Fluency Illusion: A Review on Influence of ChatGPT in Classroom Settings

Abstract

The rapid adoption of generative artificial intelligence tools such as ChatGPT in educational settings has generated both enthusiasm and concern regarding their influence on student learning. While several studies report improvements in efficiency, confidence, and perceived understanding, evidence for durable conceptual learning and knowledge transfer remains mixed. This article examines these tensions through the concept of fluency illusion, a cognitive phenomenon in which information that is easy to process is mistakenly judged as being well understood. Using a narrative conceptual review approach, this study synthesizes findings from 41 publications identified through searches of Google Scholar, Scopus, Web of Science, and ERIC covering the period from 2022 to early 2026. The reviewed literature includes 28 empirical studies, nine conceptual or theoretical analyses, and four review articles addressing the use of ChatGPT in educational contexts. Across domains such as writing and language learning, STEM problem solving, feedback and tutoring, and assessment, the literature shows a recurring pattern in which fluent AI-generated responses increase learners’ confidence without consistently improving deeper conceptual understanding. Drawing on research from cognitive psychology and metacognition, this review proposes an integrative conceptual account of how fluent AI output may shape learners’ judgments of understanding and influence their engagement with learning tasks. The paper concludes by discussing implications for instructional design, assessment practices, and metacognitive scaffolding, and outlines directions for future research aimed at empirically examining the proposed framework and identifying strategies to reduce fluency-driven misjudgments while preserving the potential benefits of generative AI in education.

1. Introduction

The rapid emergence of large language models has introduced a new class of tools into educational contexts, with ChatGPT among the most widely adopted examples [1]. Capable of generating coherent, contextually appropriate, and stylistically polished responses across a wide range of academic tasks, ChatGPT is increasingly used by students and educators for writing support, problem explanation, idea generation, and feedback [2]. Its accessibility and apparent versatility have accelerated its integration into classroom settings, often preceding the development of well-established pedagogical frameworks or shared understandings of its influence on learning processes.

While early discussions have highlighted the potential benefits of ChatGPT for personalized learning and instructional support, growing attention has begun to focus on a more subtle challenge [3]. The fluency of AI-generated text frequently creates an impression of correctness, depth, and understanding that may not correspond to genuine learning [4]. Well-structured and linguistically polished responses can appear authoritative even when they contain inaccuracies or simplified reasoning [5]. This divergence between surface quality and underlying understanding raises important questions about how learners interpret and engage with AI-generated content in educational environments [6].

From a cognitive perspective, this phenomenon aligns closely with what has been described as fluency illusion [7]. Research in cognitive psychology has long demonstrated that information that is easy to process is often judged as more accurate, more credible, and better understood than information that requires greater cognitive effort [8]. In classroom contexts, such fluency-based judgments can lead learners to overestimate their own comprehension, reduce critical evaluation, and prematurely conclude that learning has been achieved. When fluency is externally generated by an AI system rather than arising from the learner’s own cognitive processing, the risk of misattributing understanding becomes particularly pronounced.

The rapid adoption of ChatGPT in education has, in many cases, outpaced systematic pedagogical reflection on these cognitive dynamics [9]. Existing studies often focus on discrete outcomes such as writing quality, task efficiency, or student attitudes, and frequently report mixed or context-dependent findings. However, these investigations are rarely connected through a shared conceptual framework capable of explaining why fluent AI assistance may simultaneously enhance performance while undermining deeper learning. As a result, the literature remains fragmented, with limited integration between empirical observations and established theories of learning and metacognition.

At the same time, the empirical evidence regarding learning outcomes remains uneven. Several experimental and classroom-based studies report improvements in task completion, perceived understanding, or writing quality when students use generative AI tools. However, evidence concerning deeper conceptual learning and knowledge transfer is more limited and sometimes inconsistent. Some studies indicate that while AI-assisted explanations can support immediate performance, learners may still struggle to reproduce or apply the underlying reasoning independently. These findings are consistent with research on self-regulated learning and metacognitive calibration, which shows that learners often rely on subjective cues such as the ease with which information is processed when judging their own understanding. In AI-supported learning environments, fluent responses generated by systems such as ChatGPT may therefore increase learners’ confidence even when underlying understanding remains uncertain.

Recent empirical studies also illustrate this tension more directly. Experimental and classroom-based investigations have reported that the use of ChatGPT can improve task completion, writing fluency, and students’ perceived understanding. For example, studies examining AI-assisted writing tasks report that students often produce more fluent and well-structured text, while improvements in argument quality or conceptual reasoning remain limited [10,11,12]. Similarly, research in STEM learning contexts suggests that students may successfully follow AI-generated explanations yet struggle to reproduce or adapt the same reasoning independently in new problems [13,14]. These findings indicate that fluent AI support may increase confidence and short-term performance while leaving deeper learning outcomes less certain.

Viewing ChatGPT through the lens of fluency illusion offers a unifying perspective for understanding these tensions [15]. Rather than framing AI use as inherently beneficial or inherently problematic, this perspective emphasizes how the perceptual qualities of AI-generated content shape learners’ judgments, behaviors, and learning strategies. It also highlights the difference between fluent academic performance and genuine conceptual understanding, a distinction that is central to education but often blurred when students use AI tools [16].

Against this background, the present study adopts a narrative and conceptually oriented review approach to examining the influence of ChatGPT in classroom settings, with particular attention to how fluent AI-generated responses may shape learners’ perceptions of understanding. It brings together empirical findings, theoretical discussions, and emerging pedagogical perspectives to clarify how fluency illusion operates in AI-supported learning. Specifically, the review seeks to synthesize existing work on the use of ChatGPT in classrooms, to advance a conceptualization of fluency illusion in AI-mediated learning, and to outline implications for pedagogy and assessment. By situating current research within a coherent theoretical framework, the review aims to contribute to a more nuanced understanding of how generative AI reshapes not only what students produce, but also how they perceive and evaluate their own learning.

The objectives of this review are threefold. First, it integrates research on the adoption of ChatGPT across diverse classroom contexts. Second, it links insights from cognitive psychology with emerging evidence on AI-mediated learning to elaborate the concept of fluency illusion. Third, it examines how this phenomenon challenges prevailing assumptions about student understanding, instructional design, and assessment practices. Throughout, the review prioritizes theoretical coherence and explanatory insight over exhaustive coverage, with the goal of illuminating recurring mechanisms that shape learners’ experiences with fluent AI systems.

2. Methodology

This study adopts a narrative review approach to examine how ChatGPT (OpenAI) and related generative AI tools are being used in educational settings, with particular attention to patterns of perceived fluency, learning confidence, and evidence of actual learning outcomes. The review was designed to prioritize conceptual coherence and pedagogical relevance rather than exhaustive coverage of all published studies.

2.1. Search Strategy

A targeted literature search was conducted using Google Scholar, Scopus, Web of Science, and ERIC. The searches were carried out between January 2022 and January 2026. Core search terms included combinations of “ChatGPT”, “generative AI”, “large language models”, “education”, “learning”, “writing”, “assessment”, “feedback”, “student confidence”, and “learning outcomes”. Searches were performed primarily in titles, abstracts, and keywords. The review focused on studies published between 2022 and early 2026, reflecting the period during which ChatGPT and similar tools entered widespread educational use.

2.2. Inclusion and Exclusion Criteria

Studies were included if they examined the use of ChatGPT or closely related generative AI systems in formal educational contexts, including secondary and higher education. Eligible publications included empirical studies reporting classroom implementation, student perceptions, learning outcomes, or assessment-related effects, as well as conceptual or theoretical papers that provided analytical discussion of the pedagogical or cognitive implications of generative AI in education.

