Factors shaping the adoption of large language models among hospital administrative staff: A cross-sectional survey study

Abstract

Objective

The use of large language models (LLMs) in hospital administration is increasing, yet the factors influencing LLM adoption among administrative staff remain insufficiently explored. This study aims to identify the key determinants of LLM adoption among hospital administrative staff.

Methods

An integrated model was developed by incorporating constructs from the Technology Acceptance Model (TAM) along with task–technology fit (TTF), trust (TR), social influence (SI), and work stress (WS). A questionnaire survey was carried out among administrative staff in 40 public hospitals in Guangxi, China, and 1,377 valid responses were collected using a stratified cluster sampling approach. The integrated model and hypothesised relationships were then tested using partial least squares structural equation modelling (PLS-SEM).

Results

TTF was the strongest predictor of perceived usefulness (PU). Perceived ease of use (PEOU) influenced behavioral intention (BI) primarily indirectly through PU. SI showed positive associations with both PU and BI. TR was an important predictor of BI and also of use behavior (UB). WS was positively related to UB; however, it did not moderate the relationship between BI and UB.

Conclusion

This study suggests that the adoption of LLMs among hospital administrative staff is shaped not only by the core elements of the TAM, but also by TTF, TR, SI and WS. These findings provide practical insights for public hospitals seeking to support the effective use of LLMs in administrative work.

Keywords

large language models (LLMs)hospital administrative staff technology acceptance model (TAM)adoption

1 Introduction

Large language models (LLMs), such as ChatGPT, Gemini, and Claude, are a major form of generative artificial intelligence (GAI) and have contributed to the digital transformation of knowledge-intensive work. These models can process natural language, combine information, and produce coherent outputs with a high degree of fluency because they were trained on large-scale corpora.^1–3 Since OpenAI released ChatGPT in November 2022, LLMs have rapidly expanded across sectors, including education, healthcare, finance, and public administration.^4–6 An increasing corpus of research indicates that LLMs may enhance efficiency in information processing and decision support, transform human-AI collaboration, and impact practices in organizational management, research, and education.^7–9 The availability of locally developed LLMs in China, such as DeepSeek, Qwen, ERNIE Bot, and SparkDesk, has made access even easier and encouraged users to look into how they can be used in specific situations. The Interim Measures for the Management of Generative Artificial Intelligence Services (2023) were put in place to help different industries, especially healthcare, develop GAI in a safe and responsible way.

In this context, there has been growing interest in the application of LLMs in healthcare practice. Previous research has demonstrated the potential value of LLMs in clinical decision support, scientific writing enhancement, automated documentation, and medical text processing.^3,6,10 Notably, beyond clinical applications, LLMs may also play an important role in hospital administration.¹ In general, LLM use in hospital administrative work can be summarised into four functional categories: ideation support and first-draft generation, interactive text refinement, information retrieval and synthesis, and data analysis and report drafting (Figure 1). Specifically, ideation support and first-draft generation refers to helping administrative staff obtain a structured outline and rapidly produce an initial draft when preparing administrative documents such as notices, meeting minutes, and briefing reports. Interactive text refinement involves iterative, multi-turn interactions with LLMs to polish language and improve structure and logic in existing texts. Information retrieval and synthesis refers to using LLMs to quickly locate relevant policy documents or institutional clauses, extract key points, and reorganise dispersed information into structured content. Data analysis and report drafting supports preliminary interpretation of hospital operational indicators (e.g., outpatient volume and admission rates) and the generation of narrative reporting text. Through these functions, LLMs can, to some extent, support knowledge-intensive administrative tasks related to text processing, information integration, and operational management.

Figure 1.

Functional categories of LLM-supported hospital administrative work.

For example, when administrative staff need to prepare a new notice, they can provide key inputs (e.g., topic, target audience, and requirements), and the LLMs can rapidly generate a well-structured draft, enabling staff to obtain an administrative editable text within a short time and reducing the cost of initial drafting.

Despite the substantial potential of LLMs in hospital administrative work, existing studies have largely focused on clinicians and researchers, and empirical research on the adoption of LLMs among hospital administrative staff remains limited.^11–13 Thus, a systematic examination of the mechanisms shaping LLM adoption among hospital administrative staff is warranted.

