From credibility to reliance: Trust in generative AI as an academic information source

Generative AI in higher education

Generative AI is now part of routine academic work. Students use these systems to summarise readings, clarify concepts, generate ideas, refine writing, and obtain academic assistance (Baig and Yadegaridehkordi, 2024; Blahopoulou and Ortiz-Bonnin, 2025). Much of the existing literature has interpreted this expansion primarily through adoption-oriented constructs such as perceived usefulness, ease of use, attitudes, and behavioural intention (Faraon et al., 2025; Shuhaiber et al., 2025). The adoption perspective is useful for explaining students’ willingness to engage with Generative AI as a tool, but it is less effective for explaining how students evaluate the information quality of AI-generated content once those outputs enter academic work. The analytical problem, therefore, shifts from technology acceptance to information evaluation.

Generative AI as an academic information source

Treating Generative AI as an academic information source changes the unit of analysis. When students request Generative AI to explain a theory, compare viewpoints, summarise a concept, or suggest a direction for an assignment, they are engaging in academic information seeking (Holland and Ciachir, 2025). The key question is not only whether the system is used, but whether the resulting information is judged suitable for academic purposes.

This distinction matters because AI-generated outputs differ from conventional academic sources. Journal articles, textbooks, and institutional materials usually provide cues of authorship, publication context, references, and evidentiary basis. By contrast, Generative AI often produces fluent responses without clearly revealing the provenance, certainty, or source basis of its claims (Hwang and Jeong, 2025; Schilke and Reimann, 2025). Students may therefore use coherence and convenience as proxies for quality even when the informational basis of a response is uncertain.

This difference changes the unit of evaluation, not merely the technology being used. With conventional academic sources, students often evaluate information through visible markers of authorship, publication legitimacy, and traceable evidence. With Generative AI, those source-level cues are often obscured or absent, so evaluation shifts more directly to the plausibility of the output itself and to the user's willingness to validate it before it is used for academic purposes. The issue is therefore not simply whether a new tool has been adopted, but how unverified yet plausible content is judged, checked, and incorporated into academic work. This framing aligns the study with scholarship on information services and use, particularly work on information quality judgement, digital source evaluation, and responsible information behaviour (Bearman et al., 2024; Kunz et al., 2024; Lao et al., 2025).

Theoretical grounding: credibility, trust, verification, and reliance

This study is grounded primarily in Source Credibility Theory, originally developed in communication research by Hovland and colleagues, which explains how users evaluate information in terms of perceived believability, reliability, and trustworthiness of a source or message (Hovland et al., 1953; Appelman and Sundar, 2016; Lao et al., 2025). In AI-mediated environments, credibility judgements are especially important because source cues are often incomplete.

Within this study, perceived credibility refers to a plausibility judgement: whether AI-generated academic information appears believable, accurate, and dependable enough to merit academic consideration. Trust is conceptually distinct. Whereas credibility concerns the assessment of informational quality, trust refers to the willingness to depend on that information under uncertainty. A student may judge an AI-generated response to be plausible yet still withhold trust if its evidentiary basis remains unclear (Appelman and Sundar, 2016; Nazaretsky et al., 2025). Reliance is also distinct from both credibility and trust because it refers to the academic incorporation of AI-generated information into study-related tasks. In this framework, credibility is an evaluative judgement about the information, trust is a willingness to depend on that information under uncertainty, and reliance is the use-oriented response. A student may therefore perceive an AI-generated output as credible without fully trusting it, and may trust it without necessarily relying on it for academic work.

Source Credibility Theory alone, however, does not fully explain what students do after that initial plausibility judgement. For this reason, the study also draws on an information literacy and digital verification perspective. This perspective emphasises that responsible academic use of digital information requires checking, comparing, and interpreting information before integrating it into scholarly work (Bearman et al., 2024; Kunz et al., 2024). In the present context, verification behaviour refers to uncertainty management: the extent to which students cross-check AI-generated outputs against textbooks, peer-reviewed sources, lecture materials, or instructor guidance before relying on them for academic purposes (Martín-Moncunill and Alonso Martínez, 2025).

