Executive summary
This paper advances a single axiomatic claim — the Human Agency Anchoring (HAA) axiom — and derives from it a competency framework for educators in the age of AI. The axiom has two clauses. First, in educational contexts, AI systems function as extensions of human cognition rather than as autonomous agents alongside human cognition. Second, the agency that makes such extensions purposive — the capacity for goal-setting, self-regulated action, and responsible judgement — remains anchored in the human. What AI systems exhibit as apparent agency is inherited human agency: judgement crystallised in training data at design time and re-played at inference time.
The two clauses operate at different levels and do not conflict. Cognition can be distributed across human and artefact (Clark & Chalmers, 1998); agency, in the strict sense of intentional, forethought-capable, self-reactive, and self-reflective action (Bandura, 2001), cannot. The first clause permits extension; the second constrains what extension can coherently do in a domain whose very object is the formation of human agency.
The axiom is the paper's organising principle. From it we derive four corollaries — ontological, epistemological, ethical, and pedagogical — a six-mode typology of extension (Types A through F), an eight-dimension competency framework, and twenty-four sub-indicators. The framework is positioned against three international reference frameworks: UNESCO's AI Competency Framework for Teachers (Miao & Cukurova, 2024), the European Commission's DigComp 3.0 (Cosgrove & Cachia, 2025), and the ISTE Standards for Educators (ISTE, 2024). The paper closes with limitations and an agenda for empirical work.
A terminological note. Among candidate expansions of HAA — "Human-as-Anchor," "Human Agency Augmentation," "Human Agency Anchoring" — we adopt the last because it specifies both what is anchored (agency, in Bandura's sense) and how (as the purposive anchor of the extended activity, not merely as one contributor among others). The expansion also echoes the sub-indicator "Anchoring of Teacher Agency" in Dimension I, reinforcing internal coherence.
1Introduction: why educational AI needs its own axiomatic foundation
Policy and practitioner frameworks for AI in education have proliferated since 2023. Most borrow — explicitly or tacitly — either from general digital-literacy frameworks or from enterprise and industrial framings of AI adoption. The first treats AI as one digital technology among many; the second treats AI as a productivity multiplier whose virtue is the displacement of costly human effort. Both fail to address what is constitutive about education as a domain, and both consequently mis-specify what educators need to know and do.
The argument of this paper is that education occupies a distinctive position among AI application domains. Its telos — the cognitive, moral, and social development of the learner — is itself an intelligent agent under formation. What the activity produces is not a thing but a mind capable of its own purposive action. Any deployment of AI that displaces the learner's cognitive effort, or that displaces the teacher's purposive direction of that effort, is not a more efficient instance of the activity but a disfiguration of it. The activity is constituted by the exercise of the agency that AI might otherwise be tempted to substitute for. Hence: extension is permitted; displacement is not. The anchor of the activity — its purposive centre — must remain human.
We name this commitment the Human Agency Anchoring (HAA) axiom. Taking it as given, we construct, in what follows, a competency framework that the axiom constrains and motivates. The paper proceeds in ten steps. Part 2 states the axiom and locates it in a longer literature on cognition, tools, and agency. Part 3 defends the claim of constitutive difference by contrasting education with three comparison domains. Part 4 draws four corollaries from the axiom. Part 5 introduces a six-mode typology of extension. Part 6 derives the eight-dimension structure. Part 7 presents the framework. Part 8 positions it against three international reference frameworks. Part 9 notes limitations and open questions. Part 10 concludes.
2The Human Agency Anchoring axiom
2.1Statement
The HAA axiom consists of two clauses:
In educational contexts, AI systems function as extensions of human cognition. They can constitutively participate in cognitive activities by supplying representational, computational, or inferential resources reliably coupled to the human user.
The agency that makes such extensions purposive — the capacity to set goals, select means, evaluate outcomes, and bear responsibility — remains anchored in human agents. What AI systems exhibit as apparent agency is inherited human agency: judgement crystallised in training corpora at design time and re-played at inference time.
The two clauses operate at distinct levels. Clause I concerns cognition — the informational and representational substrate of thought. Here distribution across human and artefact is unproblematic; it is the standard finding of distributed-cognition research. Clause II concerns agency — the purposive, self-directed dimension of action. Here distribution is not available: agency is something only an intentional being can have, and AI systems, as currently constituted, are not intentional beings. The two clauses are therefore complementary rather than in tension. An AI-extended cognitive activity is one in which cognition is distributed across human and system while agency remains anchored in the human.
2.2What is being anchored: a brief account of agency
By "agency" we mean what Bandura (2001) calls agentic perspective: the capacity to "intentionally make things happen by one's actions." Bandura identifies four core features of agency — intentionality, forethought, self-reactiveness, and self-reflectiveness — that cohere only in an intentional being and operate through phenomenal and functional consciousness. To these he adds three modes through which agency can be exercised: direct personal agency, proxy agency (acting on another's behalf), and collective agency. The distinctions matter for our argument. A teacher exercises proxy agency on behalf of the learner-in-formation, and this proxy exercise is part of what teaching is; an AI system, lacking the phenomenal and functional consciousness through which agency's core features operate, cannot do the same.
The four features are directly relevant to educational competencies. Intentionality and forethought underwrite the teacher's goal-setting function (what are we extending toward?). Self-reactiveness and self-reflectiveness underwrite the teacher's quality-evaluation function (is the extension enlarging the right capacities?). Together, these features underwrite the teacher's responsibility function (who answers for this?). We return to these in the pedagogical corollary (§4.4), and to the collective mode in Dimension VI (§7).