The Technology Acceptance Model (TAM) is widely recognised as one of the most influential theoretical frameworks for explaining information technology adoption,¹⁴ and it has been extensively applied across studies of information systems and emerging technologies. However, in hospital administration, adopting LLMs is not simply a matter of selecting an efficiency-enhancing tool; instead, it is shaped by the interplay between LLM-specific technological attributes and the workflow constraints of hospital administrative practice. Therefore, relying solely on the perceived usefulness (PU)- and perceived ease of use (PEOU)-centred pathways emphasised in TAM may be insufficient to account for LLM adoption among hospital administrative staff. Specifically, administrative staff routinely handle a large volume of formal and standardised text-based tasks, including notices and bulletins, official documents and institutional policies, meeting minutes, and operational analysis reports. Such work places high demands on informational accuracy, document standardisation, and traceable review processes. By contrast, LLM outputs are produced through probabilistic generation mechanisms and may exhibit uncertainty; in some circumstances, they may also generate content inconsistent with source information (hallucinations). These characteristics can create tension with hospital administrative requirements for accuracy, consistency, and traceability, thereby shaping administrative staff’s adoption decisions. This mismatch creates a structural limitation for TAM, as the model does not incorporate mechanisms that capture uncertainty management, task alignment, or institutional constraints.

To address these limitations, this study introduces complementary constructs that capture these missing dimensions. Specifically, whether LLMs can effectively support concrete administrative tasks becomes a salient consideration, suggesting that task–technology fit (TTF) may be a key determinant of LLM adoption.¹⁵ Meanwhile, the opacity of LLM reasoning may affect confidence in generated outputs, making trust (TR) an important psychological mechanism facilitating adoption under uncertainty.^16,17 In public hospitals with clear organisational hierarchies, leadership and peer behaviors may shape individual technology choices; thus, social influence (SI) may influence LLM adoption among administrative staff.^18,19 Finally, hospital administrative staff often face high workload and time pressure, which may drive them to seek efficiency-enhancing tools, making work stress (WS) a key contextual factor influencing their actual use of LLMs.^20,21 Building on TAM, this study therefore incorporates TTF, TR, SI, and WS to develop an integrated research model (Figure 2) that systematically explains the mechanisms underpinning LLM adoption among hospital administrative staff. Taken together, this study conceptualises LLM adoption in hospital administration as a process jointly shaped by technological uncertainty, task requirements, and organisational context, rather than being driven solely by PU and PEOU as suggested by TAM.

Figure 2.

Proposed research model and hypotheses.

Using survey data from administrative staff in 40 public hospitals in Guangxi, China, we investigate the factors shaping behavioral intention (BI) and use behavior (UB) in the adoption of LLMs among hospital administrative staff. The hypothesised relationships specified in the integrated model were tested using partial least squares structural equation modelling (PLS-SEM). This study provides empirical evidence on the mechanisms shaping LLM adoption among hospital administrative staff and offers practical implications for supporting the effective use of generative AI in hospital administrative settings.

2 Theoretical background and hypotheses

2.1 Technology acceptance model (TAM)

TAM is widely regarded as a useful framework for explaining individuals’ adoption of information technologies.^14,22 TAM proposes that perceived usefulness (PU) and perceived ease of use (PEOU) are core cognitive antecedents of behavioral intention (BI), which in turn predicts use behavior (UB). Owing to its parsimonious structure and stable predictive performance, TAM has been widely used in research on the adoption of GAI tools.^23–25 However, LLMs exhibit features such as algorithmic opacity, uncertainty in outputs, and strong contextual dependency. These characteristics challenge the core assumptions of TAM by introducing dimensions that are not captured by PU and PEOU, particularly regarding uncertainty management, reliability assessment, and contextual appropriateness. As a result, TAM alone is insufficient to fully explain intention formation and sustained use in this context.^26–28 In addition, prior studies have advocated integrating TAM with theoretically relevant external variables to capture LLM adoption complexity.²⁹

Rather than simply applying TAM, this study extends it by incorporating TTF, TR, SI, and WS, which capture critical dimensions—such as task alignment, uncertainty management, and organisational context—not addressed by TAM. Together, these constructs form a contextually grounded framework for explaining LLM adoption among hospital administrative staff (Figure 2).

2.2 Hypothesis development

Based on the theoretical framework outlined above, the following hypotheses are proposed.

2.2.1 TTF and PU

TTF theory posits that the value of a technology depends on the degree to which its functional characteristics align with users’ task requirements. When fit is high, users are more likely to develop favourable performance expectations and stronger intentions to use the technology.^15,30,31 TTF has been widely applied in digital health and online learning contexts,^32–34 and recent GAI studies support the relevance of task-technology alignment in shaping adoption-related perceptions.^35,36

Hospital administrative work is knowledge-intensive and involves frequent document drafting, information synthesis, data compilation, and cross-departmental coordination. LLMs may support such tasks through text generation, summarisation, and report writing. When these capabilities closely match administrative task requirements, staff are more likely to perceive that LLMs improve efficiency and work performance, strengthening PU. Accordingly, this study incorporates TTF into the TAM and proposes the following hypothesis.

TTF positively affects PU.

