AIEd Knowledge BaseAIEd KB
On this page

Research Article

AI training and science student teachers’ TPACK in campus-based and distance education: a comparative study

added 2026-09-13 · updated 2026-09-19

Synthesis: Mnguni et al. (2026) ask whether the level of AI training science student teachers receive, and the institutional mode in which they receive it, explain their self-reported Technological Pedagogical Content Knowledge (TPACK). Using a quantitative, non-experimental comparative survey of 186 final-year Bachelor of Education science students at one South African distance education university (n = 97) and one campus-based university (n = 85), they report a campus advantage in self-reported TPACK (64.0% versus 47.4%) but no uniform training effect: the AI-training category was associated with self-reported TPACK only at the distance institution, where completing a five-credit short course was associated with significantly weaker reported TPACK than receiving no training at all. Across both cohorts pedagogical knowledge was the weakest component and no domain median reached the strongest-agreement Benchmark, so the authors read the pattern as evidence that readiness for AI integration in Science Education is co-constructed through institutional affordances rather than delivered by training exposure alone. Framed against SDG 4 and SDG 9, the study argues for context-responsive rather than standardized Professional Development for AI, attentive to the Digital Divide and to Equity in the Global South.

Key Findings

  1. Self-reported TPACK was higher at the campus-based university than at the distance education university. 64.0% of campus-based participants reported possessing TPACK to integrate AI into science teaching, while only 47.4% of distance education participants did so (47% reported possessing TPACK and TPK, their most commonly reported knowledge); the instrument was coded so that 1 = strongly agree and 5 = strongly disagree, meaning lower scores index stronger perceived knowledge.
  2. Formal AI training was rare in both cohorts and did not differ significantly by institutional mode. 76.4% of campus-based and 88.7% of distance education participants reported receiving no AI training at all; the Pearson chi-square test of independence found no significant association between institutional mode and level of AI training, χ²(3, N = 186) = 7.12, p = .068, although a linear-by-linear ordinal trend test was significant, χ²(1, N = 186) = 5.94, p = .015.
  3. AI training was associated with self-reported TPACK only at the distance university, and not in the expected direction. The distance education regression model was significant (F(3, 93) = 3.42, p = .021) but explained only a modest share of variance (R² = .099, adjusted R² = .070, f² = .11, a small-to-medium effect); completing a short course was associated with significantly weaker self-reported TPACK than no training (B = 1.83, p = .039), full-course completion showed a non-significant tendency toward stronger TPACK (B = −1.51, p = .088), and attending a few lectures was not significant (B = −1.01, p = .112).
  4. The campus-based model produced a null result. The level of AI training did not significantly predict self-reported TPACK at the campus-based university (F(3, 85) = 0.615, p = .607), with the training variables explaining only 2.1% of the variance (R² = .021) and a negligible effect size (f² = .02); none of the individual training variables approached significance (p = .431 for a full course, p = .916 for a short course, p = .312 for a few lectures).
  5. Pedagogical knowledge was the weakest TPACK component in both institutions. PK was reported by 53% of campus-based participants — the lowest of any domain in that cohort — and was the least reported knowledge in the distance cohort as well; the authors connect this to prior findings that teachers often overestimate technological abilities while expressing lower confidence in pedagogical capacity.
  6. No domain median reached the strongest possible endorsement. One-sample Wilcoxon signed-rank tests against a fixed test value of 1.00 found that all seven TPACK components differed significantly from the benchmark at both institutions (p < .001 throughout), indicating that participants' self-reported knowledge did not consistently reach the strongest level of agreement.
  7. Institutional differences were concentrated in technological and content-related domains, not pedagogical ones. Independent-samples Kruskal-Wallis tests found significant differences in Technological Knowledge (H(1) = 4.29, p = .038), Pedagogical Content Knowledge (H(1) = 4.48, p = .034) and Technological Content Knowledge (H(1) = 4.13, p = .042), but not in PK (H(1) = 2.52, p = .112), CK (H(1) = 1.00, p = .317), TPK (H(1) = 2.51, p = .113) or the integrated TPACK construct (H(1) = 3.59, p = .058).