Empirical studies were included when they reported original data collected through experimental, survey-based, classroom-based, or mixed-methods designs. Conceptual and theoretical contributions were included when they offered substantive analysis relevant to AI-mediated learning, metacognition, or instructional design. Studies that did not address educational contexts or that focused exclusively on technical model development were excluded.

Studies were excluded if they focused purely on technical model development without educational application, discussed generative AI only in non-educational contexts, or consisted solely of opinion pieces without analytical or empirical grounding. Only English-language publications were considered.

2.3. Screening and Selection Process

An initial set of articles was identified through database searches and reference list screening. Duplicates were removed, after which titles and abstracts were reviewed for relevance. Full texts were then examined to confirm alignment with the study’s focus on learning processes, perceived fluency, and educational impact.

During the full-text screening stage, attention was also given to basic indicators of methodological quality. Although the review did not apply a formal scoring system, studies were evaluated with reference to commonly used appraisal principles in educational research, such as clarity of research design, transparency in data collection procedures, and adequacy of the reported analysis. These criteria were informed by widely used appraisal frameworks, including the Critical Appraisal Skills Programme (CASP) and the Mixed Methods Appraisal Tool (MMAT).

Publications lacking sufficient methodological description or analytical grounding were excluded during the eligibility assessment stage. This approach allowed the review to maintain methodological transparency while remaining consistent with the interpretive and conceptually oriented nature of a narrative synthesis.

2.4. Data Extraction and Synthesis

For each included study, key information was extracted, including educational setting, participant group, task type, reported benefits, reported concerns, and the type of evidence used to evaluate learning. Particular attention was paid to whether studies relied on self-reported confidence, perceived ease of use, or measurable performance outcomes.

The synthesis followed a thematic narrative approach. Studies were grouped by instructional context such as writing, language learning, STEM education, feedback, and assessment. Recurring patterns across these contexts were used to develop the conceptual framework presented in the figures, linking fluency cues to confidence, reduced cognitive friction, and potential mismatches between perceived and actual learning.

2.5. Scope and Limitations

This review does not claim to be exhaustive or fully systematic. Its aim is to clarify emerging patterns and conceptual tensions in the literature rather than to quantify effect sizes or rank interventions. While this approach allows for integrative interpretation across diverse study designs, it also means that conclusions should be read as analytically grounded rather than statistically definitive.

2.6. Study Selection Overview

To improve transparency in the selection process, a structured screening procedure was applied during the literature search. The initial search across the selected databases produced a broad set of publications related to generative AI and education. After removing duplicate records, titles and abstracts were screened to identify studies that examined the use of ChatGPT or closely related generative AI systems in educational contexts.

Articles focusing exclusively on technical model development or non-educational applications were excluded at this stage. The remaining publications were then evaluated through full-text review to determine their relevance to learning processes, student perceptions, or instructional outcomes associated with generative AI use.

The overall study selection process is summarized in Figure 1. The final review corpus consisted of 41 publications that met the inclusion criteria and directly addressed classroom applications of ChatGPT or comparable generative AI tools. Of these, 28 were empirical studies reporting original classroom-based investigations or survey-based analyses of ChatGPT use in educational settings. Nine publications were conceptual or theoretical papers that examined the pedagogical, cognitive, or metacognitive implications of generative AI in learning environments. In addition, four review articles were included because they provided broader syntheses of emerging research on generative AI in education and helped contextualize the empirical findings discussed in this study.

3. Theoretical Foundations of Fluency Illusion

3.1. Processing Fluency and Judgments of Understanding

Fluency illusion is grounded in a well-established body of cognitive and educational research showing that learners’ judgments of understanding are frequently shaped by subjective experiences rather than by objective indicators of learning [17]. A central construct in this literature is processing fluency, which refers to the perceived ease with which information is read, interpreted, or mentally manipulated (see Figure 2 for the proposed AI-mediated pathway).

The model presented in Figure 2 should be interpreted as a conceptual framework designed to organize insights from cognitive psychology and emerging research on AI-supported learning. It does not represent an empirically validated causal model. Rather, it offers a theoretical proposal that highlights potential relationships between processing fluency, metacognitive judgments, and learning outcomes in AI-mediated learning environments.

Information that is processed fluently is consistently judged as more accurate, more comprehensible, and better learned, even when such judgments are misaligned with performance on transfer or application tasks [18]. In instructional contexts, this tendency can lead learners to equate clarity of presentation with conceptual mastery.

Research on illusions of competence further elucidates how fluency can distort self-evaluations of learning [19]. When instructional materials are well structured, linguistically polished, or immediately responsive, learners often report high confidence despite limited ability to independently reproduce or apply the content. This pattern reflects long-standing findings in educational psychology showing that exposure and recognition are frequently mistaken for genuine learning. From a metacognitive perspective, such misjudgments arise because learners rely on salient and cognitively economical cues when evaluating their own understanding. Judgments of learning are often informed by how easily information is processed in the moment rather than by diagnostic evidence of learning, such as the ability to retrieve, explain, or transfer knowledge [20].

These dynamics can also be considered in relation to research on retrieval practice and cognitive offloading. A large body of work on retrieval practice shows that actively recalling information strengthens long-term learning more effectively than passive exposure to explanations or summaries. When learners rely heavily on AI-generated explanations, opportunities for retrieval-based learning may diminish, especially if students consult the system before attempting to articulate their own understanding. At the same time, generative AI tools can function as a form of cognitive offloading by allowing learners to delegate portions of explanation or reasoning to an external system. While such support may improve efficiency and accessibility, it may also reduce the cognitive effort required for durable learning when it replaces rather than complements active engagement with the material.

3.2. Metacognitive Monitoring and Judgments of Learning

Insights from research on metacognitive monitoring further clarify why fluency-driven misjudgments occur in learning contexts [21]. Cue-utilization theories of judgments of learning suggest that learners often rely on heuristic signals when evaluating their own understanding, particularly when more diagnostic evidence is not immediately available. Koriat’s cue-based framework [22] proposes that judgments of learning are influenced by both intrinsic cues related to the material itself and experiential cues such as processing fluency or familiarity. Because these cues are only imperfect indicators of actual knowledge, learners may develop inflated confidence when information is easy to process. Related work on metacognitive calibration by Dunlosky and Metcalfe [23] similarly demonstrates that learners frequently overestimate their level of understanding when evaluation relies on subjective feelings of comprehension rather than retrieval-based evidence. When fluent explanations are generated by AI systems, these metacognitive cues may become even more salient, increasing the likelihood that processing ease will be interpreted as genuine learning.

A related phenomenon discussed in the human–automation interaction literature is automation bias, which refers to the tendency to accept recommendations generated by automated systems without sufficient verification. Although automation bias originates in research on decision support systems rather than metacognitive judgment, it may interact with fluency illusion in AI-mediated learning environments. Fluent and confident system responses can encourage both uncritical acceptance of the output and inflated perceptions of understanding.

3.3. Distinguishing Fluency Illusion from Related Metacognitive Illusions

Several related constructs in the metacognition literature describe situations in which learners misjudge their own understanding. Although these concepts are closely related, they originate from different theoretical traditions and describe distinct mechanisms of miscalibrated learning judgments. The illusion of competence refers broadly to situations in which learners overestimate their level of mastery despite limited ability to recall or apply knowledge independently. Similarly, the illusion of knowing arises when familiarity or recognition is mistaken for genuine understanding.

A related but conceptually distinct phenomenon is the illusion of explanatory depth, which describes individuals’ tendency to believe that they understand complex systems or concepts more deeply than they actually do. When learners are asked to produce detailed explanations of such systems, this perceived understanding often collapses [24]. This effect has been widely observed in domains involving causal reasoning and conceptual explanation.

