{"resourceId":"du-yuan-epistemic-dependence-review-2026","versions":[{"version":"external-fdbf3eb2e7843c09d9919ef450db4d58a27be4310a16aedc20b79e69b011eda7","resource":{"id":"du-yuan-epistemic-dependence-review-2026","title":"Critical review offers questions for preserving learner judgment, not a validated dependency scale","organization":"Yiran Du and Yijia Yuan, University of Cambridge","sector":"Higher education teaching and learning","geography":"UK-authored review drawing on international literature; no study population","publishedAt":"August 29, 2026","publicationDate":"2026-08-29","eventDate":null,"sourceName":"Epistemic dependence in AI-mediated learning","sourceLabel":"Original critical-integrative academic review; conceptual and normative evidence","sourceUrl":"https://link.springer.com/article/10.1007/s00146-026-03294-1","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["knowledge-work","data-security","governance-procurement","accessibility-workforce","operating-model"],"finding":"The review distinguishes useful assistance from delegation that displaces learner judgment; it does not establish that frequent AI use causes harm.","sledRelevance":"Interpretation: Useful for U.S. college assessment design discussions, with no local prevalence or causal effect to transfer.","evidence":"Purposive synthesis of 60 works, searches updated July 2026. No new dataset, experimental baseline or effect estimate. Proposed diagnostic criteria require validation.","architectureImplications":"Interpretation: Pilot traceable evidence links and a student-visible route to challenge feedback. Test retrieval permissions and provenance separately from model fluency. No cloud, on-premises or hybrid architecture is validated; autonomous agents receive no deployment endorsement.","governanceImplications":"Interpretation: Define which judgments a course expects students to perform, then make permitted assistance explicit. Evaluate policy effects before attaching penalties to a new rubric.","securityPrivacyImplications":"Interpretation: Use minimal assessment artifacts rather than comprehensive surveillance. Protect personal reflections and accommodation records from unnecessary model processing.","caveats":"Nonexhaustive conceptual review; hypotheses from adjacent domains are not demonstrated educational effects. Detailed table endpoint failed, so no table-only claims are used; main-text methods and limitations were accessible.","streamIds":["student-success"],"roles":{"sales":"Interpretation — Academic leaders may struggle to distinguish legitimate assistance from work that conceals missing understanding. Convene faculty, students, disability services, librarians and assessment leads. Ask which learning outcomes require an explanation, whether students can challenge automated feedback and which current practices already support that skill. A bounded engagement could review one course's AI-assisted assessment and develop a pilot rubric. The value hypothesis is clearer and more defensible assessment, to be tested locally. This conceptual review is a discussion aid rather than evidence of institutional harm. Do not claim that high usage identifies dependence, that a new rubric is validated, or that restricting tools necessarily improves equity.","engineering":"Interpretation — Fit the proposal to a course workflow in which students can inspect evidence and record a reasoned response to feedback. Prerequisites include accessible source materials, clear permissions and faculty-approved examples. Build a small test collection containing unsupported claims, conflicting sources and ambiguous feedback; verify that citations resolve and access controls hold. Proposed proof of value: students can challenge a wrong suggestion and complete a separate explanation task using agreed accommodations. Record unresolved cases without creating a permanent behavioral profile. The review cannot select a model or prove interface changes cause learning gains. Developer copilots may be examined with the same task-specific questions, but production agents require separate assurance.","delivery":"Interpretation — An assessment lead should own the pilot, with faculty judging disciplinary reasoning and disability staff reviewing equivalent participation. Map permitted assistance, prepare student examples and train markers before the assignment opens. Dependencies include moderation time, approved data handling and a clear appeal route. Proposed acceptance criteria: all sampled marks can be explained against published criteria; students can access a challenge route; agreed accommodations work; and a follow-up task is reported with missing responses. These are proposed service and evaluation gates, not observed effects. Risks include inconsistent marking, extra student burden and turning a conceptual framework into an unsupported misconduct detector. Review feedback before extending the approach."},"retrievedAt":"2026-09-11T03:01:47Z","enrichedAt":"2026-09-11T03:03:23Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Interpretation: Preserve assistive technologies when testing independent judgment, and compensate staff for rubric development and moderation.","procurementImplications":"Interpretation: Request demonstrations of citation inspection, student challenge routes, configuration control and export before committing to a platform.","operatingModelImplications":"Interpretation: Place responsibility for educational decisions with course owners rather than expecting students alone to detect every system error.","updateExplanation":"New in all 182 archive records and identifier-specific search. August 29 review is newly archived evidence backfill, not a new September 10 event.","sourceVerification":{"openedUrl":"https://link.springer.com/article/10.1007/s00146-026-03294-1","referenceExcerpt":"No datasets were generated or analysed during the current study.","promptVersion":"sled-research-v3.1","model":null,"basis":"agent-reported inspection"}}}]}