Beyond access and algorithms: a curricular model for inquiry-driven student-AI engagement

The integration of Artificial Intelligence (AI) into education represents a pivotal moment for global educational equity. AI technologies offer unprecedented potential to democratise high-quality instruction through personalised learning and automated support for diverse learners, with proponents envisioning a partnership in which AI automates routine tasks while educators focus on mentorship and understanding student needs (Chen et al., 2020; Garzón et al., 2025; Zawacki-Richter et al., 2019; Cardona et al., 2023).

Alongside this potential, the rapid proliferation of AI also threatens to create new vectors of exclusion. Current discourse on AI and educational equity has converged on two main interventions: closing the material access gap, often described as the digital divide, and ensuring algorithmic fairness through the mitigation of coded bias (Dastyari and Jose, 2024; Boateng and Boateng, 2025; Gándara et al., 2024). Material access is shaped by structural inequalities that cannot be solved through device distribution alone (Selwyn, 2016; Williamson, 2017), and algorithmic systems can encode and amplify social bias, with harms falling disproportionately on already marginalised groups (Noble, 2018; Eubanks, 2018; Benjamin, 2019).

This paper addresses a different but related question. Within learner-side work, valuable research has examined AI literacy as a set of competencies (Long and Magerko, 2020; Ng et al., 2021; European Commission and OECD, 2025; UNESCO, 2022), evaluative judgement as a response to generative AI's destabilisation of traditional assessment (Bearman et al., 2024), and curiosity as a driver of self-directed learning (Loewenstein, 1994; Litman, 2005; Engel, 2015). These conversations have, however, developed largely in parallel. AI literacy frameworks describe what learners should be able to do but say less about the pedagogical mechanism that supports their progression toward critical engagement. Work on evaluative judgement focuses on how learners assess AI outputs but does not extend to the structure of inquiry behaviour itself. Curiosity research provides theoretical resources but has not been systematically brought to bear on student-AI engagement.

These conversations are drawn together to argue that the cultivation of inquiry-driven agency, the capacity to ask productive questions of AI rather than merely receive its answers, is a worthwhile and currently underdeveloped object of pedagogical design. It proposes the Student-AI Learning Trajectory (SALT) Model as a four-phase curricular sequence for scaffolding this development, grounded in a curiosity-driven account of how learners deepen engagement with any new domain, and inquiry-sequence assessment as an operational mechanism through which inquiry capacity can be made visible and assessable in classrooms.

The contribution is pedagogical, not a solution to the structural conditions of educational inequity. The locus of responsibility sits with the institutions and educators designing learning environments, not with students themselves, and pedagogical design cannot substitute for structural reform of material access, algorithmic governance, or institutional resourcing. What schools and universities scaffold and assess in this space nonetheless matters for the kinds of learners we send into a world increasingly shaped by automated systems.

This paper employs an integrative conceptual approach to develop a pedagogical model for student-AI engagement. Conceptual papers provide theoretical scaffolding on which later empirical inquiry can be built (Jaakkola, 2020). Among the four approaches to conceptual articles identified by Jaakkola (theory synthesis, theory adaptation, typology, and model paper), this paper most closely aligns with the model paper category that articulates relationships between constructs to explain a phenomenon. Elements of theory synthesis are also present, as the model brings together perspectives from previously disconnected literature.

Three guiding questions shape the conceptual development:

1. How do learners develop inquiry-driven agency in AI-mediated learning environments?

2. What pedagogical structures support the development of this agency across distinguishable phases of student-AI engagement?

3. How might inquiry-driven agency be assessed in ways that remain meaningful when AI can substitute for many traditional learning outputs?

The paper synthesises five bodies of literature: Scholarship on AI in education and emerging AI literacy frameworks; critical scholarship on educational technology and the structural dimensions of digital inequity; foundational and contemporary work on curiosity, dual-process cognition, and adaptive expertise; constructive-developmental theory and adjacent traditions in cognitive development; and research on question-asking in learning and authentic assessment.

Sources were identified through targeted forward and backward citation tracing from key reference points in each literature, supplemented by database searches in Google Scholar, ERIC, and Web of Science. Peer-reviewed scholarly work was prioritised. International policy documents (UNESCO, OECD, European Commission, World Bank) were retained where they represent consensus competency frameworks or supranational evidence syntheses. Sources from the past seven years were prioritised for empirical claims about AI in education, while foundational theoretical works were included regardless of date where they provide the original articulation of constructs central to the model. The paper draws purposefully on the literature most relevant to the contribution being advanced, in line with the conventions of integrative conceptual scholarship (Jaakkola, 2020).

