Perception of AI tool adoption and training: initial validation using GSEM method Open Access

The integration of AI tools is crucial in professional work, driving digital transformation and necessitating ongoing skill updates [1]. Mastering AI has become essential for enhancing efficiency, decision-making, and professional practice [2]. Universities support this by providing comprehensive AI training that blends theory with practical applications, helping students apply skills in complex real-world contexts [3]. AI training should be interdisciplinary, addressing both technical and broader impacts on professional roles. Assessing AI training must consider not only instruction but also student perceptions of AI’s impact on their careers and the role of institutions in meeting these demands [4]. Instruments to measure these perceptions are therefore essential, with students needing to show openness toward new technologies.

This study aimed to validate the “Perception of the Adoption and Training in the Use of Artificial Intelligence Tools in the Profession” instrument, designed to assess perceptions of AI knowledge, training, use, adoption, and teaching in professional training. Adapted from the “Attitudes and Perceptions of Students Towards Artificial Intelligence” instrument [5], it was tailored for the Latin American higher education context across multiple disciplines. The validation process involved three stages: conceptual validation, expert review, and validation using GSEM with Quasi-Maximum Likelihood (QML).

Preparing students for future work underscores the value of AI in higher education. The evolving job market demands adaptability and emerging tech proficiency, making AI integration in curricula essential for developing skills in innovation and problem-solving [2, 6]. AI transforms pedagogy by enabling personalized learning, which tailors content to students' individual needs, enhancing engagement and motivation [3, 7–9]. AI also democratizes education by expanding access beyond geographic and socioeconomic barriers, crucial for equitable vocational training and skill development [1, 10]. Additionally, AI-powered simulations provide practical, low-risk environments for students to build real-world skills, supporting comprehensive and accessible professional preparation [11].

Integrating AI tools in university programs is transforming higher education, aligning with modern demands and preparing students for a dynamic job market [3, 4]. Key applications include personalized tutoring systems [12], 24/7 chatbot support [13], and AI-based simulations that enhance experiential learning and critical skills [14]. Effective AI integration requires strategic planning, investment in technology, and teacher training [15]. Collaboration among educators, students, and tech professionals ensures AI meets educational needs and is well-received [16]. Positive student perceptions are crucial for successful adoption, making it essential to consider these attitudes when introducing AI in education.

Recent developments in Latin America emphasize the importance of AI education for adapting to the evolving technological landscape. The Inter-American Development Bank (IDB) [17] underscores the rising demand for AI skills, urging universities to integrate AI to foster innovation and economic growth. National AI strategies in Brazil [18] and Mexico [19] promote specialized programs to bridge skills gaps and establish regional AI leadership.

The Economic Commission for Latin America and the Caribbean (ECLAC) [20] highlights challenges like resource constraints, digital divides, and the need for teacher training, calling for supportive policies to ensure equitable AI integration. Cross-cultural studies [21], including Reyes et al. [22], show that varying infrastructures and policies affect AI adoption, pointing to the need for tailored strategies. UNESCO [23] advocates for context-specific AI education that considers student perceptions, promoting inclusive knowledge societies and reducing technological gaps.

Initiatives like AI4D-LAC [24] focus on research and capacity building, underscoring regional commitment to AI literacy and innovation. AI training is essential for Latin American universities to meet these challenges, improve higher education, and boost economic competitiveness.

There is growing interest in how AI affects educational and vocational training environments. While studies have examined AI acceptance, perception, and training, most focus on specific contexts, highlighting the need for broader, cross-cultural research [6, 11, 12]. Studied focusing on Attitudes and Perceptions of the students of health sciences in UK [5] underscore a need for more practical AI training, while the same instrument was translated for Spanish business and education students, though cultural differences widens its relevance to Latin America. Another review on AI in education takes stock of the trends, however, lacks in empirical data on student perceptions. Further reviews give a picture of AI applications without focusing on student perceptions [6, 9]. Similarly, other studies report AI’s role in education but not giving insight into the student attitudes [11, 12, 16]. Another research employs the use of AI chatbot-based learning but again without exploring student perceptions [14], though addressing these gaps by validating an instrument to measure Mexican university students' perceptions of AI in education, using generalized structural equation modeling (GSEM) with quasi-maximum likelihood (QML). Findings emphasize the need for targeted AI training to promote effective adoption, offering empirical data to guide future research and educational practices in Latin America.