2.3What "anchoring" does
Anchoring is a structural rather than defensive claim. It is not the assertion that humans should resist AI encroachment on agency; it is the assertion that, in the educational setting, human agency is the purposive centre of the activity whether we attend to it or not — and that designs, deployments, and professional practices that obscure this fact produce incoherence, not efficiency. An anchor is (i) a stable point of reference from which extensions reach out, (ii) a constraint against drift, and (iii) the seat of responsibility. The metaphor captures all three: a ship without its anchor is not faster; it is adrift.
2.4Intellectual lineage
The axiom consolidates and specialises converging lines of work in the learning sciences and philosophy of mind:
- Extended cognition. Clark and Chalmers (1998) argued that cognitive processes can constitutively include extra-cranial resources when these resources are reliably coupled to the agent and function as internal resources would. Clause I of the axiom applies this thesis to AI.
- Partners in cognition. Salomon, Perkins, and Globerson (1991) distinguished effects of technology (what a technology does while in use) from effects with technology (what the human can do in enduring ways as a result of working with it). The educationally decisive quantity is the second: a well-designed extension must leave the human agentically larger after withdrawal.
- Reorganising mental functioning. Pea (1985) argued that cognitive technologies do not merely amplify pre-existing capacities but reorganise them, producing qualitatively different activity. Under HAA, the admissible reorganisations are those that reorganise cognition without relocating agency.
- Agentic perspective on action. Bandura (2001) supplies the theory of agency underwriting Clause II: intentionality, forethought, self-reactiveness, self-reflectiveness, and the three modes (personal, proxy, collective).
To these we add a Vygotskian consideration: tools are not neutral intermediaries between agent and task but mediators that reshape the activity. When the tool is AI, the activity is re-composed; the question is by what principles the re-composition proceeds. HAA supplies a principle: cognition may be redistributed, but agency must remain anchored.
2.5What the axiom does and does not commit to
HAA is an ontological claim: it specifies what AI systems are and are not in the educational setting. It does not, by itself, imply that AI is always beneficial, that teachers should adopt any particular tool, or that concerns about AI are overstated. It commits only to this: the appropriate evaluative question for any educational AI deployment is not "Did the AI perform the task?" but "Did the human — teacher or learner — come out of the activity with enlarged cognitive reach and intact agency, in ways that persist when the AI is withdrawn?" A "yes" to the first without a "yes" to the second is a failure of educational AI, however fluent the output.
3Why educational AI is constitutively different
HAA is not a preference for human-centred design. It is a claim that education is structurally unlike the domains from which AI-adoption vocabulary has mostly been imported, and that the structural difference forces anchoring rather than merely recommending it.
3.1A general principle: the telos of the domain fixes where agency can reside
In any domain in which AI is deployed, three elements are arranged: the human, the AI system, and the object of the activity. What AI does in the domain depends on how the object of the activity stands to the human. If the object is external to the human — a physical good, a delivered package, a correctly categorised image — then AI can in principle substitute for human effort and the relevant metric is some efficiency ratio. If the object is the human, or an attribute of the human that only agentic activity can constitute, then substitution is either incoherent or harmful.
3.2Contrast: manufacturing and logistics
In manufacturing, the object of the activity is a physical good. The human worker stands in an instrumental relation to that good: the worker is a means of producing it, and the worker's own cognitive or agentic development during the process is an externality. Paradigmatic uses of AI in manufacturing — robotic assembly, defect detection, route optimisation — are therefore substitutive, and the metric of success is displacement of human effort per unit of throughput. The welder replaced by a robotic arm marks success, not failure; so does the dispatcher replaced by an optimiser. There is no educational analogue of this structure: a classroom in which AI has "replaced" the learner is not a more efficient classroom but an absurdity, because the learner's agentic participation is what the classroom is for.
3.3Contrast: medicine
Medicine is a more revealing comparison because it shares with education the centrality of a human object — the patient's health — and the centrality of professional judgement — the clinician's. Medical AI is accordingly closer to educational AI than to industrial AI. But the two diverge on one decisive point: the patient's agentic development is not, in general, the telos of medical care. A patient who recovers without having understood the treatment is still a successful case. In education, by contrast, a learner whose assignment is completed without the learner's cognitive engagement is not a successful case but a failed one: what was supposed to happen — the learner's agentic engagement — did not happen. Tellingly, psychotherapy — the branch of medicine whose telos most resembles education's — is also the branch where displacement of the therapist's judgement by AI is most obviously problematic.
3.4Contrast: creative production
Creative industries offer a mixed case. A professional using AI image generation to produce assets for a client is engaged in substitutive use; the client's interest is in the artefact, not in the professional's development. But a student using the same tool in an art course is engaged in educational use, and the educational interest is precisely in the student's development as an artist — which can be undermined by substitutive use even when the artefact produced is visually indistinguishable. The same technology plays different roles in different domains, not because of any property of the technology itself but because of how the object of the activity relates to human agency.