The fluency illusion, as used in this review, refers more specifically to misjudgments that arise from the subjective ease with which information is processed. When explanations are clear, coherent, and easy to follow, learners may interpret this processing ease as evidence of genuine understanding. In AI-mediated learning environments, fluent explanations generated by systems such as ChatGPT may amplify this effect. In particular, in STEM learning contexts where students frequently request step-by-step conceptual explanations, fluency illusion may interact with the illusion of explanatory depth by encouraging learners to equate the ability to follow an explanation with the ability to construct or apply that explanation independently.

3.4. Surface Coherence and Deep Learning

The distinction between surface coherence and deep learning provides an important lens for interpreting these effects [25]. Surface coherence refers to the internal consistency, readability, and rhetorical smoothness of information, whereas deep learning involves conceptual integration, inferential reasoning, and flexible application. Highly coherent explanations can therefore create a sense of understanding without engaging the cognitive processes required for durable learning. This distinction helps explain why learners may feel confident while remaining unable to apply or extend knowledge beyond familiar contexts.

These dynamics can also be interpreted through the lens of Cognitive Load Theory, which highlights the role of cognitive effort in effective learning [26,27]. While reducing unnecessary cognitive load can improve learning efficiency, highly fluent instructional materials may also reduce the depth of learners’ engagement with the underlying concepts. In such situations, AI-generated explanations that minimize perceived difficulty may inadvertently reduce the level of productive cognitive effort invested by learners. A related perspective is offered by the Desirable Difficulties framework [28], which suggests that learning conditions involving manageable challenges, such as retrieval practice, problem generation, or effortful explanation, often lead to stronger long-term retention and transfer. From this viewpoint, the fluency of AI-generated responses may create learning conditions that feel cognitively efficient in the short term but limit opportunities for the effortful processing that supports durable learning.

Generative AI systems introduce a qualitatively distinct form of fluency that is externally generated and dynamically adaptive [29]. Unlike traditional instructional resources, these systems produce context-sensitive responses that closely align with user prompts, creating interactions that feel personalized and authoritative [30]. As a result, the source of fluency may shift from the learner’s own cognitive activity to the system itself. In learning situations where such output is readily available, learners may experience clarity and resolution that reflect the fluency of the response rather than their own evaluative competence.

The conversational nature of large language models further amplifies this effect by fostering a sense of dialogue and understanding [31]. The coherence and plausibility of generated responses can create strong impressions of preparedness and comprehension, even when learners have not engaged in the reasoning processes required to articulate, defend, or adapt those ideas independently. This highlights how fluency illusion can extend beyond private study to influence participation, confidence, and performance in learning environments.

By integrating insights from processing fluency and metacognition, this review conceptualizes fluency illusion as a key explanatory mechanism shaping learners’ experiences with fluent external supports [32]. Rather than positioning such systems as inherently beneficial or detrimental, this perspective emphasizes how fluent outputs interact with established cognitive biases, potentially undermining accurate self-assessment. Recognizing this mechanism provides a theoretical foundation for interpreting empirical findings on AI-supported learning and for designing pedagogical and assessment practices that promote active engagement, metacognitive awareness, and genuine understanding.

The strength of fluency-driven misjudgments is unlikely to be uniform across learners. Factors such as prior domain knowledge, metacognitive monitoring ability, epistemic motivation, and instructional context may moderate the relationship between processing fluency and perceived understanding. Learners with stronger prior knowledge or more developed metacognitive regulation may be better able to recognize the limitations of fluent AI-generated explanations and therefore less likely to equate processing ease with genuine understanding [33].

4. ChatGPT in Classroom Settings: What the Literature Shows

The growing body of literature on ChatGPT in education reflects a wide range of classroom applications, methodological approaches, and disciplinary perspectives [34]. Rather than converging on a single narrative of benefit or risk, existing studies reveal patterned tensions between perceived usefulness and concerns about learning depth, learner autonomy, and assessment validity. An integrative synthesis of this work highlights how ChatGPT’s pedagogical affordances are frequently accompanied by cognitive effects linked to fluency, even when these effects are not explicitly named or theorized by the original authors.

Table 1. Key Gaps and Future Research Directions in AI-Mediated Learning.

Research Gap	Rationale/Implications	Suggested Research Approaches
Lack of longitudinal studies	Short-term studies cannot capture the development of fluency illusion or long-term learning outcomes	Longitudinal designs tracking metacognitive calibration, cognitive strategies, and dependence on AI across semesters
Overreliance on self-reported learning	Self-reports may reflect perceived understanding shaped by fluent AI output rather than actual learning	Triangulate with performance-based assessments, error detection tasks, and delayed transfer measures
Limited attention to metacognition	Few studies measure how students monitor and regulate their learning with AI	Incorporate explicit metacognitive measures such as confidence calibration, reflective prompts, and think-aloud protocols
Need for classroom-level observational research	Controlled studies may not capture interactional and contextual factors affecting AI use	Mixed-methods approaches combining observation, artifact analysis, and instructor interviews in authentic classroom settings
Theory-driven integration	Many studies lack a unifying framework to interpret findings	Use fluency illusion as an organizing construct to connect cognitive, metacognitive, and educational research

summarizes recurring applications, benefits, concerns, and associated fluency-related patterns observed across classroom contexts.

Figure 3 illustrates the conceptual distinction between surface performance and deep learning in AI-mediated environments. In many AI-supported tasks, fluent explanations can help learners complete problems quickly and produce well-structured responses. However, this apparent success may reflect surface-level performance rather than deeper conceptual integration. The figure highlights how learners may successfully follow or reproduce AI-generated reasoning while still lacking the capacity to independently construct or adapt similar explanations. This distinction helps clarify why students may perform adequately on routine tasks yet struggle with problems requiring conceptual transfer or independent reasoning.

Figure 4 summarizes the different contexts in which fluency-driven misinterpretations of learning may emerge. Across writing, STEM problem solving, feedback interactions, and assessment contexts, the common mechanism involves the tendency to interpret coherent and well-structured AI output as evidence of understanding. While the specific instructional situations differ, the underlying cognitive dynamic remains similar: learners rely on the perceived clarity of external explanations rather than on demonstrable reasoning or knowledge generation. The figure therefore emphasizes that fluency illusion is not limited to a single domain but represents a cross-contextual challenge for AI-mediated learning environments.

The final review corpus consisted of 41 studies that met the inclusion criteria described in Section 2. The studies were distributed across four major instructional contexts discussed in this review. Seventeen studies examined the use of ChatGPT in writing and language learning settings, eleven focused on STEM-related problem solving and conceptual explanation, seven addressed the use of generative AI for feedback or tutoring purposes, and six investigated issues related to assessment and academic integrity. This distribution reflects the current concentration of research activity in writing-related applications while also highlighting the growing attention to AI-supported learning across other instructional domains.

4.1. ChatGPT in Writing and Language Learning

Research on writing and language learning constitutes one of the most active areas of inquiry into ChatGPT’s classroom use [10]. Studies consistently report that students perceive AI-generated text as helpful for brainstorming, drafting, paraphrasing, and improving linguistic accuracy [11]. In second language contexts, ChatGPT is frequently described as a source of accessible input, vocabulary expansion, and grammatical support, particularly for learners with limited confidence or exposure to proficient language models [35]. From an instructional perspective, these uses are commonly framed as scaffolding mechanisms that reduce cognitive load and support engagement.

Although a growing number of studies have examined the use of generative AI tools in writing instruction, the research varies considerably in its methodological design and evaluation approaches.