Likert scale for assessing medical students’ attitudes toward AI in their training [5] modified for Spanish research [25] was used in the present investigation.

The “Perception of the Adoption and Training in the Use of Artificial Intelligence Tools in the Profession” instrument was administered to 238 students at a Mexican university to assess their perceptions of AI use and training. The sample included students from various semesters across six disciplines: (1) Architecture, Art, and Design; (2) Health Sciences; (3) Social Sciences and Government; (4) Humanities and Education; (5) Engineering and Science; and (6) Business.

Data was collected via a digital Google Forms survey in October and November 2023. Adhering to ethical research guidelines, all participants provided informed consent, understanding that their responses would be used for academic and research purposes.

Given that GSEM (Generalized Structural Equation Model) estimation with QML (Quasi-Maximum Likelihood) model will be employed for the pilot test, it became imperative to identify two types of variables: latent ^[1] and indicator ^[2].

The latent variables correspond to:

• (1)

  The perception of the value (or lack thereof) of having training for the proper use of these AI tools.

• (2)

  Understanding of AI tools and professional implications.

• (3)

  Attitude towards including this knowledge as part of the professional training process.

The Shapiro-Wilk test was used to determine whether the data followed a normal distribution, but the result was non-normal. Since the data did not follow a normal distribution, traditional validation models such as Exploratory Factor Analysis and Confirmatory Factor Analysis would be less reliable because the violation of normality can affect the model fit. Therefore, we decided to use a GSEM model in Stata 18 [26].

The GSEM, an advanced statistical technique used to analyze complex relationships between variables in social sciences, behavior, and other fields was employed given that it allows for modeling relationships between variables that are direct, indirect, and reciprocal [27] unlike the linear regression models.

According to Skrondal and Rabe-Hesketh [28] in model Y, which takes values from 1 to n, an ordinal family with n outcomes has cutoff points c0, c1, …, cn, where c0 = −∞, cy < cy+1, and cn = +∞.

If we have a linear prediction s, the probability that a random response Y takes the value y is:

where Ya refers to the underlying stochastic component of Y, and its distribution is given by GSEM. The logit link is assigned to Ya with the extreme value distribution.

In this study, a generalized structural equation model (GSEM) was used to examine relationships between latent variables and observed indicators, with model quality assessed using the Akaike Information Criterion (AIC) [29] and Bayesian Information Criterion (BIC) [30]. The GSEM produced an AIC of 4382.872 and a BIC of 4532.179. When comparing models, those with the lowest AIC and BIC are preferable. Here, these values indicate that the proposed model offers a good fit without unnecessary complexity, confirming its suitability.

The model’s good fit, indicated by AIC and BIC metrics, supports the hypothesized relationships among latent variables like understanding AI tools, valuing AI training, and attitudes toward AI in professional training. This validation confirms that the model effectively represents students' perceptions and attitudes, ensuring the instrument’s reliability in measuring these constructs. This is valuable for educators and curriculum designers aiming to enhance AI integration in education.

When conducting a reflective measurement model, assuming that the observed variables (indicators) reflect an underlying latent variable (construct), as shown in Figure 1, it was found that the standardized loadings, representing the correlation between each indicator and the latent construct.

The Figure 1 shows the following:

• (1)

  The latent variable “Understanding of AI tools and professional implications” has standardized loadings between 0.903 and 0.529, indicating that its indicators measure the concept adequately. For “Attitude toward including AI knowledge in professional training,” loadings range from 0.920 to 0.816, showing strong correlations with the indicators. “Perception of the value of AI training” has loadings from 0.930 to 0.828, confirming a robust correlation with its indicators. These high values suggest that all indicators effectively measure their respective latent constructs (see Table 1).