3.5The constitutive claim
Education's distinctive structural feature is that the object of the activity — the formation of an agent capable of intentional, self-regulated action — cannot be attained by any process that displaces that agent's own engagement. This is not a contingent preference of educators; it follows from what education is. An "education" in which the learner has not exercised agency is a contradiction in terms, in a way that a "manufacture" in which the worker has not exercised agency is not. The same holds, mutatis mutandis, for the teacher: a teacher who has delegated pedagogical judgement to an AI system has not merely been more efficient but has ceded the purposive centre of the practice, because the practice is constitutively one of purposive formation of another's agency.
It follows that the appropriate framing of educational AI is not "AI as partial substitute for teaching or learning labour" but "AI as extension of the cognitive capacities whose exercise is the activity itself, with agency anchored in the human." Anchoring is not a style choice; it is what consistency with the nature of the domain requires.
4Four corollaries of the HAA axiom
Granting the axiom, four further propositions follow. Each states what educational AI cannot coherently be; together they supply the design space within which any competency framework for educators must be constructed.
4.1Ontological corollary: AI has no independent agency
Descriptions of AI systems as "autonomous agents," "intelligent tutors," or as "deciding" between options are, strictly speaking, misdescriptions in the educational setting. What the system exhibits as apparent agency is the residue of its training — human judgement crystallised at design time and re-played at inference time. This corollary defuses a common contradiction in policy documents, which simultaneously affirm teacher agency and describe AI as though it had agency of its own. Within the HAA frame, the contradiction disappears: what appears as the system's agency is in fact inherited human agency, temporally displaced. There is no second agent in the room.
4.2Epistemological corollary: from operation to enlargement
If AI is an extension and agency is anchored in the human, the evaluative question is not whether the teacher can operate AI, nor whether the AI produces fluent output, but whether the human agent with whom the extension is coupled has had his or her agentic capacities enlarged by the coupling. Writing literacy supplies the natural analogy: the point of writing is not proficiency with a pen but the externalisation, accumulation, and reorganisation of thought that writing affords — and the agentic control the writer thereby gains over her own thinking. Analogous questions for AI literacy: what is being accumulated? what is being reorganised? what agentic control is being gained? and — the mirror question — what is being offloaded that should have been retained?
4.3Ethical corollary: responsibility traces to the anchoring agent
If AI has no independent agency, it cannot bear responsibility. Responsibility for AI outputs traces to the humans who chose to extend themselves in a particular way: the teacher who selected the tool and configured its use, the institution that approved the tool, the designers who built it, and the data curators whose choices shaped what the tool "knows." Algorithmic bias, on this view, is not a property of AI but the encoded amplification of prior human judgement; and the teacher's ethical obligation is not to "supervise the AI" but to take responsibility for the extension choices being made.
4.4Pedagogical corollary: the teacher's irreplaceability acquires precise content
Three functions cannot, in principle, be performed by the extension itself because they require the exercise of agency in Bandura's strict sense — features that presuppose an intentional being:
- Goal-setting (intentionality + forethought). What are we extending toward? Which learner capacities should be enlarged? The answer requires intentional projection of a future state, which is not available to systems that lack phenomenal consciousness.
- Quality evaluation (self-reactiveness + self-reflectiveness). Is the extension enlarging the right capacities? Judging this requires the capacity to reflect on one's own practice and to regulate it in response, in light of values that cannot be fully specified in advance.
- Responsibility-bearing (all four features jointly). Who answers for the consequences of the extension? Only an agent capable of all four features can intelligibly bear responsibility, because responsibility requires that the agent could have acted otherwise and knew that it could.
The teacher's role, in the age of AI, is not primarily to operate AI but to perform these three functions competently — often, in Bandura's terms, as proxy agent for the learner. Everything else that teachers do can, in principle, be redistributed across extensions; these three functions cannot.
5Six modes of AI extension: Types A through F
The HAA axiom is consistent with multiple modes of extension, each of which imposes distinctive demands on the anchoring agent. We distinguish six types as organising categories for the competencies that follow. The typology is derived from the pattern of human–AI coupling rather than from the underlying technology: the same large language model may instantiate several types depending on how the teacher engages with it.
The six types are not points on a single scale. They occupy a space defined by at least three axes:
Types A, B, and C (introduced in earlier iterations of this framework) cluster around explicit, user-initiated couplings. Types D, E, and F extend the typology into three phenomena that the original three did not cleanly name: generative couplings in which agency is exercised at authorship time and then withdrawn (D); delegated couplings in which the human authorises a policy but does not observe individual acts (E); and ambient couplings in which the infrastructure itself extends cognition silently and persistently (F). Each of the three additions carries distinctive HAA risks that the original typology could not directly address.