At the same time, the literature raises concerns about students’ reliance on fluent AI-generated prose and the implications for writing-related learning outcomes [36]. Several studies report that while students express increased confidence after using ChatGPT, this confidence does not consistently translate into improvements in argumentation quality, rhetorical control, or independent writing performance [12,36,37,38]. Linguistic polish can obscure weaknesses in organization or reasoning, leading students to overestimate their writing competence. Even when studies focus primarily on authorship or originality, fluency-related effects are evident in the tendency to treat well-formed output as a proxy for quality and learning.

Although these studies generally report improvements in perceived writing quality or argument organization, the methods used to evaluate such outcomes vary considerably across the literature. In some cases, argumentative quality is assessed through rubric-based evaluations focusing on elements such as claim clarity, evidence integration, coherence, and rhetorical structure. Other studies rely on instructor grading, automated text analysis tools, or self-reported perceptions of writing improvement. Differences in participant populations, task types, and outcome measures further contribute to methodological heterogeneity. For example, some investigations examine short argumentative essays written under controlled classroom conditions, while others analyze revisions produced during longer writing assignments or AI-assisted drafting processes. These variations make direct comparison across studies difficult and highlight the need for more standardized evaluation frameworks when assessing the impact of generative AI on writing development.

An additional distinction emerging in the literature concerns the different ways in which students use generative AI during writing tasks. In some cases, ChatGPT functions as a partial support tool for brainstorming, editing, or improving linguistic clarity. In other situations, students rely on the system to generate substantial portions of the written content. These different forms of use can lead to substantially different learning outcomes. Improvements in linguistic fluency or grammatical accuracy do not necessarily translate into stronger argumentative structure or deeper conceptual reasoning. Distinguishing between linguistic support and authorship substitution is therefore important when interpreting reported improvements in writing performance.

To clarify these distinctions,

Table 2. Common forms of ChatGPT use in writing tasks reported in the literature and their potential learning implications.

Type of AI Use	Typical Student Activity	Immediate Benefit	Potential Learning Risk
Linguistic support	Grammar correction, paraphrasing, vocabulary suggestions	Improved language fluency and readability	Surface-level improvement may be mistaken for stronger writing competence [10,39]
Partial assistance	Brainstorming ideas, outlining arguments, revising drafts	Faster idea generation and clearer structure	Students may rely on AI for organization without strengthening reasoning [11,36]
Substantial drafting support	AI generates paragraphs or sections based on prompts	Faster completion and more polished text	Reduced engagement in constructing arguments independently [11,12]
Full authorship substitution	AI generates most or all of the final text	High-quality surface output with minimal effort	Limited development of writing, reasoning, and conceptual understanding [12,37]

summarizes common forms of ChatGPT use in writing tasks and their potential learning implications.

This comparison shows that not all uses of ChatGPT have the same learning implications. When the tool is used for limited support such as grammar correction or idea generation, it can function as a scaffold that helps students improve their work. However, when students rely on the system to generate large portions of text, the cognitive work involved in organizing arguments and developing ideas may be reduced.

4.2. Problem-Solving and STEM Support

In STEM-oriented classroom settings, ChatGPT is commonly examined as a tool for explaining concepts, outlining solution procedures, or supporting debugging and code comprehension [13]. Students frequently report that AI-generated explanations help clarify abstract ideas, translate formal notation into natural language, and provide timely assistance outside scheduled instructional time. In introductory courses, such support is often associated with reduced frustration and increased persistence.

The interpretation of AI-supported problem solving in STEM contexts can also be informed by broader instructional research on cognitive load and mathematical problem-solving strategies. Ngu and Phan [40] examine how instructional guidance and cognitive load influence the development of mathematical proficiency. Although their study does not directly investigate the use of ChatGPT or other generative AI tools, it offers relevant theoretical insights into how external instructional supports may shape learners’ engagement with complex problem-solving tasks.

Several studies show that students who rely heavily on AI assistance perform adequately on routine or near-transfer tasks but struggle with novel or conceptually demanding problems. These findings suggest that recognizing coherent reasoning is not equivalent to developing the ability to generate or adapt solutions independently. While such outcomes are often discussed in terms of dependency or surface learning, they align closely with fluency-related misjudgments observed across instructional contexts.

These findings are also consistent with a long-standing distinction in learning research between studying worked examples and generating solutions independently. Studies in cognitive load theory have shown that worked examples can support initial comprehension by reducing problem-solving complexity, particularly for novice learners. However, deeper learning often requires opportunities for learners to generate explanations or solutions themselves. AI-generated step-by-step explanations may function similarly to worked examples by supporting immediate task completion while reducing the need for learners to actively construct the reasoning process. As a result, students may perform adequately during guided tasks but encounter difficulties when attempting to transfer the same reasoning to new problems.

Recent studies and systematic reviews examining the use of generative AI in education report similar patterns. Reviews of ChatGPT use in STEM higher education indicate that students often rely on AI-generated explanations to understand procedures or verify intermediate steps in problem solving [13]. While this support can improve immediate task completion and reduce frustration, it may not always strengthen independent reasoning skills. Related work on over-reliance on AI dialogue systems also suggests that frequent dependence on AI-generated solutions can reduce opportunities for learners to engage in the active reasoning processes that support durable conceptual learning [14]. These findings reinforce the distinction between recognizing correct explanations and developing the ability to generate or adapt problem-solving strategies independently.

Recent classroom studies in programming and quantitative subjects show similar patterns. Students who use ChatGPT often complete routine exercises more quickly and report that explanations appear clearer. However, some studies also report that students experience difficulties when they are asked to solve similar problems independently or explain the reasoning without AI assistance. These findings suggest that recognizing a correct explanation is not always the same as developing the ability to construct the solution independently [13,14,40].

4.3. Feedback and Tutoring Functions

Another prominent theme in the literature concerns the use of ChatGPT as a feedback provider or tutoring agent [41]. Studies highlight the perceived value of immediate, detailed, and nonjudgmental feedback, particularly in large classes where instructor feedback is limited. Students often describe AI-generated feedback as clearer and more actionable than peer feedback, and instructors note potential efficiency gains when AI is used for formative purposes.

At the same time, concerns arise regarding the epistemic authority attributed to fluent AI feedback. Because responses are typically confident, well structured, and tailored to user input, students may accept them uncritically, even when feedback is incomplete, generic, or subtly inaccurate. The conversational format can further reinforce a sense of personalized understanding, encouraging reliance on AI validation rather than reflective evaluation. Although few studies explicitly frame this phenomenon as a metacognitive issue, the literature suggests that fluency-driven acceptance may occur when students interact with highly polished AI-generated feedback.

Empirical studies on automated feedback uptake provide related evidence. Cavalcanti et al. [42] report similar patterns in research on automated feedback systems, showing that students frequently incorporate feedback suggestions into their revisions without critically evaluating the underlying reasoning, particularly when the feedback is presented in a clear and authoritative form.

These dynamics are also closely related to research on trust in automated systems and human–computer interaction. Research on trust calibration in human–AI interaction suggests that users may place excessive trust in automated systems when responses appear confident, coherent, or authoritative, which may reduce critical evaluation of AI-generated feedback [8]. In educational settings, this tendency may influence how students interpret and use AI-generated feedback. The effects may also vary depending on the type of feedback provided. Corrective feedback typically focuses on identifying specific errors or inaccuracies in a learner’s response and can support targeted revision when students actively engage with it. In contrast, elaborative feedback provides extended explanations intended to deepen conceptual understanding.

In AI-mediated learning environments, highly fluent elaborative feedback may increase the risk of passive acceptance if learners rely on the explanation without critically evaluating the reasoning behind it. Factors such as prior knowledge, digital literacy, and experience with AI tools may therefore moderate how students interpret and use AI-generated feedback. This issue can also be understood through research on trust in human–AI interaction. Studies in this field show that users may place excessive trust in automated systems when responses appear confident, clear, or authoritative. In educational settings, the fluent style of AI-generated feedback may therefore influence how students evaluate its quality. Students may accept feedback more readily because it appears well structured, even when they do not fully examine the reasoning behind it.

