Generative AI May Let Students Perform Without Learning, Review Warns

The review, published in Frontiers in Education by Walid Salameh of An-Najah National University in Palestine, synthesises 45 verified sources spanning empirical experiments, systematic reviews, conceptual works and policy guidance, and concludes that the much-discussed crisis of critical thinking is better understood as a structural failure of assessment design than as a motivational failure of students.

The review’s starting point is a historical shift. For most of the modern research university’s existence, the central intellectual challenge facing a student was access: specialised knowledge lived in libraries, archives and the minds of trained specialists, and the university’s value lay in granting curated, credentialed entry to it. Today, the barrier to obtaining an answer, correct or not, has approached zero. Multi-modal data, preprints, news feeds and generative AI systems produce content at a volume and speed with no historical precedent. Paradoxically, the review argues, this abundance has not produced a golden age of inquiry, because the cognitive bottleneck has moved from retrieval to evaluation.

That shift rests on well-established cognitive science. Cognitive load theory, formulated by John Sweller in 1988, holds that working memory is strictly limited, and dual-process theory suggests that when the volume of incoming information exceeds what deliberate processing can handle, people default to fast heuristic judgments such as surface plausibility, source popularity and visual polish. In a higher-education context, the review describes this as epistemic disorientation: students encounter peer-reviewed research, preprints, predatory journal content and algorithmically ranked commercial material on the same undifferentiated search results page, often without explicit instruction in how to tell them apart. Empirical support comes from a well-known 2019 study by Sam Wineburg and Sarah McGrew, who found that professional fact-checkers evaluated unfamiliar online sources more accurately than historians and university students, because they left the source to check it against others rather than judging it from its own presentation.

The strongest experimental evidence in the synthesis concerns what happens when students delegate reasoning itself to AI. In one randomised experiment, 91 university students researching a socio-scientific issue with either ChatGPT or a search engine reported lower cognitive load in the ChatGPT condition but produced lower-quality reasoning and justifications. Another experiment with 117 students found that ChatGPT support improved essay scores but not knowledge gain or transfer, and was associated with fewer metacognitive processes such as evaluation, a pattern the authors described as potential metacognitive laziness. A study of 1,625 students across ten courses found that students tended to rely on, rather than learn from, AI assistance: when the support was removed, the quality of their work declined.

Perhaps the most striking finding comes from a field experiment with nearly 1,000 high-school students, which the review treats as indirect evidence because the sample is pre-university. Unrestricted access to a GPT-4 tutor improved practice performance, but students performed 17 percent worse than controls once access was removed; a version of the tutor with pedagogical guardrails that withheld direct answers largely mitigated the harm. The review reads this alongside a survey of 666 adults reporting a negative association between frequent AI-tool use and critical-thinking scores, partly mediated by cognitive offloading, and a survey of knowledge workers in which higher confidence in generative AI was associated with less critical thinking. All of these are correlational or short-term, and the review is careful to flag their limits.

One influential study cited in the synthesis is treated with unusual caution. An EEG preprint comparing essay writing under unassisted, search-engine-assisted and LLM-assisted conditions reported that LLM-assisted writers showed the weakest and least distributed brain connectivity, and the lowest ability to quote their own submitted work. The review notes that the study used a small sample and has attracted published methodological criticism regarding sample size and analytic transparency, and reports its findings as suggestive rather than settled. It also cites a meta-analytic finding that students’ critical thinking generally improves during college, which cautions against any claim of a general collapse, and a phenomenological study of Bangladeshi undergraduates describing internal conflict between awareness that uncritical reliance was eroding their independent reasoning and continued reliance driven by time pressure.

The review’s third theme concerns fluency. Language models are prone to hallucination, generating confident content, including citation-like details, that is unsupported or fabricated, and studies quantifying fabricated references in AI-assisted literature reviews document this directly. Automated detection, the standard institutional response, fares no better. One evaluation found that commercial GPT detectors misclassified more than 61 percent of TOEFL essays written by non-native English speakers as AI-generated, while performing near-perfectly on native-English essays, a bias plausibly rooted in perplexity-based methods penalising standardised syntax. Another evaluation of 14 detection tools concluded that they were neither accurate nor reliable, tended to classify AI-generated text as human-written, and performed worse when text was paraphrased or machine-translated. The review interprets both errors as the same confusion: treating a statistical property of text, its fluency or predictability, as a proxy for a property it does not reliably track, whether truth or human origin.

Against this diagnosis, the review finds that the most promising responses are teachable and measurable. A meta-analysis of instructional interventions found that explicit critical-thinking instruction improves both skills and dispositions, with dialogue, authentic problems and mentoring especially effective. Lateral reading can be taught to college students in an online course, and self-regulation prompts partly, though not fully, compensated when AI assistance was withdrawn in one experiment. The literature also recommends process-visible assessment, in which student thinking is made transparent through version-tracked drafts, oral defence components, AI-use reflection logs and in-class collaborative work, grounded in assessment-validity theory: the core question is whether the evidence supports the inference that a student is capable, not whether cheating occurred. However, the review found no study that empirically compares process-visible with product-only assessment under conditions of AI availability, so this part of the argument remains largely conceptual.

The review’s principal original contribution is the Fade–Visibility Matrix, a two-dimensional diagnostic that crosses AI support structure, whether assistance is designed to persist or to fade as competence grows, against assessment design, whether the learning process or only the finished product is visible to an evaluator. This yields four states: Silent Dependency, where AI support never fades and only the product is assessed; Invisible Growth, where support fades but assessment cannot see it; Exposed Dependency, where process-visible assessment reveals offloading that has not yet changed; and Calibrated Independence, the target state combining fading support with observable process. The matrix is explicitly offered as a diagnostic tool with four testable propositions, not as an empirically validated model, and the review acknowledges that no included study tests it directly. It draws on scaffolding theory’s insight that good support fades, noting that generative AI as typically used does not fade, offering the same high level of support on the hundredth essay as on the first.

The review is candid about its constraints. It synthesises 45 purposively identified sources rather than a systematically searched literature base, no structured Scopus or Web of Science search with PRISMA-style screening was conducted, coding was performed by a single reviewer so inter-coder agreement could not be assessed, and only English-language sources were included. The strongest experimental evidence on delayed effects comes from a secondary-school sample, most studies are short-term, and critical thinking is rarely measured directly. Institutional constraints compound the problem: process-visible assessment is materially more labour-intensive than automated grading, and under-resourced institutions may lean further on scalable digital instruction, raising an equity concern the review calls plausible but empirically untested. Its conclusion is correspondingly measured: universities should teach evaluation, verification and metacognitive regulation explicitly within every discipline, design AI support to fade, and assess the visible process of reasoning rather than only its finished product, while acknowledging that whether these effects accumulate into lasting changes in critical thinking remains an open empirical question.

Subject of Research: How information abundance and generative AI use may influence critical thinking in higher education, based on a critical narrative review

Article Title: The crisis of critical thinking in higher education: rethinking what universities should teach in an age of information abundance and generative AI — a critical narrative review of international evidence

Article References: Salameh, W. (2026). The crisis of critical thinking in higher education: rethinking what universities should teach in an age of information abundance and generative AI — a critical narrative review of international evidence. Frontiers in Education, 11, 1981900. https://doi.org/10.3389/feduc.2026.1981900

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1981900

Keywords: critical thinking, higher education, generative AI, cognitive offloading, assessment redesign, information abundance, lateral reading, scaffolding theory, AI detection, metacognition, academic integrity, Frontiers in Education