The paper proceeds from a position that treats learners as agents capable of developing inquiry capacity through scaffolded pedagogy. This commitment is consequential. It locates the responsibility for what the paper describes with the institutions and educators designing learning environments, not with the student. It also does not seek to make the claim that pedagogical intervention can substitute for structural reform. It is also a position that some sociomaterial and posthuman scholars would contest, as they would foreground the entanglement of learners with technologies, institutions, and discursive structures more radically than the present argument requires. The model proposed here is offered as a pedagogical tool that complements rather than displaces those other lines of work.

This section examines the two pillars currently dominating AI equity discourse: material access and algorithmic fairness. Both represent essential infrastructure for equity, but a closer examination reveals that they do not, on their own, account for the learner-side capacities that determine how students engage with AI once access is granted and bias is mitigated. The analysis identifies the territory this paper enters.

The digital divide signifies the unequal access to digital technologies, a multifaceted issue that extends beyond mere device availability to include disparities in internet connectivity, infrastructure, skills, and affordability (Azubuike et al., 2020). Globally, approximately 2.6 billion people lack fast internet access, and two-thirds of school-aged children do not have internet access at home (International Telecommunication Union, 2023; World Bank Group, 2025).

Historically, educational policy has attempted to bridge these chasms through “material-first” frameworks such as 1:1 laptop initiatives. Contemporary research shows, however, that even when physical access is equalised, a second-level divide persists in how students utilise technology for high-level content creation versus passive consumption (Hohlfeld et al., 2017).

The COVID-19 pandemic served as a global stress test for material-first frameworks, with the mass introduction of devices frequently exacerbating rather than ameliorating existing inequalities. Students in privileged environments leveraged technology for interactive, teacher-led inquiry while those in under-resourced settings were often relegated to rote, automated activities (Reich, 2020; van de Werfhorst et al., 2022), with persistent learning consequences for lower-income households (Fallon et al., 2025).

The divide also encompasses differences in digital skills, uses, and motivations, with socio-economically advantaged schools outperforming disadvantaged schools in their preparedness for digital learning (OECD, 2024). Critical scholarship has long argued that these structural inequalities cannot be solved through device distribution alone (Selwyn, 2016; Williamson, 2017; Dastyari and Jose, 2024).

Studies repeatedly show that AI systems demonstrate various forms of bias and societal prejudice. Racial and cultural bias is reflected in systems trained on vast datasets from the internet, which often inherit a dominant white, male, and Western cultural perspective that can lead to outputs that reinforce stereotypes (Tao et al., 2024). A study by Common Sense Media (2025) found that when AI teacher assistants were prompted to create behaviour intervention plans, they recommended more punitive measures for students with black-coded names and more supportive approaches for those with white-coded names. Gender bias is observed in language models that associate roles like “nurse” and “secretary” with female pronouns, while linking “doctor” and “lawyer” with male pronouns (Kotek et al., 2023). Linguistic bias reflects the dominance of English in online content (Petrosyan, 2025), and similar disparities are evident across disability and accessibility contexts (Kane et al., 2023).

Technical interventions, such as preprocessing data to remove protected attributes or employing adversarial debiasing, can measurably reduce outcome disparities in decision-making systems (Mehrabi et al., 2019; Barocas et al., 2023). However, it remains difficult to fully eliminate biased datasets, flawed algorithm design, and the unconscious assumptions of developers (Favaretto et al., 2019), all of which influence the next evolution in technology. A telling example is the rise of AI detection tools in education, which have themselves been shown to be biased against non-native English speakers, with studies finding that these detectors falsely flagged over half of writing samples from non-native speakers as AI-generated while demonstrating near-perfect accuracy for native speakers (Liang et al., 2023).

This paper builds on the argument that algorithmic bias is a systemic feature of AI as currently developed (Boateng and Boateng, 2025), and that the harms produced by such systems fall disproportionately on already marginalised groups (Noble, 2018; Eubanks, 2018; Benjamin, 2019). A growing sociotechnical scholarship has shown that these harms are not technical problems amenable to technical fixes alone, but features of the social, economic, and institutional arrangements in which AI systems are designed, deployed, and used (Costanza-Chock, 2020).