| Type | Mode of coupling | Principal demand on the anchoring agent | Characteristic scenarios |
|---|---|---|---|
| ATask-execution | The human specifies a task; the AI executes it. | Precise intent-articulation; evaluation of output against articulated intent. Risk: agency preserved but under-specified — task executed, intent unclear. | Prompt-based lesson drafting, rubric generation, content re-levelling, translation of source materials. |
| BAnalytical | The AI produces inference or pattern over data; the human interprets and acts. | Critical interpretation of analytic output; bounded trust in algorithmic inference; agentic judgement over individual learners. Risk: agency ceded to the algorithm's framing of the situation. | Learning analytics, knowledge-state diagnostics, early-warning signals, group-level inference over assessment data. |
| CCo-cognitive | The human and the AI engage in sustained dialogue that shapes the human's own thinking. | Preservation of agentic tension under sustained interaction; recognition of echo-chamber effects; distinguishing genuine insight from fluent confirmation. Risk: agency eroded by smooth interaction that feels like thinking. | AI-assisted reflection on teaching, dialogic ideation, collaborative problem-framing, critique of one's own argument. |
| DGenerative | The AI produces a durable artefact — simulation, scenario, or environment — that the learner subsequently inhabits or re-enters over time. | Anticipatory design at authorship time (Bandura's forethought in concentrated form); awareness that agency exercised at generation is subsequently withdrawn; explicit reasoning about the pedagogical commitments embedded in the generated artefact. Risk: hidden pedagogy — teachers may not recognise what agentic capacities the generated environment cultivates or erodes. | AI-generated clinical cases and historical scenarios, auto-generated branching narratives, AI-customised virtual labs and game-based learning environments, auto-generated problem sets with embedded scaffolding. |
| EDelegated | The AI acts on the human's behalf within a loop the human authorises but does not directly supervise. | Policy-level design and audit of the delegation; construction and maintenance of meaningful oversight despite not observing individual acts; clarity about the boundary between authorised policy and emergent behaviour. Risk: the chain of agency becomes illegible — the teacher authorised the policy but not the act, the student received the act but not the policy. Responsibility is easiest to lose here. | Autonomous tutoring agents running unsupervised learner sessions, AI graders applied at scale, classroom orchestration agents that adapt instruction in real time, AI communications agents that respond to student messages. |
| FAmbient | The AI is not invoked by the user but is present in the infrastructure — recommender systems, silent re-sequencing, platform nudges — shaping cognition persistently and often invisibly. | Detection of extension one did not initiate; capacity to surface, name, and scrutinise the extension at institutional levels; resistance or principled acceptance rather than acquiescence by default. Risk: invisibility — agency cannot anchor what it does not see. This is the type where structural ethics becomes primary rather than supplementary. | LMS content-recommendation algorithms, adaptive-platform silent re-sequencing, algorithmically ordered resource libraries, notification and reminder systems that shape student attention, institutional dashboards that frame student identity. |
The six types demand different competencies and carry different risk profiles. Type A is primarily a competency of intent-articulation. Type B places the greatest demand on statistical and interpretive literacy. Type C is, on the HAA view, among the highest-risk modes because fluent interaction is most easily mistaken for cognitive collaboration. Type D shifts the locus of anchoring to authorship time and demands forethought about affordances the teacher will not subsequently control. Type E is the mode in which the chain of agency is most easily broken: pre-authorised policies executed in the teacher's absence produce acts for which responsibility is structurally ambiguous unless the policy design makes it legible. Type F is the mode in which the HAA axiom does its hardest institutional work, because the relevant agency is often not individual but collective — the capacity of teachers, as a professional community, to see and name what platforms are doing.
The framework below covers all six types, and several sub-indicators are explicitly tied to the types whose anchoring risks they principally address. The HAA-Type column in each dimension table reflects the extended typology.
Note. The A–F typology is offered here as a construct grounded in observed patterns of use; it is not yet empirically validated and, as noted in Part 9, stands in need of expert-agreement and practitioner studies. A–C represent the core user-initiated couplings addressed in earlier iterations of this framework; D–F extend the typology to generative, delegated, and ambient couplings whose HAA stakes the original three did not directly address.
6From axiom to framework: deriving the eight dimensions
The eight dimensions of the framework are not an arbitrary enumeration. Each is a competency region implied by the HAA axiom and its corollaries, partitioned so the regions are jointly exhaustive of the competencies an educator needs and minimally overlapping in content. Dimensions I–V follow directly from the four corollaries and cover the individual teacher's agentic relationship to AI extension. Dimensions VI–VIII extend the framework in three directions that earlier iterations handled only implicitly: collective agency, equity-differentiated anchoring capacity, and the assessment crisis that the epistemological corollary entails under conditions of AI. Each extension is argued for rather than appended.
| Dim. | Competency region | Derivation from HAA |
|---|---|---|
| I | Foundations of Extended Cognition | From the ontological corollary: if AI is an extension and agency is anchored in the human, teachers must understand what an extension is, what types exist, where its boundaries lie, and why their position as anchoring agent is irreplaceable. |
| II | Extension Operation Capability | From the epistemological corollary: the competence at stake is not tool operation but intent-articulation, output evaluation, and situational customisation — the skills by which a teacher uses an extension to enlarge cognitive reach without surrendering agency. |
| III | Extension-Integrated Teaching Practice | From the pedagogical corollary, applied to students: the teacher's irreplaceable functions include integrating AI into instructional structure and cultivating, in students, the metacognitive and agentic capacities that enable them to be anchoring agents for their own AI-extended learning. |
| IV | Extension-Enabled Professional Development | From the pedagogical corollary, applied to the teacher's own formation: the teacher must grow through AI extension (reflection, transfer, accountability communication) while preserving the agentic tension necessary for genuine professional learning. |
| V | Extension Ethical Responsibility | From the ethical corollary: responsibility for AI outputs traces to the anchoring agent. Ethics in this frame covers purpose (what the agent extends toward), process (how the agent extends), and structure (the institutional and commercial arrangements within which anchoring takes place). |
| VI | Collective Agentic Practice | From Bandura's (2001) three modes of agency — personal, proxy, collective — and from the Type F (ambient) risk profile: some AI extensions cannot be anchored by individual teachers acting alone. Platform-level scrutiny, institutional escalation, and policy-level engagement require the exercise of collective agency, which is not reducible to aggregate individual agency and therefore warrants its own dimension. |
| VII | Equity-Centred Extension Practice | From taking HAA seriously as a universal claim: the axiom requires that every learner be supported as an anchoring agent for his or her own AI-extended learning. This requires attention to the differential conditions under which different agents can actually anchor — conditions shaped by language, socioeconomic background, disability, cultural familiarity with AI norms, and representation in training data. |
| VIII | Assessment Redesign Under AI | From the epistemological corollary under changed conditions: if the evaluative question of educational AI is whether human cognitive capacity has been enlarged, teachers need instruments that can discriminate extension-enlarged learning from AI-substituted output. Traditional assessment forms have, in many cases, become invalid for this discrimination. Rehabilitating assessment validity is therefore not incidental to the framework but required by it. |
Each dimension is further resolved into three sub-indicators, for a total of twenty-four. The three-by-eight partition is not sacred — other partitions are defensible — but it has three practical virtues: it aligns with the level of granularity common in teacher-competency frameworks internationally; it preserves Dimensions I–V from earlier iterations unchanged, allowing comparability across versions of the framework; and it gives the three HAA-derived extensions (VI–VIII) parallel structural weight rather than treating them as cross-cutting afterthoughts.