Reliance, in turn, refers to academic incorporation: the degree to which AI-generated information is actually used for academic tasks such as explanation, planning, summarisation, or coursework completion (Stojanov et al., 2024). Verification is theoretically important because it could plausibly work in two directions. On one hand, checking AI outputs may reduce reliance by exposing errors, fabricated claims, or weak reasoning. On the other hand, in academic settings it may increase reliance when cross-checking confirms that the information is usable. The present study predicts the latter pattern because reliance in this context is conceptualised not as uncritical acceptance, but as selective academic use after evaluative scrutiny. Together, these perspectives support a four-construct framework linking plausibility judgement, willingness to depend, uncertainty management, and academic incorporation (Appelman and Sundar, 2016; Martín-Moncunill and Alonso Martínez, 2025; Nazaretsky et al., 2025; Stojanov et al., 2024).

Research gap and conceptual framework

The unresolved problem in the literature is not simply that prior studies are limited, but that the field still lacks a tested explanation of how evaluative judgement becomes consequential academic use. Existing research has documented prevalence, perceived benefits, and broad concerns, and some studies have examined trust or verification in more isolated ways (Baig and Yadegaridehkordi, 2024; Bittle and El-Gayar, 2025). However, the literature still does not adequately explain how students move from encountering plausible AI-generated content to judging whether it is trustworthy, sufficiently checked, and suitable for academic incorporation.

The present study addresses this gap by treating Generative AI as an academic information source and by linking perceived credibility, trust, verification behaviour, and reliance within one framework. In addition, the study examines whether the proposed relationships operate similarly across undergraduate and postgraduate students as a supplementary group comparison. The direct path from perceived credibility to reliance is also estimated to assess whether trust explains this association. Figure 1 presents the conceptual framework depicting the association between these constructs.

OPEN IN VIEWER

Figure 1.

Conceptual framework of students’ use of generative AI as an academic information source. Note. The dashed arrow indicates the direct path from perceived credibility to reliance estimated as part of the indirect-association assessment.

Hypotheses development

Perceived credibility is expected to be positively associated with trust in AI-generated academic information. Users are more likely to trust information they judge to be believable and dependable (Appelman and Sundar, 2016; Nazaretsky et al., 2025).

H1: Perceived credibility is positively associated with trust in AI-generated academic information.

Trust is expected to be positively associated with reliance. Trust reduces uncertainty and increases willingness to depend on information despite incomplete assurance (Visser et al., 2025; Wang and Li, 2024).

H2: Trust is positively associated with reliance on AI-generated academic information.

Verification behaviour is also expected to be positively associated with reliance. Although verification could reduce reliance when checking reveals weaknesses in AI-generated output, the present study expects a positive association in academic contexts because students who verify information successfully may become more confident in using content that withstands scrutiny. In this sense, reliance reflects selective and validated academic use rather than passive acceptance (Bearman et al., 2024; Hou et al., 2025).

H3: Verification behaviour is positively associated with reliance on AI-generated academic information.

Trust is further expected to carry the indirect association between perceived credibility and reliance. Reliance is more likely when credibility judgements are accompanied by a willingness to accept the information as dependable for academic purposes (Appelman and Sundar, 2016; Ferrario, 2025; Wang et al., 2025).

H4: Trust carries an indirect association between perceived credibility and reliance on AI-generated academic information.

In addition to H1-H4, this study also examined whether the structural relationships differ between undergraduate and postgraduate students. This comparison is treated as a supplementary robustness question rather than as a central theoretical hypothesis, given that the core contribution of the study lies in explaining the evaluative pathway from credibility to reliance.