4.4. Assessment and Academic Integrity

Assessment and academic integrity are recurring concerns in the literature, often shaped by institutional anxieties surrounding plagiarism, authorship, and fairness [43]. Reported instructional responses include redesigned assessments, greater emphasis on in-class work, and the adoption of reflective or process-oriented tasks. Some studies indicate that when assessments emphasize reasoning, explanation, or oral defense, the direct influence of ChatGPT-assisted output on performance is reduced.

Beyond integrity concerns, the literature also highlights challenges related to assessment validity. Fluent AI-generated responses can make it difficult to distinguish between independent understanding and tool-assisted performance, particularly in written or take-home assessments. Students may also interpret successful completion of AI-assisted tasks as evidence of learning, even when assessments fail to capture transferable competence. These dynamics suggest that fluency-related effects operate not only at the level of individual learning but also at the level of instructional evaluation.

These concerns are increasingly addressed through emerging institutional frameworks that seek to clarify appropriate uses of generative AI in assessment contexts. Universities and professional bodies have begun to articulate policies that distinguish between acceptable forms of AI-assisted learning and practices that undermine independent demonstration of knowledge [44]. At the same time, approaches to AI governance vary considerably across institutions, disciplinary traditions, and national educational systems, reflecting differences in assessment culture and regulatory environments [45]. An important pedagogical distinction also arises between formative and summative assessment contexts. In formative settings, AI-generated explanations and feedback may support practice, reflection, and skill development when used transparently. In contrast, in summative assessments where students are expected to demonstrate independent understanding, fluency-driven misjudgments may be more problematic because highly coherent AI-generated responses can mask gaps in reasoning or conceptual mastery.

These concerns have also encouraged renewed interest in forms of authentic assessment and assessment for learning. Authentic assessment tasks that require explanation, application, or contextual reasoning may reduce the likelihood that fluent AI-generated responses can substitute for genuine understanding. Examples include oral defenses of written work, iterative project submissions that document the development of ideas over time, or reflective commentaries in which students explain how AI tools were used during the learning process. Such approaches shift the emphasis from the final product to the reasoning and learning processes that produced it.

4.5. Integrative Perspective

Across these thematic areas, a consistent pattern becomes visible. ChatGPT is widely perceived as useful and efficient, yet these benefits are frequently accompanied by confidence gains that are not consistently matched by evidence of deep or transferable learning [46]. When viewed through the lens of fluency illusion, these findings suggest that the central challenge posed by ChatGPT in classroom settings lies in the misalignment between fluency-driven perceptions and demonstrable understanding.

By synthesizing research across writing, STEM, feedback, and assessment contexts, this review highlights fluency illusion as a unifying explanatory framework underlying diverse concerns in AI-mediated education. Recognizing this mechanism supports a more coherent interpretation of existing findings and provides a foundation for pedagogical and assessment strategies that address not only what students produce with AI, but how they evaluate and regulate their own learning in its presence.

5. Conceptualizing Fluency Illusion in AI-Mediated Learning

Despite the growing volume of research on generative AI in education, a coherent conceptual explanation of how fluent AI output shapes learners’ judgments of understanding remains underdeveloped [47]. This section advances the concept of fluency illusion in AI-mediated learning as a theoretically grounded lens for interpreting recurring patterns across classroom-based studies. By clarifying its meaning, unpacking its underlying mechanism, and outlining its broader explanatory value, the section positions fluency illusion as a central construct for understanding the educational impact of ChatGPT.

At its core, fluency illusion in AI-mediated learning refers to the misattribution of the clarity, coherence, or apparent correctness of AI-generated content to one’s own understanding or competence. What distinguishes this phenomenon from earlier accounts of learning illusions is the source of fluency itself. Rather than emerging from the learner’s cognitive processing or from instructional design alone, fluency is produced externally by an adaptive system that responds directly to the learner’s inputs, often with remarkable linguistic and conceptual polish.

This externally generated fluency initiates a subtle but powerful shift in metacognitive judgment [17]. When learners encounter responses that are easy to follow and rhetorically convincing, the subjective experience of understanding intensifies. Processing ease is interpreted as mastery, leading to elevated confidence and judgments of learning that may not reflect actual conceptual integration. As perceived understanding increases, learners may reduce cognitive effort, rely more heavily on the tool, or disengage from strategies that would otherwise promote deeper learning.

Importantly, this mechanism unfolds in ways that distinguish AI-mediated learning from more familiar instructional contexts. Traditional sources of fluency, such as well-designed textbooks or clear lectures, are relatively stable and limited in their responsiveness. By contrast, generative AI systems dynamically tailor their output to individual prompts, reinforcing fluency through personalization and conversational continuity. This responsiveness strengthens the learner’s inference that understanding has been achieved through interaction, even when key reasoning processes remain externalized.

For analytical clarity, the relationships illustrated in Figure 4 can also be interpreted in terms of a simplified causal structure. In this framework, the external fluency of AI-generated explanations functions as the primary independent variable. This fluency shapes learners’ judgments of understanding through mediating processes such as increased confidence, reduced perceived cognitive effort, and reliance on surface indicators of comprehension. The strength of these effects may vary across learners and instructional situations. Factors such as prior knowledge, metacognitive monitoring ability, epistemic motivation, and instructional context may therefore act as moderating variables that influence how strongly processing fluency affects perceived understanding and subsequent learning behavior. The model should therefore be interpreted as a theoretical representation of potential relationships rather than as an empirically validated causal mechanism.

Viewed as a process rather than a momentary bias, fluency illusion can be conceptualized as a recursive cycle [48]. Initial interaction with fluent AI output elevates perceived understanding, which in turn shapes learners’ regulatory decisions about effort, strategy use, and task engagement. Reduced cognitive investment limits opportunities for retrieval, elaboration, and transfer, resulting in learning outcomes that are shallow yet subjectively satisfying. Subsequent interactions with AI reinforce this cycle by repeatedly providing fluent resolutions, thereby stabilizing the illusion across learning episodes.

This conceptualization also underscores that fluency illusion is not solely an individual cognitive tendency but an interactional outcome shaped by the affordances of AI systems. Features such as authoritative tone, structured reasoning, and conversational turn-taking contribute to a sense of epistemic reliability that is difficult for learners to interrogate. In classroom environments characterized by time pressure and performance demands, these affordances may further encourage reliance on fluency as a heuristic for learning success.

Framing fluency illusion in this way helps reconcile apparent inconsistencies in the existing literature [49]. Reports of heightened confidence, satisfaction, and perceived learning can coexist with evidence of limited transfer, fragile understanding, or dependence on AI support. Rather than reflecting contradictory findings, these patterns are consistent with a systematic misalignment between subjective experience and learning outcomes that fluent AI output is particularly well positioned to amplify.

By articulating the role of fluency illusion in AI-mediated learning, this review offers a unifying theoretical perspective that connects cognitive psychology with emerging classroom research on generative AI. This framing shifts attention away from binary evaluations of AI use and toward the cognitive processes that shape how learners experience and interpret their own learning. In doing so, it lays the groundwork for pedagogical and assessment approaches that seek not only to leverage AI’s capabilities, but also to counteract fluency-driven misjudgments and support meaningful, self-regulated learning in AI-rich educational contexts.

