Yiran Du

1. Generative artificial intelligence (AI) is increasingly used to interpret relationships. Empirical work evaluates large language models as sources of relationship advice (Hou et al., 2024), and ...

Despite the growing use of educational technologies, including AI and generative AI, how these technologies support learners’ metacognition remains insufficiently understood. The present study synthesised 33 studies using a three-level meta-analysis. Technology-supported interventions were associated with higher metacognitive outcomes on average, Hedges’ g = 0.653, 95% CI [0.520, 0.786], although heterogeneity was substantial. Technology type did not significantly moderate effects. Exploratory analyses showed descriptively larger estimates for interventions with technology-use guidance, immediate feedback, and system-directed learner control, but none remained significant after BH-FDR adjustment (adjusted p = .095 for each). Because these features overlapped, the patterns are hypothesis-generating rather than evidence of independent moderator effects. Interventions most frequently supported evaluation/reflection, monitoring, and strategic control, whereas planning was less consistently represented. Importantly, 77 of 94 effects (81.9%) concerned metacognitive regulation, and 90 (95.7%) relied on self-report measures, indicating that the pooled association primarily reflects self-reported metacognitive regulation. Overall, the findings suggest that technology-supported interventions can support metacognitive outcomes, while these effects should be interpreted in light of the specific regulatory processes targeted and how metacognitive outcomes are measured.

Artificial intelligence governance typically focuses on whether people rely on AI appropriately during individual decisions. This Perspective addresses a different question: whether repeated AI use changes the capabilities and alternatives available to users and organisations over time. It introduces path‐induced AI dependence, defined as AI‐attributable deterioration in previously credible capabilities, alternatives or recovery routes for a valued function. This is distinguished from functional reliance, where performance depends on a useful AI system without evidence that prior options have deteriorated. A problem‐driven narrative synthesis suggests that direct longitudinal evidence remains limited. Existing studies provide some evidence that repeated AI or automation use can reduce subsequent unaided performance, but there is not yet direct longitudinal evidence establishing broader organisational mechanisms such as routinisation, technical‐economic lock‐in or dependence debt in generative‐AI settings. The framework therefore identifies five analytically distinct mechanisms: capability depletion, epistemic enclosure, relational switching costs, organisational routinisation and technical‐economic lock‐in, and treats their relationships as testable rather than established. It further proposes exit capacity as a multidimensional way to assess whether important capabilities, evidence, fallback arrangements and substitutes remain available. The proposed measures are research and governance tools rather than validated interventions. The central argument is that responsible AI governance should evaluate not only present performance, but also whether repeated AI use preserves credible future options.

Generative artificial intelligence (GenAI) is changing how students encounter and work with knowledge in higher education. Students use AI systems to obtain explanations, summarise readings, develo...

Artificial intelligence (AI) systems increasingly mediate how learners obtain explanations, synthesise sources, receive feedback and judge the quality of academic work. This critical-integrative review argues that the central educational question is not whether learners rely on AI, but whether that reliance preserves or displaces the epistemic work through which judgement develops. The review purposively bridges artificial intelligence in education, human–AI interaction, epistemic cognition, information behaviour, cognitive offloading, and social epistemology, while engaging a recent literature on epistemic authority, agency, infrastructure and over-reliance. It conceptualises epistemic dependence as reliance on AI for knowledge-related tasks and differentiates productive reliance from harmful dependence through six diagnostic criteria: contestability, recoverability, transfer, traceability, distributed responsibility and epistemic plurality. Four sociotechnical pathways are then analysed: fluent authority, frictionless delegation, opaque synthesis and institutionalised dependence. These pathways may affect verification, evaluative and disciplinary judgement, understanding, uncertainty tolerance, learner identity and epistemic justice, but the review does not assume that AI uniquely or uniformly causes these outcomes. It proposes relational epistemic agency as the normative aim: the capacity to question, verify, compare, justify and take responsibility for knowledge claims within human, technological and institutional relations. Design, pedagogical, assessment and governance implications are developed accordingly.

1. AI-mediated informal digital learning of English (AI-IDLE) has become an important form of self-directed English learning beyond formal classrooms. Through generative AI tools, learners can obta...

Generative artificial intelligence (GenAI) increasingly mediates the epistemic work of higher education: students use it not only to produce text, but also to locate information, interpret concepts, evaluate claims, obtain feedback, and regulate study. This creates a measurement problem that is not captured directly by existing measures of AI trust, technology acceptance, AI literacy, cognitive offloading, or general reliance. This article conceptualises uncritical epistemic dependence in AI-mediated learning as a context-sensitive tendency to treat AI-generated outputs as epistemically directive, such that learner verification, interpretation, justification, production, or regulation may be displaced. It reports the development and initial validation of the Epistemic Dependence in AI-Mediated Learning Scale (ED-AIL Scale), a risk-oriented instrument designed to distinguish ordinary or productive AI support from epistemically consequential delegation. The scale was developed through construct specification, expert review, cognitive pretesting, and two independent Prolific samples of English-speaking higher education students. Evidence from ordinal exploratory and confirmatory factor analyses, reliability estimates, validation measures, performance-oriented criteria, and measurement-invariance tests supported a provisional five-domain interpretation: knowledge acquisition, conceptual interpretation, epistemic evaluation, knowledge production, and epistemic regulation. Measurement evidence was strongest for acquisition, interpretation, and evaluation; more qualified for production; and weakest for regulation. ED-AIL offers an initial research tool for examining how learners allocate epistemic responsibility between themselves and AI systems, while requiring further validation across institutional, disciplinary, linguistic, platform-specific, and task-specific contexts.