What Remains Self-Directed? Revisiting Andragogy Through Cognitive Delegation in Generative AI-Mediated Adult Learning

Synthesis: Hyoung (2026) revisits Knowles's six andragogical assumptions under AI-mediated cognitive delegation. Because generative AI can participate in identifying needs, setting goals, interpreting information, solving problems, producing outputs, and evaluating performance, the paper argues that behavioral independence is insufficient evidence of meaningful self-direction. Five analytical dimensions are derived — need and goal ownership, delegation control, epistemic calibration, cognitive recoverability and transfer, and motivational autonomy — to assess whether learners remain genuinely self-directed. The work bridges Adult Learners with Cognitive Offloading, Self-Regulated Learning, and appropriate-reliance research for Adult Learners in the Generative AI era.

Key Findings

1. Behavioral independence is not evidence of meaningful Self-Directed Learning: because Generative AI can now participate in identifying needs, setting goals, interpreting information, solving problems, producing outputs, and evaluating performance, a learner can appear fully self-directed while transferring substantial cognitive functions to an external system.
2. Reanalyzing Knowles's six andragogical assumptions yields five cross-cutting analytical dimensions — need and goal ownership, delegation control, epistemic calibration, cognitive recoverability and transfer, and motivational autonomy — as conditions for evaluating whether self-direction remains meaningfully governed by the learner.
3. The amount of AI use is a contextual characteristic, not a sixth dimension: dependent and autonomous Cognitive Offloading can deliver similar immediate benefits while exhibiting different downstream correlates, so self-direction must be read as a configuration rather than a single continuum.
4. Delegation is not equivalent to dependence: functional delegation can coexist with retained recoverability, and the genuine opposite of self-direction is a progressive loss of authority over learning, not technological reliance.

Andragogy Meets Cognitive Delegation

Self-direction has long been central to Adult Learners and Lifelong Learning. Within Knowles's andragogical tradition, adults are understood as learners who increasingly assume responsibility for their decisions, bring accumulated experience as a resource, and orient learning toward practical problems. Yet Generative AI can now perform substantive parts of the learning process that andragogy traditionally assigned to the learner. The paper's central claim is non-deficit: humans have always relied on external resources, and hybrid human–AI capability may be valuable in its own right. The problem is therefore not simply whether adults use AI, but how learning-relevant functions and authority are distributed across the learner–AI relationship. This matters because independence from teachers or institutions need not imply cognitive independence — an adult can independently choose when and where to learn while allowing AI to recommend what to learn, sequence the material, interpret evidence, and judge adequacy.

From Assistance to Cognitive Delegation

The paper distinguishes assistance, Cognitive Offloading, extended cognition, and delegation rather than arranging them along a simple continuum of increasing AI use. Assistance is a descriptive baseline in which the learner retains the core interpretive or decisional function. Cognitive offloading more specifically refers to actions that alter a task's information-processing requirements to reduce internal cognitive demand (Risko & Gilbert, 2016); externalizing memory, calculation, or search is not unique to AI and is not inherently evidence of reduced learning. Delegation becomes analytically relevant when a substantive cognitive function is assigned to a system that can perform it on the learner's behalf. The extended-mind and Distributed Cognition perspectives (Clark & Chalmers, 1998; Hutchins, 1995) provide a counterweight to any deficit model in which externally supported cognition is treated as inferior. Trust Calibration adds an epistemic dimension, drawing on the human-factors finding that the design goal for automation is appropriate reliance rather than maximum trust (Lee & See, 2004).

Revisiting the Six Assumptions

Knowles's six andragogical assumptions — the need to know, self-concept, prior experience, readiness to learn, orientation to learning, and motivation — each shift under AI-mediated cognitive delegation. The paper maps each to a derived analytical dimension rather than replacing them one-for-one. GenAI can diagnose gaps and construct personalized plans, extending the need to know toward need and goal ownership. An adult can independently initiate and manage AI-supported study while delegating interpretation or judgment, so self-concept extends toward delegation control. Prior experience becomes an epistemic resource requiring calibration rather than merely accumulated capital. Readiness to learn can reflect pressured adaptation to structural change, mapping onto both need and goal ownership and motivational autonomy. Problem-centered orientation becomes vulnerable to confusing task success with capability formation, given evidence that AI-supported performance need not yield equivalent learning gains. And motivation must be read through volitional endorsement rather than behavioral initiative alone, per Self-Determination Theory (Ryan & Deci, 2000).

Five Dimensions of Self-Direction

The paper integrates Adult Learners theory with Cognitive Offloading, extended and distributed cognition, human–automation function allocation, agentic information-system delegation, and appropriate reliance, deriving five analytical dimensions that operationalize meaningful self-direction under delegation. Need and goal ownership concerns whether learners understand, critically endorse, and retain authority over why and what they learn. Delegation control concerns which functions are assigned to AI, which are retained, and when the allocation is revised — it is not simply low AI use. Epistemic calibration is the capacity to align reliance on AI with evidential adequacy, system limitations, task demands, and stakes. Cognitive recoverability and transfer concern what capability remains after delegation: whether the learner can reconstruct reasoning and use it in new contexts, consistent with Distributed Cognition and Transfer of Learning research. Motivational autonomy is the extent to which learners meaningfully endorse the reasons for learning, even under organizational or labor-market pressure. Self-direction is thus a configuration rather than a single continuum, and amount of AI use is a contextual characteristic rather than a sixth dimension.

What this means for practice

• Educators. Evaluate the distribution of cognitive work, not just the artifact: state explicitly which capability an assessment is inferring, because an AI-assisted report shows what a learner–AI configuration can do and little about what the individual understands.
• Educators. Teach the governance of delegation, not only instrumental AI Literacy: prompt writing and tool selection are insufficient without knowing what should be delegated, how delegated activity should be judged, and what capability must remain.
• Educators. Design conditions in which learners learn to govern their own cognitive participation — Scaffolding, hints, and structured verification and reconstruction — instead of supplying explanations and examples directly.
• Researchers. Treat the five dimensions as testable propositions: compare critically endorsed AI-generated goals with unreformulated ones, hold AI capability constant across learner-controlled and system-led allocation, and test whether task-specific epistemic calibration outperforms generalized trust.
• Administrators. Pair any requirement to keep learning with institutional support: the framework's motivational-autonomy dimension raises the question of what conditions make adults continuously self-directing, and what support accompanies that responsibility.

Limitations

• This is a conceptual paper with no primary data: the five analytical dimensions are derived by reanalyzing Knowles's six andragogical assumptions against the literature, so the framework is a set of claims awaiting test, as its own falsifiability statement concedes.
• Its strongest empirical anchor is a single study — Bassner et al.'s three-arm randomized controlled trial with 275 introductory-programming students — which the paper itself says is context-specific and "should not be generalized to all adult learning."
• The framework's applicability is explicitly gated (task consequentiality, learner discretion), and where goals, methods and AI functions are entirely imposed by an institution it has, in the author's words, reduced explanatory leverage for self-direction.
• The paper is an EdArXiv preprint and has not been peer reviewed.

Citation

Hyoung, J. J. (2026). What remains self-directed? Revisiting andragogy through cognitive delegation in generative AI-mediated adult learning. EdArXiv preprint.