Psychological pathways to responsible generative AI decision-making among teachers: governance clarity, perceived accountability, decision uncertainty, and AI literacy

Synthesis: Li and Su do not ask whether teachers use generative AI. They ask how a teacher regulates optional assistance at the moment a judgment about a learner is formed, and they answer with a deliberately thin context–cognition–action model. Governance clarity — rule clarity plus role clarity — is the contextual cue; perceived accountability and decision uncertainty are the psychological states it works through; AI literacy is a person-level regulatory resource; and responsible GenAI decision-making (RADM) is the regulated conduct that follows. RADM is defined as a process, explicitly not as decision accuracy, so a correct answer reached by uncritical copying counts as a failure of the construct and a defensible process can still end in a contestable judgment. The authors state plainly that these outputs are "theoretical propositions, not empirically established effects," and that the model "has not yet been empirically validated."

This page summarizes a conceptual article, published under the journal's "Hypothesis and Theory" type. It collects no data and tests no model of its own: the five propositions and one configurational expectation below are the authors' proposals, and every figure drawn from another study is attributed to that study rather than reported here as a finding.

What the paper argues

1. Move the outcome from adoption to conduct. Teacher AI competency research and technology-acceptance models can say whether teachers are willing to use a tool; they do not say what a teacher delegates, verifies, overrides, documents, or escalates. RADM is the paper's name for that regulation, defined as "selective, critical, transparent, and accountable use of GenAI input while preserving accountable human authority."
2. Separate consultation from reliance. Eliciting or reading model output is not reliance; reliance occurs when output changes an interpretation, a communication, or a decision. A teacher can consult and reject, or adopt a draft's wording after independently establishing the judgment underneath it.
3. Five behavioral dimensions. Appropriate delegation, verification, calibrated reliance with accountable human authority, traceability, and escalation or override. The authors treat these as non-interchangeable parts of a composite rather than as a reflective scale, and ask future work to compare both specifications.
4. P1–P3, the accountability path. Greater governance clarity is proposed to raise perceived accountability (P1); higher perceived accountability is proposed to raise RADM (P2); and perceived accountability is proposed to mediate the clarity-to-RADM relationship (P3). Felt accountability is defined by anticipated justification to a legitimate, competent audience, not by fear of sanction.
5. P4, uncertainty as calibration. The relationship between subjective uncertainty and RADM is proposed to be calibration-sensitive — recognized, regulable uncertainty should yield more verification and escalation than unwarranted certainty, or than uncertainty that outstrips the teacher's control strategies.
6. P5 and CE1, literacy and equifinality. AI literacy is proposed to moderate uncertainty regulation, so that unwarranted confidence produces less uncritical reliance at higher literacy. A secondary configurational expectation holds that high and low RADM each arise from several combinations of clarity, accountability, uncertainty, and literacy.

Why "responsible" is defined as a process rather than an outcome

The definitional move is the paper's sharpest decision, and it is what separates it from the compliance literature. If responsible AI meant endorsing the right principles, a survey of attitudes would settle the question. Li and Su instead make the unit of analysis the decision episode: what the model contributed, what the teacher checked, how confident the teacher was, what uncertainty remained, and where accountable human authority sat.

That choice has consequences for measurement. Because the five dimensions are not interchangeable — a teacher can verify without documenting, or protect student data without calibrating reliance — instrument development should compare a formative composite against correlated first-order factors. Validation should use scenario choices and process traces, not self-report endorsement, and convergent validity should be shown against critical evaluation and ethical competence. Discriminant validity should be shown against general AI acceptance, trust, and Self-Efficacy.

The model also keeps the institutional level from doing explanatory work it cannot do. A school can hold an extensive AI policy and a teacher can still not know whether it covers a particular student-related task, or who is meant to check a generated recommendation. Clarity matters psychologically only when it is retrievable at the point of decision and identifies both the boundary and the locus of answerability. The paper therefore declines to propose a direct path from governance clarity to decision uncertainty: clarity can remove avoidable procedural ambiguity, but it cannot remove genuine uncertainty in the learner case.

Uncertainty is a signal, not a defect

Most responsible-AI writing treats uncertainty as a technical defect to be engineered away, and treats confidence as the mark of a working system. The paper inverts that. Metacognitive research separates confidence from accuracy, and fluent generated prose can raise felt certainty without raising evidential quality — so low subjective uncertainty is the state that should worry a school, not high uncertainty that a teacher can diagnose and act on.

The distinction the authors draw inside uncertainty is the useful part. Some uncertainty belongs to the case: incomplete evidence, contestable criteria, a learner whose history is not fully known. Some is avoidable procedural uncertainty about how output should be handled within the institution. Clear governance can reduce the second and cannot touch the first, which is why the paper refuses a single "less uncertainty is better" claim. Uncertainty that exceeds the teacher's capacity to diagnose it can produce indecision, undifferentiated distrust, or deference to the model as an external authority.