Research design

This study used a cross-sectional, multi-university survey design to examine how undergraduate and postgraduate students evaluated and used Generative AI for academic purposes. It was designed to model associative patterns among perceived credibility, trust, verification behaviour, and reliance across institutional settings rather than test an intervention or establish causal ordering (Baig and Yadegaridehkordi, 2024; Blahopoulou and Ortiz-Bonnin, 2025).

Study setting, population, and sampling

The study was conducted across four universities in Nepal: Kathmandu University, Tribhuvan University, Pokhara University, and Purbanchal University, and included both undergraduate and postgraduate students. This multi-university setting provided institutional variation while supporting comparison across students with different levels of academic experience (Barber and Anderson, 2025; Nazaretsky et al., 2025).

The target population comprised students who had used any Generative AI tool for at least one academic task within the previous six months. Academic tasks included idea generation, summarisation, concept explanation, drafting support, language refinement, and assignment-related assistance (Blahopoulou and Ortiz-Bonnin, 2025; Wang et al., 2025). Students with no prior academic use of Generative AI were excluded because the focal constructs required direct experience with AI-generated academic information.

A multi-stage non-probability sampling strategy was employed. First, the four universities were selected because they are among the major university systems in Nepal and provided access to students from different academic levels and broad disciplinary areas. Second, students were approached across study levels and broad academic disciplines through accessible university channels, including online circulation and supervised in-class administration where institutional access permitted. Because recruitment was conducted through accessible university channels rather than a centralised invitation list, an exact response rate could not be calculated. This approach suited the study because the aim was to test theoretically specified relationships across a heterogeneous student sample rather than estimate nationally representative prevalence rates (Yang and Banamah, 2014). Because university-wise identifiers were not retained for institution-level analysis, the findings should be interpreted as evidence from a heterogeneous multi-university sample rather than as university-specific or nationally representative estimates.

A total of 650 valid responses were retained for analysis. Valid responses were those that met the eligibility requirement, were usable for analysis, and did not display obvious response problems during screening. The sample size was adequate for confirmatory factor analysis and structural equation modelling, given the number of latent constructs and hypothesised paths (White, 2023; Yang et al., 2025). Although a formal a priori power analysis was not conducted, the retained sample exceeded commonly used SEM adequacy expectations for models with multiple latent constructs and observed indicators. The final sample provided sufficient cases for CFA, structural path estimation, bootstrapped indirect-association testing, and supplementary multi-group comparison by study level.

Instrument development

Data were collected using a structured and self-administered questionnaire administered in English. The instrument adapted established measures from prior research on message credibility, trust in automated systems, verification behaviour, and reliance on ChatGPT for learning (Appelman and Sundar, 2016; Jian et al., 2000; Stojanov et al., 2024; Tifferet, 2021). These source measures had previously been used in studies of credibility, trust, verification, or AI-supported learning, while the adapted instrument was evaluated in the present study through expert review, pilot testing, reliability assessment, confirmatory factor analysis, and validity checks. Item wording was revised to fit Generative AI in higher education and academic use. The adapted items were reviewed for content relevance, conceptual alignment, clarity, and suitability for students’ academic use of Generative AI before full data collection.

The questionnaire comprised five sections. The first section recorded demographic and academic background information, the second captured patterns of academic use, and the remaining sections measured perceived credibility, trust, verification behaviour, and reliance.

All substantive construct items were measured in a five-point Likert format ranging from 1 = strongly disagree to 5 = strongly agree (Garratt et al., 2011). Before full administration, the instrument underwent expert review by seven reviewers to assess content relevance, clarity, conceptual alignment, and contextual appropriateness (Mokkink et al., 2025). It was then pilot-tested with 30 eligible students to identify ambiguous wording, estimate completion time, and refine difficult items (Presser et al., 2004). Revisions focused mainly on language refinement, clearer wording of source-verification items, and closer alignment of selected items with academic-use contexts. The questionnaire structure, response scale, construct item wording, item-level descriptive statistics, source notes, and standardised loadings are reported within the manuscript, particularly in Section 3.3, Section 4.2, and Table 2.

