The Ghost Learner Paradox: A Systematic Review of Metacognitive Laziness in the Age of Generative AI

Synthesis: Febriyanti, Muntoha and Uminar review post-ChatGPT literature from 2023 to 2026 and name a paradox: grades hold or improve while the cognitive work behind them disappears. The mechanism is Cognitive Offloading escalated into what they call metacognitive laziness — delegating planning, monitoring and evaluation to a model, not merely execution. They separate offloading that frees working memory for higher-order reasoning from substitution that bypasses productive struggle, argue that agentic AI breaks output-centric assessment, and locate the remedy in process-oriented assessment and "AI fluency" rather than AI literacy. The review reports no corpus count, so its evidence base cannot be audited from the text.

Key Findings

1. The window is the post-ChatGPT era. The search is bounded to 2023–2026, deliberately capturing the period after ChatGPT's late-2022 release, and uses PRISMA 2020 with Google Scholar Labs supplemented by Scopus and Web of Science for cross-validation.
2. Three themes organise the corpus. Iterative coding produced cognitive engagement and active-learning displacement, memory consolidation and retention deficits, and measurable decline in critical thinking and higher-order reasoning.
3. Offloading splits into two kinds. Preservice teachers in Iqbal et al. (2025) gained achievement when AI worked as shared metacognition, while Georgiou (2025) found significantly lower cognitive engagement among students who used ChatGPT during academic writing.
4. Agentic AI defeats the assessment proxy. Bozkurt et al. (2026) describe surrogates that couple an LLM to an agentic browser to log into institutional portals, read prompts, generate compliant responses and submit them, which similarity tools and synchronous proctoring cannot detect.
5. Confidence rises as engagement falls. Georgiou (2025) reports that heavy ChatGPT users were confident in their outputs and their own competence despite lower cognitive engagement during the task — an amplification of the Dunning-Kruger pattern.
6. One synthesis of 21 publications found a predominantly positive effect of generative AI on critical thinking where it was integrated to promote active reasoning, metacognitive monitoring and critical autonomy (Raitskaya & Tikhonova, 2025).

Where offloading stops being adaptive

The review's analytical hinge is not tool use but metacognitive agency. Beneficial offloading delegates routine work — formatting, citation, arithmetic — and frees working memory for the reasoning that builds mental models; detrimental offloading delegates the modeling itself. Fan et al. (2025) supply the definition the review leans on: metacognitive laziness is offloading planning, monitoring and evaluation, the processes that constitute learning rather than merely support it. The distinction is empirically visible. Georgiou (2025) describes a "standby mode" in which the brain reduces active processing because an external agent has assumed the inferential workload, and Yan et al. (2025) report diminished epistemic vigilance and superficial learning alongside it. But Exintaris et al. (2023) found that asking chemistry students to critique erroneous ChatGPT solutions, rather than accept correct ones, raised metacognitive engagement substantially — evidence that the task's architecture, not the presence of AI, sets the outcome.

The ghost learner is an institutional artifact

The review's most consequential move is to treat the ghost learner as a product of assessment design rather than of student character. Digital classrooms were built on the assumption that submitted work is a reliable proxy for the named student's cognitive effort, and agentic AI invalidates that assumption at each level: timed quizzes, reflective journals written in an approximate personal voice, asynchronous discussion posts with contextually appropriate replies. Bozkurt et al. (2026) call the accumulated unlearned competency "cognitive debt" — each delegated assignment adds to a balance carried into a credential that attests to competencies never exercised. The institutional record shows nothing, because completion rates do not change when the work is done by a surrogate. Yan et al. (2025) describe the same drift as erosion of epistemic agency, and the review draws the institutional conclusion: the problem is not detection but what verification of presence would require.

The illusion of competence and the atrophy it hides

Because fluent output is a high-quality artifact, students readily attribute it to their own capability, and the misattribution inflates perceived competence while the underlying skill stays undeveloped. The review traces measurable atrophy across domains: reduced epistemic vigilance and shallow retention (Yan et al., 2025); lower originality, weaker argumentation and a thinner authorial voice in frequent AI users (Tekir, 2026); passive information consumption replacing generative brainstorming (Guo & Lee, 2023); and analytical depth replaced by AI-assisted heuristics, which Suriano et al. (2025) frame as System 2 reasoning giving way to System 1 reliance. The cumulative profile is a graduate fluent in AI-mediated tasks who struggles to perform the same operations unaided — a problem for professional performance and for further learning, not only for the transcript.

From AI literacy to AI fluency

The review distinguishes AI literacy, competent operation of tools, from AI fluency, the metacognitive capacity to evaluate, selectively deploy and epistemically own an interaction with AI. Its interventions cluster into three: process-oriented assessment that evaluates how students engage with AI; Socratic dialogue that forces students to justify, extend and challenge generated content; and "AI nudges", interface or prompt features that redirect users toward reflection. The most structural proposal is to move from verifying output to verifying presence — oral defenses, in-class critical reflection, think-aloud protocols and portfolios that require students to narrate a process over time. Exintaris et al.'s SMARTCHEMPer workshop is the concrete model: students articulate reasoning while solving and critique AI output in real time, producing evidence no submitted essay can replicate. The review's conclusion is not restriction but redesign of the human role so that it is irreducibly cognitive.

What this means for practice

• Instructors. Require process evidence — a think-aloud, a draft trail, a short oral defense — on any task where the deliverable could be delegated, because the review's own distinction turns on whether the student monitors and critiques the model or merely receives its output.
• Program designers and administrators. Audit assessment regimes for delegation resistance before purchasing detection software: the review finds similarity tools and synchronous proctoring structurally blind to asynchronous agentic submission, and treats output-centric design as the incentive that produces substitution.
• Researchers. Test the proposed countermeasures rather than assuming them: the review itself calls for controlled validation of Socratic frameworks, AI-audit checklists and process-based designs, and for longitudinal tracking of critical thinking in AI-native cohorts.

Limitations

• No corpus size or PRISMA flow counts are reported. The review claims PRISMA 2020 but prints no identification, screening or inclusion numbers, and its search ran through Google Scholar Labs rather than a fixed database query, so neither the evidence base nor the result set can be reconstructed.
• No DOI is printed and the journal prints a placeholder ISSN (XXXX-XXXX), so bibliographic verification depends entirely on the publisher's site — a real caveat for citation and for any later meta-analysis.
• The authors concede that the evidence is largely cross-sectional or short-term, and that long-term or decadal effects of sustained AI reliance remain unsubstantiated.
• Several load-bearing claims rest on single studies, including a 2025 preprint (Georgiou) and a 2026 study of EFL writing in one national context (Tekir), and the shared-metacognition finding comes from one sample of preservice teachers.
• The search relied on Google Scholar Labs, which does not produce the fixed, reproducible query result set that a database-only PRISMA search would.
• The ghost-learner evidence is largely conceptual: Bozkurt et al. (2026) theorize the surrogate and the verification gap, but the review reports no measured prevalence of ghost learners in any institution.

Citation

Febriyanti, Muntoha, T., & Uminar, A. N. (2026). The ghost learner paradox: A systematic review of metacognitive laziness in the age of generative AI. Journal of Socio-Digital Integration, 1(1), 1–13.