Design, Sample and Instrument

The study used a quantitative, non-experimental, comparative survey design — an educational measurement approach that compared naturally occurring groups without manipulating the learning environment, institutional mode or training conditions. Participants were purposively sampled final-year Bachelor of Education students majoring in science education and preparing to teach school science in Grades 10–12 (ages 15–18), on the reasoning that students nearing entry into the profession offer the most informative index of perceived readiness. Data were collected through an online survey with a structured, closed-ended questionnaire in the second quarter of 2024: 97 participants from a nationwide distance education university and 85 from a conventional campus-based university (Table 1 and the regression models report n = 89 for the campus cohort), giving a total N = 186. Response rates diverged sharply between institutions — 5.3% at the distance education university versus 25% at the campus-based university — a difference the authors flag as a possible source of nonresponse bias. The cohorts also differed demographically: the campus-based group was 75.3% female and 80.9% aged 18–24, whereas the distance group was 57.7% female with broader age representation (36.1% aged 25–34 and 12.4% aged 35–44) and less teaching experience (48.5% under one year, versus 36.0% at the campus-based university).

AI training was measured as a four-level categorical variable: no training received; attending a few non-credit-bearing lectures of under 20 notional hours; completing a short course of five credits (up to 50 notional hours over a few weeks); and completing a full course of 10–12 credits (100–120 notional hours over several weeks, covering AI concepts, applications, pedagogical strategies and hands-on experience). TPACK was operationalized through seven dimensions — TK, PK, CK, PCK, TCK, TPK and the integrated TPACK construct — using items adopted from a previous AI-related TPACK questionnaire, reviewed by educational technology and Professional Development experts for face and content validity and piloted with a comparable group of student teachers. The instrument used a five-point Likert scale (1 = strongly agree to 5 = strongly disagree); noting that lower scores therefore represent stronger self-reported knowledge is essential to reading the results, because a positive regression coefficient means a weaker reported score. The PK subscale comprised only two items, and although it showed acceptable internal consistency (α = .906), the authors treat its limited length as a measurement caveat rather than a strength.

Factor-analytic results supported the instrument's structure. The data were suitable for factor analysis (KMO = .898; Bartlett's test of sphericity χ²(351) = 1730.39, p < .001); seven components with eigenvalues greater than 1 were extracted, explaining 64.79% of the total variance, with the first component accounting for 38.64% and the remaining six each explaining between 3.53% and 5.21%. The authors read this as a dominant general dimension of AI-related TPACK sitting alongside the multidimensional structure of the Technological Pedagogical Content Knowledge (TPACK) framework. Items with communalities below .30 were removed, and the retained solutions showed strong indicators for TPACK1 (.778 communality, .802 loading) and TK1 (.769, .795). Subscale reliabilities ranged from α = .881 (Technological Knowledge, 3 items) to α = .927 (Pedagogical Content Knowledge, 3 items), with α = .906 for the two-item PK scale and α = .923 for the five TPACK items — though the authors caution that high alphas may partly reflect item similarity.

AI Training as a Weak, Context-Dependent Predictor

The study's central claim is a negative one about the portability of training effects. The authors' initial assumption — that AI training would carry similar explanatory value across institutional contexts — was not supported: the relationship between training level and self-reported TPACK appeared at the distance institution but not at the campus-based one, and even there the only significant coefficient ran in the unexpected direction. Because the coding places "strongly agree" at 1, the positive coefficient for short-course completion (B = 1.83, p = .039) means those participants rated their own AI-related professional knowledge as weaker than peers who had received no training at all, while the negative coefficient for full-course completion (B = −1.51, p = .088) points toward stronger self-reported TPACK without reaching the .05 threshold.

The authors offer a tentative mechanism consistent with this asymmetry: brief, introductory exposure to AI may heighten student teachers' awareness of the complexity of AI integration — its ethical, pedagogical and practical demands — prompting more cautious self-assessment, whereas more sustained training may provide the deeper grounding needed to support stronger perceived readiness. On this reading, additional training does not translate straightforwardly into stronger perceived competence, and the depth and design of the training matter more than its presence. They are careful to add that the null result at the campus-based university and the modest, uneven result at the distance institution should not be read as evidence that one institutional mode is inherently more effective than the other; rather, the relationship is shaped by the type, depth and pedagogical design of training and by the institutional conditions that mediate how training is experienced. The training analyses are explicitly labeled exploratory, given that most participants reported no formal AI training, variation in training exposure was limited, and the short-course and full-course cells contained few participants — a combination that inflates both Type I and Type II error risk.

