The Agency Gap in AI-Assisted Higher Education: When Learner Control Shapes Reflective Practice and Critical Thinking

Jul 22

33 min read

Listen to a review of this article:

Abstract: The proliferation of generative artificial intelligence (GenAI) in higher education has intensified questions about student agency, cognitive effort, and the conditions under which AI-supported learning fosters deep engagement rather than passive consumption. This study examines the relationships between perceived human-AI agency—operationalized as students' reported initiative, monitoring of AI outputs, and decision-making control—and reflective engagement, self-reported critical thinking, and academic self-concept. Drawing on survey data from 309 university students across UK-based (n=145) and China-based (n=164) higher education contexts, partial least squares structural equation modeling (PLS-SEM) revealed that perceived human-AI agency positively predicted reflective engagement, which in turn predicted self-reported critical thinking in both samples. However, context-specific patterns emerged: reflection was associated with academic self-concept only in the UK sample, while AI literacy moderated the agency-reflection relationship exclusively in the China sample. These findings theorize an "agency gap," suggesting that students who retain perceived control when using AI also report stronger reflective and critical orientations. The study underscores the need for context-sensitive pedagogical strategies that prioritize learner agency, metacognitive scaffolding, and responsible AI integration.

Generative AI tools—from ChatGPT to Claude and beyond—have moved from experimental novelty to routine presence in university classrooms, study groups, and assessment preparation. This rapid normalization raises fundamental questions about who directs learning, how knowledge is validated, and whether cognitive offloading to algorithms undermines or augments student development (Luo & Dawson, 2025; Yang et al., 2024). The central tension is not whether students use AI—most do—but how they use it, and whether that use preserves or erodes their sense of control, judgment, and intellectual ownership.

Recent debates have often framed GenAI through binary lenses: cheating versus assistance, efficiency versus integrity, innovation versus risk. Yet these framings may miss a more foundational question: What role does perceived learner agency play in shaping whether AI-supported learning becomes a site of deep engagement or passive dependency? If students experience themselves as directing the AI—setting goals, questioning outputs, synthesizing material—they may engage in reflective and critical work. If, instead, they treat AI as an authoritative answer machine, cognitive work may be displaced rather than distributed (Darvishi et al., 2024; Krakowski, 2025).

This study approaches that question through students' self-reported experiences of agency, reflection, and critical thinking in AI-supported learning. We examine perceived human-AI agency not as an objective behavioral measure but as students' subjective sense of control, initiative, and decision-making authority when working with AI. Reflection is treated as a metacognitive practice that may interrupt uncritical acceptance and redirect attention toward evaluation and meaning-making. Critical thinking is operationalized as students' reported tendency to question, verify, and synthesize AI-generated content. Academic self-concept—students' perceived competence and confidence in their field—is included as a downstream psychological outcome that may vary by educational context.

The Contemporary Stakes

The stakes are both practical and conceptual. Practically, institutions need evidence-informed guidance on how to design curricula, assessments, and support structures that leverage AI's affordances without undermining student development. Conceptually, the rise of GenAI unsettles established theories of learner autonomy, metacognition, and distributed cognition. If learning is increasingly co-constructed with algorithmic agents, how do we theorize agency, responsibility, and growth?

Emerging frameworks—human-AI co-agency (Krakowski, 2025), hybrid intelligence (Järvelä et al., 2023), and agentic AI-enhanced collaboration (Kremantzis et al., 2025)—offer promising conceptual directions, but empirical evidence on how these dynamics unfold in diverse higher education settings remains limited. Cross-context research is particularly sparse, despite growing recognition that students' AI experiences may be shaped by educational traditions, assessment cultures, institutional policies, tool availability, and epistemic norms (Ravšelj et al., 2025; Yusuf et al., 2024).

This study addresses that gap by examining UK-based and China-based higher education contexts. These settings provide a useful comparison because they may differ in pedagogical traditions, assessment expectations, and students' prior educational socialization. However, we emphasize that "UK-based" and "China-based" refer to study locations and institutional contexts, not nationally homogeneous cultural groups. Cultural values, nationality, ethnicity, institutional policy, tool availability, and educational socialization were not directly measured, so contextual interpretations remain cautious and exploratory.

Research Objectives

Guided by distributed cognition theory and metacognitive frameworks, this study pursues four research questions:

RQ1: How does perceived human-AI agency relate to reflective engagement in AI-supported learning?
RQ2: What role does reflective engagement play in the relationship between perceived human-AI agency and self-reported critical thinking?
RQ3: How do these relationships vary between UK-based and China-based higher education contexts, particularly regarding academic self-concept?
RQ4: How does AI literacy moderate the relationship between perceived human-AI agency and reflective engagement?

By examining these questions, we aim to clarify when and how AI-supported learning may be associated with deeper cognitive engagement, and what institutional and instructional conditions might support that outcome.

The Human-AI Agency Gap: Conceptual Foundations

Reconceptualizing Agency in Algorithmic Learning

Traditional conceptions of student agency emphasize learners' capacity to act intentionally, set goals, monitor progress, and shape their educational trajectories (Jääskelä et al., 2017). In pre-AI environments, agency was primarily relational—exercised through interactions with instructors, peers, texts, and institutional structures. The rise of GenAI introduces a new category of relationship: one in which the "partner" can generate novel text, suggest framings, produce arguments, and even mimic reasoning processes.

This shift requires theoretical recalibration. Human-AI co-agency recognizes that learning may now unfold through interaction with algorithmic agents that respond, suggest, and generate in ways that shape learners' cognitive pathways (Järvelä et al., 2023; Krakowski, 2025). However, "co-agency" does not imply symmetrical partnership or shared intentionality. AI systems lack genuine understanding, intention, or epistemic responsibility. Rather, co-agency describes a distributed cognitive arrangement: AI provides material, prompts, options, and generative feedback, while learners ideally retain executive control over goal-setting, evaluation, synthesis, and final judgment.

In this framework, agency is not binary (present or absent) but gradient and relational. Students may exercise high agency—iteratively prompting, verifying claims, rejecting unsuitable outputs, revising drafts, and integrating AI-generated material into their own reasoning. Or they may exercise low agency—copying outputs verbatim, accepting fluent-sounding text uncritically, or outsourcing epistemic judgment to the algorithm. The difference may hinge on whether students perceive themselves as pilots or passengers (Darvishi et al., 2024).

The Agency Gap Hypothesis

We propose the concept of an "agency gap"—the perceived distance between students who experience themselves as retaining control, initiative, and decision-making power when using AI, versus those who experience AI-supported work as passive, dependent, or algorithmically driven. This gap is not inherent to the technology but emerges from the interaction between tool affordances, task design, instructional scaffolding, students' prior experiences, epistemic beliefs, and self-regulatory capacities.

When the agency gap is narrow—when students report high perceived control—AI may function as a cognitive prosthesis that amplifies capacity without replacing judgment (Fang & Zhou, 2026). When the gap is wide—when students report low perceived control—AI may induce cognitive offloading, automation bias, and reduced metacognitive engagement (Fan et al., 2025; Zhai et al., 2024).

