Epistemic Dependence in AI-Mediated Learning

Synthesis: Du and Yuan (2026) argue that the central educational question is not whether learners rely on AI but whether that reliance preserves or displaces the epistemic work through which judgment develops. A critical-integrative review bridging AIED, human–AI interaction, epistemic cognition, information behavior, cognitive offloading, and social epistemology, it distinguishes instrumental assistance (AI helps produce output) from judgment-bearing assistance (AI supplies the standards by which output is assessed), and separates productive reliance from harmful dependence through six diagnostic criteria: contestability, recoverability, transfer, traceability, distributed responsibility, and epistemic plurality. Four sociotechnical pathways — fluent authority, frictionless delegation, opaque synthesis, and institutionalized dependence — link AI affordances to educational consequences, and the review proposes relational epistemic agency as the normative aim of AI-mediated learning, with design, Pedagogies and Teaching Strategies, assessment, and governance implications.

From AI use to AI-mediated knowing

Generative AI has shifted from a specialist concern within AIED to a general condition of learning, with conversational systems that combine brainstorming, explanation, coding, revision, feedback, and research support in a single interface. The review argues that AI now functions as an epistemic intermediary, not merely a tool: it selects, compresses, reformulates, and evaluates knowledge-like claims while presenting the result as a direct answer. Unlike conventional search — which exposes multiple documents, authors, and domains — conversational synthesis can hide that plurality behind one voice. The learner's task therefore changes from selecting among visible sources to recovering the evidential structure of an answer that appears already integrated, making the epistemic division of labor harder to inspect.

Most educational discussion has been organized around adoption, effectiveness, academic integrity, policy, ethics, or AI literacy. These frames treat AI as an intervention, productivity aid, or compliance problem. The review instead foregrounds the epistemic relation: learners must decide not only whether to use AI, but when its output is credible, whose standards it embodies, and where responsibility for accepting it should lie.

Productive reliance vs. harmful dependence

A central move of the review is to reject the equation of dependence with failure. Learning has always depended on teachers, peers, texts, and institutions; external support can extend cognition, widen access, and reduce unnecessary load. Dependence becomes problematic when the relation is difficult to contest, obscures its evidential basis, weakens the learner's capacity to reconstruct or transfer judgment, or allocates responsibility to people who lack meaningful control. Conversely, frequent AI use can remain productive when it provokes comparison, makes uncertainty visible, and leaves the learner more capable of independent and collaborative judgment.

The paper's key analytic pivot distinguishes two forms of assistance:

• Instrumental / representational assistance helps the learner generate options, translate, format, retrieve, summarize, or re-express material.
• Judgment-bearing assistance evaluates correctness, relevance, quality, persuasiveness, or evidential sufficiency — delegating not only production but the standards by which production is assessed.

Because a summary already selects what matters, and brainstorming shapes the space of plausible ideas, the categories can overlap — but judgment-bearing assistance is normatively more demanding and is where dependence becomes educationally consequential.

The review operationalizes the productive/harmful boundary through six diagnostic criteria (Table 2): contestability (can the learner question, reject, or revise output?), recoverability (can the learner reconstruct the reasoning with reduced support?), transfer (does capability persist when the system is absent or the task changes?), traceability (can claims be connected to inspectable sources?), distributed responsibility (do responsibilities track control and knowledge across learners, teachers, institutions, and providers?), and epistemic plurality (does the interaction expose disagreement, disciplinary standards, and marginalized perspectives?). No single criterion is decisive; the pattern matters, as does the importance of the delegated judgment. These criteria are diagnostic questions for research, design, and pedagogy, not a psychometric scale.

Four pathways from affordances to consequences

A pathway is not a deterministic causal chain but a plausible relation to investigate: an affordance or institutional condition changes how epistemic work is distributed, and the resulting relation is assessed through the six criteria. The four pathways are:

• Fluent authority and personalized social presence. Polished, confident output is epistemically persuasive — users mistake ease of processing for reliability. Anthropomorphism, social presence, and personalization can transfer interpersonal trust to a system lacking human understanding, responsibility, and commitment. The risk is misplaced epistemic authority and normative displacement: learners may ask AI not only what is true but what counts as a strong argument or an ethical response. Productive reliance requires dialogue designed for contestation.
• Frictionless delegation and reduced epistemic work. Compressing search, reading, comparison, and drafting into a single request–response cycle can remove the intermediate epistemic actions a task was intended to develop. Assessment incentives intensify the pathway: if institutions reward polished products while leaving process invisible, delegating epistemic work is rational. Automated feedback supports learning when it becomes material for judgment; it displaces learning when its verdict replaces that judgment.
• Opaque synthesis and hidden evidential structure. Opacity is not only a property of model internals but an educational condition in which the learner cannot see how an answer relates to evidence. Opaque synthesis can weaken verification while improving surface quality, and a generic synthesis can flatten disciplinary standards for what counts as evidence and warranted inference.
• Institutionalized dependence and distributed authority. Universities shape dependence through procurement, platform integration, and assessment design. Once AI is built into learning management systems and feedback workflows, it becomes the default route into academic work rather than a discrete tool. This pathway makes epistemic justice central: generative systems can reproduce dominant languages and classifications while marginalizing local, minoritised, or experiential knowledge.

