📊 statistics

Generative AI Adoption in Statistics and Data Science Higher Education

A study of 72 Egyptian senior undergraduates using an adapted UTAUT2 model reveals that while ease of use and support strongly drive Generative AI adoption in statistics and data science education, responsible disciplinary application requires additional critical skills like verification and statistical judgment beyond mere behavioral intention.

Original authors: Noha Youssef

Published 2026-10-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Noha Youssef

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern classroom, a new kind of helper has arrived, one that does not just retrieve facts but writes code, solves equations, and drafts explanations in seconds. This technology, known as generative artificial intelligence, has moved quickly from novelty to necessity, particularly in fields like statistics and data science where students must constantly translate abstract concepts into working programs and clear arguments. For educators, the central question has shifted from whether students will use these tools to how they will use them. The challenge lies in the difference between simply adopting a tool and using it responsibly. A student might easily ask an AI to write a computer program, but that same ease can hide a subtle trap: the code might run perfectly while the logic behind it is flawed, or the explanation might sound confident while the math is wrong. In disciplines where the validity of a conclusion depends on rigorous checking and sound judgment, the speed of the tool can sometimes outpace the student's ability to verify the work.

To understand how this dynamic plays out in real learning environments, a researcher at the American University in Cairo surveyed seventy-two senior students enrolled in advanced courses on time series and categorical data analysis. These were not beginners; they were students already comfortable with programming languages and complex statistical models, working in a setting where artificial intelligence could be used to debug code, interpret results, or clarify difficult concepts. The researcher asked these students to describe their habits and feelings regarding the technology, using a framework that breaks down the reasons people adopt new tools into a few core ideas: how easy the tool is to use, how much support the school provides, and how confident the student feels in their own ability to handle the technology. The goal was to see which of these factors actually drove students to use the AI, and whether that intention translated into actual usage.

The study found that the single strongest reason students intended to use generative AI was simply how much effort it saved them. When the technology made it easier to move from a confusing concept to a working solution, students were far more likely to incorporate it into their daily workflow. This factor, which the researcher calls effort expectancy, was a much more powerful driver than the students' general confidence in their own skills or the specific support they received from instructors. While a supportive environment and personal confidence did play a role, the ease of integration was the primary gatekeeper. Once a student formed the intention to use the tool, that intention strongly predicted whether they actually used it, creating a clear path from feeling that the tool is convenient to making it a regular part of their study routine.

However, the research also revealed a significant gap between using the tool and using it well. The students who participated in the survey expressed a clear appreciation for the immediacy of the AI, noting how it helped them simplify complex ideas and write code faster. Yet, in their own words, they also voiced deep concerns about the reliability of the answers they received. They described instances where the AI provided incorrect calculations, generated code that looked right but failed to work, or offered explanations that contradicted what they had learned in class. Many students worried about becoming too dependent on the technology, fearing that the convenience might erode their own critical thinking skills. They emphasized that while the tool was useful for getting started, it could not replace the need for them to verify every step, check the assumptions behind a model, and ensure the final output was statistically sound.

The study concludes that while making a technology easy to use is the most effective way to get students to adopt it, ease of use does not guarantee responsible use. In the specific context of advanced statistics and data science, the factors that encourage a student to pick up a tool are not the same factors that ensure they will use it correctly. The research suggests that for education to keep pace with this technology, instructors must look beyond simple adoption metrics. The real educational challenge is not just getting students to use the AI, but teaching them how to maintain their own judgment while using it. This means creating assignments that require students to find errors in AI-generated work, justify their own model choices, and explicitly state how they verified the results. The study indicates that without these specific safeguards, the very ease that makes the tool so popular could lead students to accept answers without the necessary scrutiny that defines their field.

Technical Summary: Generative AI Adoption in Statistics and Data Science Higher Education

Problem Statement
While Generative AI (GenAI) is increasingly integrated into quantitative STEM education for tasks ranging from conceptual explanation to coding and model interpretation, the specific conditions driving its adoption in advanced statistics and data science courses remain underexamined. Unlike general academic settings, these disciplines require tight coupling of conceptual reasoning, programming, model diagnostics, and evidence evaluation. A critical gap exists in understanding whether the factors that drive adoption (intention and use) also ensure responsible disciplinary use (verification, statistical judgment, and transparent disclosure). Current literature often stops at general perceptions or integrity concerns, lacking discipline-specific evidence on how students integrate AI into complex analytical workflows where procedural ease might mask incorrect assumptions or interpretations.

Methodology
The study employed a mixed-methods approach grounded in an adapted Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) framework, analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM).

Sample: Data were collected from 72 senior undergraduates (approximately 70% of the eligible population) enrolled in two advanced quantitative courses (Time Series Analysis and Categorical Data Analysis) at a private university in Egypt. Participants had substantial background in statistics, mathematics, and programming (R/Python).
Instrument: A survey adapted from validated UTAUT2 measures was used, featuring a 5-point Likert scale for closed-ended items and four open-ended prompts. The conceptual model initially included Performance Expectancy (split into learning-oriented and technical/task-oriented dimensions), Effort Expectancy (EE), Facilitating Conditions (FC), Hedonic Motivation (HM), and Self-Efficacy (SE) as predictors of Behavioral Intention (BI), which in turn predicted Use Behavior (UB).
Analysis: The analysis utilized the seminr package in R. Procedures included:
- Measurement Model Assessment: Evaluation of indicator reliability (outer loadings), internal consistency (Cronbach's $\alpha$ , $\rho_A$ , composite reliability), convergent validity (AVE), and discriminant validity (Fornell-Larcker criterion, HTMT ratios with bootstrapped confidence intervals).
- Structural Model Estimation: Path coefficients ( $\beta$ ) were estimated using 3,000 bootstrap resamples with 95% percentile confidence intervals. Effect sizes ( $f^2$ ) and $R^2$ values were calculated.
- Predictive Assessment: Out-of-sample predictive relevance was tested using PLSpredict (comparing RMSE/MAE against a linear model benchmark) and $Q^2_{predict}$ .
- Robustness Checks: Sensitivity analyses were conducted by removing specific indicators (e.g., FC2, UB3, BI3) to assess trade-offs between Average Variance Extracted (AVE) and HTMT overlap.
- Qualitative Component: Open-ended responses were analyzed descriptively to contextualize quantitative findings regarding benefits, risks, and policy expectations.

