In Ecuador, there are currently no formal institutional policies or official regulations governing the academic use of generative artificial intelligence tools such as ChatGPT. Its implementation in higher education settings remains largely informal, with usage patterns monitored primarily through online activity and anecdotal evidence. Despite this regulatory gap, university students have rapidly incorporated AI tools into their academic routines, employing them for tasks such as literature searches, essay composition, test preparation, and practical assignments. Based on direct academic observation, the presence and frequency of such use are increasingly evident in classroom and virtual learning environments. Accordingly, this study aims to explore students’ actual experiences and perceptions regarding the use of ChatGPT, with a particular focus on identifying the sociodemographic and perceptual factors that influence their intention to use it for academic purposes.

The study sample consisted of undergraduate students enrolled in various Ecuadorian universities who voluntarily and anonymously completed an online survey. Data collection was conducted between October 17 and November 2, 2024, resulting in a total of 210 valid responses. The gender distribution was predominantly female (75%), with male students representing 25% of the sample. Four additional participants initiated but did not complete the survey and were therefore excluded from the final analysis. Respondents represented seven different universities and a range of academic disciplines, including medicine, nutrition, systems engineering, and education, among others. The sample composition reflects voluntary participation from students enrolled in faculties with greater female representation, particularly in health sciences. While this limits generalizability, it provides insight into populations more actively adopting AI tools. This study involved anonymous voluntary surveys from students and did not involve sensitive personal data or intervention. As such, it was exempt from full ethical board review but adhered to institutional guidelines for research ethics. Informed consent was obtained electronically prior to survey completion (Table 1).

Data were collected using a structured and previously validated survey instrument designed to assess students’ opinions and experiences related to the use of ChatGPT for academic tasks. The questionnaire was adapted from a study developed by Yu et al. (2024), which focused on measuring factors influencing the use of ChatGPT in educational contexts.

The survey instrument was adapted from Yu et al. (2024) and translated into Spanish. Linguistic and contextual adjustments were made to reflect the Ecuadorian higher education environment. Content validity was confirmed through expert review. Factor analysis assumptions were tested: Kaiser-Meyer-Olkin measure = 0.882 and Bartlett’s Test of Sphericity: χ² = 1643.2, p < 0.001. Multicollinearity was assessed using variance inflation factors (VIF < 2), indicating no significant collinearity between predictors. Internal consistency was assessed via Cronbach’s alpha: compatibility (α = 0.81), efficiency (α = 0.84), perceived ease of use (α = 0.79), perceived usefulness (α = 0.85), satisfaction (α = 0.82), and intention (α = 0.88). Factor loadings from EFA exceeded 0.60.

The key variables analyzed included six constructs: compatibility, efficiency, perceived ease of use, perceived usefulness, satisfaction, and intention of continued use. Grouping variables used for comparisons included gender, area of residence (urban/rural), and academic program.

The instrument consisted of six key constructs: compatibility (3 items), efficiency (4 items), perceived ease of use (3 items), perceived usefulness (3 items), satisfaction (3 items), and continued use intention (3 items). All items were measured on a five-point Likert scale, ranging from “strongly disagree” (1) to “strongly agree” (5). The survey included an informed consent section at the beginning, ensuring that participation was entirely voluntary and that respondents had the option to discontinue at any point. The dichotomization of the 5-point Likert scale into high vs. low intention (cut-off at >3) was adopted to align with common practice in behavioral intention modeling, where values above the midpoint reflect meaningful agreement with intent to act (Pellegrino et al., 2024). While this reduces variability, it enhances interpretability and logistic modeling efficiency (Table 2).

The statistical analysis was conducted using Python programming language within the Google Colab environment to ensure transparency, reproducibility, and computational efficiency. In addition to descriptive statistics and exploratory factor analysis, independent samples t-tests were conducted to evaluate differences in key variables (e.g., perceived ease of use, usefulness) across demographic groups such as gender and place of residence. Assumptions of normality and homogeneity of variances were assessed using the Shapiro–Wilk test and Levene’s test, respectively. When these assumptions were violated, non-parametric Mann–Whitney U tests were used. Effect sizes were calculated using Cohen’s d for t-tests and Rosenthal’s r for non-parametric comparisons. Core libraries included pandas and numpy for data preprocessing and descriptive statistics, factor_analyzer for exploratory factor analysis (EFA), scikit-learn for clustering, and statsmodels for regression modeling (Kuroki, 2021).

Descriptive analyses were first performed to summarize the participants’ sociodemographic characteristics and their responses to each item, including measures of central tendency and dispersion. Subsequently, an Exploratory Factor Analysis (EFA) was conducted using principal axis factoring with varimax rotation to evaluate the construct validity of the perception-based items. The Kaiser-Meyer-Olkin (KMO) test and Bartlett’s test of sphericity were used to assess sampling adequacy and the factorability of the data matrix (Luo et al., 2019).