Distributed Cognition and Epistemic Responsibility

Distributed cognition theory (Salomon, 1997) provides a foundation for understanding these dynamics. Cognition is not confined to individual minds but distributed across people, tools, artifacts, and representations. In AI-supported learning, knowledge construction unfolds through interaction: students pose questions, AI generates responses, students evaluate those responses, refine prompts, verify claims, and integrate material into their own understanding.

Crucially, executive control and epistemic responsibility remain with the learner. AI may suggest, but students must judge. AI may generate, but students must evaluate. AI may synthesize, but students must integrate. When these responsibilities are exercised, cognitive offloading may become cognitive extension—leveraging external resources to think with rather than think less (Krakowski, 2025; Nandagopal, 2025).

However, this outcome is not automatic. GenAI's fluency, confidence, and speed may seduce students into passive acceptance. Without reflective interruption, students may treat AI outputs as authoritative rather than provisional, adopt algorithmic suggestions uncritically, and mistake retrieval for understanding (Chan & Hu, 2023; Polyportis & Pahos, 2025).

Reflection as Cognitive Brake and Epistemic Gateway

From Habitual Action to Critical Reflection

If perceived agency shapes whether students retain control, reflection shapes how they use that control. Reflection is not mere repetition or review but a metacognitive process that involves pausing, questioning, evaluating, and revising beliefs in light of new information or experiences (Kember et al., 2000).

Kember and colleagues distinguish four levels of engagement:

Habitual action: Routine, unreflective application of familiar procedures (e.g., copying AI output verbatim)
Understanding: Surface-level comprehension without questioning assumptions
Reflection: Active evaluation of ideas, consideration of alternatives, integration with prior knowledge
Critical reflection: Deep interrogation of assumptions, epistemic vigilance, recognition of limitations

Applied to AI-supported learning, habitual action might involve treating ChatGPT as a search engine or answer key—prompting once, accepting the first response, and moving on. Reflection, by contrast, involves iterative prompting, verification of claims, comparison with other sources, identification of gaps or contradictions, and synthesis into the learner's evolving understanding.

Reflection as Mediator Between Agency and Critical Thinking

We hypothesize that reflection mediates the relationship between perceived human-AI agency and self-reported critical thinking. That is, students who report retaining control when using AI are more likely to engage reflectively because they perceive the AI output as material to be evaluated rather than truth to be accepted. That reflective engagement, in turn, predicts stronger critical thinking—questioning assertions, seeking supporting evidence, considering alternatives, and developing independent perspectives (Pintrich et al., 1991).

This pathway aligns with evidence that metacognitive monitoring and strategic regulation predict deeper learning outcomes (Darvishi et al., 2024; Yuan & Hu, 2024). It also resonates with warnings that passive AI use—"prompt and paste"—may weaken critical faculties (Rudolph et al., 2024). By positioning reflection as a necessary condition for critical engagement with AI outputs, we suggest that critical thinking is not a direct byproduct of AI use but emerges when students reflectively interrogate what AI produces.

Epistemic Vigilance and Trust

Students' epistemic beliefs about AI also matter. Those who view AI as fallible, probabilistic, and prone to error may be more inclined to verify, question, and integrate cautiously. Those who anthropomorphize AI, trust it excessively, or perceive it as an "omniscient epistemic authority" may accept outputs uncritically, reducing both reflection and critical thinking (Martín-Moncunill & Alonso Martínez, 2025; Urhahne et al., 2026).

Epistemic vigilance—awareness that AI outputs require verification—may therefore function as a prerequisite for reflective engagement. Instructional strategies that explicitly frame AI as a probabilistic text generator rather than a knowledge oracle may help students maintain epistemic distance and critical stance (Urban et al., 2025; Wittig McPhee & Jerowsky, 2025).

Context Matters: Self-Concept, Self-Enhancement, and Self-Improvement

Why Academic Self-Concept May Vary Across Contexts

Academic self-concept—students' beliefs about their competence, capability, and academic potential—is not only an outcome of learning but also a motivational driver that shapes persistence, effort, and future achievement (Liu et al., 2005; Wu et al., 2021). In traditional educational psychology, self-concept and performance operate reciprocally: success enhances self-concept, which supports future success (Bandura, 2009).

However, the relationship between reflection and self-concept may be contextually shaped by educational traditions, assessment cultures, and self-evaluative norms. In some Western higher education settings, pedagogical cultures often emphasize individual achievement, self-advocacy, and the validation of competence through independent critique. Reflective engagement—especially when it reveals growth, mastery, or insight—may therefore reinforce students' sense of academic capability.

In contrast, some Confucian-heritage educational traditions have been described as foregrounding self-improvement, effort attribution, and moral cultivation over self-validation (Chen et al., 2009; Fwu et al., 2018). Reflection may be oriented more toward identifying inadequacies and areas for further growth than toward celebrating progress. This does not imply deficiency or lower capability, but rather a different interpretive frame for self-evaluation (Min et al., 2016).

Reference-Group Effects and Modesty Bias

Cross-cultural self-report comparisons are further complicated by reference-group effects and response styles. Students compare themselves to peers, and those reference standards vary by context. In highly competitive settings, even high-performing students may report lower self-concept because their comparison group is equally capable (Heine et al., 2002). Modesty bias may also lead students in some contexts to understate their abilities, not because they lack confidence but because cultural norms discourage self-promotion (Min et al., 2016).

Because this study did not directly measure cultural values, educational socialization, or self-enhancement/self-criticism orientations, we treat contextual differences as tentative patterns requiring cautious interpretation. The UK-based sample reflects study location, not national identity, and includes diverse nationalities, ethnicities, and prior educational experiences. Similarly, the China-based sample should not be essentialized as culturally homogeneous. Both samples are embedded in specific institutional contexts with unique AI policies, tool access, disciplinary norms, and assessment regimes (Ravšelj et al., 2025; Taras et al., 2016).

Hypothesis: Reflection and Self-Concept

We hypothesize that reflective engagement will be positively associated with academic self-concept (H3), and that reflection will mediate the relationship between perceived human-AI agency and self-concept (H4b). However, we expect this pathway to be contextually contingent, potentially stronger in the UK-based sample than in the China-based sample, based on the interpretive frameworks outlined above. This hypothesis is theory-driven and exploratory, not a definitive claim about national or cultural mechanisms.

AI Literacy as Moderator: Competence as Gateway or Prerequisite?

Defining AI Literacy

AI literacy encompasses the knowledge, skills, and dispositions required to understand AI systems, use them appropriately, evaluate their outputs critically, and act ethically in AI-mediated contexts (Wang et al., 2023). It includes technical understanding (how AI works), operational competence (how to use it effectively), critical evaluation (recognizing limitations and biases), and ethical awareness (navigating privacy, fairness, and transparency concerns) (Lintner, 2024).