7The eight-dimension, twenty-four-sub-indicator framework
The tables that follow present the framework. For each dimension, a descriptor states the competency region; a table enumerates three sub-indicators, each with its characteristic extension type(s) and a specification of what the competency requires. The HAA-Type column reflects the six-mode typology introduced in Part 5. Dimension I contains a sub-indicator — Anchoring of Teacher Agency — whose name echoes the HAA axiom directly; this is intentional. The axiom and the framework are co-designed so that the axiom's organising commitment has a first-class position in the competency structure itself.
Dimensions I–V (Tables 7-1 through 7-5) address the individual teacher's agentic relationship to AI extension. Dimensions VI–VIII (Tables 7-6 through 7-8, marked ★ as new in v5) address competencies that the HAA axiom implies but that cannot be reduced to individual, context-general, or pre-AI educational practice: collective agency under platform-level extension, equity-differentiated anchoring capacity, and assessment validity under AI.
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Extension-System Understanding | Foundational across Types A–F | Understand AI as the materialised form of prior human agency and cognition; grasp the basic operating principles of generative AI, recommender systems, and intelligent agents; distinguish the nature and boundary conditions of the six extension types (task-execution, analytical, co-cognitive, generative, delegated, ambient); recognise that AI outputs are products of training on human knowledge rather than independent creations. |
| Recognition of Extension Boundaries | Foundational across Types A–F | Accurately identify the effective range and failure conditions of AI extension: technical limits (hallucination, bias), context-dependent limits (cultural understanding, relational construction), and value-judgement limits (setting pedagogical goals, ethical choice-making); recognise that these limits stem from how extension systems are constituted rather than from independent deficiencies of AI. |
| Anchoring of Teacher Agency | Especially critical in Type C | Amid pervasive AI extension in teaching, clearly recognise one's irreplaceable role as the anchoring agent of the extension system — setting extension goals, evaluating extension quality, and bearing extension responsibility; cultivate awareness of, and the disposition to actively resist, erosion risks such as cognitive outsourcing, emotional detachment, and weakening of agency. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Extension Design Capability | Core for Type A; foundational for B / C | Precisely translate pedagogical intentions into AI-executable task instructions: prompt design (task specification, role setting, constraints), task chaining, multi-turn interaction adjustment. What matters is not operational fluency but the clarity of pedagogical-intent expression — prompt engineering is, in essence, the externalisation of agentic goals. |
| Extension Evaluation Capability | General across Types A–F; especially important for C and E | Critically evaluate the quality, accuracy, and educational appropriateness of AI outputs: identify factual hallucinations, algorithmic bias, and over-confident output; judge whether an AI extension genuinely expands pedagogical judgement rather than merely providing convenience; within Type C co-creation, identify echo-chamber effects and refuse to mistake fluent AI output for shared insight; for Type E delegated extension, evaluate outputs at both policy and sampled-act levels, since no single output is fully representative of the underlying delegation. |
| Extension Customisation Capability | Core for Types B / C; central for Type D | Perform situational configuration and personalised adaptation of AI tools: system-prompt customisation, workflow orchestration, low-code tool configuration, and — where Type D extensions are used — authoring-time design of generated environments. The emphasis is on meaningful extension design aligned with pedagogical goals, so that the tool (or generated environment) adapts to the learning situation rather than vice versa. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Implementation of Human–AI Collaborative Teaching | Integrated use of Types A–F | Appropriately select across the six extension types (A–F) at each phase of instruction, using AI as scaffolding to optimise instructional structures, enable differentiated support, and provide immediate assessment feedback. The evaluative criterion is whether the extension has structurally changed the teaching process (rather than merely inserting an AI tool into it); actively regulate the boundary between AI and human judgement, maintaining agentic control when AI is involved in critical decisions — especially where Type E (delegated) or Type F (ambient) couplings place acts outside the teacher's direct supervision. |
| Guidance and Support for Students' Extension Use | Core scenario for Type C; critical for Type F | Guide students to understand AI as cognitive extension anchored by their own agency; cultivate students' metacognitive capacity to consciously use, evaluate, and regulate AI extensions — recognising applicability boundaries, critically evaluating output quality, maintaining agentic control, and guarding against cognitive outsourcing; develop students' capacity to detect ambient (Type F) extension in the platforms they use, so that agency can anchor what would otherwise remain invisible; create authentic opportunities for students to exercise extension control rather than merely completing AI-assisted tasks. |
| Application of Learning Analytics | Core capability for Type B | Read, interpret, and critically respond to learning-analytics data generated by AI systems (learning trajectories, knowledge diagnostics, early-warning signals); let data drive instructional improvement without letting data replace situational understanding of individual students; recognise the risk that aggregate data obscure individual differences, and maintain reflective balance between algorithmic recommendations and professional judgement. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| AI-Assisted Professional Reflection | Core scenario for Type C | Conduct pedagogical reflection, peer collaboration, and resource co-construction with the aid of AI extension; critically integrate AI insights with one's own professional judgement — accepting AI analyses that genuinely revise one's views while rejecting outputs that merely confirm pre-existing thinking; identify echo-chamber effects in Type C extension, ensuring that professional reflection preserves the agentic tension that makes reflection genuine. |