Material access and algorithmic fairness are necessary interventions but not, on their own, sufficient to account for what happens once students encounter AI in their learning. A focus only on external technical factors risks producing the Matthew Effect: the sociological observation that initial advantages tend to beget further advantages (Merton, 1968). The dynamic has been demonstrated in reading development, where early success builds cumulative advantage while early struggles widen the gap (Stanovich, 2000), and in mathematics, where parental socioeconomic status and prior educational experiences shape outcomes (Walberg and Tsai, 1983).

A similar dynamic is now visible in students' engagement with AI. Students who arrive with stronger foundational reasoning skills are more likely to use AI as a thinking partner, asking better questions, interrogating answers, and using the tool to build something new (van de Werfhorst et al., 2022). Conversely, students without these foundational skills are more likely to use AI for passive consumption or basic task completion, engaging in copy-paste behaviours that can lead to cognitive disengagement or a diminished capacity for independent thought (Elshall and Badir, 2025). In this scenario, AI becomes a multiplier of existing educational divides rather than a neutralising force, regardless of how equally distributed the underlying technology is.

This observation aligns with a substantial cognitive psychology literature on cognitive offloading: the use of external aids to reduce cognitive load, which can produce reliable metacognitive illusions about what learners actually know and can do (Sparrow et al., 2011; Risko and Gilbert, 2016). When learners offload demanding cognitive work to AI without scaffolding for critical engagement, they may produce fluent outputs while developing little of the underlying capacity that the work was meant to build. The risk is not that AI will be used; it is that AI will be used in ways that hollow out the cognitive activities that schooling has historically been designed to develop.

The analysis points toward a different kind of question: not what AI does to learners, but what learners do with AI, and what enables some learners to engage productively while others use the technology in ways that hollow out their own cognitive development. This is a question about learner-side capacities; the dispositions, motivational states, and cognitive resources that determine how a tool is used. The question has not been systematically addressed in the AI-in-education literature.

Building the case for a pedagogical model of student-AI engagement requires an account of how learners deepen engagement with a new domain in the first place. This section draws together four bodies of cognitive psychology research that, taken together, describe this process: epistemic curiosity, dual-process cognition, the development of adaptive expertise, and the cyclical character of learning. These literature have developed in parallel and have not previously been integrated into a coherent account of inquiry-driven learning. The synthesis offered here is the analytical foundation on which the SALT Model in Section 4 is built.

Two terms warrant brief working definitions before the argument proceeds, since each appears in a substantial scholarly literature whose internal debates are not directly required here. Curiosity is used in this paper in the sense of epistemic curiosity, following Litman (2005): a motivational state that arises when a learner perceives a gap or possibility in their understanding and is moved toward engagement that might close or explore it. The broader literature contains related distinctions – between trait and state curiosity, diversive and specific curiosity, perceptual and epistemic forms – that matter for empirical research but are not adjudicated here (Loewenstein, 1994; Kashdan and Silvia, 2009). Inquiry refers to the effortful, deliberative engagement with a question or problem that curiosity may, under the right conditions, produce. The distinction matters for the argument that follows: curiosity is the motivational engine, inquiry is the work the engine drives, and the gap between them is the pedagogical territory this paper is concerned with.

Several adjacent constructs in educational psychology might equally be drawn upon to theorise the motivational dimension of inquiry-driven learning. Self-determination theory (Deci and Ryan, 2000) addresses intrinsic motivation; interest research (Hidi and Renninger, 2006) addresses the development of sustained engagement with content; engagement research (Fredricks et al., 2004) addresses behavioural, emotional, and cognitive components of involvement in learning. Curiosity is selected here on two grounds: it uniquely centres the perception of a gap as the trigger for engagement, which is the foundational cognitive move in the interrogation of AI outputs that this paper is concerned with; and it is the construct most directly connected to question-asking behaviour.

Curiosity, in this sense, is the motivational foundation of self-directed learning. Loewenstein's (1994) influential information-gap account holds that curiosity arises when a learner perceives a discrepancy between what they know and what they want to know, with the resulting motivational tension driving the search for information that closes the gap. Litman (2005) refined this account by distinguishing two types of epistemic curiosity: interest-type curiosity characterised by the anticipated pleasure of new discovery, and deprivation-type curiosity characterised by the discomfort of a recognised gap in understanding. Both types motivate inquiry, but they operate differently across the learning process. Interest-type curiosity tends to broaden the scope of a learner's engagement; deprivation-type curiosity tends to deepen it. Engel (2015) argues that curiosity is the foundational disposition schooling must cultivate if learners are to become genuinely independent thinkers, and that it is fragile, easily eroded by pedagogies that prioritise correct answers over productive questions.