2.2.2 PEOU, PU, BI, and UB

Prior studies indicated that PEOU and PU are positively linked to BI in the adoption of GAI tools, and PU frequently mediates the relationship between PEOU and BI.^25,37–39 These findings are consistent with the core propositions of the TAM. When a technology is viewed as easy to learn and use, anticipated effort costs are reduced, which can increase usefulness appraisals and, subsequently, intention.

In hospital administration, when LLMs are perceived as user-friendly and require limited learning effort, staff may be more likely to view them as useful for improving efficiency and output quality, reinforcing BI. In addition, TAM further posits that BI is the most proximal predictor of UB—a relationship consistently supported in GAI adoption research.^40–42 Accordingly, this study proposes the following hypotheses.

PEOU positively affects PU.

PEOU positively affects BI.

PU positively affects BI.

BI positively affects UB.

2.2.3 SI, PU, and BI

SI is rooted in subjective norms and refers to an individual’s perception that important others (e.g., supervisors, colleagues, or the organisation) expect and endorse a behavior.^22,43 In organisational settings, SI may operate through leadership advocacy, peer modelling, and the normative climate, shaping value judgements and adoption intention.

Evidence indicates that SI is a key factor influencing the adoption of GAI by strengthening performance expectations and BI.^19,44–46 This mechanism may be particularly salient in public hospitals characterised by formal rules, clear hierarchies, and strong normative governance. In such contexts, leadership advocacy and peer modelling may signal legitimacy and utility, thereby enhancing PU and reinforcing BI. Accordingly, this study proposes the following hypotheses.

SI positively affects PU.

SI positively affects BI.

2.2.4 TR, BI, and UB

TR is a critical psychological mechanism in technology adoption and is commonly defined as users’ belief in a system’s reliability, predictability, and controllability under uncertainty and risk.^16,17 In GAI contexts, TR has been conceptualised as a bridge between technological uncertainty and willingness to adopt, with higher TR associated with stronger intention.^47,48

TR is particularly salient for LLMs. Compared with conventional information systems, LLM outputs are not always transparent or readily verifiable, and their use often involves privacy-sensitive information and requires professional judgement. Moreover, LLM applications may introduce further risks such as hallucinated outputs, privacy breaches, and unclear accountability.^49–51 In high-responsibility, low-tolerance-for-error settings, TR is more likely to influence both BI and UB. Prior evidence suggests that TR can directly increase BI and predict UB.^47,48,52

In hospital administration, staff handle formal documentation, internal reporting, and sensitive information under stringent accuracy and confidentiality requirements. Therefore, LLMs are unlikely to be used in routine work unless they are perceived as reliable, secure, and controllable. Based on this, trust is incorporated into the TAM as a key factor influencing both BI and UB, leading to the following hypotheses.

TR positively affects BI.

TR positively affects UB.

2.2.5 WS, BI, and UB

Drawing on the Job Demands-Resources (JD-R) model, WS emerges when job demands challenge or exceed available resources.^21,53 The JD-R model allows divergent behavioral responses under stress: stress may deplete resources and reduce change-oriented behaviors, but it may also trigger compensatory coping, whereby individuals mobilise external resources (including technologies) to maintain task performance.

In digitally mediated workplaces, AI-enabled systems can function as job resources by reducing effort for routine activities and enhancing information-processing efficiency, potentially buffering demand-related strain.⁵⁴ Related evidence indicates that work-related pressure may be associated with stronger willingness to use GAI tools.⁵⁵ In hospital administration, documentation-heavy tasks under tight deadlines may prompt staff experiencing higher WS to rely on LLMs as pragmatic resources, suggesting a positive association with UB. In addition, WS may strengthen the translation of BI into UB, implying a moderating effect. Accordingly, this study incorporates WS into the TAM and proposes the following hypotheses.

H10

WS positively affects UB.

H11

WS positively moderates the BI-UB relationship, such that the BI-UB association is stronger under higher WS.

3. Methods

3.1 Study design and participants

The investigation was conducted in the Guangxi Zhuang Autonomous Region, China, between October and December 2025, using a cross-sectional survey design. Public hospitals were first grouped by level (tertiary and secondary) to create a stratified cluster sampling framework, and 40 hospitals (20 tertiary and 20 secondary) were selected to ensure coverage across different hospital types and regions. Within each selected hospital, administrative staff from key departments (e.g., hospital office, human resources, medical affairs, and information management) were invited to participate. As tertiary hospitals usually employ larger administrative teams, our recruitment target was approximately 50 administrative staff per tertiary hospital and 35 per secondary hospital. Questionnaires were completed voluntarily by administrative staff.

Inclusion criteria included being ≥18 years old, having worked for at least one year in an administrative role, and routinely undertaking administrative documentation, coordination or data-reporting duties. Written informed consent was obtained from all participants. We excluded interns, temporary or rotating staff, individuals not engaged in administrative work and senior executives, as well as those who did not consent to participate.