Conceptualizing fluency illusion as a central mechanism in AI-mediated learning has direct implications for how ChatGPT is integrated into classroom practice and how learning is evaluated [50]. If fluent AI output systematically shapes learners’ perceptions of understanding, then pedagogical and assessment decisions cannot rely solely on observable performance or polished products as indicators of learning. Instead, instructional design must account for the ways in which AI-generated fluency can obscure gaps in understanding and distort metacognitive judgment. This perspective shifts the focus from whether generative AI is used to how its use is structured, monitored, and assessed within educational settings. The following section builds on this conceptual framework to examine how teaching practices and assessment strategies can be adapted to mitigate fluency-driven misjudgments while preserving the potential benefits of AI-supported learning.

6. Pedagogical and Assessment Implications

Recognizing fluency illusion as an important explanatory factor in AI-mediated learning requires reconsidering how generative AI tools such as ChatGPT are integrated into instructional and assessment practices [51]. Rather than framing AI use as inherently beneficial or inherently problematic, a fluency-based perspective highlights the conditions under which fluent AI output supports learning and the conditions under which it may distort learners’ judgments of understanding. The pedagogical implications discussed in this section therefore focus on instructional and assessment strategies that mitigate fluency-driven misjudgments while preserving the learning benefits of AI-supported activities. Accordingly, Figure 5 summarizes the instructional, assessment, and metacognitive design strategies discussed in this section for reducing fluency-driven overconfidence while maintaining the advantages of AI-supported learning.

6.1. Managing the Risk of Over-Reliance

One of the most immediate pedagogical concerns associated with fluency illusion is the risk of over-reliance on AI-generated content [14]. When students experience repeated success through interaction with fluent AI responses, they may increasingly defer cognitive responsibility to the tool. This shift can reduce engagement in effortful processes such as problem formulation, hypothesis testing, and self-explanation. Over time, learners may come to associate understanding with access to fluent output rather than with their own capacity to reason or apply knowledge.

Instructional design can address this risk by deliberately structuring AI use as a support rather than a substitute for cognitive effort [52]. For example, activities that require students to generate initial responses before consulting ChatGPT, or to compare their own reasoning with AI-generated explanations, can help maintain the learner’s role as the primary agent of understanding. Such approaches do not restrict access to AI but instead reposition it within a learning sequence that prioritizes active engagement.

Early classroom experiences suggest that such strategies can often be implemented without major changes to existing course structures. For example, some instructors ask students to submit an initial attempt at solving a problem or outlining an argument before consulting ChatGPT. Others incorporate brief comparison activities in which students examine differences between their own reasoning and AI-generated explanations. Although systematic evaluations of these practices are still limited, preliminary classroom reports suggest that they can encourage more active engagement with the material while preserving the practical advantages of AI-supported assistance.

In practice, several simple strategies can help instructors integrate AI tools while maintaining active learning. For example, students may first attempt a problem or write a draft before consulting ChatGPT. Instructors can also ask students to explain how they used AI assistance or to compare their own reasoning with the AI-generated explanation. These approaches help keep students actively engaged while still allowing them to benefit from AI-supported guidance [53,54].

6.2. Rethinking Feedback Practices

ChatGPT’s ability to provide immediate, detailed, and linguistically polished feedback has clear appeal in classroom contexts, particularly where instructor feedback is limited [55]. However, from the perspective of fluency illusion, the clarity and confidence of AI-generated feedback can be misleading. Students may interpret fluent feedback as authoritative and complete, reducing their inclination to question, revise, or seek alternative perspectives.

To counter this tendency, feedback practices should emphasize interpretation rather than acceptance. Instructors can encourage students to treat AI feedback as a starting point for reflection by requiring them to explain how they used the feedback, which aspects they accepted or rejected, and why. Embedding reflective prompts or requiring justification of revisions can shift attention from the fluency of feedback to its alignment with learning objectives and disciplinary standards.

Emerging classroom experiences provide initial examples of how such strategies can be implemented in practice. In some courses, instructors ask students to submit an initial solution or written draft before consulting ChatGPT, followed by a short reflection comparing their own reasoning with AI-generated responses. Other implementations include brief in-class explanations, oral follow-up questions, or reflective prompts that require students to justify how they used AI assistance. These approaches can encourage active engagement with the material while still allowing students to benefit from AI-supported guidance. However, such practices also raise practical considerations related to instructor workload and institutional policies.

Implementing these practices also requires attention to practical constraints faced by instructors. In large classes, closely reviewing AI-assisted revisions may increase instructor workload if additional monitoring or reflection tasks are introduced. Some instructors have therefore experimented with short reflection prompts in which students explain how they interpreted and applied AI-generated feedback. These brief reflections can provide insight into students’ reasoning without substantially increasing grading demands. At an institutional level, clearer guidance on appropriate uses of AI-generated feedback may also help instructors design assignments that encourage critical engagement rather than passive acceptance.

6.3. Designing Assessments to Counter Fluency Illusion

Assessment design plays a critical role in either amplifying or mitigating fluency illusion [56]. Traditional assessments that emphasize polished written responses or procedural correctness may inadvertently reward fluency-driven performance, particularly when AI assistance is available. In such cases, high-quality output may mask limited understanding, making it difficult to distinguish between independent competence and tool-supported performance.

Assessments that foreground explanation, reasoning, and transfer are better positioned to counter fluency illusion [57]. Tasks that require students to articulate their thinking, apply concepts in novel contexts, or respond to follow-up questions can reveal the extent to which understanding has been internalized. Oral assessments, reflective components, and process-oriented submissions can further support this goal by shifting emphasis from final products to learning trajectories. Importantly, these approaches do not require the exclusion of AI but rather make learners’ engagement with content and tools more transparent.

At the same time, implementing such assessment approaches requires consideration of practical constraints. Large-enrollment courses, standardized assessment formats, and institutional policies may limit the use of extensive oral examinations or complex project-based evaluations. In such cases, hybrid approaches may offer a workable alternative. For example, instructors may combine AI-supported practice activities with shorter in-class explanations, brief oral follow-up questions, or reflective summaries in which students explain the reasoning behind their responses. These formats can maintain logistical feasibility while still encouraging students to demonstrate independent understanding.

6.4. Supporting Metacognitive Scaffolding

Given that fluency illusion operates primarily through distorted metacognitive judgment, explicit metacognitive scaffolding is essential [58]. Students cannot be expected to recognize fluency-driven misjudgments without guidance, particularly in AI-rich learning environments where fluent output is the norm. Instructional interventions that make the distinction between perceived understanding and demonstrated understanding explicit can help learners develop more accurate self-monitoring skills.

Metacognitive scaffolding can take the form of structured self-assessment prompts, prediction and reflection activities, or guided comparisons between AI-assisted and unassisted performance [59]. For instance, asking students to predict their performance before assessment or to reflect on discrepancies between confidence and outcomes can surface fluency effects that would otherwise remain implicit. Over time, such practices can foster more calibrated judgments of learning and reduce dependence on fluency as a heuristic.

An important conceptual tension arises when generative AI tools themselves are used to deliver metacognitive scaffolding. While reflective prompts and self-assessment activities are intended to help students monitor and evaluate their own understanding, AI-generated reflection templates may introduce a secondary layer of fluency. When prompts appear pedagogically sophisticated and easy to complete, students may experience a sense of metacognitive engagement without necessarily performing the underlying evaluative processes. In such cases, learners may follow the structure of AI-generated reflection rather than actively interrogating their own reasoning. This possibility suggests that metacognitive scaffolding supported by AI should be designed in ways that require explanation, justification, or independent articulation rather than passive completion of reflection prompts.

Assessing the effectiveness of these interventions also requires appropriate ways of measuring metacognitive calibration. Future studies could combine traditional learning assessments with measures that capture the relationship between confidence and performance. For instance, students may be asked to estimate how well they expect to perform before completing a task and later compare those estimates with their actual results. Such approaches can help identify situations in which fluent AI-generated explanations increase perceived understanding without corresponding improvements in learning outcomes. Experimental studies that compare AI-assisted and non-assisted learning conditions could also provide useful evidence on how different forms of scaffolding influence students’ judgments of understanding.