Curiosity does not automatically produce inquiry. Whether it does so depends on whether the learner engages in the more effortful kind of thinking that inquiry requires. Dual-process accounts of cognition (Kahneman, 2011; Evans and Stanovich, 2013) describe two modes of mental operation: a fast, automatic, intuitive mode (often called Type 1 or System 1) and a slower, more effortful, deliberative mode (Type 2 or System 2). Most cognition operates in the automatic mode, which is efficient but vulnerable to bias and superficial pattern-matching. Deliberative cognition is what enables critical evaluation, considered judgement, and the interrogation of intuitive responses. Inquiry, in the sense relevant to this paper, requires the deliberative mode.

The pedagogical question this raises is one of conditions: under what circumstances does a learner shift from automatic engagement with a domain into deliberative engagement? Curiosity, in both its interest-type and deprivation-type forms, is one such condition. The recognition of a gap, the experience of surprise, or the perception of a promising lead all create motivational pressure to engage deliberatively rather than to accept the automatic response.

Curiosity and deliberation, however, do not by themselves constitute learning. What turns episodes of inquiry into developing capability is the cumulative consolidation of insights across repeated cycles of engagement. Research on adaptive expertise describes how learners develop the capacity to apply domain knowledge flexibly to novel situations rather than only to familiar problems, and scholars in this tradition have argued that effective pedagogy prepares learners not only to perform within a domain but to transfer their understanding to future learning situations (Schwartz et al., 2000). Adaptive expertise develops through repeated exposure to the kind of inquiry-driven engagement described above, with each cycle deepening the learner's flexible command of the domain.

Synthesising these threads, this paper articulates an inquiry-driven learning cycle with four moments, and summarised in Figure 1.

1. First, exposure: A learner encounters new material or a novel system, often through automatic engagement that takes the material at face value.

2. Second, curiosity: Something in the encounter produces interest-type or deprivation-type curiosity, creating motivational pressure for deliberative engagement.

3. Third, investigation: The learner moves into deliberative cognition, asking questions, examining limits, testing the boundaries of what they have encountered.

4. Fourth, consolidation and transfer: The learner integrates what they have investigated into their developing adaptive expertise, becoming better equipped to engage productively with the next encounter, whether in the same domain or an adjacent one.

The cycle is not a stage theory of individual development; it is a description of the mechanism through which learners deepen engagement with new domains, regardless of the specific domain involved. The same cycle is at work when a child learns to use a calculator, when an adult learns a new programming language, and when a student engages with generative AI for the first time. What varies is the content of the engagement, the depth of the cycles, and the pedagogical conditions that scaffold the cycles. It is to those pedagogical conditions, applied to student-AI engagement, that this paper now turns.

The SALT Model proposes a four-phase curricular sequence for scaffolding student-AI engagement, reflected in Table 1 (see Section 4.2 below). Each phase describes a pedagogically designed environment in which the inquiry-driven learning cycle described in Section 3 is given a particular shape, focused on a particular dimension of the student's engagement with AI. Movement between phases is not a developmental transition occurring inside the learner; it is a curricular design decision made by educators, who scaffold increasingly sophisticated engagement as students develop adaptive expertise from earlier phases. This pedagogical framing is consequential and is explained further below.

In addition to the curiosity-driven cycle developed in Section 3, the SALT Model draws on the constructive-developmental theory for its articulation of the subject-object shift: the cognitive movement by which an element previously fused with the self becomes available for reflection and evaluation (Kegan, 1983). This mechanism speaks directly to a central question in student-AI engagement: how a learner moves from a position in which the AI's output is accepted as authoritative (the AI is, in Kegan's language, “subject” to the learner and fused with the learner's sense of what is true) to a position in which the output is interrogable from a stance of independent judgement (the AI has become “object” and available for reflection). The subject-object move is the conceptual heart of what Phase 3 of the SALT Model aims to scaffold.

Several adjacent developmental traditions describe how learners develop more sophisticated epistemic stances, including dualism-to-relativism (Perry, 1970), reflective judgement (King and Kitchener, 1994), and transformative learning (Mezirow, 1991). Kegan was selected over these alternatives because the subject-object mechanism provides the most direct conceptual tool for the specific transition this paper concerns: the movement from accepting AI outputs to interrogating them. Vygotskian and sociocultural traditions (Vygotsky, 1978) and posthuman scholarship (Braidotti, 2013) would generate a different paper, since they would treat the learner-AI relationship as inseparable rather than as a relationship between distinct entities, and would foreground the cultural and material mediation of cognition more radically than the present argument requires. Their critiques are taken seriously but are not adopted as the analytical frame here.