Recent reviews suggest that AI literacy supports responsible engagement and reduces uncritical adoption (Yang et al., 2025). Students with higher AI literacy may be better equipped to recognize hallucinations, verify claims, and interrogate outputs, thereby exercising stronger agency and engaging more reflectively.

Moderation Hypothesis

We hypothesize that AI literacy moderates the relationship between perceived human-AI agency and reflective engagement (H5). Specifically, we expect this moderation to vary by context. In settings where AI is relatively unfamiliar, regulated, or viewed instrumentally, AI literacy may function as a gatekeeper—students may need technical confidence before feeling authorized to exercise agency and engage reflectively. In settings where AI use is more exploratory, experimental, or normalized, students may exercise agency regardless of formal literacy, relying on trial-and-error and iterative learning.

This hypothesis aligns with evidence that technical competence can reduce technology anxiety and increase willingness to engage (Salhab & Aboushi, 2025). It also resonates with educational traditions that prioritize mastery-before-experimentation versus learning-by-doing approaches. However, because we did not directly measure educational norms, risk orientation, or institutional AI policies, this hypothesis remains exploratory and context-sensitive.

Research Model and Hypotheses

Figure 1 presents the hypothesized model, which integrates the theoretical relationships outlined above.

Figure 1: Hypothesized Research Model

H1: Perceived human-AI agency is positively associated with reflective engagement.
H2: Reflective engagement is positively associated with self-reported critical thinking.
H3: Reflective engagement is positively associated with academic self-concept.
H4a: Reflective engagement mediates the relationship between perceived human-AI agency and self-reported critical thinking.
H4b: Reflective engagement mediates the relationship between perceived human-AI agency and academic self-concept (with expected contextual variation).
H5: AI literacy moderates the relationship between perceived human-AI agency and reflective engagement (with expected contextual variation).

These hypotheses are tested using cross-sectional survey data and PLS-SEM, which allows for examination of associative patterns but does not establish causal ordering or temporal precedence.

Methodology

Research Design

This study employed a comparative quantitative design using an online survey administered to university students in UK-based and China-based higher education contexts. PLS-SEM was selected as the analytical approach because it is well-suited for exploratory theory-building with non-normal data, does not require strict distributional assumptions, and prioritizes predictive accuracy over model fit (Hair et al., 2019; Lowry & Gaskin, 2014).

Participants and Sampling

Participants were undergraduate and postgraduate students enrolled at universities in the UK and China. Recruitment targeted modules with documented AI usage, such as research methods, data analytics, and business strategy. Inclusion criteria were: current enrollment, age ≥18, and self-reported use of AI tools (e.g., ChatGPT, Claude, or similar) in coursework during the current academic term.

The final sample comprised 309 valid responses: 145 from UK-based institutions and 164 from China-based institutions. The UK sample was defined by study location and institutional context, not nationality; no within-UK nationality-based subgroup analysis was conducted. Demographic details:

UK sample: 50.3% women, 48.3% men, 2 non-binary; 62.8% aged 18–24, 27.6% aged 25–29
China sample: 55.5% women, 43.9% men, 1 preferred not to state; 39.0% aged 18–24, 20.1% aged 30–34, 20.7% aged 35–39

Prior GenAI experience was common. In the UK sample, 42.1% used AI more than once daily, 26.9% daily, 20.7% a few times per month, and 10.3% once or twice in the past six months. In the China sample, the corresponding figures were 26.8%, 39.6%, 31.1%, and 2.4%. Common uses included coursework support (UK 65.5%, China 78.7%), writing assistance (UK 60.7%, China 71.3%), coding/technical tasks (UK 49.0%), and personal productivity (China 75.6%).

These data contextualize prior AI exposure but do not capture specific tools used, regional location, institutional AI policies, local tool availability, disciplinary norms, or prior educational pathways—factors treated as limitations rather than controlled variables.

Ethical Considerations

Ethical approval was granted by the participating universities. Participants provided informed digital consent and were assured of anonymity to reduce social desirability bias around AI use in assessment contexts. Data were collected between September and December 2025.

Instruments

The survey used five established constructs, adapted for AI-supported learning contexts and measured on 5-point Likert scales (1 = Strongly Disagree, 5 = Strongly Agree):

1. Perceived Human-AI Agency (8 items; adapted from Jääskelä et al., 2017)

Operationalized as students' reported initiative, monitoring, strategic use, and control when learning with AI. Sample items:

"I took the initiative in directing my own learning process in this course."
"When using the AI tool (e.g., ChatGPT), I decided how to use its answers in my work rather than just copying them."
"I used the AI tool in a responsible way (as a means to learn and explore) rather than using it to do the work for me."

Conceptual note: While theorized as human-AI co-agency, the scale primarily captures perceived learner control and self-regulation, making it closer to AI-assisted individual agency than fully reciprocal co-agency.

2. Reflective Learning Engagement (8 items; adapted from Kember et al., 2000)

Captures students' tendencies to pause, question, evaluate, and revise beliefs when working with AI. Sample items:

"During this course, the AI tools led me to discover faults in what I had previously believed to be right."
"I often reflect on my actions when using AI in this course to see whether I could have improved what I did."
"Using AI in this course has challenged some of my firmly held ideas about learning or the subject."

3. Self-Reported Critical Thinking (5 items; adapted from Pintrich et al., 1991)

Measures students' reported tendencies to question, verify, and develop independent perspectives on AI-generated content. Sample items:

"I often find myself questioning things I read from AI to decide if I find them convincing."
"When a theory, interpretation, or conclusion is presented in AI, I try to decide if there is good supporting evidence."
"I treat AI content as a starting point and try to develop my own ideas about it."

Limitation note: This scale captures self-reported critical orientation, not objective critical thinking performance or specific behaviors such as source verification or revision practices.

4. Academic Self-Concept (9 items; adapted from Liu et al., 2005)

Assesses perceived competence, confidence, and effort in one's field. Sample items:

"I can follow lectures in my modules/units easily."
"If I work hard, I believe I can get better grades in this unit."
"I do not give up easily when faced with a difficult question in my coursework."

Note: This scale includes both competence beliefs and academically oriented engagement items (e.g., effort, persistence, attention) from the original validated scale.

5. AI Literacy (8 items; adapted from Wang et al., 2023)

Measures technical competence, conceptual understanding, evaluative capacity, and ethical awareness. Sample items:

"I can operate AI applications in everyday life."
"I know the most important concepts of the topic 'artificial intelligence.'"
"I can assess what the limitations and opportunities of using an AI are."
"I always comply with ethical principles when using AI applications or products."

Translation and Cultural Adaptation

For the China-based sample, translation/back-translation was conducted with bilingual academics, semantic equivalence was tested (n=12), and discrepancies resolved by bilingual educational researchers to preserve intended meanings of terms such as "agency" and "critical reflection" (Brislin, 1970).