| Extension-Transfer Capability | General meta-competence across Types A–F | Against a backdrop of rapid tool iteration, rapidly assess the extension type of emerging AI tools — distinguishing task-execution, analytical, co-cognitive, generative, delegated, and ambient couplings — along with their boundary conditions and educational potential; transfer existing extension experience to new tools; possess the metacognitive capacity to independently judge whether a new tool is suitable for a given educational scenario, without relying on external evaluations or commercial claims. |
| Extension Accountability Communication | Applied across Types B / C; critical for Type E | Clearly explain to students, parents, and administrators the role AI plays in pedagogical decisions (extension type), the basis of those decisions (judgement process), and their limits (boundary conditions). Where extension is proxy / delegated (Type E), articulate the pre-authorised policy that produced an individual act and take agentic responsibility for the policy's consequences — making legible the chain of authorisation that was invisible to affected parties. Where Type B or Type C extensions have visible effects on individual students, articulate the logic of AI intervention intelligibly. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Ethics of Extension Purpose | High-risk zone for Type C | Uphold the position that the sole legitimate purpose of extension is to advance students' cognitive development and holistic agentic formation — not efficiency optimisation, workload reduction, or performance attainment; identify and resist drift in extension purpose toward cognitive outsourcing, emotional detachment, or weakening of agency; remain alert to tendencies to effectively outsource pedagogical decisions to AI systems under the guise of "co-creation." |
| Ethics of Extension Process | General across Types A–F | Actively identify and respond to specific risks in the extension process: algorithmic bias (human biases encoded and amplified through training data), content hallucination (a failure mode of the extension system), tacit disciplining (the extension system's unintended shaping of student cognition), and privacy erosion (risks in the flow of student data); recognise that the responsible agent is the teacher who chose the mode of extension. |
| Ethics of Extension Structure | Deep-level risk zone for Types B / C; primary site for Type F | Recognise the structural risks of commercial AI platforms: platform dependence and vendor lock-in, student data held by commercial entities, and potential conflicts between algorithmic interests and educational goals. Where extension is ambient (Type F), structural scrutiny becomes primary rather than supplementary — teachers often are not the users of the extension but are affected by it, making detection and institutional escalation the principal competencies. Individual teachers bear the professional responsibility to identify, name, and surface these risks to their institutions. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Professional-Community Scrutiny | Primary site for Type F; relevant to E | Participate actively in peer-teacher communities that collectively scrutinise AI platforms and tools; share evaluations and failure cases; contribute to and draw from shared professional vocabulary for naming what platforms are doing; build collective knowledge about extension types, risk profiles, and good practice that no individual teacher could build alone. Treat collective evaluation as a core professional obligation under conditions where individual evaluation is structurally incomplete. |
| Institutional Voice and Escalation | Core for Types E / F; applies across A–F | Identify AI-related risks at the institutional level and escalate through appropriate channels; participate in institutional decisions about AI adoption, procurement, policy, and governance; represent learner interests — especially those of learners not positioned to represent themselves — in deployment decisions; hold institutions accountable for aligning AI deployments with educational purposes rather than merely administrative convenience. |
| Policy-Level Agency | Cross-cutting, with primary salience for Type F | Contribute to local, national, or professional-association policy formation on AI in education; engage substantively with regulatory consultation where available; develop working familiarity with the policy landscape within which institutional and classroom decisions are constrained; understand the teacher's role as a political agent in shaping the conditions under which AI enters education, not merely as an end-user of conditions set by others. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Recognition of Differential Extension Effects | General across Types A–F | Identify how AI extensions differentially affect learners across lines of language, socioeconomic status, race, gender, disability, neurodivergence, and cultural background; recognise training-data representation gaps and their downstream pedagogical consequences; distinguish "all learners access this AI" from "all learners benefit equally from this AI," and treat the second as the relevant educational criterion. |
| Inclusive Extension Design | Central for Types A / C / D | Design and select AI extensions with attention to differential impact; provide alternatives, adjustments, and scaffolds where extensions systematically disadvantage particular learners; resist defaulting to the modal learner as implicit design target; for Type D generative extensions, examine generated environments for cultural, linguistic, and representational assumptions before deploying them with diverse learners. |
| Distributed Anchoring Capacity | Cross-cutting; primary for Type C | Actively cultivate the conditions under which all learners can be anchoring agents for their own AI-extended learning; address gaps in prior AI exposure, cultural familiarity with AI interaction norms, and confidence in exercising agency over AI tools; treat anchoring capacity itself as unevenly distributed and subject to pedagogical cultivation, rather than as a given background condition for AI use. |