The paper applies the subject-object mechanism as a heuristic resource for pedagogical design, not as a claim that learners pass through Kegan's developmental stages in their engagement with AI. Where Kegan's stages are slow claims about individual psychological maturation spanning years to decades, SALT's phases are pedagogically scaffolded environments sequenced over weeks to terms. The subject-object move that Phase 3 aims to scaffold is analogous to the one Kegan describes; it is not a literal instance of it.

In the first phase, the curricular focus is on the mechanics of AI use for immediate task completion. The learner treats the technology as a discrete instrument, analogous to the early introduction of handheld calculators, where students performed arithmetic without understanding circuit logic.

Illustrative activities include using AI as a basic retrieval tool to find historical facts or definitions, performing simple arithmetic, translating individual words or short phrases, and generating basic visual ideas. The pedagogical goal of this phase is operational fluency: students develop the mechanical familiarity that subsequent phases will build upon.

The second phase shifts curricular focus toward demystification. Pedagogy is designed to help learners understand AI as a probabilistic prediction engine with specific logical constraints, rather than as an oracle. This mirrors the calculator user learning the order of operations: they realise that incorrect inputs or a lack of logical structure will yield erroneous outputs.

Illustrative activities include refutation-based fact-checking, in which students cross-reference AI claims against primary sources to identify factual errors and address misconceptions (Schroeder and Kucera, 2022); iterative prompting designed to reveal how refinement of instructions shapes generated responses and exposes the tool's underlying constraints (Mollick and Mollick, 2023); and STEM applications involving the manual verification of suggested code or the identification of “garbage in, garbage out” results in data processing.

The pedagogical goal is technical demystification, helping students develop a more accurate mental model of what AI can and cannot do.

The third phase is the pivotal point of the curricular sequence. Pedagogy here scaffolds the subject-object move: students are supported to step outside the algorithm and interrogate its outputs from a position of independent judgement. This is the phase in which critical sovereignty over synthetic outputs is developed.

Illustrative activities include the social and ethical interrogation of AI outputs, in which students intentionally audit synthetic outputs for encoded bias and name the ideologies present in generated text (Boateng and Boateng, 2025); learning-by-teaching, in which students attempt to “teach” a complex concept to an AI chatbot to identify where its synthetic logic fails or reverts to generic patterns (Tan et al., 2025); and the interrogation of AI as a medium, using the historical analogy of photography to evaluate AI-generated artistic media and to question whether synthetic outputs expand aesthetic diversity or merely reproduce common cultural norms found in the training data (Epstein et al., 2023).

The pedagogical goal of this phase is the cultivation of critical sovereignty and the capacity to distinguish one's own voice and values from the synthetic narratives the machine produces. Curiosity drives deep investigation, investigation issues in critique, and critique consolidates into an interrogative stance that students can take into novel encounters with AI.

In the final phase, pedagogy positions the student as the primary architect of inquiry, with AI as a provocative extension of their cognition. The machine is no longer an instrument for task completion but a partner used to amplify and challenge the learner's own thinking. This mirrors the use of graphing calculators to visualise complex functions or to test hypotheses that would be difficult to explore manually.

Illustrative activities include AI-assisted Socratic questioning, in which students task the AI with generating sophisticated counter-arguments or alternative perspectives to their own thesis, sharpening their analytical skills in response (Mollick and Mollick, 2023; Tang et al., 2024); collaborative ideation, in which student and machine refine a vision together, shifting the focus from manual execution to high-level conceptual intent and creative direction (Epstein et al., 2023); and cognitive mirroring, in which AI is prompted to identify logical blind spots or synthesise conflicting datasets to broaden the learner's perspective (Yatani et al., 2024).

The pedagogical goal of this phase is integrative agency in which AI provokes and deepens human cognition rather than replacing it.

Two features of the SALT Model warrant explicit articulation. First, movement between phases is a curricular design decision, not a developmental trigger inside the learner. Phase progression is the educator's choice about which engagement to scaffold next, informed by the adaptive expertise students have developed in earlier phases. This is what distinguishes SALT from a stage theory: it is not a description of what happens inside students; it is a description of what curricula and pedagogy can scaffold.

