Acceptance of AI-gamified adaptive learning in EFL: bifactor evidence for essential unidimensionality and the limits of subscale scoring

Ainur Zhorabekova, Marzhan Moldakassymovna Duisenova, Moldir Kurmasheva, Assel Yesnazar

Frontiers in Education, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3389/feduc.2026.1929123
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Artificial intelligence (AI) and gamification are increasingly combined in language learning, and learners' acceptance of such tools is typically measured with multidimensional questionnaires whose dimensional structure is rarely tested. This study asked whether the commonly assumed subdimensions of acceptance of AI-gamified adaptive learning are empirically separable in English-as-a-foreign-language (EFL) learners, or whether acceptance is better understood as a general evaluative orientation. Using a 13-item instrument, survey data from 401 EFL students at a Kazakhstani university were analyzed with reliability analysis, exploratory factor analysis with parallel analysis, competing confirmatory models (one-factor, four-factor, second-order, and bifactor), the heterotrait–monotrait (HTMT) ratio, bifactor dimensionality indices, ordinal (WLSMV) sensitivity analysis, and independent confirmation on a held-out subsample. Parallel analysis (exploratory half) retained a single factor. Although a strict one-factor model fitted poorly, the four-factor, second-order, and bifactor models fitted better; however, the subdimensions lacked discriminant validity (HTMT up to .94; latent correlations up to .95 under WLSMV). A bifactor analysis showed a dominant general factor (ECV = .82; ωh = .93; PUC = .81), and the bifactor model was not identifiable under WLSMV, consistent with very weak specific factors (although non-identification can also reflect model instability). Formal comparisons (χ 2 difference tests and log-likelihood-based information criteria) favored the bifactor representation, and an S-1 bifactor diagnostic implicated the effectiveness-specific factor as a major contributor to the WLSMV non-identification. A Harman test (first factor = 66.5%), together with the uniformly positive, single-occasion, self-report design, suggests that common-method variance may also contribute to this generality. The evidence supports scoring this 13-item measure as a single, possibly method-influenced general factor and warns against interpreting fine-grained subscales, with perceived gamification the only partly distinguishable facet. These findings, obtained from a predominantly female sample of prospective English teachers at one Kazakhstani university, provide structural validity evidence only; nomological and external-criterion validity were not examined and remain the necessary next step.