Data Analysis

Data were analyzed in SmartPLS v.4.1.1.6 using the two-step PLS-SEM procedure (Hair et al., 2019):

Measurement model assessment: Internal consistency (Cronbach's α, composite reliability), convergent validity (AVE ≥ 0.50), and discriminant validity (HTMT < 0.85).
Structural model assessment: Path coefficients (β), confidence intervals, and significance via bootstrapping (5,000 subsamples), alongside R² for endogenous constructs, mediation/moderation tests, and SRMR as an approximate fit index (< 0.08).

Measurement invariance was assessed using the MICOM procedure (Henseler et al., 2016) to support UK–China comparisons. MICOM evaluates configural invariance, compositional invariance, and equality of composite means/variances.

Common method variance was assessed using the full collinearity VIF approach (Kock, 2015). All VIF values ranged from 1.513 to 2.694, well below the conservative threshold of 3.3, indicating minimal common method bias. Procedural remedies included validated scales, mixed item ordering, and anonymity assurance.

Results

Measurement Model Validation

All item loadings exceeded 0.708, indicating that each construct explained >50% of indicator variance. Internal consistency was strong across all constructs:

Cronbach's α ranged from 0.825 (Critical Thinking) to 0.905 (AI Literacy)
Composite reliability (ρc) ranged from 0.877 to 0.939

Convergent validity was confirmed: all AVE values exceeded 0.50, ranging from 0.500 (Academic Self-Concept) to 0.590 (AI Literacy).

Discriminant validity was established using both Fornell-Larcker criterion (square root of AVE exceeded all correlations) and HTMT ratios (all < 0.85, ranging from 0.258 to 0.841). The HTMT ratio between Human-AI Agency and Reflection was 0.841, confirming they are empirically distinct despite being theoretically related.

Measurement Invariance (MICOM)

MICOM results indicated partial measurement invariance:

Configural invariance: Established (identical constructs, indicators, specifications, and algorithm settings)
Compositional invariance: Established for Reflection (c=0.998) and Critical Thinking (c=0.999); not established for AI Literacy (c=0.916), Human-AI Agency (c=0.990), and Academic Self-Concept (c=0.991)

Interpretation: Path comparisons involving invariant constructs (Reflection, Critical Thinking) are substantively comparable across samples. However, cross-group comparisons involving non-invariant constructs (AI Literacy, Human-AI Agency, Academic Self-Concept) should be interpreted cautiously and treated as tentative rather than definitive evidence of contextual differences.

Descriptive Statistics

Table 1 reports means, standard deviations, skewness, and kurtosis by context. Descriptively, UK respondents reported higher scores on perceived Human-AI Agency, Reflection, Critical Thinking, and Academic Self-Concept, while AI Literacy was slightly higher in the China sample.

Table 1: Descriptive Statistics

Critical caveat: Because MICOM established only partial invariance, mean comparisons involving non-invariant composites (especially AI Literacy, Human-AI Agency, Academic Self-Concept) should not be treated as definitive between-context differences. They reflect sample-specific patterns requiring further investigation.

Structural Model: Hypothesis Testing

Three models were estimated: China-based sample, UK-based sample, and combined sample. Results are summarized below.

H1: Perceived Human-AI Agency → Reflective Engagement

Supported in all models.

China: β = 0.554, p < .001
UK: β = 0.721, p < .001
Combined: β = 0.739, p < .001

Interpretation: Students who reported higher perceived control when using AI also tended to report stronger reflective engagement. This relationship was robust across both samples, with a permutation test indicating a significant difference in path strength (p = .020). However, because Human-AI Agency did not establish compositional invariance, this cross-group difference should be interpreted as descriptive rather than definitive.

H2: Reflective Engagement → Self-Reported Critical Thinking

Supported in all models.

China: β = 0.243, p = .022
UK: β = 0.317, p = .003
Combined: β = 0.270, p < .001

Interpretation: Reflective engagement was positively associated with self-reported critical thinking in both samples. The permutation test showed no significant cross-group difference (p = .719). This finding suggests that reflection is a common correlate of critical thinking across contexts, though the cross-sectional design prevents causal claims.

H3: Reflective Engagement → Academic Self-Concept

Supported in UK and combined models; not supported in China model.

China: β = 0.109, p = .260 (not significant)
UK: β = 0.371, p < .001
Combined: β = 0.216, p = .001

Interpretation: Reflection was associated with academic self-concept in the UK sample but not in the China sample. However, the permutation test was marginally non-significant (p = .073), and Academic Self-Concept did not establish compositional invariance. Therefore, this pattern should be treated as tentative and indicative rather than as confirmed evidence of contextual difference.

Possible interpretations (exploratory, not tested directly):

Reflection may be oriented more toward self-improvement and gap-identification in the China sample (consistent with effort-attribution and self-critical frameworks; Chen et al., 2009; Fwu et al., 2018)
Self-concept scores may reflect reference-group effects and modesty bias rather than true competence differences (Heine et al., 2002; Min et al., 2016)
Institutional, technological, or disciplinary factors not measured may shape the reflection–self-concept relationship

H4a: Mediation via Reflection to Critical Thinking

Supported in all models.

China: Indirect effect = 0.134, p = .021
UK: Indirect effect = 0.228, p = .003
Combined: Indirect effect = 0.199, p = .001

Interpretation: Reflective engagement mediated the relationship between perceived Human-AI agency and self-reported critical thinking in both samples. This supports the theorized pathway: students who report retaining control also tend to report stronger reflection, which in turn predicts higher self-reported critical thinking. However, the cross-sectional design prevents causal claims; these are theory-consistent associative patterns.

H4b: Mediation via Reflection to Academic Self-Concept

Supported in UK and combined models; not supported in China model.

China: Indirect effect = 0.060, p = .259 (not significant)
UK: Indirect effect = 0.267, p < .001
Combined: Indirect effect = 0.160, p = .002

Interpretation: Reflection mediated the agency–self-concept relationship in the UK sample but not in the China sample. The permutation test was marginally non-significant (p = .051). Given the lack of compositional invariance for Academic Self-Concept, this finding should be interpreted cautiously as a sample-specific pattern rather than a definitive contextual difference.

H5: Moderation by AI Literacy

Supported in China model; not supported in UK or combined models.

China: Interaction = 0.067, p = .033
UK: Interaction = 0.002, p = .984 (not significant)
Combined: Interaction = −0.007, p = .827 (not significant)

Interpretation: AI Literacy moderated the agency–reflection relationship exclusively in the China sample, where higher AI literacy slightly strengthened this association. No moderation was observed in the UK sample. The permutation test showed no significant cross-group difference (p = .328).

Possible interpretations (exploratory):

In the China sample, technical competence may function as a confidence-building prerequisite for exercising agency and engaging reflectively (consistent with mastery-before-exploration norms; Wang et al., 2023)
In the UK sample, exploratory, trial-and-error approaches may allow students to exercise agency regardless of formal literacy (consistent with learning-by-doing traditions)
However, because AI Literacy did not establish compositional invariance, this moderation finding should be treated as tentative

Model Fit and Explanatory Power

SRMR = 0.069 (below the 0.08 threshold; Hu & Bentler, 1999), indicating good approximate fit
R² for Reflection: China = 0.556, UK = 0.539 (strong explanatory power)
R² for Critical Thinking: China = 0.518, UK = 0.283 (higher variance explained in China sample)
R² for Academic Self-Concept: China = 0.574, UK = 0.442

Effect sizes (f²):

Human-AI Agency → Reflection: f² = 1.000 (very large effect)
Human-AI Agency → Critical Thinking: f² = 0.189 (medium effect)
Human-AI Agency → Self-Concept: f² = 0.296 (medium-to-large effect)
Reflection added small incremental effects to downstream constructs

PLSpredict indicated acceptable predictive relevance for all endogenous constructs (Q²predict > 0.50).