Second, the model is recursive in a specific sense. When learners encounter a new domain of knowledge – a new subject, a new topic within a subject, or a new kind of problem – the curricular sequence can be productively re-applied, even by learners who have previously reached Phase 4 with respect to other domains. The recursion is not a developmental regression but a redeployment of a pedagogical scaffold to a new content area, with the learner bringing increased adaptive expertise to each successive iteration.

The SALT Model is offered as a complement to, not a replacement for, existing frameworks in the AI literacy and learning sciences literature. Several frameworks describe relevant terrain, and clarifying the relationship between SALT and these neighbours sharpens what SALT distinctively contributes. The AILit Framework (European Commission and OECD, 2025) and adjacent AI literacy work (Long and Magerko, 2020; Ng et al., 2021; UNESCO, 2022) map the competencies that learners require for productive engagement with AI; SALT articulates a pedagogical sequence through which such competencies might be cultivated.

Self-regulated learning frameworks (Zimmerman, 2002) describe how learners regulate their own learning processes; SALT describes the pedagogical environments that scaffold inquiry-driven engagement, with self-regulation operating within and across phases. Table 2 summarises the principal points of distinction.

Critical digital literacy work (Pangrazio, 2014) addresses the political and economic conditions under which digital technologies operate; SALT does not engage these conditions directly but is designed to be compatible with the kind of critical engagement this tradition seeks to cultivate. Metacognition research (Veenman et al., 2006) describes the regulatory awareness learners bring to their own thinking; SALT's Phase 3 in particular draws on this capacity without claiming to develop it independently. These adjacent traditions describe capacities that SALT recruits and complements rather than replaces.

The principal contribution of SALT to this conversation is the articulation of phase progression as a curricular sequence informed by a cyclical model of inquiry learning. Where AI literacy frameworks describe what learners should know and self-regulated learning frameworks describe how learners govern their own learning, SALT describes how curricula can be designed to scaffold the inquiry-driven engagement that turns knowing about AI into knowing how to engage critically with it.

The pedagogical model developed in Sections 4 and 5 requires an assessment counterpart. Traditional output-based assessment (essays, problem solutions, written examinations, etc) has been destabilised by generative AI, which can produce on demand many of the outputs that traditionally indexed learning (Dawson, 2021; Bearman et al., 2024). Cultivating inquiry-driven agency requires assessment that remains meaningful under these conditions. This section proposes inquiry-sequence assessment, which evaluates the questions students ask of AI rather than the answers they produce with it, as one such form.

One promising response in the literature is to cultivate students' evaluative judgement: the capacity to assess the reliability and relevance of AI-generated outputs (Bearman et al., 2024). Evaluative judgement is an important and necessary capability, but it focuses on the learner's relationship to outputs rather than on the structure of the learner's own inquiry behaviour. There is space alongside it for an assessment approach that examines the inquiry sequence itself: the questions a learner asks and the way those questions develop over the course of an engagement with a problem.

The educational research literature has long treated question-asking as a meaningful indicator of cognitive engagement. Foundational work on question-asking during tutoring established that the questions students ask reveal what they have understood and how deeply they are engaging with material (Graesser and Person, 1994), and review of research in science education has demonstrated that student questioning is a diagnostic resource for both teaching and learning that answers alone do not provide (Chin and Osborne, 2008). These capacities are what the inquiry-driven cycle in Section 3 describes and what the SALT Model in Section 4 aims to scaffold.

Generative AI changes the assessment calculus for question-asking in a specific way. Where answers can now be produced by AI on demand, questions remain meaningfully attributable to the learner. A student may use AI to generate an essay, but the sequence of prompts they enter, the iterations of questioning they pursue, the points at which they push back on a response and the points at which they accept it, are observable expressions of their cognitive engagement. Inquiry-sequence assessment proposes that this observable expression is itself assessable.

A methodological precedent for assessing inquiry sequences exists in medical education. The Objective Structured Clinical Examination (OSCE) assesses clinical competence by structured observation of how trainees approach standardised patient scenarios (Harden and Gleeson, 1979). A central feature of the OSCE format is that it assesses not only what the trainee concludes but how they get there: the sequence of history-taking questions, the order of clinical reasoning, the moments at which the trainee revises their working diagnosis in light of new information. OSCE assessors use structured rubrics that score inquiry behaviour along defined dimensions, with established reliability across raters.

