitating Conditions

Recent studies have consistently emphasized the importance of institutional support in the acceptance and intended use of ChatGPT by university faculty and students. According to Baek et al. (2024), institutional policies that authorize the use of ChatGPT are predictive of higher usage rates. Supporting this view, the study by Romero-Rodríguez et al. (2023), conducted using the UTAUT2 model, found that facilitating conditions—specifically access, availability of resources, and institutional support—play a crucial role in both the intention to use and the actual use of ChatGPT among university students.

Similarly, in Malaysia, Poobalan and Latip (2023) demonstrated that facilitating conditions directly affect the adoption of ChatGPT. When students have access to adequate resources, technical guidance, and appropriate infrastructure, the likelihood of using the tool increases significantly. This trend is also evident in faculty perceptions. A study conducted in Poland reported that facilitating factors—such as institutional support and access to pedagogical and technological resources—have a direct impact on the use, acceptance, and adoption of ChatGPT in higher education settings (Strzelecki et al., 2024).

This conclusion is further supported by the systematic review conducted by Kovari (2024), who argue that the ethical and effective use of ChatGPT by students and faculty largely depends on the existence of clear institutional guidelines and policies designed to support—and regulate—its use (Alzahrani & Alzahrani, 2025). Likewise, H. Wang et al. (2024) found that the intention and purpose of using ChatGPT increase significantly when explicit institutional policies are accompanied by pedagogical integration strategies and effective institutional support.

Collectively, these studies—both from the perspectives of students and faculty—underscore the need to incorporate structural and organizational dimensions into technology adoption models, particularly when analyzing the use of generative artificial intelligence tools in higher education contexts.

2.3.5. Accuracy and Reliability

ChatGPT users frequently express concern regarding the accuracy and reliability of the information it provides, as the model can sometimes generate inaccurate or biased content while presenting it as truthful (Golden, 2023; Limna et al., 2023). This phenomenon, known as algorithmic “hallucination,” negatively impacts users for two reasons. First, it undermines their trust; second, it forces them to spend a considerable amount of time identifying inaccuracies in the information received (Sun et al., 2024). From a reliability perspective, Balaskas et al. (2025) conducted a study using a partial least squares structural equation model (PLS-SEM) and demonstrated that trust in the tool not only directly influences the intention to use it, but also mediates the relationship between perceived usefulness and that intention. Additionally, this trust translates into greater ease of use and a more positive attitude toward accepting responses provided by generative artificial intelligence, which is ultimately crucial for its sustained adoption.

This study aimed to adapt a theoretical model previously supported by replicable empirical evidence to a faculty sample by validating the CHASSIS model (Pereira-González et al., 2025a), originally developed with student data. The IMPACT model proposed in this research integrates empirically derived dimensions through exploratory factor analysis, grounded in the theoretical frameworks of prior technology acceptance models. Recognizing the specificities of the teaching role, professional maturity, and institutional context, the results reveal a five-factor structure that differs from that found among students. Notably, it highlights the merging of behavioral intention and performance expectancy into a single functional factor and the absence of the social influence component. These transformations reflect a more pragmatic and autonomous appropriation of technology by faculty members.

Furthermore, the study confirms the critical relevance of perceived reliability and institutional conditions in ensuring the sustainable adoption of this technological innovation in higher education. By integrating theoretical foundations with empirical findings, this research contributes both to a deeper understanding of the acceptance of generative artificial intelligence technologies in higher education and to the advancement of the scientific field by providing a robust theoretical framework for analyzing their adoption in diverse educational contexts. Additionally, it offers a valid and reliable instrument to support decision-making and the design of educational policies aimed at optimizing the ethical and effective integration of ChatGPT in university teaching.

2.4. Related Work

Recent studies have examined the adoption of generative AI tools, such as ChatGPT, in higher education; however, most have focused on students rather than faculty. In general, empirical research has relied on technology acceptance frameworks such as the Technology Acceptance Model (TAM; Davis, 1989) and the Unified Theory of Acceptance and Use of Technology (UTAUT; Venkatesh et al., 2003). These models emphasize behavioral intention, ease of use, and perceived usefulness (Dwivedi et al., 2023; Alshammari & Alshammari, 2024). More recent investigations have modified these frameworks to account for the integration of generative AI in educational settings, highlighting the importance of perceived ethics, trust, and cognitive load (Đerić et al., 2025; Acosta-Enriquez et al., 2024).

While student attitudes toward ChatGPT have been extensively explored (Kasneci et al., 2023; García-Peñalvo, 2024; Uppal & Hajian, 2025), studies validating instruments that address teacher-specific adoption factors remain scarce and have been conducted in contexts that may not reflect the reality of Latin America (Alzahrani & Alzahrani, 2025; Enang & Christopoulou, 2024). Particularly, two studies were conducted with professors from two public universities in the United States (Shata, 2025; Shata & Hartley, 2025). These studies employed t-tests and multiple regression to test hypotheses about generative AI use among faculty. The study found that, although users expressed more concerns than non-users, the latter reported less comfort with technology—comfort being the only significant predictor of intention to use—while specific concerns had no significant effect. This suggests that familiarity and confidence with technology are key to its educational adoption.

Despite these contributions, no psychometric instrument has yet combined functional appropriation, ethical concerns, and infrastructural conditions within a unified factorial model. The proposed IMPACT model, designed to capture the multidimensional nature of ChatGPT adoption among university faculty, bridges the empirical gap between student-centered frameworks and the practical realities of faculty integration of AI in teaching and assessment.

Although numerous studies have examined the adoption of generative artificial intelligence tools by students, the topic has been scarcely explored among university faculty. The IMPACT model represents an extension of previous theoretical frameworks addressing technology acceptance but incorporates additional dimensions such as ethical and academic concerns, functional appropriation, and perceived reliability. Moreover, it provides a theoretical foundation specifically adapted to the Latin American context, contributing to a deeper understanding of the dimensions influencing ChatGPT adoption among higher education instructors.

3. Materials and Methods

A cross-sectional study was conducted, corresponding to an explanatory investigation with emphasis on the structural component, framed within the relational–explanatory level (Hurtado de Barrera, 2012), with a quantitative orientation and an analytical observational design.

The population consisted of faculty members from the five undergraduate schools of Universidad Técnica del Norte (Ibarra, Ecuador) during the March 2025–August 2025 academic term: 142 from the Health Sciences School; 150 from Education, Arts, and Technology; 91 from Applied Science and Engineering; 77 from Agro-Environmental and Biotechnology Sciences; and 81 from Economic, Legal, and Administrative Sciences, for a total population of N = 541 faculty members.

For this study, a probabilistic sample (n) was drawn using stratified sampling, with each school representing a stratum. The sample size was calculated using a 5% significance level (Z = 1.96), an allowable sampling error (e) of 5.5%, and a 50% success probability (p), which represents the most conservative condition that maximizes sample size. The formula used for calculating the sample size, applicable to a finite population (Mukti, 2025), was:

$$\mathrm{n}= {\displaystyle \frac{{\mathrm{Z}}^2\mathrm{p}\mathrm{q}\mathrm{N}}{{\mathrm{e}}^2\left(\mathrm{N}-1\right)+{\mathrm{Z}}^2\mathrm{p}\mathrm{q}}}$$

(1)

In Equation (1), n is the number of faculty members who took part in the study, and N is the number of professors at the university. For this study, the sampling error tolerance for the estimation process, represented by the variable e, was set at 5%. At the 95% confidence level, Z is the critical value. p and q are the estimated percentages of people who have and do not have the relevant trait, respectively. Using these criteria, the smallest sample size was determined to ensure that the faculty population was statistically representative. p and q denote the estimated proportions of individuals who possess and do not possess the characteristic of interest, respectively. This approach guarantees that the resulting sample adequately reflects the variability within the target population, providing a reliable basis for subsequent statistical analyses.