Discussion

The Agency Gap: Perceived Control as Prerequisite for Reflective Engagement

The most robust finding across both samples was the strong positive association between perceived human-AI agency and reflective engagement (H1). Students who reported greater initiative, monitoring, and decision-making control when using AI also tended to report stronger reflective practices—questioning AI outputs, identifying contradictions, revising beliefs, and integrating material into their own understanding.

This finding supports the agency gap hypothesis: AI-supported learning appears most educationally meaningful when students experience themselves as retaining executive control. When students perceive themselves as pilots—setting goals, tweaking prompts, evaluating outputs, and making final judgments—they engage reflectively. When they perceive themselves as passengers—accepting fluent outputs uncritically, copying verbatim, or outsourcing judgment—reflection may be bypassed.

This pattern challenges deterministic narratives that AI inherently induces cognitive atrophy or passive consumption (Lindebaum et al., 2025; Yang et al., 2024). Instead, it suggests that how students use AI matters more than whether they use it. The technology itself is neither inherently empowering nor disempowering; rather, the learning outcomes depend on the cognitive and metacognitive processes students engage when working with AI.

Reflection as Mediator: The Cognitive Brake Between Agency and Critical Thinking

The significant mediation pathway (H4a) further clarifies this relationship. Perceived human-AI agency predicted self-reported critical thinking not directly, but through reflective engagement as an intermediary process. This finding supports the theoretical claim that reflection functions as a cognitive brake—interrupting automatic acceptance, redirecting attention toward evaluation, and enabling epistemic vigilance (Kember et al., 2000; Wittig McPhee & Jerowsky, 2025).

In practical terms, this suggests that critical thinking is not a spontaneous byproduct of AI use but emerges when students reflectively interrogate what AI produces. Students may prompt effectively, receive fluent outputs, and complete assignments efficiently without ever engaging critically—unless reflection is explicitly scaffolded. This finding resonates with warnings that AI's speed and fluency can foster automation bias and excessive trust (Fan et al., 2025; Urban et al., 2025).

Instructional implications are clear: educators must design tasks, prompts, and assessments that require students to pause, question, verify, and synthesize rather than merely retrieve and submit. Process-oriented assessments—reflective audit trails, annotated AI transcripts, comparative source analyses—may support this goal more effectively than traditional product-oriented assessments, which AI can now complete with minimal human input (Dawson et al., 2024).

Contextual Patterns in Academic Self-Concept: Tentative Interpretations

The non-significant reflection–self-concept path in the China sample (H3), alongside the significant association in the UK sample, presents a more complex interpretive challenge. Because Academic Self-Concept did not establish compositional invariance and the cross-group comparison was marginally non-significant (p = .073), this finding should be treated as a sample-specific pattern requiring cautious interpretation rather than definitive evidence of cultural difference.

Several plausible interpretations emerge, all exploratory:

1. Self-Enhancement vs. Self-Improvement Orientations

Educational psychology research suggests that self-concept formation may be shaped by culturally mediated self-evaluative norms. In some Western contexts, achievement is often interpreted as confirming competence, reinforcing self-concept, and validating capability (Bandura, 2009; Sedikides et al., 2003). Reflective engagement—especially when it reveals growth, mastery, or insight—may therefore strengthen students' sense of academic capability.

In contrast, some Confucian-heritage educational traditions have been described as foregrounding self-improvement (自我提升), effort attribution, and moral cultivation (修养) over self-validation (Chen et al., 2009; Fwu et al., 2018). Reflection—particularly when it reveals gaps, contradictions, or inadequacies—may be interpreted not as evidence of competence but as a prompt for further effort and self-correction. This does not imply lower actual capability or deficiency, but rather a different interpretive script for self-evaluation.

However, this interpretation remains speculative because cultural values, self-enhancement orientation, and self-criticism were not directly measured. It is also essential to recognize intra-cultural heterogeneity: not all UK students are self-enhancing, and not all Chinese students are self-critical. Educational traditions vary by institution, discipline, language background, and prior educational socialization (Taras et al., 2016; Yusuf et al., 2024).

2. Reference-Group Effects and Modesty Bias

Cross-cultural self-report comparisons are complicated by reference-group effects: students compare themselves to peers, and those comparison standards vary by context. In highly competitive settings, even high-performing students may report lower self-concept because their reference group is equally capable (Heine et al., 2002). Modesty bias may also lead students in some contexts to understate abilities, not because they lack confidence but because cultural norms discourage self-promotion (Min et al., 2016).

Since this study did not measure reference-group standards, modesty bias, or self-criticism norms directly, we cannot definitively attribute the non-significant path to these mechanisms. However, they remain plausible contributing factors.

3. Technological, Institutional, and Disciplinary Factors

Students' GenAI experiences in China may have been shaped by access restrictions, cost, language preferences, disciplinary differences, institutional AI policies, and the affordances of domestic versus global AI tools (Xie et al., 2025). These factors were not captured in the survey, so they cannot be separated from cultural or psychological explanations.

For example, if some China-based students primarily used domestic AI platforms with different interaction designs or language capabilities, their reflective experiences may have differed systematically from UK-based students using ChatGPT or Claude. Similarly, if institutional AI policies were more restrictive or ambiguous in certain China-based contexts, students may have approached reflective work with greater caution or perceived risk.

4. Measurement Concerns

Because Academic Self-Concept did not establish compositional invariance, the construct may have been interpreted or weighted differently across the two samples. For instance, items capturing effort and persistence (e.g., "I study hard for my coursework") may have been endorsed more strongly relative to competence items (e.g., "I can outperform most of my friends") in the China sample, leading to divergent latent variable structures. This possibility underscores the importance of rigorous measurement invariance testing before making cross-context claims.

Implications for Educators and Institutions

Given these uncertainties, educators should exercise caution when interpreting self-reported academic self-concept scores across contexts. Lower self-reported self-concept does not necessarily indicate weaker learning, lower capability, or poorer educational outcomes. It may reflect modesty bias, reference-group effects, effort-oriented self-evaluation, or measurement non-equivalence. Educators should triangulate self-concept data with behavioral evidence (e.g., assignment quality, revision patterns, source verification), reflective artifacts (e.g., annotated AI transcripts, metacognitive logs), and performance data (e.g., exam results, project outcomes) to form a holistic picture of student learning.