Measures

Perceived credibility captured students’ judgements about whether AI-generated academic information was believable, accurate, dependable, and worthy of academic consideration (Appelman and Sundar, 2016).

Trust referred to students’ confidence in, and willingness to accept, AI-generated academic information as sufficiently dependable for academic purposes (Jian et al., 2000). Trust was treated as distinct from credibility: credibility reflected judgement of quality, whereas trust reflected readiness to depend on that information under uncertainty.

Verification behaviour referred to the extent to which students cross-checked or validated AI-generated academic information before using it in study tasks, assignments, or academic decision-making (Tifferet, 2021).

Reliance referred to the degree to which students depended on Generative AI outputs in academic information seeking and task completion (Stojanov et al., 2024; Wang et al., 2025). It therefore captured academic dependence rather than the simple frequency of use.

Control variables included frequency of Generative AI use, prior AI experience, level of study, and broad academic discipline (Blahopoulou and Ortiz-Bonnin, 2025; Nazaretsky et al., 2025). Accordingly, reliance was treated as the academic incorporation of AI-generated information, whereas trust was treated as a willingness to depend on information under uncertainty rather than as another label for perceived credibility.

Data collection procedure

Data were collected through a combination of online distribution and supervised in-class administration, depending on institutional access and feasibility. Approximately 40% of responses were collected online, and 60% were collected through supervised in-class administration. Participation was voluntary and anonymous. Before participation, respondents received an information sheet explaining the study purpose, eligibility criteria, completion time, confidentiality protections, and the voluntary nature of participation. Only those who indicated informed consent proceeded to the survey (Sagitova et al., 2024; Yang et al., 2025).

To reduce common method concerns, the questionnaire used clear, non-leading wording, followed a logical sequence, emphasised anonymity, and informed respondents that there were no right or wrong answers (Yao and Xu, 2024). Screening items confirmed prior academic use of Generative AI and ensured eligibility.

Data analysis

Data analysis was conducted in R statistical software using lavaan, semTools, and psych for measurement evaluation, covariance-based structural equation modelling, indirect-association testing, and measurement invariance assessment. The first stage involved data screening, including checks for incompleteness, excessive missingness, patterned responding, and distributional irregularities (Stosic et al., 2024; Yao and Xu, 2024). After eligibility and quality screening, 650 responses were retained for analysis.

The second stage reported descriptive statistics and evaluated reliability using Cronbach's alpha and composite reliability. Construct validity was examined using confirmatory factor analysis, including standardised factor loadings, average variance extracted, and discriminant validity indicators (Cheung et al., 2024).

The third stage tested the hypothesised relationships using structural equation modelling. Model fit was evaluated using the chi-square to degrees-of-freedom ratio, comparative fit index, Tucker-Lewis index, root mean square error of approximation, and standardised root mean square residual (Hu and Bentler, 1999). The indirect association between perceived credibility and reliance through trust was examined using bootstrapped indirect-association analysis. In the controlled SEM, trust was regressed on perceived credibility and the control variables, while reliance was regressed on perceived credibility, trust, verification behaviour, and the same control variables. Because the data were cross-sectional, indirect-association results were interpreted as associative rather than causal.

Finally, multi-group analysis was used to examine whether the structural relationships differed between undergraduate and postgraduate students, with measurement invariance assessed before group comparison (Sterner et al., 2025).

Ethical considerations

This study adhered to the ethical principles set out in the Declaration of Helsinki. The study involved an anonymous minimal-risk educational survey of adult participants, collected no personally identifiable or sensitive data, and did not involve intervention, deception, or vulnerable participants. In the participating context, formal ethical review was not required for this type of anonymous minimal-risk survey. Participation was voluntary, informed consent was obtained before the survey, and all data were used only for academic research purposes and stored securely (Sagitova et al., 2024; Yang et al., 2025).