To predict the academic intention to use ChatGPT, a binary logistic regression model was employed (Mahmud et al., 2024). The dependent variable was defined by dichotomizing the Likert-based intention item into two categories: low intention (≤ 3) and high intention (> 3). Independent variables included factor scores derived from the EFA (e.g., compatibility, perceived usefulness, ease of use, satisfaction) and sociodemographic variables such as gender, residence, and parental education.

In addition, a k-means clustering analysis was performed using the standardized factor scores to explore latent student profiles based on their perception patterns. The optimal number of clusters was determined through the elbow method and silhouette score. A significance level of p < 0.05 was adopted for all inferential procedures (Kanungo et al., 2002).

A total of 210 valid responses were analyzed from university students in Ecuador. The majority of respondents were female (75%) and resided in urban areas (69%). Participants represented seven universities and various academic disciplines.

Descriptive statistics indicated generally positive student perceptions of ChatGPT. The item “ChatGPT improves the quality of my results” received a mean score of 3.89 (SD = 0.92), while “I plan to continue using ChatGPT” had a mean of 3.73 (SD = 0.98). Conversely, the lowest score was observed in the item “I find it easy to make ChatGPT do what I want” (mean = 3.23), suggesting potential usability challenges. Overall, the Likert-scale responses displayed consistency and internal reliability across constructs such as perceived usefulness, compatibility, ease of use, and satisfaction.

A correlation matrix revealed strong associations between items related to perceived usefulness, satisfaction, and continued use intention. For instance, “ChatGPT improves the quality of my results” correlated strongly with “I am satisfied with ChatGPT’s performance” (r = 0.68) and “I intend to recommend ChatGPT” (r = 0.71). A moderate positive correlation was observed between ease of use and compatibility constructs.

Independent samples t-tests were conducted to compare perception scores across grouping variables. Urban students reported significantly higher perceived ease of use (t = 2.32, p = 0.021, Cohen’s d = 0.42) compared to rural students. Although female students scored slightly higher in perceived usefulness and satisfaction, these differences were not statistically significant (p > 0.05, d < 0.2). All t-tests were conducted after confirming assumptions of normality and homogeneity of variances; in cases where assumptions were violated, Mann–Whitney U tests were applied.

To identify predictors of high academic intention to use ChatGPT (defined as Likert score > 3), a binary logistic regression model was estimated. Four perception variables were included as predictors:

The intercept was −2.013. The model was statistically significant (χ² = 48.23, df = 4, p < 0.001) and explained 39.5% of the variance (Nagelkerke R²). The two strongest predictors were perceived usefulness and compatibility.

The logistic regression model predicts the probability P(Y = 1), where Y = 1 denotes a high academic use intention of ChatGPT. The predictor variables included are: X₁, representing the perception that “ChatGPT improves the quality of my results”; X₂, “ChatGPT fits my learning style”; X₃, “I find it easy to use ChatGPT”; and X₄, “I am satisfied with ChatGPT’s performance.” The model coefficients (β) are expressed in log-odds, and their exponentials correspond to odds ratios. Specifically, a one-unit increase in X₁ is associated with a 2.37-fold increase in the odds of high use intention. Similarly, the odds increase by 1.87 for X₂, 1.41 for X₃, and 1.68 for X₄, indicating that perceived usefulness and compatibility are the strongest predictors of continued academic use of ChatGPT.

Logistic regression results showed the following: “ChatGPT improves the quality of my results” (β = 0.864, OR = 2.37, CI: 1.45–3.89, p < 0.001), “ChatGPT fits my learning style” (β = 0.627, OR = 1.87, CI: 1.15–3.04, p = 0.011), “I find it easy to use ChatGPT” (β = 0.576, OR = 1.78, CI: 1.02–3.11, p = 0.041), and “I am satisfied with ChatGPT’s performance” (β = 0.499, OR = 1.65, CI: 1.04–2.61, p = 0.036).

Model fit was acceptable: Nagelkerke R² = 0.395; Hosmer-Lemeshow test = 0.423 (Table 3).

K-means clustering (k = 3) was conducted using standardized scores of all Likert items. A silhouette score of 0.52 indicated moderate internal consistency. The resulting profiles were:

Principal Component Analysis (PCA) was used to visualize the clusters, confirming clear separations based on students’ perceptions of ChatGPT. Despite the moderate silhouette score (0.52), cluster interpretability was supported by significant differences in external variables such as satisfaction and usage frequency. Stability was further tested by rerunning k-means on 80% bootstrap samples, yielding consistent grouping patterns (Figure 1).