The application of the formula yielded a sample size of 201, which was increased to 206 for practical purposes related to stratum distribution. Stratification was conducted using a two-stage sampling design. In the first stage, proportional allocation was applied using the following formula (Paliz Sánchez et al., 2024; Sánchez Espejo, 2019):

$$\mathrm{Stratum} \mathrm{sample} \mathrm{size}= \left({\displaystyle \frac{\mathrm{n}}{\mathrm{N}}}\right)\times \mathrm{Stratum} \mathrm{size}$$

(2)

and subsequently, the participants were selected through simple random sampling.

In Equation (2), n stands for the total number of samples that were already chosen for the study, and N stands for the total number of faculty members in all strata. Stratum size is the number of faculty members in each stratum, such as a department, academic area, or faculty. The term (n/N) shows the sampling fraction used for each stratum to get a fair distribution. This procedure ensures that each subgroup of the population is represented in the sample according to its actual proportion in the total population, thereby improving the accuracy and representativeness of the estimates.

Participants were categorized into generational cohorts based on their year of birth, following internationally recognized classifications (

Table 1. Generational Cohort Classification Criteria.

Generation	Birth Years	Key Characteristics	Authors & Years
Traditionalists/Silent	1928–1945	Experienced pre-technology era; value discipline and hard work.	Ismail and Shaari (2019), Ismail et al. (2025); V. H.-C. Wang et al. (2022)
Baby Boomers	1946–1964	Experienced post-WWII economic boom; value loyalty and work ethic.	Ismail and Shaari (2019), Ismail et al. (2025); Fernandes et al. (2025)
Generation X	1965–1980	Adapted to both pre- and post-digital worlds; value independence and work–life balance.	Ismail and Shaari (2019), Ismail et al. (2025); Fernandes et al. (2025); Yurtseven and Karadeniz (2020)
Millennials/Generation Y	1981–1996	Digital natives; value flexibility and social consciousness.	Ismail and Shaari (2019), Ismail et al. (2025); Fernandes et al. (2025); Srivastava (2024)
Generation Z	1997–2012	Grew up with digital technology; value instant access to information and social media.	Ismail et al. (2025); Fernandes et al. (2025); Srivastava (2024)
Generation Alpha	2013–2024	Known as “tech thumbs,” highly integrated with technology from birth.	Srivastava (2024); V. H.-C. Wang et al. (2022)

Note. Generational cohort classification adapted from Ismail and Shaari (2019), Ismail et al. (2025), Fernandes et al. (2025), Srivastava (2024), V. H.-C. Wang et al. (2022), and Yurtseven and Karadeniz (2020).

). This framework allows for the analysis of adoption patterns considering generational differences in technological familiarity and attitudes toward innovation.

Table 2. Sociodemographic and Professional Characteristics of the Participants (n = 206).

Variable	Categories/Values	%/M (SD)
Academic area	Health	26.2
	Education and Technology	27.7
	Applied Engineering	17.0
	Agro-Environmental Sciences and Biotechnology	14.1
	Administration and Economics	15.0
Age (years)	Range: 28–77	M = 45.0 (SD = 9.5)
Sex	Male	57.3
	Female	42.7
Generation cohort	Baby Boomers	6.0
	Generation X	43.2
	Generation Y/Millennials	49.5
	Generation Z	1.0
Years of teaching experience	<6 years	2.4
	6–10 years	20.4
	11–15 years	32
	16–20 years	24.8
	21–25 years	14.1
	>25 years	6.3
Self-identification (ethnicity)	Mixed-race	93.7
Nationality	Ecuadorian	97.6

summarizes the sociodemographic and professional characteristics of the participating university faculty members (n = 206). The participants represented a diverse academic workforce spanning five disciplinary areas and a wide age range, indicative of an experienced and mature teaching population. Most respondents were in mid-career stages, corresponding primarily to Generations X and Y, which aligns with the predominance of educators who have navigated both pre- and post-digital academic environments. Overall, the sample reflects a balanced gender distribution and a high level of professional experience, consistent with the demographic profile of Ecuadorian higher education faculty.

In addition to sociodemographic data, participants reported how frequently they use ChatGPT for academic and research purposes. As summarized in

Table 3. Frequency of ChatGPT Use in Academic and Research Activities (n = 206).

Usage Level 1	%
Never	0
Rarely (less than once a month)	14.1
Occasionally (1–3 times per month)	29.6
Frequently (1–3 times per week)	44.7
Very frequently (almost every day)	11.7

¹ Categories were derived from a self-reported five-point frequency scale.

, the results indicate that most faculty members engage with generative AI tools on a regular basis, reflecting a moderate-to-high level of integration of ChatGPT into their professional routines. This behavioral pattern underscores the growing familiarity of university educators with AI-assisted academic practices and supports the need to examine the factors underlying their adoption through the IMPACT model.

The CHASSIS model, developed and validated by Pereira-González et al. (2025a), and originally composed of 39 items, was adapted to construct the ChatGPT Adoption Intention Questionnaire for university faculty. This theoretical framework is grounded in a strategic matrix integrating constructs from established technology acceptance theories (UTAUT, UTAUT2, TAM, and TAM2), supported by empirical evidence and by recent adaptations of validated instruments incorporating generative artificial intelligence (Davis, 1989; Venkatesh et al., 2003, 2012; Venkatesh & Davis, 2000; Bolívar-Cruz & Verano-Tacoronte, 2025; Dwivedi et al., 2023; García-Alonso et al., 2024; Menon & Shilpa, 2023; Romero-Rodríguez et al., 2023; Sallam et al., 2023).

Since the original items were in English, two independent bilingual translators with expertise in educational technology conducted a forward–backward translation. A team of three educational psychologists reviewed the Spanish adaptation to ensure technical, semantic, and conceptual equivalence (Ponce Gea & Serrano Pastor, 2022), adjusting terminology to the Spanish-speaking university context and replacing generic references to “technological tools” with “ChatGPT.” The experts evaluated each item in terms of relevance, clarity, and theoretical coherence, removing repetitive or ambiguous items that did not meet the agreement criteria (CVC > 0.80) (Beaton et al., 2000; Boateng et al., 2018).

Due to the change in population between the CHASSIS model and the instrument proposed in this study, a new exploratory factor analysis (EFA) was carried out to examine the structure and identify possible differences in response patterns based on faculty characteristics, professional experience, and institutional environment. Two experts in higher education methodology and educational technology further refined the items to ensure contextual and linguistic appropriateness for faculty respondents. Cognitive validity was assessed with 25 university professors selected across faculties using the think-aloud technique to evaluate comprehension and interpretation. Problematic items were restructured prior to administering the final version of the instrument.