Respondent profile and patterns of generative AI use

As shown in Table 1, the sample comprised 650 students drawn from four universities, including 339 undergraduate students (52.15%) and 311 postgraduate students (47.85%). The sample was broadly distributed across age groups, gender, and disciplines, with the largest disciplinary group from Science/Engineering (27.23%), followed by Management/Business (21.38%), Social Sciences/Humanities (20.46%), Health/Medical (19.08%), and Other disciplines (11.85%). Respondents also had prior exposure to Generative AI: most reported moderate or high prior AI experience, and the largest group reported using Generative AI often for academic purposes. Explaining concepts was the most common academic purpose, followed by summarising readings and idea generation or brainstorming. Overall, the profile indicates that the analysis concerns students with practical experience of Generative AI rather than first-time or purely hypothetical users.

OPEN IN VIEWER

Table 1.

Respondent profile and generative AI use characteristics.

Variable	Category	n	%
Level of study	Undergraduate	339	52.15
	Postgraduate	311	47.85
Age group	18–20	126	19.38
	21–23	218	33.54
	24–26	177	27.23
	27–29	80	12.31
	30 or above	49	7.54
Gender	Female	364	56.00
	Male	286	44.00
Broad academic discipline	Science/Engineering	177	27.23
	Management/Business	139	21.38
	Social Sciences/Humanities	133	20.46
	Health/Medical	124	19.08
	Other	77	11.85
Prior AI experience	Very low	29	4.46
	Low	132	20.31
	Moderate	233	35.85
	High	182	28.00
	Very high	74	11.38
Frequency of GenAI academic use	Rarely	71	10.92
	Occasionally	101	15.54
	Sometimes	143	22.00
	Often	244	37.54
	Very often	91	14.00
Most common academic purpose	Explaining concepts	174	26.77
	Summarising readings	121	18.62
	Idea generation/brainstorming	114	17.54
	Drafting or writing support	86	13.23
	Assignment direction/planning	74	11.38
	Language refinement/editing	61	9.38
	Other	20	3.08
Familiarity with GenAI	Very unfamiliar	17	2.62
	Unfamiliar	57	8.77
	Neutral	82	12.62
	Familiar	262	40.31
	Very familiar	232	35.69

Descriptive statistics and measurement quality

Table 2 reports the descriptive statistics and measurement properties of the four latent constructs. Verification behaviour recorded the highest mean (M = 3.93, SD = 0.72), indicating relatively frequent checking or cross-validation of AI-generated academic information before use. Reliance showed a moderate mean (M = 3.33, SD = 0.90), while perceived credibility (M = 3.22, SD = 0.90) and trust (M = 3.13, SD = 0.89) were also moderate. This pattern suggests that although respondents did not rate AI-generated academic information as uniformly highly credible or highly trustworthy, they still reported substantial engagement with verification and a moderate level of reliance in academic contexts.

OPEN IN VIEWER

Table 2.

