Research Article
Teacher competence and institutional governance for learner data privacy in AI mediated education
Synthesis: Kaziya and Lungu (2026) develop the Learner Data Protection Competence and Governance Framework to explain a familiar split in learner data privacy: a teacher can recognize that a disclosure is risky yet lack the authority, advice or time to avoid it, conversely, an institution can supply restricted accounts and contracts while a teacher still discloses unnecessary detail or accepts an unsupported inference. This is a conceptual framework, a purposive theory synthesis connecting contextual integrity with distinctions between demonstrated competence, confidence, perceived control and actual provision. It reports no data of any kind. Four components — risk recognition, defensible decisions, institutional enablement and accountable review — are analyzed at the decision-episode level. Two central predictions follow: support should complement judgment in discretionary tasks but substitute for particular skills in constrained ones, and reporting, correction and recurrence are separate events. Three supporting propositions, a five-domain rubric and a bounded application to teacher education and school placements in Zambia complete the proposal.
Key Findings
- Two contrasting failures motivate the framework. A teacher recognises that a learner's narrative should not go to an external service, but a supervisor demands AI feedback by day's end and no approved alternative or adviser exists. Or the institution provides restricted accounts, contracts and responsive advice, yet a teacher misses that a distinctive narrative reveals a disability and uploads it. The same disclosure can arise from constrained implementation or deficient judgment, and the remedies differ.
- The decision episode, not the teacher, is the unit of analysis. Each proposal to collect, transmit, infer, record, disclose or correct learner information is one episode with its own measures. Competence is estimated from repeated standardized tasks; enablement has an institutional face and an episode-specific one, such as whether advice arrived before the deadline.
- Complementarity is predicted for discretionary tasks (P3a). Where a task requires interpreting purpose, spotting contextual identifiers or evaluating a novel inference, timely authoritative advice and usable alternatives should strengthen the relationship between independently demonstrated competence and protective implementation. Support makes a sound judgment feasible; it does not make it.
- Substitution is predicted for narrowly constrained actions (P3b). Where a reliable safeguard blocks a defined disclosure without requiring the teacher to identify it, the benefit should be larger among teachers with lower baseline competence for that action, flattening the competence–implementation gradient even as average protection improves.
- Accountable review is a three-stage pathway that can break at any link (P5a/P5b). Accessible reporting should raise detection and submission of concerns; an implemented corrective response should reduce recurrence of that specific problem per comparable decision opportunity. Neither is assumed from the other.
- Three supporting propositions specify recognition, calibration and authorisation. P1 predicts that contextual recognition improves independently scored decision quality on disjoint tasks. P2 predicts that calibration, the correspondence between confidence and observed accuracy, adds value beyond mean confidence when predicting advice seeking. P4 predicts that mapping authorisation to purpose, data category and processing arrangement improves discrimination between permitted and unpermitted uses, including challenge to a questionable approval.
- Justification is assessed separately from implementation and legality. A defensible decision has a specified educational purpose, uses necessary information, meets applicable law, weighs the learner's interests and feasible alternatives, and assigns responsibility for checking and correcting output. Five rubric domains are scored 0–2, but a serious breach of a pre-specified constraint cannot be offset by a high aggregate score.
Why competence and governance are not substitutes
The problem is divided responsibility: the professional who initiates a transfer carries immediate responsibility for its educational justification without authority over the infrastructure. Treating privacy as individual competence lets institutions hold teachers accountable for risks they cannot resolve; treating it as compliance lets the situated decisions through which information enters a service escape scrutiny. The evidence supports the distinction, not the proposed interaction. Marín et al. (2023) found preservice teachers in Germany, New Zealand, Spain and the United States enthusiastic about educational social media while barely engaging with privacy policies. Michos et al. (2023), surveying 1,059 upper-secondary teachers in a German-speaking Swiss canton, linked self-reported data literacy and technology access to reported data use — perceived capability, not tested performance. Prinsloo and Kaliisa (2025) interviewed 13 African higher-education experts and found privacy meanings varying with circumstance, which the authors read as a case for local inquiry, not a continental definition. None of these designs shows that training or governance produces protection.
Four usually blurred constructs are separated. Demonstrated competence is performance on independently scored tasks; confidence is a belief about capability; perceived control is a belief that an action can be carried out; enablement is available authority, advice, time and alternatives. A confident teacher may be incompetent, and a competent one may correctly perceive that the necessary action is unavailable. From contextual integrity the framework takes the question of whether a flow suits its context, given actors, attributes and transmission principles, and adds what that account leaves open: whether an appropriate flow is feasible to implement, and whether approval can be challenged.