The final questionnaire was digitized using Microsoft Forms, maintaining the same five-point Likert scale employed in the CHASSIS model: 1 = Strongly disagree, 2 = Disagree, 3 = Neither agree nor disagree, 4 = Agree, and 5 = Strongly agree. The exploratory analysis conducted in SPSS v.30 led to a redistribution of factors and the elimination of six items that did not meet psychometric criteria (factor loadings < 0.50, cross-loadings, or low communalities), resulting in a 33-item instrument with high validity and reliability for research in higher education.

The selection of items for the IMPACT questionnaire was based on theoretical and empirical foundations. Each item was selected to reflect constructs considered in existing technology acceptance theories (UTAUT, UTAUT2, TAM, TAM2) and to capture the specific characteristics of academic use of generative artificial intelligence, particularly those associated with ethical and academic dimensions.

To ensure semantic consistency in the Likert-scale statements, items expressing concerns, risks, or negative perceptions about ChatGPT use were conceptually reverse-coded. The inversion was performed using the transformation X_(reverse-coded) = k + 1 − X, where k is the number of response categories (in this case, 5). This recoding ensured that higher scores across all items consistently reflected a more favorable attitude toward ChatGPT use. In total, prior to conducting EFA, items from the following factors were reverse-coded: (a) ethical and academic concerns, (b) cost and financial accessibility, and (c) social influence/anxiety.

McDonald’s (2013) ω was used as an indicator of internal consistency, following theoretical foundations and current best methodological practices (Malkewitz et al., 2023; Rogers, 2022), as it does not assume tau-equivalence and is more appropriate for multifactorial structures. To calculate McDonald’s ω, the closed-form algorithm for estimating factor loadings proposed by Hancock and An (2020), ω_ha, was used, as implemented in the SPSS macro developed by Hayes and Coutts (2020). However, Cronbach’s α was also reported to facilitate comparisons with previous literature and provide boundary estimates.

To determine the number of factors to extract in adapting the model to the faculty context, four criteria were employed: (a) Kaiser’s a, with eigenvalues greater than one; (b) the scree plot (Cattell, 1966); (c) parallel analysis (Horn, 1965); and (d) the minimum average partial (MAP) test (Velicer, 1976). For the last two methods, O’Connor’s (2000) syntax for SPSS was used.

For the univariate normality assumption, skewness and kurtosis values were considered acceptable when below an absolute value of 1.5 (George & Mallery, 2010). To assess multicollinearity, the inverse of the correlation matrix was calculated, and a direct inspection of the diagonal elements was conducted.

An oblique rotation (direct Oblimin) was employed based on the theoretical assumption that the instrument’s latent dimensions could be correlated (J. D. Brown, 2009; Costello & Osborne, 2005; Dai et al., 2025; Lloret-Segura et al., 2014), which is common in models of technology perception and use. Oblique rotation enables the identification of complex, non-orthogonal structures with greater fidelity. Kaiser normalization, the default option in the rotation algorithm, was applied.

Pearson’s correlation matrix was used for EFA due to two advantages offered by SPSS: (1) it allows extraction using principal axis factoring (Rogers, 2022), a recommended method given the behavior of the collected data (Fabrigar et al., 1999), under the assumption that underlying factors explain the observed covariances among items; and (2) it allows the specification of the delta parameter in the direct Oblimin rotation method. Additionally, polychoric correlations are more complex estimations that attempt to better capture the relationships among ordinal variables, such as Likert-scale responses. Since they are not directly estimated (but derived from the assumption of underlying normally distributed continuous variables), they tend to introduce greater variability and sampling error compared to the Pearson matrix, which is calculated directly from continuous or treated-as-continuous data. Consequently, a larger dataset is needed to stabilize estimates and reduce error in factor extraction when using polychoric matrices (Lloret-Segura et al., 2014).

Oblique rotation was selected because, unlike orthogonal rotation which assumes latent variables are independent, some correlation among factors was expected. A delta value of 0.2 was chosen, as it yielded a more interpretable and coherent factor solution and accommodated moderate correlations without forcing strong dependencies—an approach consistent with factorial structures in educational or psychological contexts where constructs tend to be correlated but not redundant (Costello & Osborne, 2005; Jennrich, 1973; Lloret-Segura et al., 2014).

To enhance the external and contextual validity of the model being adapted for university faculty, six additional questions were included to obtain a broader characterization of participants’ current experiences and attitudes toward ChatGPT. These questions capture key dimensions not directly covered by the model’s latent factors, such as current frequency of ChatGPT use—measured on an ordinal scale (1 = Never, 2 = Rarely [less than once a month], 3 = Occasionally [1–3 times a month], 4 = Frequently [1–3 times per week], 5 = Very frequently [almost every day])—as well as concrete perceptions regarding perceived academic usefulness, the reliability of the information provided, ethical concerns about authorship, future willingness to use the tool, and an overall evaluation of the platform. The latter five self-reported questions reflect participants’ self-perceptions and were measured using a continuous scale from 0 to 10, allowing greater sensitivity in capturing variability in perceptions compared to categorical scales. Their inclusion is methodologically justified as a complementary strategy for assessing related constructs that may influence the intention to use or reflect significant variations among user profiles. These variables also enable the subsequent development of descriptive, correlational, and ordinal regression analyses to explore relationships between current use and future expectations, as well as potential moderating effects on the intention to use as measured by the structural model. This approach strengthens the study design by integrating experiential and attitudinal data of an empirical nature, providing additional evidence for a more comprehensive and contextually grounded interpretation of the exploratory factor analysis results.

In the final version of the manuscript in English, the generative artificial intelligence tool ChatGPT (GPT-4o) was employed as a supportive resource for stylistic, grammatical, and editorial refinement. This assistance was strictly limited to improving the language and style of the text and did not influence the scientific content, the data analysis, or the interpretation of the results. The authors have reviewed and edited the final version and take full responsibility for the content of this publication.

4. Results

An exploratory factor analysis (EFA) was conducted to assess whether the CHASSIS model (comprising seven factors proposed by Pereira-González et al., 2025a) could be replicated in a faculty sample. The analysis was based on a questionnaire consisting of 39 items, applied to a sample of 206 participants, with ordinal data measured on a five-point Likert scale: Strongly disagree (1), Disagree (2), Neither agree nor disagree (3), Agree (4), Strongly agree (5).

4.1. Normality

Univariate normality was assessed using skewness and kurtosis statistics. All item values were found within the acceptable range of ±1.5 (skewness: –1.21 to 0.25; kurtosis: −1.06 to 1.70) (Burdenski, 2000; Demir, 2022; Ferrando & Anguiano-Carrasco, 2010; Tabachnick & Fidell, 2013), indicating an approximately normal univariate distribution. Only item IU_I3 showed elevated kurtosis (4.02), but it was deemed tolerable since all other items exhibited acceptable behavior.

For the multivariate normality test, Mardia’s coefficients were calculated using the SPSS macro adapted from DeCarlo (1997), Appendix A. This test revealed significant kurtosis values for all analyzed factors, indicating a violation of the assumption of multivariate normality (b₂ₚ > expected value; p < 0.001). Consequently, estimation methods based on maximum likelihood were ruled out, and a more robust extraction method was chosen—specifically, principal axis factoring—as recommended by Cain et al. (2017), Fabrigar et al. (1999), and Mardia (1970).