Measurement items, descriptive statistics, and standardised loadings

Construct	Code	Item wording	Mean	SD	Loading
Perceived credibility	PC1	The academic information generated by Generative AI is accurate.	3.17	1.14	.745
Perceived credibility	PC2	The academic information generated by Generative AI is believable.	3.48	1.16	.688
Perceived credibility	PC3	The academic information generated by Generative AI appears authentic rather than misleading.	3.00	1.19	.674
Perceived credibility	PC4	The academic information generated by Generative AI is dependable enough to be taken seriously in academic work.	3.22	1.14	.671
Trust	TR1	I feel confident using academic information generated by Generative AI.	3.16	1.13	.745
Trust	TR2	Academic information generated by Generative AI is reliable for routine academic use.	3.27	1.15	.646
Trust	TR3	I can depend on academic information generated by Generative AI when working on my studies.	2.90	1.17	.622
Trust	TR4	Overall, I trust academic information generated by Generative AI.	3.18	1.13	.719
Verification behaviour	VB1	Before using information from Generative AI, I check it against other reliable sources.	4.13	0.87	.686
Verification behaviour	VB2	I compare Generative AI responses with textbooks, journal articles, or lecture materials.	3.98	0.95	.650
Verification behaviour	VB3	I verify unfamiliar claims produced by Generative AI before using them academically.	3.90	0.96	.668
Verification behaviour	VB4	I check whether examples, facts, or references provided by Generative AI are correct.	3.61	1.06	.671
Verification behaviour	VB5	I do not use academic information from Generative AI without first validating it.	4.04	0.96	.717
Reliance	RL1	I rely on Generative AI to obtain initial explanations of academic topics.	3.46	1.13	.717
Reliance	RL2	I rely on Generative AI to summarise or organise academic information.	3.37	1.16	.743
Reliance	RL3	I rely on Generative AI when I need direction for assignments or study tasks.	3.35	1.15	.690
Reliance	RL4	I rely on information generated by Generative AI as part of completing academic work.	3.38	1.16	.711
Reliance	RL5	Generative AI is one of my regular academic information sources.	3.08	1.21	.679

Note. All items were measured in a five-point Likert format from 1 = strongly disagree to 5 = strongly agree. PC items were adapted from Appelman and Sundar (2016), TR items from Jian et al. (2000), VB items from Tifferet (2021), and RL items from Stojanov et al. (2024). PC = perceived credibility; TR = trust; VB = verification behaviour; RL = reliance.

The measurement properties were acceptable, although convergent validity required cautious interpretation for three constructs. Cronbach's alpha values ranged from 0.775 to 0.834, composite reliability values ranged from 0.778 to 0.834, and standardised factor loadings ranged from 0.622 to 0.745. AVE was above 0.50 for reliance and slightly below 0.50 for perceived credibility, trust, and verification behaviour. Following Fornell and Larcker (1981), these marginal AVE values were retained with caution because composite reliability was acceptable, all standardised loadings were statistically significant, discriminant validity was supported, and the proposed four-factor CFA model outperformed competing models. To examine whether the modest AVE limitations affected the SEM results, a post-hoc sensitivity analysis was also conducted and is reported in Table 3.

OPEN IN VIEWER

Table 3.

Additional validity and robustness checks

Check	Key evidence	Conclusion
Fornell–Larcker criterion	√AVE values ranged from .679 to .709, exceeding the highest inter-construct correlations for all constructs.	Discriminant validity supported.
HTMT ratios	HTMT values ranged from .050 to .685, all below .85.	Discriminant validity supported.
Competing CFA models	The proposed four-factor model showed the best fit, robust CFI = .993, robust TLI = .992, robust RMSEA = .018, and outperformed the one-factor, two-factor, and alternative three-factor models.	Proposed measurement structure supported.
Harman's single-factor test	The first unrotated factor explained 27.89% of total variance.	No strong indication that common method variance dominated the results.
Single-factor CFA diagnostic	The single-factor model showed poor fit, robust CFI = .548, robust TLI = .488, robust RMSEA = .140, SRMR = .128.	Preliminary diagnostic evidence did not suggest that common method variance dominated the observed pattern of results.
AVE sensitivity analysis	The weakest-loading indicators from the three constructs with AVE below .50 were removed: PC4, TR3, and VB2. The sensitivity model showed excellent fit, CFI = .999, TLI = .998, RMSEA = .008, SRMR = .026. AVE improved for perceived credibility (.499) and trust (.502), while verification behaviour remained modestly below .50 (.471). The main paths remained stable: perceived credibility → trust, β = .662, p < .001; trust → reliance, β = .431, p < .001; verification behaviour → reliance, β = .174, p < .001; perceived credibility → reliance remained non-significant, β = .133, p = .063; indirect association, β = .285, p < .001.	The modest AVE limitations did not materially alter the substantive SEM conclusions.
SEM with controls	After controlling for Generative AI use frequency, prior AI experience, level of study, and discipline, the main paths remained stable: perceived credibility → trust, β = .670, p < .001; trust → reliance, β = .510, p < .001; verification behaviour → reliance, β = .186, p < .001; perceived credibility → reliance remained non-significant, β = .037, p = .609.	Main structural findings remained stable.