AI Literacy as Contextual Moderator: Confidence-Building vs. Exploratory Learning

The moderation finding—AI literacy strengthened the agency–reflection relationship in the China sample but not in the UK sample (H5)—suggests that technical competence may function differently across educational contexts.

In the China sample, higher AI literacy was associated with a stronger link between perceived agency and reflective engagement. This pattern is consistent with the hypothesis that literacy acts as a gatekeeper or confidence-building prerequisite in contexts where technology is approached instrumentally or where mastery is expected before experimentation (Wang et al., 2023). Students with stronger AI literacy may feel more authorized to question outputs, recognize limitations, and engage reflectively without fear of error or misuse.

In the UK sample, no moderation was observed, suggesting that students may exercise agency and engage reflectively regardless of formal AI literacy. This pattern aligns with exploratory, trial-and-error, and learning-by-doing pedagogical traditions often associated with Western higher education (Sharples, 2023). Students may adopt a more experimental stance, prompting iteratively and reflecting through practice rather than waiting for formal mastery.

However, because AI Literacy did not establish compositional invariance, this interpretation should be treated as tentative. Future research should directly measure educational norms, risk orientation, technology anxiety, instructional scaffolding, and institutional expectations to clarify how literacy interacts with agency and reflection.

Limitations and Future Directions

Cross-Sectional Design and Causal Claims

This study's cross-sectional design prevents causal inference. The mediation pathways tested assume temporal ordering (agency → reflection → outcomes), but cross-sectional data cannot establish that ordering (Maxwell & Cole, 2007; O'Laughlin et al., 2018). Longitudinal designs with multiple measurement occasions are needed to test whether perceived agency predicts later reflection, and whether reflection predicts subsequent critical thinking or self-concept.

Self-Report and Common Method Variance

All constructs were self-reported, introducing potential common method variance, social desirability bias, and retrospective recall errors. While procedural remedies (validated scales, mixed item ordering, anonymity) and statistical checks (full collinearity VIF < 3.3) suggest minimal bias, objective performance measures, behavioral traces, and log data would strengthen future research. For example:

Prompt logs capturing iterative refinement, verification searches, and source comparisons
Trace data from AI platforms showing revision cycles, copy-paste ratios, and interaction sequences
Performance assessments measuring actual critical thinking, not just self-reported tendencies
Qualitative interviews exploring students' epistemic beliefs, trust in AI, and reflective experiences

Broad Composite Constructs

The survey constructs are broad composite measures rather than narrowly isolated dimensions. For example:

Perceived Human-AI Agency includes initiative, monitoring, strategic use, responsibility, and self-direction—conceptually related but potentially separable facets
Reflective Engagement spans questioning, evaluation, belief revision, and identity transformation
Academic Self-Concept includes competence beliefs, effort, persistence, and attention

Future research should use GenAI-specific instruments that distinguish different forms of AI engagement:

Information retrieval (one-shot prompting for definitions or summaries)
Iterative prompting (refining queries based on initial outputs)
Fact-checking and verification (cross-referencing AI claims with external sources)
Source verification (checking citations and evidence provided by AI)
Revision and co-creation (editing AI-generated drafts, integrating material into original work)
Coding support (debugging, generating boilerplate code, learning syntax)
Brainstorming and ideation (generating alternatives, exploring perspectives)
Translation and summarization (converting between languages or formats)

By separating these activities, researchers could clarify which forms of AI use support reflection and critical thinking most strongly.

Cultural and Contextual Mechanisms Not Measured

This study did not directly measure cultural values, educational socialization, self-enhancement/self-criticism orientations, reference-group standards, modesty bias, institutional AI policies, tool availability, or prior educational pathways. Therefore, contextual interpretations remain cautious and heuristic. Future research should include:

Direct measures of cultural values (e.g., Hofstede dimensions, Schwartz values)
Educational socialization scales (e.g., mastery vs. performance orientation, effort vs. ability attribution)
Epistemic beliefs inventories (e.g., source certainty, justification of knowledge, trust in AI)
Institutional AI policy audits (e.g., restrictions, guidelines, training provided)
Tool-specific usage data (e.g., ChatGPT vs. Claude vs. domestic platforms; free vs. paid versions)
Qualitative interviews exploring students' lived experiences of agency, reflection, and AI-supported learning

Measurement Invariance and Cross-Context Comparisons

Because only partial measurement invariance was established, cross-group comparisons involving non-invariant constructs (AI Literacy, Human-AI Agency, Academic Self-Concept) should be interpreted as descriptive patterns rather than definitive contextual differences. Future research should:

Develop culturally adapted scales that establish full measurement invariance
Conduct within-context longitudinal studies to test whether the observed patterns hold over time
Use multi-method designs (survey + interviews + behavioral data) to triangulate findings
Explore within-context variation by discipline, prior AI experience, assessment type, and institutional policy

Broader Implications for Theory and Practice

Despite these limitations, this study makes three substantive contributions:

1. Theorizing the Agency Gap

We introduce the concept of an "agency gap" in AI-supported learning—the perceived distance between students who retain executive control versus those who experience AI-supported work as passive or algorithmically driven. This concept extends distributed cognition theory and human-AI co-agency frameworks by foregrounding students' subjective sense of control, initiative, and decision-making power as central to learning outcomes.

The agency gap is not inherent to the technology but emerges from the interaction between tool affordances, task design, instructional scaffolding, students' epistemic beliefs, and self-regulatory capacities. This reframes the AI-in-education debate: the central question is not "Should students use AI?" but "Under what conditions does AI use preserve or enhance learner agency?"

2. Positioning Reflection as Cognitive Brake

By demonstrating that reflection mediates the agency–critical-thinking pathway, this study positions reflection as a necessary condition for deep AI-supported learning. AI may generate material for thought, but without reflective interruption, students may mistake fluency for truth, efficiency for understanding, and retrieval for synthesis.

This finding challenges efficiency-focused narratives that frame AI primarily as a time-saving tool. If reflection is essential for critical thinking, then speed and convenience may undermine learning when they bypass metacognitive engagement. Educators must design tasks, assessments, and instructional scaffolds that slow students down, require questioning, and reward epistemic vigilance.

3. Context-Sensitive Pedagogical Strategies

The tentative contextual patterns observed—reflection associated with self-concept in the UK but not China; AI literacy moderating agency-reflection in China but not UK—underscore the need for context-sensitive rather than universalist pedagogical approaches. One-size-fits-all AI integration strategies may misalign with students' educational backgrounds, self-evaluative norms, epistemic orientations, and institutional contexts.

For example:

In contexts emphasizing self-improvement and effort attribution, educators might frame reflective tasks as opportunities for growth and mastery development rather than as validation of existing competence. Process-oriented feedback emphasizing progress, effort, and strategy adjustment may resonate more strongly than performance-oriented feedback emphasizing relative ability.
In contexts where AI literacy is low or technology anxiety is high, educators might build foundational literacy before expecting agentic AI use. Structured workshops, guided prompting tutorials, and scaffolded assignments that gradually increase complexity may support confidence-building and reduce perceived risk.
In contexts where exploratory, trial-and-error learning is normative, educators might encourage experimentation and iterative learning from the outset, framing errors as learning opportunities and rewarding process over product.
In contexts where institutional AI policies are ambiguous or restrictive, educators might co-create norms and guidelines with students, clarifying expectations, modeling ethical use, and providing transparent rationales for AI integration.