4.2. Sample Adequacy and Sphericity Analysis

The assessment of data suitability for exploratory factor analysis yielded a Kaiser–Meyer–Olkin (KMO) index of 0.89, which corresponds to the “meritorious” classification according to Kaiser’s (1974) interpretation. Additionally, Bartlett’s test of sphericity was significant (χ² = 3789, df = 378, p < 0.001). Individual sampling adequacy was also evaluated using MSA (Measure of Sampling Adequacy) indices, calculated from the diagonal of the anti-image correlation matrix. Values ranged from 0.703 to 0.936, exceeding the recommended minimum threshold of 0.50, thus supporting the appropriateness of including each variable in the exploratory factor analysis (Lorenzo-Seva & Ferrando, 2021; Sappaile et al., 2023). Taken together, the evidence from the global KMO index and Bartlett’s test confirms that the correlation matrix meets the necessary conditions to proceed with factor extraction.

4.3. Determination of the Number of Factors to Extract

Four complementary methods were applied to determine the optimal number of factors to extract: Kaiser’s rule, the scree plot, the Minimum Average Partial (MAP) method, and parallel analysis with permutations.

Kaiser’s rule identified five factors with eigenvalues greater than 1: 8.484, 5.328, 2.230, 1.806, and 1.241. The scree plot showed an inflection point at the fifth factor, with a clear drop in eigenvalues up to that point, followed by curve stabilization from the sixth factor onward. The MAP criterion yielded a minimum average squared partial correlation of 0.0174, suggesting the retention of five factors. In the parallel analysis—conducted using principal axis factoring and generating random datasets via within-variable permutations to preserve each variable’s distributional shape (randtype = 2)—the observed eigenvalues exceeded the 95th percentile of simulated data up to the fifth factor.

Table 4. Total variance explained.

	Initial Eigenvalues			Rotation Sums of Squared Loadings 1
Factor	Total	%Variance	%Cumulative	Total
EAC	8.484	30.300	30.300	7.917
PRT	5.328	19.028	49.328	4.614
CA	2.230	7.964	57.293	4.900
FACT	1.806	6.451	63.743	3.783
FC	1.241	4.431	68.175	5.884

¹ Extraction method: Principal Axis Factoring.

presents the complete breakdown of eigenvalues and explained variance. Given the consistency across all four methods, and in contrast to the original CHASSIS model, the decision was made to extract five factors, revealing a distinct structural pattern when the model is applied to faculty data.

4.4. Item Exclusion Criteria

Before initiating the item refinement process, a thorough review of the CHASSIS theoretical model (CHatGPT Adoption and Sustained usage among Students in Institutional Settings; Pereira-González et al., 2025a) was conducted, contrasting it with established conceptual frameworks such as the Technology Acceptance Model (TAM; Davis, 1989), the Unified Theory of Acceptance and Use of Technology (UTAUT; Venkatesh et al., 2003), and its extension UTAUT2 (Venkatesh et al., 2012). This review ensured conceptual consistency and content validity prior to the factor analysis.

When communalities range between 0.4 and 0.7, and each factor is measured by three to four items, a sample size of 200 participants is considered adequate (Lloret-Segura et al., 2014). Under these conditions, a factor loading is considered significant when it exceeds 0.40 (MacCallum et al., 1999; Mundfrom et al., 2005; de Winter et al., 2009). Based on these criteria, items were removed during the empirical phase if they showed factor loadings below 0.40 and communalities below 0.30, which indicate low representativeness of the latent construct and insufficient variance explained by the common factors, respectively (Costello & Osborne, 2005; Lloret-Segura et al., 2014).

In addition, items with factor loadings below 0.50 and cross-loadings on two or more factors were excluded if the difference between those loadings was less than 0.10, suggesting conceptual ambiguity and a threat to the model’s discriminant validity (Howard, 2016; Hair et al., 2019; Tabachnick & Fidell, 2013). This strategy aimed to maximize the model’s parsimony and its theoretical and statistical coherence.

The final version of the instrument included 28 items, resulting in a case-to-item ratio greater than 7. This value falls within the empirically supported range in the specialized literature, which generally recommends a minimum ratio of 3 to 10 participants per item to ensure stable factor solutions in both exploratory (EFA) and confirmatory factor analyses (CFA) (MacCallum et al., 1999; Gorsuch, 1983; Floyd & Widaman, 1995).

Although some authors advocate for more stringent criteria (e.g., 10:1 or even 20:1), recent research acknowledges that a ratio of at least 5 to 10 cases per item is adequate in contexts where factor loadings are moderate and average communalities are acceptable (Hair et al., 2019; Lloret-Segura et al., 2014). In this case, the ratio achieved provides a robust foundation for the factor analysis without compromising the validity of the results.

4.5. Correlation Between Items

Table 5. Factor correlation matrix corresponding to the IMPACT model.

Factor	1	2	3	4	5
1	1.000
2	0.015	1.000
3	−0.395	−0.485	1.000
4	0.457	0.295	−0.435	1.000
5	0.762	0.036	−0.335	0.474	1.000

Extraction method: principal axis. Rotation: Oblimin with Kaiser normalization. Significant correlations between factors (|r| ≥ 0.30) are highlighted in bold, justifying the use of oblique rotation.

presents the correlations among the resulting factors. These correlations are consistent with the obliqueness assumption of the model and allow for the assessment of the relative independence among the construct’s dimensions, in accordance with the interpretation guidelines proposed by Hair et al. (2019), Costello and Osborne (2005), and Ferrando and Lorenzo-Seva (2014).

As shown in Table 5, several factors exhibit absolute correlations equal to or greater than 0.30—highlighted in bold—which supports the use of oblique rotation (Oblimin with Kaiser normalization). This pattern of correlations indicates that, although the factors represent distinct dimensions, they share a certain degree of explanatory variance, which is to be expected given the interdependent nature of the constructs assessed in the IMPACT model.

Moreover, the high correlation between Factor 1 and Factor 5 (r = 0.762) may suggest a potential hierarchical relationship or the presence of a second-order dimension, a possibility that should be further explored in future studies using confirmatory factor analysis or structural equation modeling.

4.6. Multicollinearity

The diagonal of the inverse correlation matrix indicated collinearity values ranging from 1.85 to 4.12. These values can also be obtained through a linear regression in which the dependent variable is the probability of future use of ChatGPT, and the items are treated as independent variables. The resulting range suggests the presence of moderate, yet non-problematic, multicollinearity (Kyriazos & Poga, 2023; Vörösmarty & Dobos, 2020).

4.7. Communalities

Regarding the communalities obtained from the exploratory factor analysis, the initial values ranged from 0.460 to 0.757, suggesting that most items shared an acceptable proportion of common variance prior to extraction. After applying the extraction method, final communalities ranged from 0.392 to 0.784. These values indicate that the retained factors adequately explain the variance of the observed items, especially considering that a threshold of 0.40 is commonly accepted as the minimum value for retaining an item in the model (Ferrando & Lorenzo-Seva, 2014; Lloret-Segura et al., 2014). In this regard, the results support the appropriateness of the selected items to represent the proposed latent constructs.

Details of both initial and extracted communalities can be consulted in

Table 6. Communalities in the Exploratory Factor Analysis of the IMPACT model.