• (2)

  Cronbach’s Alpha (α) and rho_A values exceed the 0.70 threshold, confirming the internal reliability and consistency of the measured latent variables (Table 1).

• (3)

  Discriminant validity was assessed by comparing each construct’s average extracted variance (AVE) with the squared interfactor correlation (CIF) between constructs. The AVE reflects the total variance explained by the latent construct, while CIF indicates the shared variance between two constructs, supporting distinctiveness between constructs.

In Table 2, “Discriminant validity - Squared interfactor correlation vs. Average variance extracted (AVE),” the AVE for each construct is higher than the corresponding CIF across all construct combinations, indicating that each construct measures a unique concept and confirming the model’s discriminant validity.

The Shapiro-Wilk test (Table 3) was used to assess normality, revealing that variables do not follow a normal distribution, as the null hypothesis was rejected at a 1% significance level (Table 1). To account for non-normality, GSEM estimation with Quasi-Maximum Likelihood (QML) was applied. QML relaxes the conditional normality assumptions for standard error estimation, allowing GSEM to function effectively without full joint normality assumptions [29].

Figure 1 illustrates the GSEM model structure, showing that “Attitude towards including AI knowledge in professional training” is influenced by the “Perception of the value of AI training.” This perception, in turn, is shaped by the “Understanding of AI tools and their professional implications.” In GSEM, some coefficients are set to 1 as reference points for related latent variables. The model provides estimated coefficients, standard errors, z-statistics, p-values, and 95% confidence intervals for variable relationships. Detailed estimates are available in  Supplementary Material ( Appendix A) as per editorial policy.

The results indicate:

• (1)

  The latent variable “Understanding of AI tools and their professional implications” significantly impacts the “Perception of the value of AI training” (β = 0.9119261; p < 0.01).

• (2)

  “Perception of the value of AI training” has a significant positive effect on “Attitude toward including AI knowledge in professional training” (β = 0.6761926; p < 0.01).

• (3)

  Belief in AI’s importance in one’s profession (measured by the statement “Artificial Intelligence will play an important role in the teaching and development of my profession”) directly influences “Attitude toward including AI knowledge” (β = 1.043483; p < 0.01).

• (4)

  The latent variable “Understanding of AI tools and their professional implications” positively affects its indicators, such as understanding AI basics (x3), familiarity with AI terminology (x4), recognizing AI’s limitations in their field (x5), and understanding AI’s ethical implications (x6), all significant at the 1% level.

• (5)

  The variable “Perception of the value of AI training” positively impacts indicators like confidence in using AI tools with training (x7) and the ability to evaluate AI tools in their field (x8).

• (6)

  The “Attitude toward AI training” variable significantly affects indicators such as having received AI training (x10) and the belief that AI training is beneficial for professional development (x11).

• (7)

  Variables X2, X9, and X12 were excluded due to lack of statistical significance, as their inclusion would add noise without enhancing the model’s explanatory power.

GSEM analysis revealed significant relationships: a solid understanding of AI tools and their professional implications enhances the perceived value of AI training, indicating that students with greater AI knowledge are more receptive to AI training. Additionally, a positive perception of AI training’s value boosts students' attitudes toward incorporating AI in professional development.

Unlike traditional models, GSEM does not use R-squared but relies on fit criteria like AIC and BIC, which evaluate the balance between fit and complexity—lower values denote a better fit. The AIC and BIC values in Table 4 confirm that the selected model offers the best fit, validating its adequacy.

Based on these findings, the adapted instrument “Perception of Adoption and Training in the Use of Artificial Intelligence Tools in the Profession” is initially valid for its intended purpose. Its utilization aligns with Lo’s [4] emphasis on adapting educational processes to 21st-century demands and the idea that practical AI application in university education enriches learning experiences and maximizes student potential. This underscores the connection between individual perceptions of AI’s impact on professions and the importance attributed to AI in teaching and professional development.

Thus, the main findings obtained are:

• (1)

  The instrument demonstrated high standardized loading coefficients and Cronbach’s Alpha and rho_A values above 0.70, confirming its validity and reliability for measuring students' attitudes and perceptions about adoption and training in artificial intelligence (AI).