Key Results
The study underwent significant model refinement based on discriminant validity and theoretical distinctness:

Model Refinement:
- Hedonic Motivation (HM) was dropped as it showed no meaningful contribution to Behavioral Intention (BI).
- Performance Expectancy (PE) was excluded from the final model. Although a learning-oriented PE path showed nominal significance in the expanded model, discriminant validity diagnostics revealed substantial overlap with Effort Expectancy, BI, and Use Behavior. The author concluded that the incremental contribution of PE could not be cleanly separated from ease of use and intention in this sample.
- Final Model: The retained model posited EE, FC, and SE as antecedents of BI, with BI predicting UB.
Structural Path Estimates:
- Effort Expectancy (EE) was the strongest predictor of Behavioral Intention ( $\beta = 0.562, p < .001$ ), with a large effect size ( $f^2 = 0.905$ ).
- Facilitating Conditions (FC) was the second strongest predictor ( $\beta = 0.329, p = .003$ ), with a medium effect size ( $f^2 = 0.330$ ).
- Self-Efficacy (SE) had a smaller but significant positive effect ( $\beta = 0.195, p = .006$ ), with a small effect size ( $f^2 = 0.112$ ).
- Behavioral Intention (BI) strongly predicted Use Behavior (UB) ( $\beta = 0.709, p < .001$ ), with a large effect size ( $f^2 = 1.012$ ).
Explanatory and Predictive Power:
- The model explained 73.2% of the variance in Behavioral Intention ( $R^2$ ) and 50.3% of the variance in Use Behavior.
- Predictive Relevance: Out-of-sample predictive evidence was weak. $Q^2_{predict}$ values were low (0.086 for BI, 0.033 for UB), and RMSE improvements over the linear benchmark were modest. Consequently, the author interprets the model primarily as explanatory for this specific sample rather than a robust forecasting tool for other populations.
Measurement Validity Concerns:
- Discriminant validity issues were noted. Bootstrapped HTMT confidence intervals for EE-BI [0.793, 1.053] and BI-UB [0.812, 1.170] included 1, indicating that these constructs cannot be treated as empirically distinct in this sample. The author retained the paths but interpreted them with caution, noting the theoretical adjacency of these constructs.
- Facilitating Conditions (FC) had an AVE of 0.479 (below the 0.50 guideline), and Cronbach's $\alpha$ for FC and UB were below 0.70, though composite reliabilities were acceptable.
Qualitative Insights:
- Students valued AI for immediacy, coding support, and simplifying concepts.
- Significant concerns included inaccurate outputs, overreliance reducing critical thinking, and ambiguity regarding disclosure policies.
- Students distinguished between using AI for assignments (with verification) versus closed-book exams, emphasizing the need for clear verification routines.

Significance and Claims
The paper makes three primary contributions, modestly framed as discipline-specific evidence rather than universal laws:

Discipline-Specific Adoption Evidence: It provides empirical data on GenAI adoption specifically within advanced quantitative STEM courses (statistics/data science) where AI intersects with coding, diagnostics, and interpretation, moving beyond general academic perceptions.
Rigorous Methodological Evaluation: The study evaluates adoption relationships using effect sizes, predictive assessment, and sensitivity analyses rather than relying solely on path significance. It explicitly addresses the limitations of the sample size ( $n=72$ ) and measurement overlap, prioritizing theoretical interpretability over statistical significance alone.
Distinction Between Adoption and Responsible Use: The central theoretical claim is that adoption (the entry of AI into a workflow driven by ease, support, and confidence) is distinct from responsible disciplinary use (the verification of outputs, exercise of statistical judgment, and transparent disclosure). The study finds that while ease of use and support drive uptake, they do not guarantee that students will verify AI outputs or maintain disciplinary reasoning.

Limitations
The author explicitly acknowledges several constraints:

Generalizability: The study is a focused case from a single institution in Egypt; results should not be generalized to all STEM students without further replication.
Sample Size: While covering 70% of the eligible population, $n=72$ is modest for structural modeling, limiting the detection of small effects and subgroup heterogeneity.
Cross-Sectional Design: Data were self-reported at a single time point; the BI-to-UB path represents association, not causation or temporal precedence.
Measurement Issues: The inclusion of HTMT intervals containing 1 and low AVE for FC suggests that the constructs are not perfectly distinct in this context.
Predictive Weakness: The model's weak out-of-sample predictive power reinforces its role as an explanatory framework for this specific context rather than a predictive model for other populations.

In conclusion, the paper argues that for advanced statistics and data science education, the instructional challenge is not merely facilitating AI adoption, but fostering the specific practices of verification, judgment, and disclosure that constitute responsible disciplinary use.

Technical Summary: Generative AI Adoption in Statistics and Data Science Higher Education

More like this