Item	Description	Initial	Extraction *
EAC_I1	Inv_I feel that relying on ChatGPT could reduce my ability to analyze and reflect on my own.	0.678	0.701
EAC_I2	Inv_I am concerned that ChatGPT may limit the development of my critical thinking.	0.583	0.447
EAC_I3	Inv_I believe that the constant use of ChatGPT may affect my ability to think critically.	0.622	0.578
EAC_I4	Inv_I feel that by using ChatGPT, the authorship of my work may be questioned.	0.617	0.562
EAC_I5	Inv_I am concerned that my work may lose originality if I use ChatGPT.	0.621	0.643
EAC_I6	Inv_I am concerned that by using ChatGPT, the authorship of my scientific-academic work may become unclear.	0.540	0.483
EAC_I7	Inv_I feel that using ChatGPT’s responses may lead me to commit unintentional plagiarism.	0.540	0.504
PRT_I1	I consider ChatGPT’s responses to be accurate and reliable.	0.564	0.621
PRT_I2	ChatGPT’s responses are usually accurate and useful for my academic and research work.	0.651	0.660
PRT_I3	I trust the accuracy of the information provided by ChatGPT.	0.650	0.733
CA_I1	Inv_The cost of advanced ChatGPT versions may represent a barrier to its use in my academic and research work.	0.736	0.745
CA_I2	Inv_The cost of advanced ChatGPT versions is a barrier to my access and use of the tool.	0.730	0.784
CA_I3	Inv_I wish I had greater financial access to use the advanced models and features of ChatGPT.	0.533	0.486
CA_I4	Inv_The cost of advanced ChatGPT versions may influence how often I use it.	0.741	0.764
CA_I5	Inv_The cost of technology access affects my use of ChatGPT.	0.495	0.392
FACT_I1	The use of ChatGPT has had a positive impact on my work performance.	0.557	0.500
FACT_I2	Using ChatGPT helps me carry out my academic and research tasks more efficiently.	0.746	0.708
FACT_I3	ChatGPT provides me with valuable resources for my job performance.	0.622	0.554
FACT_I4	My work performance has improved thanks to the support of ChatGPT.	0.711	0.689
FACT_I5	I believe that using ChatGPT is useful for improving my job performance.	0.674	0.632
FACT_I6	ChatGPT facilitates my learning and understanding of scientific-academic topics.	0.623	0.598
FACT_I7	I am satisfied with the overall experience of using ChatGPT in my teaching and research work.	0.757	0.746
FC_I1	My institution facilitates access to tools like ChatGPT.	0.485	0.560
FC_I2	I perceive institutional support for the use of ChatGPT.	0.460	0.485
FC_I3	I receive support from my institution to use ChatGPT in my work.	0.566	0.696
FACT_I8	I consider it important to continue using ChatGPT in my teaching and research practice.	0.708	0.697
FACT_I9	I am likely to continue using ChatGPT as a support tool in my academic and research work.	0.695	0.635
FACT_I10	I plan to continue using ChatGPT in the future.	0.704	0.644

* Extraction method: principal axis factorization.

. As shown there, most items presented adequate extraction values, falling within ranges that reflect a substantive relationship with the identified latent factors. Specifically, 26 out of the 28 items showed communalities greater than 0.45, indicating that a considerable proportion of their variance is explained by the proposed factorial structure. This finding supports the model’s internal consistency and suggests a coherent factorial representation aligned with the theoretical dimensions. However, two items (EAC_I2 and CA_I5) presented communalities below 0.45, which may suggest lower integration within the general construct. Nevertheless, they did not fall below critical thresholds, warranting exclusion and were retained based on their conceptual relevance.

4.8. Latent Variables and Proposed Model

The various extraction methods applied converged in identifying a five-factor structure as the most appropriate for the model in university faculty. The complete pattern matrix, including factor loadings ≥0.40 after Oblimin rotation and supported by theoretical or empirical justification, is available in

Table 7. IMPACT Model Pattern Matrix.

Item	Factor *
	FACT	EAC	CA	FC	PRT
I believe that using ChatGPT is useful for improving my job performance.	0.929
I am likely to continue using ChatGPT as a support tool in my academic and research work.	0.913
I plan to continue using ChatGPT in the future.	0.913
My work performance has improved thanks to the support of ChatGPT.	0.905
I am satisfied with the overall experience of using ChatGPT in my teaching and research work.	0.875
Using ChatGPT helps me carry out my academic and research tasks more efficiently.	0.855
I consider it important to continue using ChatGPT in my teaching and research practice.	0.847
ChatGPT provides me with valuable resources for my job performance.	0.754
ChatGPT facilitates my learning and understanding of scientific-academic topics.	0.738
The use of ChatGPT has had a positive impact on my work performance.	0.678
Inv_I feel that relying on ChatGPT could reduce my ability to analyze and reflect on my own.		0.903
Inv_I believe that the constant use of ChatGPT may affect my ability to think critically.		0.814
Inv_I am concerned that my work may lose originality if I use ChatGPT.		0.802
Inv_I feel that by using ChatGPT, the authorship of my work may be questioned.		0.771
Inv_I feel that using ChatGPT’s responses may lead me to commit unintentional plagiarism.		0.758
Inv_I am concerned that ChatGPT may limit the development of my critical thinking.		0.692
Inv_I am concerned that by using ChatGPT, the authorship of my scientific-academic work may become unclear.		0.643
Inv_The cost of advanced ChatGPT versions is a barrier to my access and use of the tool.			0.993
Inv_The cost of advanced ChatGPT versions may influence how often I use it.			0.938
Inv_The cost of advanced ChatGPT versions may represent a barrier to its use in my academic and research work.			0.936
Inv_The cost of technology access affects my use of ChatGPT.			0.656
Inv_I wish I had greater financial access to use the advanced models and features of ChatGPT.			0.653
I receive support from my institution to use ChatGPT in my work.				0.884
My institution facilitates access to tools like ChatGPT.				0.830
I perceive institutional support for the use of ChatGPT.				0.669
I trust the accuracy of the information provided by ChatGPT.					0.982
I consider ChatGPT’s responses to be accurate and reliable.					0.911
ChatGPT’s responses are usually accurate and useful for my academic and research work.					0.771

* Extraction method: principal axis factoring. Rotation method: Oblimin with Kaiser normalization. The rotation converged in 5 iterations.

.

The new theoretical structure, validated by empirical findings, was composed of the following dimensions:

• Factor 1: Functional Appropriation of ChatGPT Technology (fusion of intention to use and perceived usefulness), FACT;
• Factor 2: Ethical and Academic Concerns, EAC;
• Factor 3: Cost and Accessibility, CA;
• Factor 4: Facilitating Conditions, FC;
• Factor 5: Perceived Reliability and Trustworthiness, PRT.

Based on the extracted factors, the resulting model was named IMPACT: Instrument for Measuring the Perceived Appropriation of ChatGPT in Teaching.

4.9. Internal Consistency

To assess the internal consistency of each subscale, Cronbach’s alpha and McDonald’s omega (hierarchical approximation version, ω_ha) coefficients were calculated based on the Pearson correlation matrix using the SPSS macro developed by Hayes and Coutts (2020). The results indicate high reliability across all subscales, with both α and ω(HA) values exceeding 0.79. The overall internal consistency of the instrument was 0.858 (α) and 0.764 (ω_ha), suggesting that while the subscales exhibit high internal homogeneity, the global construct may reflect a potentially multidimensional structure.

Table 8. Reliability estimates by subscale (α and ω_ha).