This finding has direct consequences for how curriculum designers and AI policy in teacher education treat professional development as an input. If exposure to AI training does not reliably raise perceived readiness, then widening access to training as such is not the lever; quality, contextual fit and pedagogical depth are. The study also measured readiness rather than enactment — it captured perceived readiness and not observed AI integration in real classrooms — so the results speak to confidence and perceived capability rather than to demonstrated practice.

Pedagogical Knowledge as the Persistent Weak Point

The most consistent signal in the data is not institutional but internal to the TPACK construct: pedagogical knowledge was the least-reported domain in both cohorts, reported by only 53% of campus-based participants, and every component's median fell significantly short of the strongest possible agreement in both groups. The authors tie this to evidence that teachers frequently overestimate their technological abilities while expressing lower confidence in their pedagogical capacities, and warn that when training foregrounds technical dimensions without systematically integrating pedagogical and content components, it risks producing fragmented professional identities. The practical implication they draw is that professional development should treat Pedagogies and Teaching Strategies as the anchor for technologically enhanced science teaching rather than as a secondary concern — a claim that echoes the wider argument that technology integration does not follow from tool knowledge alone.

The institutional comparison is interpreted in the same register. The divergence in TK, PCK and TCK alongside convergence in PK suggests that pedagogical instruction is more standardized across universities — plausibly through national curriculum frameworks, professional standards or shared accreditation requirements — whereas technological and content knowledge vary with access to digital tools, curriculum emphases and exposure to discipline-specific technologies. The authors therefore argue that generic training models are unlikely to address context-specific deficits or to capitalize on local strengths, and that responsive programs must engage localized needs while remaining aligned with broader pedagogical goals. For science teaching specifically, they position the integrated construct as the capacity to connect AI tools to science content, learner needs, assessment practices, ethical considerations and classroom realities — a demanding coordination problem in a subject that already requires teachers to hold abstract disciplinary content, inquiry- and modeling-based pedagogy and technological tools together, and one that makes Scaffolding and content-specific evaluation skills part of the same competence rather than separate add-ons.

Context-Responsive Training, Equity and the SDGs

The study is framed by SDG 4, on inclusive and equitable quality education and lifelong learning, and SDG 9, on resilient infrastructure, innovation and sustainable industrial development — a pairing that positions AI-supported Science Education simultaneously as a pedagogical project and an infrastructure-and-capacity project. The authors situate South Africa as a case where teacher education operates within a historically unequal schooling and higher education system and where both large-scale distance provision and conventional campus-based programs are major routes into the profession. Distance education offers flexibility, Accessibility and reduced geographical barriers, but its effectiveness depends on access to resources, digital proficiency and student interaction, and limited direct interaction may hinder the acquisition of hands-on teaching skills; reduced social presence in online environments may in turn affect communication and participation. Campus-based provision supports mentorship, immediate Feedback, concept clarification and practical experience, and is claimed to foster collaboration and academic community. Their conclusion is that sustainable AI integration depends less on widening access to training than on its quality, contextual fit and pedagogical depth, and that without parallel attention to infrastructure, ongoing support and equitable access, AI initiatives risk reproducing rather than narrowing existing disparities.

The limitations constrain how far this argument can be pushed. The sample was small and purposively drawn from two universities, so the findings should not be generalized; the unequal response rates (5.3% versus 25%) risk nonresponse bias, since participants may have differed systematically in their interest in AI or confidence with educational technologies. TPACK was self-reported rather than observed, capturing perceived readiness rather than classroom competence, and participants' responses may have been shaped by confidence, prior exposure to technology or awareness of AI-related challenges. The two-item PK subscale may not capture the full breadth of pedagogical knowledge, and the factor-analytic and reliability results are presented as preliminary evidence of structure rather than definitive validation. The cross-sectional design permits no causal conclusions and no account of change over time; the study measured broad training levels without analyzing the content, quality, duration or pedagogical design of each experience; and the small training-category cells, combined with the 76.4% and 88.7% no-training majorities, restricted variation and statistical power, making the regression coefficients indicative rather than confirmatory. The authors call for larger and more diverse samples, observational or performance-based measures of AI integration, longitudinal or mixed-methods designs, and confidence intervals around regression coefficients in future work. Ethical clearance came from the Ethics Committee of the College of Education, University of South Africa (Ref 2023_RPC_040), and the work was funded by the South African National Research Foundation (grant CPRR23032387189).