Complementarity, substitution and the correction loop
The distinctive prediction is an interaction, not a main effect. A training-only account expects better decisions from more competence; an additive one expects independent benefits from each. Kaziya and Lungu expect the return on support to depend on task discretion and safeguard coverage. Both signs must be fixed before data are seen: positive in discretionary episodes, flat or negative in reliably constrained ones, since a safeguard protects most those who would otherwise err.
The second prediction concerns time. Reporting access is measured by required steps, channels, linguistic accessibility, acknowledgement and usability. Corrective response is measured by independently verified actions — revised permissions, an amended record, changed guidance — and excludes effectiveness. Recurrence is measured later through audits over eligible episodes. A fall in reports is not improvement if reporting became harder, and a delivered remedy followed by unchanged recurrence challenges the mechanism rather than counting as a consequential response.
Each proposition carries rivals. General cautiousness, anxiety and risk aversion could mimic calibration. Leadership, workload, resources and platform design could mimic enablement. Reading comprehension and prior privacy knowledge could mimic recognition. Over-refusal and uncritical acceptance of questionable approval are specified as adverse outcomes, so P4 predicts discrimination rather than compliance.
What counts as a defensible decision
Legal permissibility, institutional authorisation and ethical justification are separate judgments. An institution cannot authorize an unlawful act, and a lawful, approved act can still be educationally unnecessary or disproportionate. Jurisdiction-specific requirements must be established for the setting, not inferred from the framework.
Two hypothetical cases carry the reasoning. A trainee adapting reading materials holds a support record with a diagnosis, family circumstances, a name and teaching strategies, while the generative service is approved for lesson planning, not sensitive records. Uploading the whole record, writing a generic description of the requirement, seeking restricted specialist help, or postponing adaptation are competing actions with different costs to disclosure, accuracy and continuity. Convenience does not justify the full upload, and a generic description is defensible only if it does not identify the learner locally. A supervisor's approval would not remove the unexplained necessity; the teacher should challenge its scope through an escalation route. In the second case, an analytics system flags a learner for withdrawal risk from attendance and activity data, and the learner explains that intermittent connectivity caused the low activity. Verifying the information and offering a limited private conversation is defensible; treating the label as a diagnosis is not, nor is refusing to use analytics.
The Zambian application is stipulated for local investigation, not offered as a finding. Four predictions follow: divided authority between university and placement school should weaken advice whose holder cannot resolve permissions across both; latency should attenuate complementarity when a response arrives after the deadline; connectivity gaps should reduce the benefit of an alternative that cannot be reached; and accessible language should improve comprehension of authorisation boundaries.
What this means for practice
These are design proposals and assessment criteria, not tested effects.
- Teacher educators. Assess competence on separate recognition and decision tasks and score the actual payload, not the explanation; a fluent rationale earns nothing for a prohibited disclosure.
- Program administrators. Publish authorisation naming purpose, data category and processing arrangement, and record who can decide across institutional boundaries and how quickly.
- Institutional leaders. Treat advice latency and connectivity as enablement failures, not teacher deficits; a pre-authorized offline alternative is an enablement intervention.
- Researchers. Fix task discretion and safeguard coverage before examining outcomes, measure reporting, correction and recurrence separately, and pre-register the interaction signs that would disconfirm the model.
- Assessment designers. Pilot the rubric and report domain-level disagreement rather than collapsing it into one score.
Limitations
- The article is conceptual and reports no empirical validation: no sample, no intervention, no measured effect, and no demonstration that the framework reduces privacy harm.
- The source base is a purposive synthesis without a search log, reproducible screening or formal risk-of-bias assessment, so coverage cannot be reconstructed as a systematic review.
- The evidence is indirect and heterogeneous — social media, learning analytics and generative AI, several studies with adults — and self-reported knowledge or use cannot validate demonstrated competence.
- The worked cases and the Zambian application are hypothetical; divided authority, latency, connectivity and language are stipulated predictions requiring local evidence, not observations about Zambia or Africa.
- The rubric anchors, interaction signs and recurrence measures are research specifications, not validated instruments or estimated effects; cognitive interviews, measurement invariance and power planning remain to be done.
- Appropriate conduct cannot guarantee protection against provider misconduct, security failure or downstream use, so proximal implementation must be distinguished from measured learner-level harm; blanket restrictions may protect information while impeding support.
Citation
Kaziya, K., & Lungu, N. (2026). Teacher competence and institutional governance for learner data privacy in AI mediated education. Frontiers in Education, 11, Article 1999453.