Subscale	α (Pearson)	ω (McDonald HA)
EAC	0.894	0.893
PRT	0.852	0.855
CA	0.882	0.886
FACT	0.940	0.941
FC	0.795	0.797
General	0.858	0.764

summarizes the reliability indices obtained for each subscale and for the instrument as a whole.

5. Discussion

5.1. Main Findings

The factorial structure of the IMPACT model, adapted for university faculty, shows a significant reorganization compared to the original CHASSIS model developed for students (Pereira-González et al., 2025a). Five factors emerged in the IMPACT model: (1) Functional Appropriation of ChatGPT Technology (FACT), which integrates items related to intention to use and performance expectancy; (2) Ethical and Academic Concerns (EAC); (3) Cost and Accessibility (CA); (4) Facilitating Conditions (FC); and (5) Perceived Reliability and Trustworthiness (PRT).

In contrast, the CHASSIS model included seven distinct factors: Ethical and Academic Concerns (EAC), Performance Expectancy (PE), Cost and Financial Accessibility (CFA), Intention to Use (IU), Social Influence/Anxiety (SIA), Perceived Confidence and Reliability (PCR), and Facilitating Conditions (FC). The comparison between both models reveals two key transformations: the merging of the IU and PE factors into a more robust component (FACT), and the non-emergence of the SIA component in the faculty sample. These results are consistent with findings reported by Bolívar-Cruz and Verano-Tacoronte (2025), and Valaidum and Mahat (2024), who also observed that among experienced professionals, behavioral intention tends to integrate with perceived usefulness when technology use becomes instrumental to performance goals.

The FACT dimension was supported by high factor loadings (ranging from 0.678 to 0.929), strong internal consistency (McDonald’s ω_ha = 0.941), an in-depth semantic analysis (see Appendix B), and solid theoretical backing. This factorial convergence aligns with prior findings in technology acceptance models such as TAM (Davis, 1989) and its extension TAM2 (Venkatesh & Davis, 2000), which established that perceived usefulness becomes the dominant predictor of behavioral intention once users gain expertise. More recent studies (Dwivedi et al., 2019; Barakat et al., 2025; Strzelecki et al., 2024) corroborate this pattern, emphasizing that experienced users internalize perceived utility as part of their professional practice rather than as a preliminary motivational stage. Moreover, prior studies have documented practical differences in technology adoption attitudes between students and faculty (Bhaskar & Rana, 2024; Bhat et al., 2024; Scherer et al., 2019; van Raaij & Schepers, 2008).

This reconfiguration therefore suggests that, for faculty, behavioral intention to use is mediated by perceived usefulness and integrate into a functional dimension oriented toward professional performance. It reflects a more pragmatic, internalized, and sustained appropriation of the tool, consistent with evidence from Bhaskar and Rana (2024) and Scherer et al. (2019), who found that faculty members’ technology adoption is shaped more by cognitive evaluations of functionality and less by affective or social influences. The concept of functional appropriation also parallels findings by van Raaij and Schepers (2008), who demonstrated that in professional environments, usage intention and performance expectancy converge as interdependent processes that guide sustained adoption.

In this context, the functional appropriation of ChatGPT can be understood as the extent to which university faculty internalize its use as a valuable, efficient, and necessary resource for their academic and research activities, demonstrating a sustained willingness to integrate it into their professional practice. This appropriation implies not only behavioral intention but also functional valuation, perceived usefulness, and satisfaction with the outcomes. While this factor has not been specifically defined in previous studies, the present findings are consistent with those reported by Barakat et al. (2025), Dwivedi et al. (2019) and Strzelecki et al. (2024). This finding cannot overlook the growing concerns about ChatGPT’s limitations in academic contexts, where responses often appear credible but may be formulaic, outdated, or lack precise substantiation (Mahama et al., 2023). The notion of functional appropriation thus extends beyond mere technical adoption but cannot ignore the risk of what Martín-Crespo Rodríguez (2024) describes as a form of pseudo-instrumental rationality—a technocratic mindset that prioritizes efficiency and procedural performance while paradoxically constraining critical engagement and human creativity.

Although faculty may perceive ChatGPT as a tool that reduces cognitive workload and improves efficiency, this instrumental dependence risks fostering a form of cognitive somnolence —a condition in which users uncritically accept AI-generated results without subjecting them to rigorous evaluation (Martín-Crespo Rodríguez, 2024). This observation aligns with the warnings of Mahama et al. (2023), who emphasize that AI-assisted academic writing can undermine intellectual autonomy and academic integrity by normalizing uncritical dependence on algorithmic text generation.

The convergence between these findings on functional appropriation and broader critiques suggests that patterns of generative AI adoption in higher education should be examined not only through behavioral and usage metrics but also through their epistemic and ethical implications—specifically, how such technologies may reshape academic labor, critical reasoning, and the very notion of intellectual authorship within universities.

On the other hand, the theoretical factor of Social Influence/Anxiety (SIA) did not emerge as a distinct dimension in the faculty model. Its exclusion was due to factorial collapse and weak loadings, suggesting a lack of structural coherence within this population. This absence may be interpreted considering faculty members’ professional profiles, which generally reflect greater autonomy and lower dependence on social judgment. This finding aligns with the UTAUT model (Venkatesh et al., 2003), which posits that social influence weakens as users gain autonomy and experience. Similar patterns have been reported in higher education contexts, where professional independence reduces the role of social judgment in decision-making (Booyse & Scheepers, 2024; Hakimi et al., 2024). Furthermore, in institutional settings where the use of AI tools is still emerging and lacks formal regulation, the limited visibility of social or institutional pressure may further explain the non-significance of this construct—consistent with the observations of Palm (2022) and Takahashi et al. (2024) regarding innovation diffusion in educational systems.

5.2. Practical Implications

The IMPACT model provides a theoretically grounded framework for the adoption of generative artificial intelligence tools in educational institutions, addressing the growing interest in understanding the factors that determine the successful integration of emerging technologies in educational settings (Bhat et al., 2024; Yu et al., 2024). Beyond offering a psychometric instrument, the model contributes to institutional policy and pedagogical design by identifying the dimensions most relevant to educators—functional appropriation, ethical and academic concerns, perceived reliability, cost and accessibility, and facilitating conditions—these dimensions highlight that adoption is not only a technical process but also a pedagogical and ethical one.

The emergence of the Functional Appropriation of ChatGPT Technology (FACT) factor suggests that institutional training should move beyond technical proficiency to foster reflective and critical use of AI, aligning with prior findings that sustained adoption depends on perceived pedagogical utility rather than social pressure (Dwivedi et al., 2019; Scherer et al., 2019). Moreover, the inclusion of the Ethical and Academic Concerns dimension underscores the need for policies that prevent what Martín-Crespo Rodríguez (2024) calls pseudo-instrumental rationality—an efficiency-driven mindset that may limit critical engagement—and guard against overreliance or cognitive somnolence in AI use.

Consistent with Mahama et al. (2023), these results suggest that effective integration of generative AI in universities requires balancing functional benefits with ethical oversight to preserve academic integrity and intellectual autonomy.