• (2)

  A direct and significant relationship was found where a greater understanding of AI tools and their professional implications increases the perceived value of receiving training in their use (β = 0.912; p < 0.01).

• (3)

  Positive perception of the value of AI training significantly influences a favorable attitude toward the inclusion of this knowledge in the professional training process (β = 0.676; p < 0.01).

• (4)

  Students who recognize that AI will play an important role in their profession show a greater willingness to include AI training in their professional training (β = 1.043; p < 0.01).

• (5)

  The findings suggest that fostering a positive understanding and perception of the value of AI is essential to promote proactive attitudes towards its learning and integration in professional training, which facilitates its effective adoption in educational and work environments.

This strong model fit confirms the structural validity of the instrument, demonstrating its effectiveness in measuring students' perceptions and attitudes regarding AI adoption and training. The accurate data representation provided by the model enables reliable conclusions that can inform the development of educational and curricular strategies.

The instrument offers a robust framework for understanding how student perceptions affect their learning and adoption of emerging technologies, enabling exploration of theories on technological acceptance, adaptation, and resistance within the context of AI in education. Aligning with González-González’s [10] vision, it supports the democratization of quality education through AI-based tools.

Practically, validating this instrument benefits educators, curriculum designers, and administrators by providing insights into students' attitudes toward AI. This data can inform the development of curricula and training programs that effectively integrate AI into higher education, emphasizing its practical applications in various professional fields. Incorporating AI content and hands-on experiences enhances student understanding and confidence.

Additionally, fostering collaboration among educators, students, and technology professionals ensures that educational offerings meet student expectations and industry demands. Addressing resistance and encouraging positive perceptions of AI helps develop essential digital and AI competencies, preparing students for a dynamic technological landscape. Regularly using the validated instrument to assess student perceptions allows for the continuous improvement of educational programs, maintaining their relevance and effectiveness. Furthermore, the instrument can identify knowledge and perception gaps in integrated AI projects across different student groups.

This study conducted an initial statistical validation of the “Perception of the adoption and training in the use of artificial intelligence tools in the profession” instrument, designed to assess knowledge, training, usage, adoption, and pedagogy of AI tools in professional education across diverse Latin American populations and academic disciplines. Using generalized structural equation modeling (GSEM), the research confirmed the instrument’s validity and reliability in evaluating students' attitudes and perceptions toward effective AI tool usage, demonstrating its potential as a diagnostic tool for technological acceptance and adaptation in educational settings. However, the study’s generalizability is limited by a sample of 238 students from a single Mexican institution and the exclusion of variables X2, X9, and X12 due to insignificant results. Future research should involve larger and more diverse samples to enhance the instrument’s applicability across different cultural and academic contexts. In conclusion, the study successfully validated the instrument, providing a valuable tool for measuring students' perceptions and attitudes toward AI adoption and training in vocational education.

This study opens avenues for future research on AI perceptions and adoption in professional education. Expanding sample size and diversity across institutions and cultures is essential for consistent findings and cross-cultural insights on AI attitudes. Broader application across universities and disciplines would improve model robustness and generalizability.

Longitudinal studies are also promising, as they could track trends in AI acceptance and usage over time, revealing how ongoing exposure impacts professional readiness. Cross-disciplinary studies could identify unique needs in fields like engineering, business, health sciences, and humanities, enabling more tailored educational interventions.

Lastly, addressing ethical implications and responsible AI use is vital. Integrating these aspects into future research would contribute to a well-rounded AI education that balances technical proficiency with ethical awareness.

The authors acknowledge the technical and financial support of Writing Lab, Institute for the Future of Education, Tecnologico de Monterrey, Mexico, and Fundación Universitaria Konrad Lorenz, Colombia in the production of this work.

The authors declare that generative AI or AI-assisted technologies, specifically ChatGPT, were utilized to assist with paraphrasing, synthesis, and translation tasks during the preparation of this manuscript. However, all essential authoring tasks, including the development of original ideas, data analysis, and interpretation of results, were conducted solely by the authors without AI assistance.

The supplementary material for this article can be found online.