The OSCE approach demonstrates that inquiry sequences can be assessed systematically and can be scored reliably with structured rubrics, and that doing so produces educationally meaningful information about the assessed competence. The proposal here applies this methodological precedent, adapted for the secondary and tertiary classroom contexts in which AI is increasingly present.

However, inquiry-sequence assessment is not proposed as a replacement for all other forms of assessment. There are dimensions of learning that are better assessed in other ways, and there are practical constraints, such as the scoring labour required, that make inquiry-sequence assessment unsuitable for many contexts in its current form. Nor is the proposal that AI cannot, in principle, produce sequences of plausible questions.

Inquiry-sequence assessment makes the structure of a student's engagement visible in a way that AI-produced answers do not, and that visibility is diagnostically useful for educators designing within the SALT Model. The cultivation of inquiry-driven agency that the model aims to support requires forms of assessment that can index how a student enquires, not only what they produce. This section has proposed one direction such assessment might take; empirical work is needed to develop, test, and refine specific instruments.

The most direct implication of the SALT Model is for curriculum design. The model proposes that AI literacy, in the sense relevant to inquiry-driven engagement, is not adequately addressed by adding a unit on AI to an existing curriculum or by listing competencies students should acquire. It is addressed by sequencing students' encounters with AI across phases of increasing sophistication, with each phase scaffolding the inquiry-driven cycle in a particular shape. This has practical consequences. A SALT-informed curriculum at secondary level might begin with operational uses of AI in Year 9 or 10 (Phase 1), introduce systematic interrogation of AI's logical structure in Year 11 (Phase 2), develop critical sovereignty over AI outputs in Year 12 (Phase 3), and cultivate integrative use of AI as a partner in tertiary or post-secondary contexts (Phase 4). The specific year-level mappings are illustrative; the principle is that the sequence matters.

Such sequencing is in tension with current curriculum design in many jurisdictions, which prioritises breadth of content coverage and leaves limited curricular time for the iterative inquiry that the model assumes. The OECD has documented this “curriculum overload” as a global challenge (OECD, 2020), and the present argument does not pretend to resolve it. What the argument does suggest is that the integration of AI into education is itself an opportunity to reconsider where in the curriculum the cultivation of inquiry capacity sits, and that AI literacy is poorly served by approaches that treat it as content to be covered rather than as a developmental sequence to be scaffolded.

This tension is itself shaped by the broader performativity regime that has come to dominate contemporary schooling, in which measurable outputs are valued over the more diffuse capacities the SALT Model proposes to cultivate (Ball, 2003); a SALT-informed curriculum is at risk of being absorbed into the same metric-driven logic it might otherwise help displace.

The SALT Model has consequences for how educators are themselves prepared. The most demanding pedagogical work the model envisions, particularly in Phases 3 and 4, requires educators who can model the inquiry-driven engagement they are seeking to cultivate. The educator's role in this conception is not principally that of a transmitter of content but that of a role model of the enquiring learner: someone who, in front of students, demonstrates what it looks like to engage productively with an AI output, to identify its limits, to push back where it has been glib, and to extend it where it has been useful. This is a meaningful shift in the educator's professional self-conception, and it cannot be achieved through brief professional development sessions alone.

Emerging data suggests a significant readiness gap within the teaching workforce: only 28% of teachers reported receiving formal training on how to respond to student use of generative AI (Dwyer and Laird, 2024), and educators in well-resourced environments are more likely to develop the pedagogical capacity that the SALT Model assumes than those in under-resourced settings. The risk is the “pedagogical Matthew Effect” identified earlier: educators in resource-rich environments scaffold students into Phase 4, while those in resource-poor environments remain stalled at Phase 1.

The inquiry-sequence assessment proposition developed in Section 5 has implications for assessment policy and practice. Most concretely, it suggests that institutions should treat the redesign of assessment for the AI era as a pedagogical question rather than an integrity question. The dominant response to generative AI in many institutions has been to invest in detection tools and proctoring technologies designed to catch students using AI inappropriately. These responses are reactive and increasingly fragile (Liang et al., 2023). A more durable response is to redesign assessment around what AI cannot easily substitute for: the structure of a learner's inquiry. Such redesign requires investment in rubric development, assessor training, and the kind of process-oriented assessment that authentic assessment scholars have been advocating for (Abdul Karim et al., 2025). It is more demanding than maintaining current practice. It is also more likely to produce assessment that remains diagnostically useful as AI capabilities continue to develop.

