Reclaiming Epistemic Agency: A Critical Framework for Human-Generative AI Co-Agency in Education

Synthesis: Poudyal (2026) argues that generative AI in education does more than add a new tool: it reassigns epistemological authority from teachers to students to machines, redistributing Learner Agency across humans and non-human systems in a process the paper terms co-agency. After evaluating Distributed Agency, Self-Determination Theory, Society 5.0, and integration frameworks such as Technological Pedagogical Content Knowledge (TPACK) and SAMR — none of which address equitable power, data ownership, or accountability — the paper introduces the Ecological Co-Agency Framework, defining agency through relational, regulatory, and pedagogical processes bounded by a non-negotiable condition of human epistemic accountability. The framework offers educators and policymakers a more precise vocabulary than "balance" for deciding how tasks, evaluation, and responsibility are distributed between learners and AI in higher education.

From tool to co-agent: the epistemic shift

Generative AI has largely been framed as an impartial educational tool, and most empirical research has examined its effects on student interest, teacher workload, and assessment validity. Poudyal argues this framing overlooks a deeper shift: the reassignment of who produces knowledge, validates claims, and creates evidence of learning. Because GenAI can generate content, evaluate its own output, and adapt without direct human involvement, it participates in the evaluation process and thereby acts within the context of the Learner Agency it represents. This redistribution of agency from human to non-human entities is what the paper calls co-agency (drawing on Godwin-Jones, 2024). The epistemic concern is not that GenAI may produce unreliable content, but that its ease of use may obscure whose judgment — human or machine — ultimately underlies an assertion, determination, or product.

Why existing frameworks fall short

The literature review evaluates four theoretical traditions that have been applied to AI-mediated agency:

• Distributed Agency and social-material views treat competency, Creativity, and judgment as products of relationships among student, tool, and context — yet do not address who owns the interaction data or who is accountable.
• Self-Determination Theory (Ryan & Deci, 2000) identifies autonomy, competence, and relatedness as drivers of Motivation; GenAI changes the circumstances under which these occur but does not resolve equity of power.
• Society 5.0 offers a "human-centered" model of digital–physical symbiosis, but Harari's caution that GenAI autonomously creates and assesses knowledge claims warns against assuming its equilibrium will be automatic.
• Technological Pedagogical Content Knowledge (TPACK) and SAMR integration frameworks (including contextual-knowledge extensions and substitution-level adoption critiques) focus on teacher knowledge and adoption levels rather than on power, data management, and ethical responsibility.

None of these treats GenAI integration as a primary issue of Learner Agency or articulates the ethically justifiable conditions under which the dimensions can be combined — the gap the framework fills.

The Ecological Co-Agency Framework

The framework treats co-agency as the relationship between three interdependent, co-constituted dimensions, all bounded by a non-negotiable condition of human epistemic accountability:

• Relational co-agency defines agency as resulting from interaction rather than residing in person or machine. It requires (1) a transparent description of which tasks are assigned to AI and which to students, and (2) ethical co-agency — a joint responsibility for an educational outcome in which humans retain primary accountability and act in a monitoring capacity.
• Regulatory co-agency maps Zimmerman's three phases of self-regulated learning (forethought, performance, reflection) onto AI-mediated environments. GenAI can assist with objectives and sources in forethought, offer feedback and prompts during performance (what Dede terms "augmented intelligence" or Scaffolding), and produce progress summaries in reflection — but the point in the SRL cycle at which GenAI enters by default determines whether it amplifies or erodes the learner's sense of control. The paper notes that strategic Cognitive Offloading can support transformative learning when the offloading decision is intentional rather than routine.
• Pedagogical co-agency places the teacher at the center of integration, positioned along a continuum from observer to adopter to collaborator to innovator (Zhai, 2025) and shaped by institutional support in time, training, and technology. Without structural support, the continuum becomes less a developmental tool and more a means of identifying teachers who already possess resources to experiment. Teachers report being assigned substantial invisible labor reviewing, editing, and rewriting AI-generated materials — an example of professional judgment exercised precisely where AI cannot.

The ethical boundary condition constrains all three dimensions through three requirements: (1) contestability — learners and teachers must be able to question and cross-check AI output, requiring some degree of explainability; (2) provenance — institutions should describe the sources of training data and how they introduce bias; and (3) non-delegation of moral and intellectual credit — decisions about student welfare, academic standing, and high-stakes judgments must not be determined solely by GenAI.

Redesigning assessment and governance

The framework's concrete implications reshape both practice and governance. In Assessment, instructors move from evaluating a completed essay as a product toward evaluating the "decision trail" visible throughout the writing process — which tasks were delegated, how suggestions were accepted, modified, or rejected. In institutional procurement, transparency requirements turn purchasing into an exercise of epistemological governance, a demand that is particularly difficult for low-income districts already facing a digital divide and uneven AI literacy, so GenAI may widen the gaps it intends to close. For policymakers, the framework translates vague principles of transparency, fairness, and oversight into specific questions about division of labor, placement of GenAI in the learning cycle, institutional investment in teachers, and whether learners are permitted to challenge AI outputs — connecting to Equity, AI Governance, and the philosophy of ethical AI design.

What this means for practice

• Instructors. Grade the decision trail rather than the finished product — which tasks were delegated to GenAI and how suggestions were accepted, modified, or rejected — so that the locus of authorship stays visible in the Assessment.
• Instructors. Choose deliberately where GenAI enters the SRL cycle: the point of entry during forethought, performance, or reflection determines whether the tool amplifies or erodes the learner's sense of control.
• Faculty developers. Read the observer-to-innovator continuum as a development path rather than a ranking of teachers, and budget time and training for the invisible labor of reviewing, editing, and rewriting AI-generated materials.
• Administrators. Turn procurement into epistemological governance by requiring vendors to describe training-data sources and how they introduce bias before adoption, and keep decisions about student welfare and academic standing out of machine-only decision paths.
• Researchers. Take up the framework's measurable claims, since the authors state that no empirical evidence yet links its three dimensions to student achievement or perceived autonomy.

Limitations

• This is a theoretical paper: the authors state there is still no empirical evidence on whether the Ecological Co-Agency Framework improves student achievement, equity among students, or the time teachers spend on their jobs.
• The literature review was a targeted rather than systematic search, so the authors acknowledge they may have missed articles written in other languages, in regional journals, or in related fields.
• The framework sets conditions without resolving the tensions it identifies: the authors state it does not reconcile the pedagogical utility of AI with the problems raised throughout the paper, and offers a basis for deliberation rather than a tested remedy.
• Its equity claims are argued rather than measured — the potentially inequitable impacts for under-resourced institutions are named as a hypothesis the authors propose testing comparatively across socio-economic contexts.

Citation

Poudyal, B. (2026). Reclaiming epistemic agency: A critical framework for human-generative AI co-agency in education. arXiv:2608.26937.