6.5. Implications for Instructional Culture

Beyond individual practices, addressing fluency illusion requires a broader shift in instructional culture [60]. When classrooms prioritize efficiency, coverage, or output quality without attending to how understanding is constructed, fluency-driven misjudgments are more likely to persist. By contrast, environments that value explanation, questioning, and intellectual struggle create space for learners to recognize the limits of fluent assistance.

In this sense, the pedagogical challenge posed by ChatGPT is not unique but rather an amplification of long-standing tensions between performance and learning [61]. Fluency illusion provides a lens through which educators can critically examine these tensions and design learning environments that leverage AI’s capabilities while safeguarding the development of genuine understanding.

6.6. Implementation Considerations Across Instructional Contexts

The feasibility of these strategies may vary across instructional contexts and institutional settings. Although the empirical evidence base remains limited, emerging studies on AI-supported learning environments provide initial indications of how such approaches can be implemented in practice. In large-enrollment courses, structured comparison tasks and reflective prompts have been used to encourage students to critically evaluate AI-generated explanations without requiring extensive additional grading by instructors. In resource-constrained settings, simple interventions such as requiring students to explain or justify AI-assisted answers have been shown to promote deeper engagement with the material while preserving the efficiency benefits of AI support.

In educational systems with high-stakes assessment cultures, instructors have increasingly experimented with hybrid approaches that combine AI-supported practice activities with supervised or oral assessments designed to verify independent understanding. Although systematic evaluation of these interventions remains limited, early classroom implementations suggest that pedagogical designs emphasizing explanation, comparison, and reflection can mitigate some of the fluency-related risks associated with generative AI use.

7. Gaps and Future Research Directions

Future research on AI-supported learning should focus on several key priorities. One important direction is examining whether fluent AI explanations increase the gap between students’ perceived understanding and their actual performance. Another area concerns identifying the conditions under which AI support contributes to deeper learning rather than simply enabling faster task completion. In addition, research is needed to evaluate instructional approaches that allow students to benefit from AI tools while maintaining critical thinking and independent reasoning.

Although research on generative AI in education has expanded rapidly, the empirical evidence base remains relatively limited and uneven across methodological approaches [9,34]. In the short term, an important priority for future research is to develop controlled experimental and classroom-based studies that directly examine how AI-generated fluency influences learners’ judgments of understanding, confidence, and learning outcomes. Over a longer time horizon, research should move toward broader investigations that explore how these cognitive effects interact with instructional design, disciplinary contexts, and students’ developing metacognitive skills. Establishing such a progression of research priorities may help clarify which questions can be addressed immediately and which require more sustained empirical investigation.

While the existing literature offers valuable insights into how ChatGPT is being adopted in classroom settings, it remains fragmented in ways that limit cumulative understanding of its cognitive and pedagogical effects [34]. In particular, the phenomenon of fluency illusion, though implicit in many findings, has yet to be systematically examined as a central explanatory construct. Addressing this gap requires a research agenda that moves beyond short-term evaluations and surface-level outcomes toward designs that capture learning processes, metacognitive dynamics, and classroom realities over time. Table 1 illustrates the key gaps and future research directions in AI-mediated learning.

7.1. Moving Beyond Short-Term and Cross-Sectional Designs

A prominent limitation of current research lies in its reliance on short-term and cross-sectional studies. Many investigations examine students’ perceptions, performance, or satisfaction following brief exposure to ChatGPT, often within a single task or instructional unit [46,62]. While such studies are useful for identifying immediate affordances and concerns, they provide limited insight into how fluency-driven misjudgments develop, stabilize, or attenuate over time. Longitudinal research is needed to examine whether repeated interaction with fluent AI output leads to enduring changes in learners’ metacognitive calibration, study strategies, or dependence on external support.

Longer-term designs would also allow researchers to explore how instructional interventions influence the trajectory of fluency illusion. For example, studies could examine whether early exposure to metacognitive scaffolding alters how students interpret fluent AI output across a semester or academic year. Such work would provide a more robust empirical foundation for pedagogical recommendations.

7.2. Reducing Overreliance on Self-Reported Learning

Another recurring feature of the literature is the heavy reliance on self-reported measures of learning, confidence, and perceived usefulness [63]. While these measures are informative, they are particularly susceptible to fluency effects, as learners’ judgments may be directly shaped by the clarity and coherence of AI-generated responses. As a result, positive self-reports may reflect perceived understanding rather than demonstrable learning.

Future research would benefit from triangulating self-reports with performance-based and process-oriented measures. Designs that compare perceived understanding with independent task performance, delayed transfer, or error detection can help disentangle subjective experience from learning outcomes. Such approaches are especially important for studying fluency illusion, as the construct itself concerns the divergence between perception and competence.

7.3. Examining Metacognition as a Central Variable

Although metacognition is often invoked in discussions of AI-supported learning, it is rarely treated as a primary object of empirical investigation [53]. Few studies directly measure how students monitor, evaluate, and regulate their learning when interacting with ChatGPT. This represents a significant gap, given that fluency illusion operates through metacognitive judgment rather than through content acquisition alone.

Future work should more explicitly examine how learners form judgments of understanding in AI-mediated contexts and how these judgments influence subsequent learning behavior. Experimental and mixed-methods studies that capture metacognitive processes, such as confidence calibration, prediction accuracy, and reflective reasoning, would deepen understanding of when and for whom fluency illusion is most likely to occur. Attention to individual differences, such as prior knowledge or epistemic beliefs, may further refine this line of inquiry.

7.4. Advancing Classroom-Level and Ecologically Valid Research

Much of the existing research on ChatGPT in education is conducted under controlled or simulated conditions, often outside of authentic classroom settings [54]. While such designs offer methodological clarity, they may overlook interactional and contextual factors that shape how fluency illusion manifests in practice. Classroom-level observational research is needed to capture how students actually integrate AI tools into their learning routines and how instructors respond to emerging patterns of use.

Naturalistic studies that combine observation, artifact analysis, and instructor interviews can shed light on how fluency-driven perceptions influence participation, feedback uptake, and assessment performance. Such work would also help clarify how institutional norms, assessment practices, and disciplinary cultures interact with AI-generated fluency to shape learning experiences.

7.5. Toward Theory-Driven and Cumulative Research

Finally, future research would benefit from more explicit theoretical integration. Many studies implicitly reference concepts related to fluency, confidence, or overreliance, yet few situate their findings within a coherent cognitive or educational framework. Treating fluency illusion as an organizing construct can facilitate cumulative knowledge building by providing shared terminology and explanatory mechanisms.

By aligning empirical designs with theory-driven questions about perception, regulation, and learning, future research can move beyond isolated findings toward a more coherent understanding of ChatGPT’s role in education. Such an approach not only advances scholarship but also supports the development of pedagogical practices that are responsive to the cognitive realities of AI-mediated learning.

Future research would also benefit from more systematic experimental designs that isolate the cognitive mechanisms underlying fluency illusion in AI-supported learning environments. For example, studies could compare learning outcomes across conditions in which students receive fluent AI-generated explanations, less polished explanations, or no external assistance. Such comparisons could examine how differences in fluency influence confidence judgments, perceived understanding, and performance on transfer tasks. Longitudinal studies may also help clarify whether repeated reliance on fluent AI output affects the development of self-regulated learning strategies over time. Framing these investigations around clearly defined research questions will help move the field beyond descriptive observations toward a more cumulative understanding of AI-mediated learning.