Five limitations of the argument warrant explicit acknowledgement. First, the model is conceptual rather than empirically validated. The four phases of SALT, the four moments of the inquiry-driven learning cycle, and the inquiry-sequence assessment proposition are all theoretically derived and have not been tested in classroom settings. Empirical validation is needed before the model can be confidently recommended for adoption. Section 6.5 discusses the research that could test it.

Second, a model that focuses on the cultivation of learner inquiry capacity can be misread as placing the burden of educational equity on individual students. The argument here has been explicit that the locus of responsibility sits with institutions and educators designing learning environments, not with students themselves, but this framing must be maintained in implementation. If a SALT-informed curriculum is deployed without the institutional resources, educator preparation, and curricular space the model assumes, it can replicate the pedagogical Matthew Effect described above, in which inquiry-driven pedagogy becomes available primarily to already advantaged students.

Third, the model's applicability to younger learners, neurodiverse learners, and learners in low-literacy contexts has not been fully explored. The illustrative activities in Section 4 assume a level of language facility and metacognitive readiness that is most appropriate to upper-secondary and tertiary contexts. Adapting the model for primary settings, or for learners with significant language or attention differences, would require further conceptual and empirical work. The same applies to contexts in which curriculum is heavily standardised or examination-driven, where the curricular space the model assumes may not be available.

Fourth, the model is at risk of pedagogical determinism if its phases are read as a fixed progression. The phases describe pedagogically designed environments, not a developmental sequence learners must pass through, and educators treating Phase 4 as an end-state for all learners would be misreading the model. The four phases offer a curricular vocabulary for thinking about what is being scaffolded at each stage; they do not require any particular learner to reach any particular phase.

Fifth, the model addresses the pedagogical dimension of AI in education and does not address the structural dimensions of educational inequity that the manuscript's analysis acknowledged in Section 2. Pedagogical work alone cannot substitute for the structural reforms that critical scholarship has shown to be necessary (Selwyn, 2016; Williamson, 2017; Noble, 2018; Eubanks, 2018; Benjamin, 2019).

The argument advanced here opens several lines of empirical inquiry. Longitudinal studies are needed to investigate whether SALT-informed curricula produce demonstrable differences in students' inquiry-driven engagement with AI, compared to alternative pedagogical approaches. Studies of educator practice are needed to identify what pedagogical moves most effectively scaffold the subject-object shift the model describes in Phase 3. Methodological work is needed to develop and validate inquiry-sequence assessment instruments, drawing on the OSCE tradition for rubric design and assessor training (Harden and Gleeson, 1979) and on emerging authentic-assessment scholarship (Bearman et al., 2024; Dawson, 2021).

Comparative studies across different institutional contexts and learner populations will be needed to test the model's applicability beyond the upper-secondary and tertiary contexts that have shaped its initial formulation.

The integration of AI into education has prompted substantial scholarly and policy attention to the technological and structural dimensions of equitable AI use. The cultivation of learner-side capacities, the inquiry-driven engagement that determines whether AI becomes a tool for independent thought or a multiplier of existing divides, has received comparatively less sustained attention. This paper has argued that this is a worthwhile object of pedagogical thought in its own right, and has offered three connected contributions toward its development.

The first contribution, conceptual, is the cyclical model of inquiry-driven learning developed in Section 3, drawing on epistemic curiosity, dual-process cognition, and adaptive expertise scholarship to describe how learners deepen engagement with new domains. The second, pedagogical, is the Student-AI Learning Trajectory Model in Section 4, a four-phase curricular sequence in which phase progression is a curricular design decision informed by the inquiry-driven cycle and by the subject-object mechanism drawn heuristically from constructive-developmental theory. The third, operational, is inquiry-sequence assessment in Section 5, which evaluates the questions students ask of AI rather than the answers they produce with it.

These contributions are conceptual rather than empirically validated, and they invite the empirical research described above. Within these limits, the argument offers a vocabulary and a set of tools for educators, curriculum designers, and assessment scholars who are responding to the question of how AI should be integrated into the work of cultivating independent learners. The question of what schools and universities scaffold and assess in this space matters for the kinds of learners we send into a world increasingly shaped by automated systems.

AI tools were utilised in the production of this paper, primarily for the purposes of conducting research and finding credible sources. Some copy-editing was applied via Gemini in the refinement and clarity of the language.

Perplexity (2026). Perplexity [Large language model]. Link to the website

Google (2026). Gemini (Version 2.5 Pro) [Large language model]. Link to the website