5.2.1. Factorial Structure and Theoretical Foundation

The five-factor structure of the model aligns with contemporary trends toward more comprehensive and context-sensitive approaches compared to classical models such as TAM and UTAUT2 (Davis et al., 1989; Venkatesh et al., 2012). The Functional Appropriation of ChatGPT Technology (FACT) factor integrates intention to use and performance expectancy—elements that have consistently demonstrated strong predictive power in educational contexts (Feng et al., 2025; Hakimi et al., 2024). This integration is consistent with studies identifying performance expectancy as the strongest predictor of behavioral intention in higher education (Abdi et al., 2025; Faraon et al., 2025; Wedlock & Trahan, 2020) and reflects a pragmatic, outcome-oriented perspective that emerges among experienced faculty (Bhaskar & Rana, 2024; Scherer et al., 2019).

The Ethical and Academic Concerns (EAC) dimension addresses a critical gap in the literature on AI adoption in education. Ethical considerations have increasingly been recognized as determinants of user trust and acceptance, encompassing issues such as academic integrity, algorithmic bias, and data privacy (Bozkurt, 2024; Fowler, 2023). These results reinforce recent arguments that technology adoption models should integrate moral and epistemic responsibility as core constructs (Martín-Crespo Rodríguez, 2024; Mahama et al., 2023).

The Cost and Accessibility (CA) dimension incorporates economic and availability-related barriers, which have been widely documented as critical factors in institutional technology adoption (Capraro et al., 2024; Xiao et al., 2024), particularly in public universities in developing countries, where budgetary constraints can restrict equitable access (Complete College America, 2025; Hughes et al., 2025). This finding aligns with global calls for inclusive AI integration that bridges the digital divide across socioeconomic contexts (Dwivedi et al., 2019).

Facilitating Conditions (FC) represent a well-established construct in UTAUT, with prior research demonstrating that support infrastructure, training, and regulatory frameworks play a crucial role in technology acceptance (Venkatesh et al., 2003; Dysart & Weckerle, 2015; Kong et al., 2024). Within the context of ChatGPT, this includes not only technical support but also AI literacy and governance policies that ensure responsible and transparent implementation (Çer, 2025; Jaipal-Jamani et al., 2018).

Finally, the Perceived Reliability and Trustworthiness (PRT) dimension constitutes a significant contribution to the model, as it captures one of the most pressing concerns surrounding the use of generative artificial intelligence systems in educational environments. Specifically, perceived reliability has been closely linked to the willingness to integrate such tools into teaching and academic practices, since errors, biases, or inconsistent responses can compromise both the quality of learning and user trust (Giannakos et al., 2024; Panda & Kaur, 2024; Miraz et al., 2025). Recent studies have documented performance limitations of ChatGPT in specialized domains that require complex cognitive tasks—such as law, medicine, the hard sciences, or empirical sciences—where high levels of precision and contextual understanding are essential (Cong-Lem et al., 2024). Such findings reinforce the importance of this dimension as an independent determinant influencing both acceptance and sustainable use of generative AI in higher education.

Figure 1 illustrates the final factorial configuration of the IMPACT model, derived from the exploratory factor analysis and grounded in prior technology acceptance frameworks (TAM, TAM2, UTAUT, UTAUT2, and CHASSIS). The figure highlights the theoretical and structural evolution leading to the proposed model, particularly the integration of Intention to Use and Performance Expectancy into Functional Appropriation and the non-significance of Social Influence/Anxiety in faculty contexts.

5.2.2. Implications for Institutional Implementation

The approach proposed in the IMPACT model is consistent with research on the diffusion of innovations in higher education, which highlights the importance of considering multiple levels of analysis and contextual factors (Palm, 2022; Stasewitsch et al., 2022; Takahashi et al., 2024). This multi-layered perspective reinforces the idea that faculty adoption of generative AI cannot be understood in isolation from institutional culture, governance structures, and pedagogical norms, as previously noted in studies on digital transformation in universities (Dwivedi et al., 2019; Scherer et al., 2019).

The model’s capacity to support differentiated strategies based on faculty profiles addresses the well-documented differences in attitudes and readiness among educator groups. Variables such as teaching experience, disciplinary background, and attitudes toward technology have been shown to moderate adoption processes (Abedi & Ackah-Jnr, 2023; Nikoçeviq-Kurti & Bërdynaj-Syla, 2024). This aligns with prior findings indicating that personalized and discipline-sensitive implementation strategies enhance the sustainability of technological innovation in academic contexts (Herodotou et al., 2020; Yu et al., 2024).

The emphasis on continuous evaluation grounded in empirical evidence aligns with best practices in institutional technological change management. Successful implementation of emerging technologies requires iterative monitoring and adaptive feedback mechanisms that ensure sustained alignment with pedagogical objectives and ethical standards (Herodotou et al., 2020; Strielkowski et al., 2025). In this regard, the IMPACT model contributes a structured framework for evaluating not only adoption rates but also the depth and quality of integration, supporting the transition toward evidence-based and ethically grounded AI governance in higher education (Mahama et al., 2023; Martín-Crespo Rodríguez, 2024).

5.2.3. Contribution and Applicability

The IMPACT model provides a framework that enables educational institutions to navigate the inherent complexity of integrating generative AI tools, balancing potential benefits with the minimization of risks and resistance. This approach is particularly relevant in the current context, where institutions face growing pressure to adopt emerging technologies while maintaining educational quality standards and addressing ethical considerations (Booyse & Scheepers, 2024; Feng et al., 2025; Stasewitsch et al., 2022). Similarly to prior frameworks emphasizing responsible digital transformation (Dwivedi et al., 2019; Mahama et al., 2023), the IMPACT model advances a holistic vision of institutional AI adoption that integrates pedagogical effectiveness, transparency, and ethical accountability.

The model offers an evidence-based roadmap that can support informed decision-making processes and increase the likelihood of successful technology implementation in complex and dynamic educational environments. By combining empirical validation with theoretical coherence, it contributes to institutional strategies that promote sustainable and equitable AI integration (Kong et al., 2024; Yu et al., 2024).

Additionally, the IMPACT model advances theoretical research on the emerging adoption of generative artificial intelligence by demonstrating that, in faculty attitudes toward ChatGPT acceptance, social influence is not a determining factor. Instead, adoption behaviors are structured around functional, ethical, and infrastructural considerations. This finding aligns with prior studies reporting that professional autonomy and self-regulated decision-making moderate the role of social influence in technology adoption (Scherer et al., 2019; Bhaskar & Rana, 2024). Consequently, it highlights the need to adapt the conceptualization of existing theoretical models of technology use intention, such as UTAUT and TAM, and suggests that, in academic contexts, the intention to use generative AI is shaped by a more integrated perception of usefulness and professional ethics (Martín-Crespo Rodríguez, 2024; Mahama et al., 2023).

5.3. Methodological Strengths

The present study incorporates several methodological considerations aimed at enhancing the robustness and replicability of the IMPACT model, proposed to evaluate the functional appropriation of ChatGPT among university faculty. First, multiple convergent strategies were employed to determine the optimal number of factors to extract, including Kaiser’s criterion (Kaiser, 1960), the scree plot, parallel analysis, and the Minimum Average Partial (MAP) method, following the recommendations of O’Connor (2000) and Velicer (1976). This methodological triangulation strengthened the empirical validity of the resulting five-factor structure, overcoming limitations associated with the exclusive use of heuristic criteria (Hayton et al., 2004; Fabrigar et al., 1999). Similar approaches have been recommended in psychometric validation studies in education to ensure replicability and theoretical precision (Ferrando & Lorenzo-Seva, 2014; Lloret-Segura et al., 2014).