Relational epistemic agency as the normative aim

The review's normative core is relational epistemic agency: the capacity to question, verify, compare, justify, and know responsibly within relations of interdependence. The value at stake is not independence from others or tools — ideals of self-sufficient knowing are unrealistic and exclusionary — but whether the human–technology relation preserves meaningful epistemic participation. AI systems may shape and extend reasoning without possessing reciprocal responsibility or legitimate authority; the norm is compatible with substantial AI use and rejects both technological solutionism and a simple anti-dependence position.

Responsibility and justice are integral to this account. Learners should take responsibility for claims they submit, but responsibility is fair only when they have sufficient knowledge and control. A relational account replaces the fiction of the isolated user with distributed responsibility, and it incorporates formative epistemic injustice — learning arrangements that wrong students by restricting the knowledge, practice, and accurate self-assessment through which they develop as knowers. This connects to calibrated trust, human oversight, and the wider question of how Learner Agency is preserved when AI systems act with increasing autonomy.

Connections to the knowledge base

This review converges with several existing threads. It resonates with Poudyal's (2026) Ecological Co-Agency Framework, whose boundary condition of human epistemic accountability parallels the review's relational account; with the PEARLS framework for verifying AI output; and with the distinction between instrumental and epistemic assistance underlying Shaw & Nave's (2026) cognitive surrender. Its six diagnostic criteria offer a more granular vocabulary than "over-reliance" for judging when offloading crosses from strategic support into learning displacement, complementing the Cognitive Offloading concept's treatment of over-reliance. Its treatment of AI feedback as material for judgment connects to AI Feedback Quality and Feedback Literacy.

What this means for practice

• Instructors. Separate instrumental from judgment-bearing assistance when you set task rules — permit AI for generating options, translating, formatting, retrieving, and summarizing, but require learners to supply the standards by which correctness, relevance, and evidential sufficiency are judged — and make that epistemic work visible in the assignment itself: annotate an AI response, identify its unsupported assumptions, document why a suggestion was accepted or rejected, and assess through staged drafts, source maps, oral defense, or reflective decision logs rather than the final product alone.
• Instructors. Test recoverability by withdrawing the system: if a learner can reconstruct and perform the relevant reasoning with reduced support, the reliance was productive; collapse when support is removed is the diagnostic signal of harmful dependence.
• Researchers. Ask which epistemic actions are preserved, transformed, or displaced rather than whether AI helps or harms on average, and use withdrawal and transfer designs to test whether learners can explain and perform the reasoning when support is reduced.
• Researchers. Disaggregate outcomes by prior knowledge, language, disability, socioeconomic context, and discipline rather than reporting aggregate AI effects, since the same affordance may scaffold one learner while displacing another's judgment.
• Designers and governance. Build systems for contestability rather than fluency — claim-level provenance, uncertainty shown through alternatives and explicit unknowns, and user control to revise prompts, compare outputs, and reject recommendations — and have procurement and policy evaluate provenance, accessibility, bias, data governance, and auditability, distinguish legitimate reliance from prohibited substitution, and allocate responsibility across stakeholders rather than placing every verification burden on individual learners.

Limitations

• The authors state it is a purposive critical-integrative review, not a systematic review or meta-analysis, so it claims neither exhaustive coverage nor prevalence estimates; the documented search (last updated July 2026) improves auditability but is not equivalent to systematic screening.
• Direct evidence on long-term epistemic development is limited, so the four pathways are presented as theoretically grounded hypotheses rather than established causal effects, with mechanisms sometimes rendered intelligible from evidence in Internet search, smartphones, medicine, or aviation.
• The field moves faster than the search cutoff, and several cited works are preprints, editorials, or conference papers whose status and arguments may change after publication.
• The framework abstracts across systems and disciplines that differ in retrieval, provenance, personalization, and governance, and no datasets were generated or analyzed: the six criteria are diagnostic questions, not a validated psychometric scale.

Citation

Du, Y., & Yuan, Y. (2026). Epistemic dependence in AI-mediated learning. AI & Society.