| Sub-indicator | HAA Type | Competency content |
|---|---|---|
| Assessment Validity Under AI | General across Types A–F; acute for A and E | Recognise where traditional assessment forms (take-home essays, standard problem sets, report writing) have become invalid signals of the capacities they aim to measure under conditions where learners have access to Type A extensions; distinguish capacities that AI can demonstrate from capacities that require agentic exercise by the learner; treat assessment validity as a property to be actively maintained rather than assumed, revisiting instruments regularly as AI tools evolve. |
| Authentic Agentic Assessment | Core for Type C; general across A–F | Design assessment tasks that require and make visible the learner's own agentic engagement: oral examination, process portfolios, in-class writing, scaffolded real-time construction, iterative review with explained reasoning, think-aloud protocols. Treat agentic evidence — the visible exercise of the learner's own cognitive agency — as the primary assessment target, rather than product quality alone; design so that AI-substituted output is not merely discouraged but structurally non-evidential for the construct being assessed. |
| AI-Inclusive Assessment Policy | Cross-cutting; primary for A and C | Develop and communicate clear, educationally grounded policies on AI use in assessment that distinguish legitimate extension from prohibited substitution; work with students to co-articulate the reasons for the policies rather than imposing them as external rules; revise policies iteratively as tools and learner practice evolve; recognise that policy legitimacy depends on its alignment with the HAA purpose — supporting learners as anchoring agents — rather than on administrative enforceability alone. |
8Positioning relative to existing frameworks
Three frameworks supply the most likely points of comparison for the present proposal: UNESCO's AI Competency Framework for Teachers, the European Commission's DigComp 3.0, and the ISTE Standards for Educators. The HAA framework does not displace these; it occupies a different conceptual slot. Where they are structured to be broadly applicable and, in varying degrees, axiomatically agnostic, the HAA framework is structured to be axiomatic and domain-specific — and it is the axiom that does the work.
8.1UNESCO AI Competency Framework for Teachers (Miao & Cukurova, 2024)
UNESCO's framework, published in August 2024, is a close structural analogue for Dimensions I–V of the present framework: five aspects (human-centred mindset, ethics of AI, AI foundations, AI pedagogy, AI for professional development) crossed with three progression levels (Acquire, Deepen, Create), yielding fifteen competency blocks. The first five HAA dimensions preserve the five-aspect partition because it captures something real about the space of individual-teacher competencies and because superficial structural compatibility facilitates dialogue with a framework that shapes national policies.
Two differences are consequential. First, UNESCO's framework affirms human agency as a principle but does not commit to a specific ontology of AI in the educational setting; its progression levels (Acquire / Deepen / Create) are skill-maturity levels, not modes of human–AI cognitive coupling. The HAA framework commits to an ontology — AI as extension, agency as anchored — and derives its dimensions from that commitment. Second, the HAA framework is now broader than UNESCO's: Dimensions VI–VIII (collective agency, equity-centred practice, assessment redesign) cover ground that UNESCO either treats as cross-cutting values or does not address at first-class. This breadth reflects the argument that HAA, taken seriously, entails collective and distributive competencies alongside the individual-teacher competencies UNESCO addresses.
8.2DigComp 3.0 (Cosgrove & Cachia, 2025)
DigComp 3.0, published by the European Commission's Joint Research Centre in November 2025, is the fifth edition of the European Digital Competence Framework. Its scope is general digital competence for citizens; it retains the 21-competence, 5-area, 4-proficiency-level structure of previous editions and integrates AI systematically across all competences rather than as a separate area. It supplies over 500 learning outcomes.
DigComp 3.0 and the HAA framework operate at different levels. DigComp 3.0 describes the digital competence a citizen needs across a life-span; HAA describes the competence an educator needs for the specific activity of educating. Two consequential differences follow. First, DigComp 3.0 is axiomatically agnostic: it describes what digital-and-AI competence looks like without committing to what AI is in the relevant setting. HAA makes the commitment and reasons from it. Second, DigComp 3.0's ethical content is primarily citizen-facing (rights, protection, responsible use); HAA's structural-ethics layer — platform dependence, data sovereignty, alignment between algorithmic interests and educational goals — does not fit naturally within a citizen-competence frame.
8.3ISTE Standards for Educators (ISTE, 2024, v4.02)
The ISTE Standards for Educators are organised by role: Learner, Leader, Citizen, Collaborator, Designer, Facilitator, Analyst. ISTE's 2024 revision took an incremental approach to AI: rather than add AI-specific standards, the organisation updated select indicators across the role-based standards to cover the use of AI safely, responsibly, and innovatively.
The contrast with HAA is sharper than with the other two. ISTE carves the competency space by the role a teacher plays in a given activity; HAA carves it by the mode of cognitive coupling between teacher and AI and by the anchoring of agency. The two structures cross-cut: a single role (say, Facilitator) will involve multiple extension types, and a single extension type (say, Type C) will cross multiple roles. Both carvings can be defended; the case for the HAA carving is that, in a domain where the AI's role depends on the nature of cognitive coupling and on where agency is anchored, these are the better dimensions along which to organise teacher competence.