Second, the assumptions of univariate and multivariate normality were rigorously assessed. While individual items displayed acceptable distributions (skewness and kurtosis ranging from −1.21 to 1.70), Mardia’s test revealed significant deviations from multivariate normality. This finding led to the selection of principal axis factoring, a method more robust to such violations (Cain et al., 2017; Ferrando & Anguiano-Carrasco, 2010). This decision aligns with best practices in psychometric analysis for non-normally distributed data, ensuring that factor solutions remain stable even when multivariate assumptions are not met (Kyriazos & Poga, 2023).

In addition, McDonald’s hierarchical omega (ω_ha) was computed for each subscale using the closed-form algorithm proposed by Hancock and An (2020), implemented in SPSS via the Hayes and Coutts (2020) macro. This approach overcomes the limitations of Cronbach’s alpha by not requiring tau-equivalence and provides a more accurate estimate of internal consistency in multifactorial structures (Zinbarg et al., 2005; Malkewitz et al., 2023). The inclusion of ω_ha as a reliability index strengthens the psychometric rigor of the study and aligns with current recommendations for evaluating internal structure consistency in social sciences (Deng & Chan, 2017; Valencia Londoño et al., 2025).

Ultimately, a rigorous item refinement process was carried out, grounded in both theoretical and statistically supported criteria. Item removal was based on the combination of low factor loading (<0.40), insufficient communalities (<0.30), and ambiguous cross-loadings, following thresholds established by Costello and Osborne (2005), Lloret-Segura et al. (2014), and Hair et al. (2019). This strategy allowed for the construction of a parsimonious, theoretically coherent instrument with adequate discriminant validity, ensuring the conceptual integrity of each factor. Moreover, the combination of empirical filtering and theoretical validation represents a methodological advance over traditional exploratory approaches, supporting the internal consistency and construct validity of the IMPACT model (Ferrando & Lorenzo-Seva, 2014).

5.4. Study Limitations

Although the results support the psychometric strength of the proposed model, certain methodological limitations must be acknowledged when interpreting the findings.

First, although the sample size used (n = 206) met established empirical criteria—exceeding the minimum 5:1 participant-to-item ratio (Memon et al., 2020) and the recommended threshold of 20 participants per independent variable (Hair et al., 2019)—and was validated using a priori analysis with the A priori Sample Size Calculator for Structural Equation Models (Soper, 2025), recent literature suggests that larger samples may lead to more precise estimates and more stable factor solutions, particularly in studies targeting diverse populations (Lorenzo-Seva & Ferrando, 2024). The a priori analysis, assuming a medium effect size (f² = 0.30), statistical power of 0.80, significance level of 0.05, five latent variables, and 28 observed variables, indicated a minimum of 150 participants to detect significant effects and 148 to adequately represent the model structure.

Second, the study relies exclusively on exploratory factor analysis and requires validation through confirmatory factor analysis using an independent sample. This limitation prevents drawing definitive conclusions regarding the stability and replicability of the identified factorial structure.

Finally, potential biases related to self-selection and socially desirable responding cannot be ruled out. In studies involving academic expectations, professional ethics, or the adoption of tools such as ChatGPT, faculty members may report more favorable attitudes or responses aligned with institutional expectations, even if their actual opinions or behaviors differ from those reported.

5.5. Future Research Directions

The findings of this study open several avenues for further research aimed at strengthening and expanding the IMPACT model toward broader and more contextualized applications. First, it is recommended to validate the factorial structure through confirmatory factor analysis (CFA) on an independent sample stratified by key variables such as gender, age, and academic discipline. This approach would allow for the assessment of factorial invariance across different faculty subpopulations using multigroup techniques (G. T. L. Brown et al., 2015; Fischer & Karl, 2019), thereby ensuring that the instrument’s structure remains stable and generalizable in diverse contexts. Additionally, it would be pertinent to explore second-order hierarchical models that group the identified dimensions under broader latent constructs, allowing for more integrative and parsimonious theoretical interpretations (Gould, 2015).

It is also advisable to compare the instrument with other psychometrically validated questionnaires grounded in related theoretical frameworks. Such cross-validation would enhance the empirical evidence of construct validity and reinforce the scientific foundation of the IMPACT model (Strauss & Smith, 2009; Grand-Guillaume-Perrenoud et al., 2023).

Another important line of inquiry involves examining potential moderation effects by contextual and personal variables, such as teaching experience, academic field, familiarity with artificial intelligence tools, and faculty digital competencies (Dunn et al., 2015; Garrido-Ruso & Aibar-Guzmán, 2022). These comparisons could be explored through multigroup structural equation models or moderation analyses, advancing a more nuanced understanding of university faculty’s adoption of emerging technologies like ChatGPT.

Finally, future studies should prioritize implementing more diversified and rigorous sampling strategies to minimize self-selection bias and enhance sample representativeness, thereby improving the external validity and generalizability of the findings to other educational settings.

6. Conclusions

This study developed and validated, through Exploratory Factor Analysis, the IMPACT model (Instrument for Measuring the Perceived Appropriation of ChatGPT in Teaching), a psychometric tool specifically designed to assess the adoption of generative artificial intelligence—ChatGPT—by university faculty. The findings reveal fundamental differences compared to the model previously applied to students, indicating factorial patterns that reflect the distinctive professional and institutional roles of teaching staff.

Within the resulting five-factor structure of the IMPACT model, a factor emerged describing the Functional Appropriation of ChatGPT Technology (FACT), which integrates two constructs—intention to use and performance expectancy—that appeared as separate dimensions in the student model. This suggests that university faculty experience a more cohesive integration between perceived usefulness and behavioral intention toward generative AI technologies.

The absence of the Social Influence/Anxiety factor, present in CHASSIS, aligns with prior literature documenting increased professional autonomy and reduced dependence on social judgment as users gain experience. This implies that in ChatGPT adoption, faculty members tend to value intrinsic criteria of usefulness and efficiency over social pressure or external expectations.

The four factors common to both models—Ethical and Academic Concerns (EAC), Cost and Accessibility (CA), Facilitating Conditions (FC), and Perceived Reliability and Trustworthiness (PRT)—maintained their conceptual structure, reaffirming their theoretical and empirical relevance in teaching contexts.

The use of principal axis factoring, oblique rotation, and triangulation through multiple extraction criteria resulted in a stable and conceptually coherent factorial structure. Internal consistency values (α between 0.795 and 0.940; ω_ha between 0.797 and 0.941) confirm the reliability of the instrument’s subscales. The factorial solution explained 68.2% of the total variance and yielded a Kaiser–Meyer–Olkin index (KMO = 0.89), classified as meritorious (Kaiser, 1974), with high communalities (0.460 to 0.757) and strong factor loadings (0.678 to 0.929).

The identification of the factorial structure of the IMPACT model provides a foundation for higher education institutions to design specific interventions that strengthen facilitating and accessibility conditions while addressing ethical concerns and implementing strategies to help faculty enhance trust and evaluate the reliability of outcomes generated by such tools.

While this research contributes to the emerging theoretical corpus on ChatGPT adoption in universities, it should be recognized as an initial approach requiring further validation through confirmatory factor analysis and replication across diverse contexts.

Finally, the IMPACT model provides not only a reliable psychometric tool but also a theoretical contribution that serves as a bridge between general technology acceptance models and the ethical–functional realities underlying the adoption of generative artificial intelligence in higher education.