Building on the research gaps identified in this review, several research directions emerge that may guide future investigation. Future research should examine the extent to which exposure to fluent AI-generated explanations increases the gap between perceived understanding and actual learning performance. It should also investigate the instructional conditions under which AI assistance supports conceptual transfer rather than only immediate task completion. Another important direction is to explore how learner characteristics, such as prior knowledge, digital literacy, and metacognitive monitoring ability, influence susceptibility to fluency-driven misjudgments. Finally, future studies should evaluate which instructional and assessment strategies are most effective in reducing overconfidence while preserving the potential learning benefits of generative AI tools.

8. Descriptive Bibliometric Overview

The purpose of this bibliometric overview is not to provide a comprehensive scientometric analysis but rather to offer a descriptive snapshot of the rapidly emerging literature on ChatGPT in educational contexts. The analysis aims to illustrate publication trends and the disciplinary spread of research on generative AI in education, thereby situating the present conceptual review within the broader scholarly landscape.

Because the objective of this section is to provide a descriptive overview rather than a full bibliometric mapping, advanced bibliometric network analyses such as co-citation analysis, bibliographic coupling, or keyword co-occurrence mapping were not conducted.

The Dimensions database [64] was used to retrieve publication metadata using a keyword-based search applied to titles and abstracts. The search query used in the database was: (“ChatGPT” OR “generative AI” OR “large language model”) AND (education OR learning OR classroom OR teaching). The search focused primarily on peer-reviewed journal articles published between 2022 and 2025, corresponding to the period following the public release of ChatGPT and the early development of related educational research. The database query was conducted in January 2026, which resulted in the retrieval of a small number of records indexed in early 2026. Only articles indexed under education- and learning-related Fields of Research were included in order to maintain relevance and consistency across the dataset. The initial search retrieved 112 publications. After screening titles and abstracts according to the inclusion and exclusion criteria described in Section 2, duplicates and non-relevant records were removed, resulting in a final review corpus of 41 studies. Details of the data collection settings, search criteria, and applied filters are summarized in

Table 3. Bibliometric data collection settings, search criteria, and Fields of Research (ANZSRC 2020) included in the analysis.

Category	Description
Database	Dimensions [64]
Search fields	Title and abstract
Core keywords	(“ChatGPT” OR “generative AI” OR “large language model”) AND (education OR learning OR classroom)
Publication years	2022–2025
Publication type	Articles
Language	English
Fields of Research	39 Education (3901 Curriculum and Pedagogy; 3902 Education Policy, Sociology and Philosophy;
	3903 Education Systems; 3904 Specialist Studies in Education);
	47 Language, Communication and Culture (4701 Communication and Media Studies;
	4703 Language Studies; 4704 Linguistics; 4705 Literary Studies);
	36 Creative Arts and Writing (3602 Creative and Professional Writing);
	52 Psychology (5201 Applied and Developmental Psychology;
	5204 Cognitive and Computational Psychology);
	46 Information and Computing Sciences (4601 Applied Computing;
	4608 Human-Centred Computing; 4609 Information Systems)
Date of data retrieval	15 January 2026

.

Although the Dimensions database provides broad coverage of scholarly publications, reliance on a single database introduces certain limitations and may exclude relevant publications indexed in other repositories. Some journals, particularly those published in non-English-speaking regions or those not fully indexed in major repositories, may be underrepresented. As a result, the disciplinary distribution presented in the bibliometric analysis should be interpreted as indicative rather than exhaustive.

Figure 6 presents the annual number of publications and citations related to ChatGPT in educational research. The figure shows a rapid growth of academic interest in this topic following an initial emergence phase that aligns with the public release of ChatGPT. The increase in publication volume is accompanied by a steady rise in citation counts, indicating that the topic has not only expanded quickly but has also attracted meaningful scholarly attention within a short period of time. This citation trend suggests that the educational research community is actively engaging with questions related to the pedagogical implications, opportunities, and challenges of large language models such as ChatGPT.

Figure 7 shows the distribution of publications across Fields of Research based on the ANZSRC classification. As expected, the majority of studies fall within education-related fields, reflecting the strong pedagogical focus of this research area. At the same time, notable contributions are observed in Language and Communication Studies, Information and Computing Sciences, and Psychology. This distribution highlights the multidisciplinary nature of research on ChatGPT in education and reflects growing collaboration across disciplinary boundaries.

Overall, the bibliometric analysis reveals clear and consistent patterns indicating that research on ChatGPT in education is rapidly growing. The sharp increase in publications and citations reflects strong academic interest, while the disciplinary distribution shows that this topic extends beyond education into language, computing, and psychological research. These findings support the relevance and timeliness of the present study and underline its contribution to an active and expanding research field.

The bibliometric trends observed in this analysis also provide useful context for interpreting the conceptual focus of the present review. The rapid increase in publications following the introduction of ChatGPT reflects strong interdisciplinary interest in generative AI across education, computer science, and related fields. However, the diversity of disciplinary perspectives also suggests that the theoretical mechanisms underlying AI-supported learning remain fragmented. Many studies focus primarily on practical applications or instructional outcomes, while fewer explicitly engage with cognitive or metacognitive processes that shape learners’ interpretations of fluent AI-generated output. In this sense, the growing body of literature highlights the need for clearer conceptual frameworks, such as the fluency illusion perspective developed in this review, to explain how highly coherent AI responses may influence learners’ judgments of understanding.

9. Conclusions

The growing presence of generative artificial intelligence tools such as ChatGPT in classroom settings has introduced new possibilities for learning support while also raising important questions about how students interpret and evaluate their own understanding. This review has argued that many of the mixed and sometimes contradictory findings in the literature can be more coherently explained through the concept of fluency illusion. When AI-generated output is linguistically smooth, well structured, and immediately responsive, learners may experience a strong sense of clarity that is mistakenly attributed to their own understanding rather than to the fluency of the system itself.

By integrating research from cognitive psychology, metacognition, and empirical studies on ChatGPT use in education, this article highlights the role of fluency illusion in shaping learners’ judgments, engagement, and self-regulation in AI-mediated learning environments. Across writing, problem solving, feedback, and assessment contexts, the literature consistently suggests that increased confidence and perceived learning are not always accompanied by deep or transferable understanding. These patterns are not best understood as simple misuse of technology but as predictable cognitive responses to fluent external support.

From a pedagogical perspective, the findings highlight the need to move beyond evaluating learning solely on the basis of polished outputs or efficient task completion. Instructional and assessment practices should be designed to make students’ reasoning, explanation, and transfer of knowledge visible, particularly in contexts where fluent AI assistance is available. Explicit metacognitive scaffolding can help learners distinguish between perceived understanding and demonstrated competence, reducing overreliance on fluency as a cue for learning success.

Several limitations of this review should also be acknowledged. The study adopts a narrative and conceptual approach rather than a systematic meta-analysis, and the framework proposed in this article should therefore be interpreted as a theoretical synthesis rather than an empirically validated model. In addition, the empirical literature on generative AI in education remains relatively recent and methodologically diverse. While a growing number of classroom-based studies provide valuable insights into how students interact with AI systems, many of the cognitive mechanisms discussed in this review, including the role of fluency illusion in shaping learners’ judgments of understanding, require further empirical investigation.

Future research should examine fluency illusion in AI-mediated learning through longitudinal and classroom-based designs that capture how learners’ judgments and strategies evolve over time. Greater emphasis on performance-based and process-oriented measures, alongside direct assessments of metacognitive calibration, will be essential for understanding when and for whom fluency-driven misjudgments are most likely to occur. By adopting fluency illusion as an organizing framework, future work can support the development of educational practices that leverage the benefits of generative artificial intelligence while safeguarding meaningful and durable learning.