Toward Agentic AI-Enhanced Pedagogy

Building on these findings, we propose a framework for agentic AI-enhanced pedagogy—instructional design that prioritizes learner agency, metacognitive scaffolding, and reflective engagement:

Principle 1: Design for Retained Control

Structure tasks so students must:

Set learning goals before prompting
Iteratively refine prompts based on evaluation
Compare AI outputs with course materials, peer discussions, and other sources
Synthesize AI-generated material into original arguments
Justify decisions about what to accept, reject, or revise

Principle 2: Scaffold Reflection Explicitly

Require students to:

Annotate AI transcripts with evaluative comments
Maintain reflective journals documenting what worked, what didn't, and why
Submit "AI audit trails" showing prompt evolution, verification steps, and synthesis decisions
Participate in peer review of AI-supported work, questioning decisions and suggesting alternatives

Principle 3: Assess Process, Not Just Product

Shift assessment focus toward:

Metacognitive awareness (Can students articulate their AI-supported process?)
Epistemic vigilance (Do students verify claims, check sources, and recognize hallucinations?)
Synthesis quality (Do students integrate AI outputs into coherent, original arguments?)
Critical evaluation (Do students question, compare, and develop independent perspectives?)

Principle 4: Differentiate Instructional Support by Context

Adapt strategies based on:

Prior AI experience: Provide foundational literacy for novices; challenge advanced users with complex, ill-structured tasks
Cultural and educational backgrounds: Frame reflection as growth-oriented rather than self-validating where appropriate; emphasize mastery development over relative performance
Institutional norms: Co-create clear AI policies; model ethical use; normalize productive struggle and error as learning opportunities

Conclusion

This study demonstrates that perceived human-AI agency is strongly associated with reflective engagement, which in turn predicts self-reported critical thinking in both UK-based and China-based higher education contexts. These findings support the theorization of an "agency gap"—the perceived distance between students who retain executive control when using AI versus those who experience AI-supported work as passive or algorithmically driven.

The agency gap is not technologically determined but shaped by task design, instructional scaffolding, students' epistemic beliefs, self-regulatory capacities, and contextual norms. When students report retaining control—setting goals, monitoring outputs, questioning claims, and making final judgments—they also tend to report stronger reflective engagement and critical thinking. When they report lower control—accepting outputs uncritically, copying verbatim, or outsourcing judgment—reflection and critical thinking may be bypassed.

These findings challenge deterministic narratives that AI inherently undermines learning or that it inevitably enhances it. Instead, how students use AI matters more than whether they use it. The technology is neither inherently empowering nor disempowering; rather, the learning outcomes depend on the cognitive, metacognitive, and reflective processes students engage when working with AI.

Contextual patterns—reflection associated with self-concept in the UK but not China; AI literacy moderating agency-reflection in China but not UK—suggest that one-size-fits-all approaches may misalign with students' educational backgrounds and institutional contexts. However, because cultural mechanisms were not directly measured and measurement invariance was only partial, these patterns should be treated as tentative and exploratory rather than definitive.

For educators and institutions, the core message is clear: AI-supported learning is most educationally meaningful when it preserves learner agency and scaffolds reflective engagement. This requires shifting from product-oriented to process-oriented assessment, designing tasks that require questioning and synthesis rather than mere retrieval, and providing metacognitive scaffolding that supports epistemic vigilance. By narrowing the agency gap—ensuring students experience themselves as pilots, not passengers—educators can help students leverage AI as a cognitive partner rather than a cognitive surrogate.

Future research should use longitudinal designs, behavioral data, qualitative methods, and direct measures of cultural values, epistemic beliefs, and institutional contexts to test and extend these findings. Only through such rigorous, context-sensitive inquiry can we develop evidence-informed strategies that support students in becoming critical, agentic learners in an increasingly AI-mediated educational landscape.