What this means for practice

  • Instructors. Do not count contact hours as evidence of readiness: at the distance university, completing a five-credit short AI course was associated with significantly weaker self-reported Technological Pedagogical Content Knowledge (TPACK) than receiving no training at all (B = 1.83, p = .039), so pair any introductory course with sustained follow-up practice.
  • Designers. Anchor AI training in pedagogical knowledge rather than tool demonstrations — PK was the weakest reported domain in both cohorts and was reported by only 53% of campus-based participants.
  • Instructors. Audit your own program's technology and content gaps before adopting a generic training package: Technological Knowledge, Pedagogical Content Knowledge and Technological Content Knowledge differed significantly between the two institutions (H(1) = 4.29, 4.48 and 4.13), while PK did not.
  • Administrators. Fund depth over reach when allocating professional development budgets: full-course completion trended toward stronger self-reported TPACK (B = −1.51, p = .088) while brief, introductory exposure moved self-assessment in the opposite direction.
  • Researchers. Replace perceived-readiness questionnaires with observation or performance measures when the question is whether training changed teaching, since this study measured confidence only and not classroom AI use.

Limitations

  • Two universities and 186 final-year Bachelor of Education science students were purposively sampled (n = 97 distance education, n = 85 campus-based, with the regression models reporting n = 89 for the campus cohort), and response rates diverged sharply — 5.3% at the distance university against 25% at the campus university — a gap the authors flag as a nonresponse-bias risk.
  • TPACK was measured entirely through a self-report questionnaire with no observation of practice, so the outcome indexes perceived readiness rather than actual AI integration in science classrooms.
  • The training cells were badly unbalanced: 76.4% of campus-based and 88.7% of distance education participants reported no AI training at all, leaving few participants in the short-course and full-course categories, which the authors say makes the regression coefficients (R² = .099 at the distance university; R² = .021 at the campus-based university) exploratory rather than confirmatory.
  • Pedagogical knowledge rested on only two items, and the cross-sectional, non-experimental design permits no causal conclusion about training effects and no account of change over time.

Connected Concepts

  • Technological Pedagogical Content Knowledge (TPACK) — the Technological Pedagogical Content Knowledge framework that operationalizes the study's outcome construct
  • Professional Development — the preservice preparation context whose two delivery modes are compared
  • Science Education — the disciplinary setting for AI integration, Grades 10–12 school science
  • Online Teaching and Learning — distance education as a delivery mode with distinct interaction and infrastructure conditions
  • Teacher AI Competency — the AI-specific professional knowledge indexed by self-reported TPACK
  • Self-Efficacy — perceived readiness and confidence as the proximal predictor of technology adoption
  • Pedagogies and Teaching Strategies — the weakest self-reported domain and the authors' proposed anchor for AI training
  • Digital Divide — uneven device, connectivity and infrastructure access shaping distance provision
  • Global South — the equity and infrastructure framing of teacher preparation beyond the Global North
  • Equity — the risk that AI initiatives reproduce rather than narrow disparities
  • AI Literacy — the training content whose level and depth the study tests against readiness
  • Assessment — a domain in which AI-supported science teaching is measured by the instrument's items
  • Curriculum Design — where context-responsive training programs must be specified
  • Educational AI Policy — the SDG-aligned policy frame for AI in teacher education

Connected Articles

Citation

Mnguni, L., El Islami, R. A. Z., Nuangchalerm, P., Sethole, K., Sari, I. J., Camara, J. S., & Van Bien, N. (2026). AI training and science student teachers’ TPACK in campus-based and distance education: a comparative study. Computers and Education Open, 11, 100410.

Metadata

Study designSurvey
DisciplineScience Education
LevelHigher Ed, Teacher Education
Intended audienceFaculty Developers, Instructors, Researchers, Policymakers, Curriculum Designers, Administrators
Learning and instructionOnline Teaching and Learning, Self-Efficacy
Assessment and measurementEducational Measurement
Research methods and evaluationQuantitative Research
Equity, ethics, and responsible useDigital Divide, Equity, Global South
Foundations of AI in educationTechnological Pedagogical Content Knowledge (TPACK), Teacher AI Competency
Page kindFramework, Evaluation
SourceHistoryEditChat

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.