The central structural problem the review identifies is what the author calls the verification circularity problem. A student cannot reliably check an AI output on a skill they have not yet developed, because checking requires exactly the competence that was outsourced. A student who cannot yet construct a valid argument in a domain is, by the same token, poorly positioned to recognize an invalid one presented fluently. The author links this pattern to Kruger and Dunning’s classic finding that limited competence in a domain impairs the judgement needed to recognize that limitation, and to Sperber and colleagues’ concept of epistemic vigilance, which holds that trust in a source is calibrated through cues the vigilant party can interpret. For content that is genuinely new to the learner, those cues are precisely what is missing.

The reliability backdrop matters here. Language models are prone to hallucination, producing fluent, confident, but unsupported or fabricated content, as Ji and colleagues documented. In a medical-review context, Chelli and colleagues quantified substantial rates of fabricated or inaccurate references when AI tools were asked to support literature reviews, though the review is careful to note that this evidence concerns reference accuracy in systematic reviews of medical literature and is used to illustrate a general reliability problem with AI-generated citations, not as direct evidence about higher-education coursework. The review also cites a structural analogy from the opposite direction: commercial AI-detection tools misclassified more than 61 percent of TOEFL essays by non-native English speakers as AI-generated, showing that even a system built to verify AI involvement can confuse fluency with validity. The author stresses this is an analogy, not empirical confirmation that student verification fails for the same reason.

Building on the circularity problem and on classical automation-delegation theory, the review proposes the Cognitive Delegation Boundary Framework. The framework classifies a task, for a given learner, along two dimensions: skill-formation value, meaning the extent to which performing the task oneself at the learner’s current stage builds a capability the intended learning outcomes require and that the learner has not yet consolidated; and verification independence, meaning the extent to which the learner can judge whether an AI output is correct using knowledge, criteria, or procedures that do not depend on the competence the task is meant to build. Crossing the two dimensions yields four zones: freely delegate, delegate with caution, delegate with reflection, and do not delegate yet.

Zone membership is developmental rather than fixed. A task that a novice should not yet delegate may be appropriately delegated by an advanced learner. The framework draws on Vygotsky’s zone of proximal development and Wood and colleagues’ scaffolding theory, which describe support that should fade as competence grows, and on Kirschner and colleagues’ argument that novice learners require guidance because they lack the prior knowledge needed to self-correct. Writing an argumentative essay, for example, sits in the Do Not Delegate Yet zone for a first-year undergraduate still developing argumentative structure, but plausibly shifts toward Delegate with Reflection for a doctoral candidate who has mastered argumentative writing and uses AI to stress-test a drafted position. The word “yet” in the zone label marks precisely this condition.

Applied to a stipulated profile of a typical second- or third-year undergraduate who has completed introductory academic-writing instruction but has not yet demonstrated independent competence in disciplinary argumentation, the framework places citation formatting and grammar checking in the Freely Delegate zone, since these are mechanical once the convention is understood and errors are detectable by inspection against external criteria. Initial literature searches and first-draft summaries of sources fall in the Delegate with Caution zone, where the risk is error rather than lost learning and each source must be confirmed to exist and read before use. Constructing an original thesis or argument, weighing conflicting evidence, and framing a problem or research question all land in the Do Not Delegate Yet zone, because they are both the tasks most valuable to perform independently and the ones the learner is least equipped to verify once delegated. Checking one’s own completed argument with AI assistance falls in the Delegate with Reflection zone, where the learner’s own prior attempt makes verification possible. The author emphasizes that these classifications are reasoned illustrations, not findings of the reviewed literature.

The empirical evidence base for the framework’s central prediction remains limited, and the review is unusually explicit about this. The reviewed studies differ in design, population, and evidentiary weight, and the author separates three kinds of outcome: reduced cognitive effort during AI-assisted work, weaker task performance or recall associated with AI assistance, and lasting effects on learning once AI assistance is removed. Kosmyna and colleagues, using EEG to compare essay writing under unassisted, search-engine-assisted, and LLM-assisted conditions, found the weakest and least distributed brain connectivity in the LLM-assisted group and a reduced ability to quote their own submitted essays accurately. That study is a preprint with a small sample, and EEG connectivity is not a validated measure of learning, so it is treated as suggestive, preliminary evidence. Lee and colleagues surveyed 319 knowledge workers and found self-reported reductions in cognitive effort concentrated among users with the highest confidence in the AI tool’s output, though the sample was not students and relied on self-report.

Further evidence is similarly qualified. Zhai and colleagues’ systematic review of 14 studies found that over-reliance on AI dialogue systems is associated with reduced decision-making, critical-thinking, and analytical-reasoning performance, but the primary studies are heterogeneous and the associations are not causal. Hu and colleagues’ qualitative interviews with 20 pre-service science teachers found that participants scaled their verification effort with stakes, applying lighter scrutiny to low-stakes personal use and escalating to scholarly verification when preparing material to teach; this study describes verification practice rather than measuring learning outcomes. Crucially, no reviewed study provides evidence of the third outcome type: none tested whether delegating tasks classified as Do Not Delegate Yet causes lasting skill loss. The classification of those tasks therefore rests on the theoretical argument that effortful generation builds durable skill and that unverifiable delegation removes the learner’s means of catching errors, not on direct empirical demonstration.

The framework’s implications for institutions are deliberately narrow. The author argues that disclosure is necessary but not sufficient: a fully disclosed, fully permitted delegation of original argument construction still falls in the Do Not Delegate Yet zone, whereas an undisclosed use of AI to check citation formatting would incur comparatively little developmental cost despite violating a strict policy. The verification circularity problem also reframes investment in AI-detection technology as addressing one symptom of the underlying risk rather than the risk itself, a reading reinforced by the finding that detection tools are unreliable and biased against non-native English writers. Because a single fixed policy is unlikely to be well calibrated across a population spanning first-year undergraduates to doctoral candidates, the review suggests delegation guidance may need to be tied to demonstrated competence checkpoints within a course rather than only to course level.

The review also acknowledges its own limitations candidly. It synthesizes 31 purposively identified sources rather than the results of an exhaustive database search, stage-level screening counts were not recorded, and sources were screened and coded by a single researcher without double-coding or inter-rater checks. The Cognitive Delegation Boundary Framework is a proposed conceptual framework, not an established theory: its indicators have not been validated, its inter-rater reliability has not been assessed, and its central prediction about lasting skill loss remains untested. The author lays out a validation agenda that includes refining the indicators, developing a rubric for instructors and students, Delphi studies with domain experts, and longitudinal or quasi-experimental designs comparing partial and complete delegation on retention, transfer, and critical-thinking outcomes measured after AI assistance is removed. The review’s conclusion is that students should not yet delegate the judgements they are still learning to make, and that testing this developmental logic empirically is the work that comes next.

Subject of Research: A conceptual framework for determining which academic tasks higher education students should not yet delegate to generative AI

Article Title: What should students not yet delegate to AI? Toward a theory of cognitive delegation boundaries in higher education

Article References: Salameh, W. (2026). What should students not yet delegate to AI? Toward a theory of cognitive delegation boundaries in higher education. Frontiers in Education, 11, Article 1982690. https://doi.org/10.3389/feduc.2026.1982690

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1982690

Keywords: generative AI, higher education, cognitive offloading, verification circularity, skill formation, automation trust, self-determination theory, academic integrity, cognitive delegation, desirable difficulty, learning theory, AI detection