8.4Summary comparison
| Framework | Organising principle | Principal differences from HAA |
|---|---|---|
| UNESCO AI CFT (2024) | Five aspects × three progression levels (Acquire / Deepen / Create); "human-centred" values. | Skill-maturity progression rather than modes of cognitive coupling; "human-centred" affirmed as principle rather than derived from an ontology; HAA covers collective agency, equity-differentiated anchoring, and assessment redesign at first-class, which UNESCO treats as cross-cutting or not at all. |
| DigComp 3.0 (2025) | Twenty-one competences across five areas; four proficiency levels; AI integrated transversally. | General citizen digital competence, not education-specific; axiomatically agnostic about AI's role; citizen-facing ethics rather than institutional-structural ethics. |
| ISTE Standards for Educators (2024 v4.02) | Seven role-based standards (Learner, Leader, Citizen, Collaborator, Designer, Facilitator, Analyst); incremental integration of AI across roles. | Carves the space by teacher role rather than by mode of cognitive coupling and agency anchoring; AI added incrementally rather than as organising constraint; limited structural-ethics treatment. |
The relationship is best understood as complementarity rather than rivalry. National and international bodies will likely continue to work with UNESCO, DigComp, and ISTE as reference frameworks; HAA supplies an axiomatic grounding that any of them could, in principle, adopt as an interpretive layer.
9Limitations and open questions
Six limitations deserve explicit statement, each of which identifies a direction for subsequent work.
- Behavioural anchors are absent. All twenty-four sub-indicators are currently stated in "able to" form. A four-level set of behavioural anchors (Novice / Developing / Proficient / Distinguished) should be developed for each sub-indicator, with concrete descriptors at each level. Until this is done, the framework is better suited to orienting teacher education and policy than to formal assessment.
- The expansion from five to eight dimensions is argued but not empirically validated. Earlier iterations of this framework used a five-dimension partition (I–V). Dimensions VI–VIII — collective agency, equity-centred practice, and assessment redesign — are introduced here on theoretical grounds derived from HAA and from the Type A–F typology. Whether the eight-dimension structure produces better assessment discrimination or teacher-education outcomes than the five-dimension structure is an empirical question for future work. Practitioners and institutions adopting the framework may find that a subset of dimensions suits their immediate purposes.
- The A–F typology awaits empirical validation. Three studies are needed: an expert-agreement study testing inter-rater reliability when coding AI use into the six types; a teacher-practice study testing whether the distinctions cohere with practitioners' own phenomenology of AI use; and, given the conceptual distance of Types D–F from the original three, a theoretical-coherence study examining whether the six types genuinely partition the extension space or whether sub-types should be collapsed or further split. The typology's extension from three types to six is motivated but not itself validated by practitioner research.
- Extension-transfer capability is under-specified. The meta-competence of transferring extension experience to newly arriving tools is currently specified at a level too abstract for reliable operationalisation. Concrete case studies of tool-to-tool transfer are needed to sharpen the construct.
- The concept of agency is philosophically contested. HAA rests on a particular theoretical account of agency (Bandura, 2001; broadly consistent with philosophy-of-action orthodoxy). Those working from eliminativist, radically situated, or post-humanist accounts of agency will find the axiom less compelling. The framework should be explicit about the theoretical commitment whenever deployed; and agency as an empirical construct — specifically, how it is shaped by extended AI interaction — is itself a productive topic for research.
- Cultural specificity versus international comparability. The sub-indicator "Anchoring of Teacher Agency" aligns closely with the Chinese teacher-as-moral-exemplar tradition (师道), which supplies a particularly deep version of the purposive-agent commitment. A globalised version of the framework should provide explicit cross-cultural translation of this positioning; whether the sub-indicator should be re-written or simply footnoted is a matter of subsequent deliberation.
- The constitutive-difference argument depends on a specific reading of education's telos. If one holds that education's primary function is credentialing rather than agentic formation, the constitutive-difference argument weakens considerably; credentialing is closer to a manufacturing-style telos (a document is produced) than to a formation-style telos. The framework as a whole presupposes the formation reading. Those who hold competing views of education's purpose will find the framework correspondingly less compelling, and this presupposition should be made explicit whenever HAA is deployed.
10Conclusion
The argument of this paper is that AI in education is not a local instance of a general pattern but a domain whose structure forces a particular framing. Because education's object is an agent in formation, no deployment of AI that displaces that agent's own purposive engagement can be said to have served the activity. Extension of cognition is permitted; displacement of agency is not. This is the Human Agency Anchoring axiom.
From the axiom, four corollaries, a six-mode typology of extension, and an eight-dimension competency framework follow. The framework's practical aim is to orient teacher education, institutional policy, and professional self-evaluation toward the question that the axiom makes central: is the human — teacher and learner, individually and collectively — coming out of this with enlarged cognitive reach and intact agency, across all learners rather than only the most advantaged? Where the answer is yes, the extension has done its work. Where the answer is no, the presence of sophisticated technology is not a ground for reassurance but a warning.
The framework is offered as an axiomatic contribution to a space populated by pragmatic and descriptive frameworks. Its value, if any, lies in being wrong precisely rather than broadly: in specifying a commitment that can be disagreed with, refined, or overturned by argument, rather than in stating what everyone already grants. The invitation of the paper is accordingly both to adoption and to productive disagreement.