Note. AVE = average variance extracted; HTMT = heterotrait–monotrait ratio; CFA = confirmatory factor analysis; SEM = structural equation modelling. The full-item model was retained as the primary model because it preserved the theoretically specified construct coverage, while the reduced-indicator model was used only as a sensitivity check.

Additional validity and robustness checks

As shown in Table 3, additional checks supported the distinctiveness and robustness of the measurement and structural models. First, discriminant validity was supported using both the Fornell–Larcker criterion and HTMT. For all constructs, the square root of AVE exceeded the corresponding inter-construct correlations, and all HTMT ratios were below 0.85. Second, the proposed four-factor CFA model fitted the data substantially better than the alternative one-, two- and three-factor models, supporting the empirical distinctiveness of perceived credibility, trust, verification behaviour, and reliance. Third, common method diagnostics were treated as preliminary checks rather than definitive evidence and did not suggest that common method variance dominated the observed relationships. Harman's single-factor test showed that the first unrotated factor accounted for 27.89% of total variance, and the single-factor CFA model showed poor fit. Fourth, because AVE values for perceived credibility, trust, and verification behaviour were marginally below .50, a reduced-indicator sensitivity analysis was conducted by removing the weakest-loading indicators from these constructs. The main SEM conclusions remained substantively stable. Finally, the structural relationships also remained stable after including controls for Generative AI use frequency, prior AI experience, level of study, and academic discipline. Taken together, these checks support the proposed four-factor structure and indicate that the main SEM findings were not materially dependent on measurement or control-variable specifications.

Structural model and hypothesis testing

The structural model showed a very good fit to the data, with CFI = 0.993, TLI = 0.992, RMSEA = 0.017, and SRMR = 0.028. The model explained 46.6% of the variance in trust and 34.0% of the variance in reliance, indicating that the proposed framework accounted for a meaningful share of students’ evaluative and use-related responses toward Generative AI as an academic information source. Table 4 presents the structural path estimates and hypothesis-testing results, while Figure 2 summarises the supported structural relationships visually. As shown in Table 4 and Figure 2, perceived credibility was positively associated with trust (β = 0.683, p < .001), supporting H1. Trust was also positively associated with reliance (β = 0.497, p < .001), supporting H2. Verification behaviour was positively associated with reliance (β = 0.177, p < .001), supporting H3. By contrast, the direct association of perceived credibility with reliance was not statistically significant (β = 0.069, p = .327), indicating that credibility alone did not show a direct association with reliance once trust was included in the model. The indirect-association analysis showed that the indirect association of perceived credibility with reliance through trust was positive and statistically significant (β = 0.339, p < .001), thereby supporting H4. Because the indirect association was significant while the direct association was not, the findings are consistent with an indirect-only association pattern. The total standardised association of perceived credibility with reliance was 0.409 (p < .001).

OPEN IN VIEWER

Figure 2.

Structural model showing standardised path coefficients and factor loadings. Note: Only statistically supported structural paths are displayed for visual clarity. The non-significant direct association between perceived credibility and reliance is reported in Table 4. PC1–PC4 = perceived credibility items; TR1–TR4 = trust items; VB1–VB5 = verification behaviour items; RL1–RL5 = reliance items.

OPEN IN VIEWER

Table 4.

Structural model results and hypothesis testing.