AI literacy enters here as a boundary condition rather than an antecedent of adoption. A literate teacher can tell case uncertainty from model-introduced uncertainty, distinguish plausibility from support, recognize that rephrasing an answer does not validate it, and know when specialist review is required. That is a regulatory repertoire, and it is the mechanism by which monitoring becomes control.

What the model adds, and what its evidence can carry

The authors' own contribution is integration: they bring felt-accountability research, metacognitive monitoring and confidence calibration, and teacher AI-literacy frameworks into one process model, and they specify what an empirical program would have to measure. They list four contributions — conduct instead of adoption as the outcome, a psychological link between institutional expectations and teacher behavior, the connection of uncertainty to calibration, and the configurational extension.

The evidence they cite is thinner than the framework, and they say so. Steiss et al. (2024) compared ChatGPT feedback with trained human evaluators on secondary writing and found human feedback better on four of the five assessed quality dimensions — which the authors read as support for treating generated text as a fallible draft. Kinder et al. (2025) studied 269 German pre-service teachers and found adaptive ChatGPT feedback improved the quality of written justification without significantly improving decision accuracy; that is a training-setting result and the paper does not stretch it to in-service referral or grading. Henderson et al. (2026) is cited only as background on selective reliance toward algorithmic course recommendations, explicitly not as evidence about conversational GenAI.

The constructive claim is therefore narrow by design. GenAI may help organize information the teacher is already authorized to hold, draft candidate feedback, generate alternative explanations, and propose questions. It cannot turn an unsupported interpretation into evidence about a particular student, and appropriate non-use counts as responsible delegation. Human-in-the-loop authority, on this account, is not a checkpoint bolted onto a workflow; it is the teacher's retained power to accept, modify, reject, or proceed without the output.

What this means for practice

• School leaders. Write fewer, more usable boundaries. The paper's practical argument is that teachers need a small set of consequential rules they can retrieve mid-decision: which tasks may receive assistance, what student information may not be entered, which outputs require independent checking, and where accountable authority sits.
• Anyone designing a workflow. Answer four questions for the teacher before the tool goes live — what may I delegate, what am I responsible for verifying, where does human authority reside, and whom do I consult when risk exceeds my expertise. Diffuse responsibility across teacher, leader, vendor, and system is the failure mode the model targets.
• Professional development. Train calibration, not prompting. Practice in distinguishing fluency from evidential strength, noticing missing context, protecting student data, and deciding when to escalate transfers better to consequential decisions than productivity training does.
• Researchers. Take the measurement agenda seriously before testing the paths. The paper specifies a validation study, a three-wave temporally separated design, a staged experiment comparing unaided, fixed-advice, and conversational conditions, and fsQCA only after the constructs hold up.
• Teachers. The framework legitimizes refusal. Deciding that a task is unsuitable, that sensitive context cannot be supplied, or that verification costs more than the assistance saves is a competent response, not a gap in agency.

Limitations

• The model is untested. The authors state that it "has not yet been empirically validated" and that its relationships are propositions rather than established causal effects, so every path here is a hypothesis awaiting data.
• The calibration-sensitive prediction for decision uncertainty is, by the authors' own account, the most provisional element, and it cannot be evaluated without pairing subjective confidence with objective task ambiguity or evidence quality.
• The evidence base is secondhand and does not cover the target setting. Direct evidence on GenAI in real, consequential teacher decisions is described as limited; the feedback-quality study and the pre-service experiment inform possible assistance functions without establishing effectiveness in referral, placement, or sustained student support.
• Task examples and cross-task expectations are theoretical elaborations rather than observations. Nothing in the paper reports how often teachers consult GenAI, what they do with output, or whether any benefit exceeds verification costs.
• The constructs remain to be discriminated empirically: rule and role clarity may not separate, perceived accountability overlaps with conscientiousness, AI literacy overlaps with self-efficacy and ethical competence, and the higher-order structure of RADM is provisional.
• The framework is built for in-service K–12 judgment and is offered as a psychological account rather than a governance design. Accountability audiences and the meaning of uncertainty vary across systems and cultures, and the authors treat generalization to higher education or advising as something to test rather than assume. The authors also disclose that ChatGPT and Codex were used to support drafting and revision, which they state was not treated as a source of evidence.

Citation

Li, W., & Su, Y. (2026). Psychological pathways to responsible generative AI decision-making among teachers: governance clarity, perceived accountability, decision uncertainty, and AI literacy. Frontiers in Psychology, 17, 1989652.