Dataset & Analysis Code: AI-Hallucination Awareness Study (Indonesian PGSD, N=181)

PURPOSE AND BACKGROUND

Generative AI tools have penetrated teacher education globally,
yet their tendency to produce fluent but factually incorrect
outputs — a phenomenon widely termed AI hallucination — poses
an underexamined threat to the epistemic integrity of primary
classroom instruction. Pre-service primary teachers who
uncritically relay AI-generated content to 6-to-12-year-old
learners create an epistemic harm chain whose compounding cost
across a child's formative years implicates teacher professional
ethics rather than AI engineering alone.

Existing empirical research on pre-service teachers and
generative AI has prioritized technology acceptance, self-
efficacy, and general AI literacy. Almost none has modeled
the causal pathway that converts hallucination awareness into
ethical instructional action. The field has assumed — without
testing — that awareness produces practice change. This study
opens that mechanism empirically for the first time in a
Global South pre-service teacher population.

THEORETICAL FRAMEWORK

The study draws on three bodies of theory:

(1) Ecological model of teacher agency (Priestley, Biesta, &
Robinson, 2015; Biesta & Tedder, 2007): agency is achieved
in context, not possessed as a trait. The generative-AI
classroom constitutes exactly such a context-in-formation.

(2) Teacher professional ethics (Campbell, 2008; Rest et al.,
1999; Noddings, 2013): ethical pedagogical agency (EPA)
is operationalized as a higher-order construct comprising
four dimensions — moral responsibility, professional
courage, reflective practice, and autonomous decision-
making — each theoretically grounded in distinct strands
of moral philosophy and teacher education research.

(3) AI governance and institutional theory (Dignum, 2019;
Jobin et al., 2019; UNESCO, 2021): institutional AI
governance (IAG) is theorized as a direct determinant of
ethical agency, providing the external scaffolding that
supports ethical action independently of individual
awareness levels.

RESEARCH MODEL

The study tests a formative higher-order PLS-SEM model in
which:

— AI-Hallucination Awareness (AHA, 14 items, 3 dimensions)
serves as the exogenous predictor.

— Ethical Pedagogical Agency (EPA, 18 items, 4 dimensions)
serves as the higher-order mediator.

— Institutional AI Governance (IAG, 12 items, 3 dimensions)
is tested both as a first-stage moderator (H5) and as a
direct determinant of EPA.

— Instructional Integrity (II, 14 items, 3 dimensions)
serves as the endogenous outcome.

All four constructs are operationalized as formative higher-
order composites (Diamantopoulos & Winklhofer, 2001; Jarvis
et al., 2003), a novel specification for AI-ethics research
in education that treats these constructs as multidimensional
composite capacities rather than unidimensional reflective
traits.

Five directional hypotheses are tested:

H1. AHA directly predicts II.
H2. AHA predicts EPA.
H3. EPA predicts II.
H4. EPA mediates the AHA → II relationship (complementary
partial or full mediation).
H5. IAG moderates the AHA → EPA pathway, producing a
conditional indirect effect stronger at higher
governance levels (first-stage moderated mediation).

SAMPLE AND CONTEXT

A cross-sectional survey was administered to N = 181 pre-
service primary teachers enrolled in the PGSD (Pendidikan
Guru Sekolah Dasar) program at Universitas Negeri Malang,
one of Indonesia's leading LPTK (Lembaga Pendidikan Tenaga
Kependidikan) institutions. The single-institution design
is treated as a representative critical case providing
upper-bound estimates for the PGSD population nationally.
Indonesia trains more than 100,000 new PGSD graduates per
year; each will mediate AI-generated content for 28–30
primary learners across a multi-decade career.

METHODOLOGICAL APPROACH

Data were analyzed using partial least squares structural
equation modeling (PLS-SEM) via the seminr R package.
The disjoint two-stage approach (Sarstedt et al., 2019)
was employed for higher-order construct estimation. The
structural model was assessed via 10,000 bootstrap
resamples (bias-corrected accelerated 95% CIs). Mediation
was analyzed following Nitzl et al. (2016); moderated
mediation following Memon et al. (2018). Out-of-sample
predictive relevance was evaluated via PLSpredict (Shmueli
et al., 2019). Common-method bias was screened via
Harman's single-factor test and Kock's (2015) full-
collinearity VIF procedure.

EXPECTED AND OBSERVED OUTCOMES

Based on theory and prior literature, the following outcomes
were hypothesized and observed:

H1 (AHA → II direct): Supported (β = 0.173, p = .026)
H2 (AHA → EPA): Supported (β = 0.558, p < .001)
H3 (EPA → II): Supported (β = 0.624, p < .001)
H4 (Mediation): Supported — complementary partial
mediation; indirect effect = 0.348,
VAF = 66.9%; 95% CI excludes zero
H5 (Moderation): Not supported (β = −0.073,
p = .154; CI crosses zero)

A theoretically important unexpected finding emerged: although
IAG did not moderate the AHA → EPA pathway as hypothesized,
IAG functions as a strong and significant DIRECT determinant
of EPA (β = 0.359, p < .001). This reveals a dual-pathway
structure in which awareness-driven and governance-driven
routes to ethical agency operate in parallel and independently,
rather than conditionally. The structural model explains 66.2%
of variance in EPA and 58.1% of variance in II.

CONTRIBUTIONS

(1) Theoretical: First empirical operationalization of ethical
pedagogical agency as a formative higher-order construct
for the generative-AI era, extending Biesta-Priestley
agency theory into a domain it has not previously
addressed.

(2) Methodological: First demonstration that AI-ethics
constructs in teacher education are best specified as
formative rather than reflective, with all indicator VIFs
below 3 and weights statistically significant.

(3) Empirical: First Global South documentation of the
awareness-agency-integrity pathway among pre-service
primary teachers, revealing a dual-pathway structure
where individual awareness and institutional governance
build ethical agency through independent mechanisms.

(4) Policy: Direct empirical input for Kemendikbud-Ristek's
current drafting of generative-AI guidance for Indonesian
higher education, reframing hallucination from an
engineering problem to a teacher-ethics problem with a
clear curricular intervention point.

OPEN SCIENCE

The anonymized dataset (N = 181), complete 70-item instrument
(English and Bahasa Indonesia versions), analysis code
(R/seminr), and codebook are openly deposited at OSF under
a CC-BY 4.0 license. This registration serves as a
retrospective record of the pre-specified hypotheses,
analysis pipeline, and decision rules, submitted to promote
transparency and reproducibility.