Hypothesis/Effect	Path	Standardised β	SE	z	p	Decision
H1	Perceived Credibility → Trust	0.683	0.035	19.560	< .001	Supported
H2	Trust → Reliance	0.497	0.070	7.090	< .001	Supported
H3	Verification Behaviour → Reliance	0.177	0.044	4.042	< .001	Supported
Direct association	Perceived Credibility → Reliance	0.069	0.071	0.977	.327	Not supported
H4	Indirect association: Perceived Credibility → Trust → Reliance	0.339	0.073	5.726	< .001	Supported

Note. Standardised coefficients are reported. The direct association is included to clarify the indirect-association pattern. Because the data are cross-sectional, the indirect association should not be interpreted as evidence of causal mediation.

Overall, the results are consistent with an indirect-association pattern in which trust statistically links credibility judgements to reliance, while verification behaviour is also independently associated with students’ use of Generative AI in academic contexts. Because the data are cross-sectional, this pattern should be interpreted as a concurrent association rather than evidence of causal or chronological mediation.

Multi-group analysis by level of study

A multi-group analysis was conducted as a robustness and generalisability check to examine whether the measurement and structural relationships differed between undergraduate and postgraduate students. As shown in Table 5, the measurement invariance tests supported configural, metric, and scalar invariance, as changes in robust fit indices were negligible across increasingly constrained models. The configural model showed good fit, and imposing equality constraints on factor loadings and then on intercepts did not materially reduce model fit. The robust chi-square difference tests for both the configural-to-metric comparison (p = .659) and the metric-to-scalar comparison (p = .525) were non-significant. These findings indicate that the four latent constructs operated comparably across the two study levels. Table 5 also shows that constraining the structural paths to be equal across undergraduate and postgraduate students did not significantly worsen model fit (Δχ² = 2.166, Δdf = 4, p = .705). The group-specific path estimates showed closely similar patterns in both groups. Therefore, the supplementary group comparison did not indicate meaningful structural differences between undergraduate and postgraduate students, suggesting that the proposed model operates in a broadly stable and comparable manner across the two study levels.

OPEN IN VIEWER

Table 5.

Measurement invariance and multi-group comparison by level of study.

Panel A. Measurement invariance
Model	Robust CFI	Robust TLI	Robust RMSEA	SRMR	ΔCFI	ΔRMSEA	χ2 difference test
Configural	0.990	0.988	0.021	0.037	—	—	—
Metric	0.991	0.990	0.020	0.039	0.001	−0.001	Δχ2(14) = 11.342, p = .659
Scalar	0.991	0.990	0.019	0.040	0.000	−0.001	Δχ2(14) = 13.021, p = .525

OPEN IN VIEWER

Panel B. Structural invariance and group-specific paths.

Path	Undergraduate β	Postgraduate β	Structural difference
Perceived Credibility → Trust	0.710***	0.654***	No significant difference
Trust → Reliance	0.564***	0.447***	No significant difference
Verification Behaviour → Reliance	0.220***	0.149*	No significant difference
Perceived Credibility → Reliance	0.006	0.129	No significant difference

Note: Structural invariance test: Δχ²(4) = 2.166, p = .705. * p < .05, *** p < .001.

Practical implications for information literacy and academic governance

The findings have practical implications for academic libraries, information professionals, and university governance. For academic libraries and information professionals, Generative AI should be incorporated into information literacy instruction as an academic source-evaluation issue, not only as a tool-use issue. Training should help students distinguish fluent AI-generated text from evidence-based information, verify claims against textbooks, journal articles, lecture materials, and trusted databases, and document how AI-generated information has been checked before being used for academic purposes. At the institutional level, universities should move beyond individual student responsibility and develop clear guidance on acceptable AI use, verification expectations, citation or disclosure practices, and discipline-specific standards for academic work. Administrators, curriculum planners, and academic leaders thus have a role in embedding AI verification expectations into assessment policies, course guidance, and information-literacy provision. Such guidance can support responsible use while reducing the risk that students rely on plausible but insufficiently verified AI-generated information.