Research Infographic

The AI Agency Gap Slide Deck

References

Bandura, A. (2009). Social cognitive theory of mass communication. In J. Bryant & M. B. Oliver (Eds.), Media effects: Advances in theory and research (3rd ed., pp. 94–124). Routledge.
Brislin, R. W. (1970). Back-translation for cross-cultural research. Journal of Cross-Cultural Psychology, 1(3), 185–216.
Chan, C. K. Y., & Hu, W. (2023). Students' voices on generative AI: Perceptions, benefits, and challenges in higher education. International Journal of Educational Technology in Higher Education, 20(1), 43.
Chen, S. W., Wang, H. H., Wei, C. F., Fwu, B. J., & Hwang, K. K. (2009). Taiwanese students' self-attributions for two types of achievement goals. The Journal of Social Psychology, 149(2), 179–194.
Darvishi, A., Khosravi, H., Sadiq, S., Gašević, D., & Siemens, G. (2024). Impact of AI assistance on student agency. Computers & Education, 210, 104967.
Dawson, P., Bearman, M., Dollinger, M., & Boud, D. (2024). Validity matters more than cheating. Assessment & Evaluation in Higher Education, 49(7), 1005–1016.
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530.
Fang, L., & Zhou, X. (2026). From tool to co-learner: Exploring student engagement with GenAI through the lens of social constructivism. Evaluation Review, 0(0). https://doi.org/10.1177/0193841X251411618
Fwu, B. J., Chen, S. W., Wei, C. F., & Wang, H. H. (2018). I believe; therefore, I work harder: The significance of reflective thinking on effort-making in academic failure in a Confucian-heritage cultural context. Thinking Skills and Creativity, 30, 19–30.
Hair, J. F., Risher, J. J., Sarstedt, M., & Ringle, C. M. (2019). When to use and how to report the results of PLS-SEM. European Business Review, 31(1), 2–24.
Heine, S. J., Lehman, D. R., Peng, K., & Greenholtz, J. (2002). What's wrong with cross-cultural comparisons of subjective Likert scales? The reference-group effect. Journal of Personality and Social Psychology, 82(6), 903–918.
Henseler, J., Ringle, C. M., & Sarstedt, M. (2016). Testing measurement invariance of composites using partial least squares. International Marketing Review, 33, 405–431.
Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal, 6(1), 1–55.
Jääskelä, P., Poikkeus, A. M., Vasalampi, K., Valleala, U. M., & Rasku-Puttonen, H. (2017). Assessing agency of university students: Validation of the AUS scale. Studies in Higher Education, 42(11), 2061–2079.
Järvelä, S., Nguyen, A., & Hadwin, A. (2023). Human and artificial intelligence collaboration for socially shared regulation in learning. British Journal of Educational Technology, 54(5), 1057–1076.
Kember, D., Leung, D. Y., Jones, A., Loke, A. Y., McKay, J., Sinclair, K., Tse, H., et al. (2000). Development of a questionnaire to measure the level of reflective thinking. Assessment & Evaluation in Higher Education, 25(4), 381–395.
Kock, N. (2015). Common method bias in PLS-SEM: A full collinearity assessment approach. International Journal of e-Collaboration, 11(4), 1–10.
Krakowski, S. (2025). Human-AI agency in the age of generative AI. Information and Organization, 35(1), 100560.
Kremantzis, M., Essien, A., Pantano, E., & Lythreatis, S. (2025). Uncovering the generative AI (GenAI) to agentic AI (AgAI) shift for business school education. Journal of Global Information Management, 33(1), 1–21.
Lindebaum, D., Nolan, E., Ashraff, M., Islam, G., & Ramirez, M. F. (2025). The transformation of epistemic agency and governance in higher education through large language models. Organization Studies, 0(0), 01708406251392002.
Lintner, T. (2024). A systematic review of AI literacy scales. npj Science of Learning, 9(1), 50.
Liu, W. C., Wang, C. K. J., & Parkins, E. J. (2005). A longitudinal study of students' academic self-concept in a streamed setting: The Singapore context. British Journal of Educational Psychology, 75(4), 567–586.
Lowry, P. B., & Gaskin, J. (2014). Partial least squares (PLS) structural equation modeling (SEM) for building and testing behavioral causal theory. IEEE Transactions on Professional Communication, 57(2), 123–146.
Luo, J., & Dawson, P. (2025). Exploring value judgements in grading: Will teachers mark down student work assisted by GenAI, and should they? Studies in Higher Education, Online first, 1–15.
Martín-Moncunill, D., & Alonso Martínez, D. (2025). Students' trust in AI and their verification strategies: A case study at Camilo José Cela University. Education Sciences, 15(10), 1307.
Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitudinal mediation. Psychological Methods, 12(1), 23–44.
Min, I., Cortina, K. S., & Miller, K. F. (2016). Modesty bias and the attitude-achievement paradox across nations: A reanalysis of TIMSS. Learning and Individual Differences, 51, 359–366.
Nandagopal, S. (2025). Transforming the self: Individual-level changes arising from collaboration with generative AI. Computers in Human Behavior: Artificial Humans, 6, 100232.
O'Laughlin, K. D., Martin, M. J., & Ferrer, E. (2018). Cross-sectional analysis of longitudinal mediation processes. Multivariate Behavioral Research, 53(3), 375–402.
Pintrich, P. R., Smith, D. A. F., Garcia, T., & McKeachie, W. J. (1991). A manual for the use of the Motivated Strategies for Learning Questionnaire (MSLQ). National Center for Research to Improve Postsecondary Teaching and Learning, University of Michigan.
Polyportis, A., & Pahos, N. (2025). Understanding students' adoption of the ChatGPT chatbot in higher education: The role of anthropomorphism, trust, design novelty and institutional policy. Behaviour & Information Technology, 44(2), 315–336.
Ravšelj, D., Keržič, D., Tomaževič, N., Umek, L., Brezovar, N., Iahad, N. A., Abdulla, A. A., et al. (2025). Higher education students' perceptions of ChatGPT: A global study of early reactions. PLoS One, 20(2), e0315011.
Rudolph, J., bin Mohamed Ismail, M. F., & Popenici, S. (2024). Higher education's generative artificial intelligence paradox: The meaning of chatbot mania. Journal of University Teaching and Learning Practice, 21(6), 14–48.
Salhab, R., & Aboushi, M. M. (2025). Influence of AI literacy and 21st-century skills on the acceptance of generative artificial intelligence among college students. Frontiers in Education, 10, 1640212.
Salomon, G. (Ed.). (1997). Distributed cognitions: Psychological and educational considerations. Cambridge University Press.
Sedikides, C., Gaertner, L., & Toguchi, Y. (2003). Pancultural self-enhancement. Journal of Personality and Social Psychology, 84(1), 60.
Sharples, M. (2023). Towards social generative AI for education: Theory, practices and ethics. Learning: Research and Practice, 9(2), 159–167.
Taras, V., Steel, P., & Kirkman, B. L. (2016). Does country equate with culture? Beyond geography in the search for cultural boundaries. Management International Review, 56(4), 455–487.
Urban, M., Brom, C., Lukavský, J., Děchtěrenko, F., Hein, V., Svacha, F., Kmoníčková, P., & Urban, K. (2025). ChatGPT can make mistakes. Check important info. Epistemic beliefs and metacognitive accuracy in students' integration of ChatGPT content into academic writing. British Journal of Educational Technology, 56, 1897–1918.
Urhahne, D., Kehle, L., Dietrich, L., & Kremer, K. (2026). The role of epistemic beliefs in predicting ChatGPT adoption and avoidance in higher education. Acta Psychologica, 263, 106334.
Wang, B., Rau, P. L. P., & Yuan, T. (2023). Measuring user competence in using artificial intelligence: Validity and reliability of artificial intelligence literacy scale. Behaviour & Information Technology, 42(9), 1324–1337.
Wittig McPhee, S., & Jerowsky, M. (2025). Beyond technical skills: A pedagogical perspective on fostering critical engagement with generative AI in university classrooms. Frontiers in Education, 10, 1593278.
Wu, H., Guo, Y., Yang, Y., Zhao, L., & Guo, C. (2021). A meta-analysis of the longitudinal relationship between academic self-concept and academic achievement. Educational Psychology Review, 33(4), 1749–1778.
Xie, Q., Li, M., & Cheng, F. (2025). Between regulation and accessibility: How Chinese university students navigate global and domestic generative AI. Globalisation, Societies and Education, Online first, 1–19.
Yang, Y., Luo, J., Yang, M., Yang, R., & Chen, J. (2024). From surface to deep learning approaches with generative AI in higher education: An analytical framework of student agency. Studies in Higher Education, 49(5), 817–830.
Yang, Y., Zhang, Y., Sun, D., He, W., & Wei, Y. (2025). Navigating the landscape of AI literacy education: Insights from a decade of research (2014–2024). Humanities and Social Sciences Communications, 12(1), 1–12.
Yuan, B., & Hu, J. (2024). Generative AI as a tool for enhancing reflective learning in students. arXiv preprint arXiv:2412.02603.
Yusuf, A., Pervin, N., & Román-González, M. (2024). Generative AI and the future of higher education: A threat to academic integrity or reformation? Evidence from multicultural perspectives. International Journal of Educational Technology in Higher Education, 21(1), 21.
Zhai, C., Wibowo, S., & Li, L. D. (2024). The effects of over-reliance on AI dialogue systems on students' cognitive abilities: A systematic review. Smart Learning Environments, 11(1), 28.

Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Professor, Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Organizational Leadership (UVU). Read Jonathan Westover's executive profile here.

Suggested Citation: Westover, J. H. (2026). The Agency Gap in AI-Assisted Higher Education: When Learner Control Shapes Reflective Practice and Critical Thinking. Human Capital Leadership Review, 36(3). doi.org/10.70175/hclreview.2020.36.3.5