Acceptance of AI-Gamified Adaptive Learning in EFL: Bifactor Evidence for Essential Unidimensionality and the Limits of Subscale Scoring

Synthesis: Zhorabekova et al. (2026) tested whether the commonly assumed subdimensions of acceptance for AI-gamified Adaptive Learning in EFL are empirically separable, using a 13-item questionnaire completed by 401 students at a Kazakhstani university. Parallel analysis, competing confirmatory models, bifactor dimensionality indices, an ordinal robustness check, and an independent held-out confirmation converge on the same answer: the scale is essentially unidimensional, dominated by a single general factor, with perceived gamification the only partly distinguishable facet. The practical message is that a general total score is defensible and fine-grained subscales are not — though the general factor may be substantially method-influenced.

Key Findings

• The assumed subdimensions are not empirically separable. Parallel analysis retained a single factor, yet a strict one-factor model also failed to reproduce the covariance structure. The data are best captured by a broad general factor running through all 13 items, with the individual facets contributing very little once that factor is taken into account.

• A dominant general factor explains nearly all the common variance. The bifactor model fitted best, and its dimensionality indices (explained common variance near .82, hierarchical omega near .93) cleared conventional thresholds for essential unidimensionality. Several specific-factor loadings were very small once the general factor was partialled out, so the fine-grained facets add little reliable measurement beyond the total score.

• Discriminant validity failed for three of the four domains. The perceived-AI, effectiveness, and engagement/satisfaction subscales were so highly inter-correlated that they could not be distinguished from one another; only perceived gamification remained consistently separable. The same pattern held on the independent confirmatory half of the sample.

• The model was not identifiable under ordinal estimation, consistent with very weak specific factors. An S-1 diagnostic that omitted the effectiveness facet converged and fitted well, implicating that facet as a major contributor to the non-identification.

• The general factor may be partly method-induced. Every item was positively worded, self-reported on a single occasion, and showed a pronounced ceiling effect, and a Harman-style screen loaded much of the variance onto one factor. Learner characteristics explained little of the variance in acceptance, and the only significant predictor — frequency of use — is itself a self-report conceptually close to acceptance.

Study Design & Method

An initial pool of 13 items was generated from the gamification, AI-in-education, and technology-acceptance literatures and organized a priori into four domains — perceived gamification, perceived AI, perceived effectiveness, and engagement/satisfaction — each rated on a five-point Likert scale and presented simultaneously in English, Kazakh, and Russian via a single Google Forms instrument. Content validation was informal (colleague review only), so the authors characterize the study as an exploratory structural validation rather than full construct validation.

Using convenience sampling, 401 complete cases were analyzed — undergraduate prospective English teachers at O. Zhanibekov South Kazakhstan Pedagogical University (Shymkent, Kazakhstan), predominantly female and first-year, with self-reported English mostly around A2–B2. The sample was randomly split into exploratory and held-out confirmatory halves so that exploration could be separated from confirmation. Analyses proceeded through reliability and sampling-adequacy checks, exploratory factor analysis with parallel analysis, competing confirmatory models (one-factor, four correlated factors, second-order, and bifactor), discriminant-validity checks using the heterotrait–monotrait ratio and the Fornell–Larcker criterion, bifactor dimensionality indices, an ordinal (WLSMV) sensitivity analysis with an S-1 diagnostic, and a preliminary regression of acceptance on learner characteristics. ML models were fitted in Python and WLSMV models in R; a large language model assisted with analysis scripting, with all results verified by the authors.

What this means for practice

• Instructors. Interpret one total acceptance score from this 13-item instrument and stop reading the AI, effectiveness, and engagement/satisfaction subscales as separate dimensions; discriminant validity failed for three of the four domains.
• Instructors. Treat gamification as a distinct design lever, since perceived gamification was the only facet that stayed consistently separable from the dominant general factor.
• Designers. Anchor items to one named platform before measuring acceptance, because respondents here may have evaluated different applications and the general factor may be partly self-report method variance.
• Researchers. Test dimensionality and discriminant validity before assuming a multidimensional structure for ad hoc TAM or UTAUT adaptations; in these data the general factor explained roughly 82% of common variance and hierarchical omega reached about .93, and strong claims about engagement or perceived effectiveness as separate mechanisms should wait on multi-method evidence, since the general factor may be partly self-report method variance.

Limitations

• No criterion validity. The preliminary regression explained very little variance, and its sole significant predictor (frequency of use) is itself a self-report conceptually close to acceptance, so shared-method variance may explain even that association; no behavioral or learning outcome (usage logs, task completion, achievement, teacher ratings) and no established acceptance instrument (e.g. UTAUT) were available for convergent comparison, and the authors' named next step is to establish criterion validity against external behavioral and learning measures.
• Common-method variance is an unresolved alternative explanation. Positively worded items, a single self-report occasion, and a pronounced ceiling effect all inflate inter-construct correlations; no theoretically unrelated marker variable was collected, so marker-based tests could not be run, and the authors argue a post hoc method factor would not provide an independent test in these data — their remedy is reverse-coded items, a marker variable, and multi-occasion data.
• Construct heterogeneity and unvalidated content. Items addressed AI-gamified tools as a class rather than one named platform, so respondents may have evaluated different applications, and no formal content-validity index, cognitive interviewing, piloting, or back-translation was run, so item validity is asserted rather than demonstrated; the authors' named next step is to anchor items to a named platform.
• Single-institution convenience sample. The sample is predominantly female and first-year, precluding gender comparison and limiting generalization; no test–retest data exist, so temporal stability is unknown, and the trilingual single-form administration precludes language-anchored modeling and measurement-invariance tests across languages. Replication in larger multi-institution samples with formal tests of measurement invariance is future work.

Connected Concepts

• Language Learning
• Game-Based Learning
• Adaptive Learning
• Technology Adoption Models
• Educational Measurement

Connected Articles

• Acceptance of AI-Assisted English Language Learning Tools in Higher Education: Psychological Correlates Across Disciplinary and Proficiency Groups — Acceptance of AI-Assisted English Language Learning Tools in Higher Education
• Factors influencing university students' intention to use and reliance on generative artificial intelligence — Intention to Use and Reliance on Generative AI: An Extended TAM with Critical Use
• Will, Skill, Not Tool: Chinese university students' acceptance of generative AI for academic writing in informal English medium instruction settings — Will, Skill, Not Tool: Acceptance of Generative AI for Academic Writing
• Enhancing AI Literacy Course Satisfaction Through Empowerment in AI Problem-Solving and Ethical Awareness: Development and Validation of an AI Project-Based Learning Scale — Development and Validation of an AI Project-Based Learning Scale
• Motivational Ergonomics in Gamified and Artificial Intelligence-Supported Learning: An ARCS Study with Implications for Workplace Training — Motivational Ergonomics in Gamified and AI-Supported Learning
• MedGame: Storytelling Gamification Empowered by Large Language Models for Medical Education — MedGame: Storytelling Gamification Empowered by LLMs

Citation

Zhorabekova, A., Duisenova, M., Kurmasheva, M., & Yesnazar, A. (2026). Acceptance of AI-Gamified Adaptive Learning in EFL: Bifactor Evidence for Essential Unidimensionality and the Limits of Subscale Scoring. Frontiers in Education, 11, 1929123.