Data were collected through an anonymous online questionnaire via Wenjuanxing (https://www.wjx.cn), a widely used web-based survey platform in China, between October and December 2025. The survey was coordinated by the Hospital Administration Professional Committee of the Guangxi Hospital Association. An official invitation and online link were sent to the Hospital Office of selected hospitals, which further distributed the link to relevant administrative departments, where staff completed the questionnaire voluntarily. A total of 1,659 questionnaires were returned. After excluding invalid responses (completion time≤180 seconds and obvious logical inconsistencies), 1,377 valid questionnaires were retained. Based on the inverse square root approach described by Hair et al. (2022), assuming path coefficients of 0.05–0.10 and a 95% significance level, the required sample size was estimated to be 619; thus, the final analytic sample exceeded this threshold. The study received ethical approval from the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (approval No. 2025-K0397).

3.2 Measurement instruments

Demographic Characteristics: age, gender, education level, work years, and hospital level.

3.2.1 Instrument development

To ensure that the instruments were linguistically accurate and suitable for the Chinese context, we adapted all measurement instruments following established cross-cultural adaptation procedures.⁵⁶ This process included translation, reconciliation, back-translation, expert review, and pilot testing. All constructs were measured using items adapted from previously published and validated scales, with appropriate citations provided (Table 1). For each construct, participants responded on a seven-point Likert scale, where 1 represented “strongly disagree” and 7 represented “strongly agree”. To further assess the applicability of the instruments in the study context, a pilot study with 150 administrative staff in a tertiary hospital confirmed good reliability for all constructs (Cronbach’s α > 0.80). All survey items used in the formal study are presented in Appendix A.

Table 1.

Measurement instruments.

Construct	Items	Source(s)
Task–Technology Fit (TTF)	4	^8,57
Social Influence (SI)	3	^18,19
Trust (TR)	4	^58,59
Work Stress (WS)	4	^20,60
Perceived Ease of Use (PEOU)	4	^14,18
Perceived Usefulness (PU)	4	^14,18
Behavioral Intention (BI)	3	^18,35
Use Behavior (UB)	3	^14,18

3.3 Data analysis

Data analysis was conducted using PLS-SEM in SmartPLS 4.1. This method is appropriate for models involving multiple latent constructs and indirect effects, especially in prediction-focused research or when the assumptions required for covariance-based SEM may not be satisfied.⁶¹ As our model integrates core TAM constructs with TTF, TR, and WS, and aims to explain and predict behavioral outcomes, PLS-SEM was considered a suitable analytical approach for the present study.

In addition to testing the hypothesised direct relationships, supplementary indirect-effect analyses were conducted to better understand how the proposed constructs may operate through intermediary paths in the structural model. These analyses were exploratory and were not treated as formal hypothesis tests.

4. Results

4.1 Participant characteristics

In total, 1,377 valid questionnaires were included in the analysis (Table 2). Among respondents, 68.41% were female and 31.59% were male. Participants were most frequently aged 26–35 years (35.88%) or 36–45 years (36.46%). The majority of participants had completed a bachelor’s degree or higher (82.06%), and more than half had over 10 years of work experience (51.05%). Regarding hospital level, 59.11% were from tertiary public hospitals and 40.89% from secondary public hospitals.

Table 2.

Characteristics of the participants.

Characteristic	Category	n	%
Gender	Male	435	31.59
Gender	Female	942	68.41
Age (years)	18-25	72	5.23
	26-35	494	35.88
	36-45	502	36.46
	46-55	288	20.92
	≥55	21	1.53
Education	College degree or below	247	17.94
	Bachelor’s degree	1008	73.20
	Master’s degree	115	8.35
	Doctoral degree or above	7	0.51
Work years	<3 years	188	13.65
	3–5 years	260	18.88
	6–10 years	226	16.41
	>10 years	703	51.05
Hospital level	Secondary hospital	563	40.89
Hospital level	Tertiary hospital	814	59.11

Overall, the sample covered a broad range of demographic characteristics and hospital levels, providing a useful basis for examining LLM adoption among administrative staff in public hospitals in Guangxi, China.

4.2 Common method bias

Since the study relied on self-reported data, we assessed common method bias (CMB). Harman’s single-factor test showed that the largest unrotated factor explained 38.06% of the total variance, which is below the commonly used 40% benchmark.⁶² We further applied the full collinearity test.⁶³ The VIFs for all latent constructs were below 3.3. Taken together, these results indicate that there was no clear evidence of CMB.

4.3 Measurement model

For the measurement model, we examined indicator reliability, internal consistency, and convergent validity (Table 3). Cronbach’s α values ranged from 0.782 to 0.879, while composite reliability (CR) values were between 0.873 and 0.917, all exceeding the recommended threshold of 0.70. Standardised outer loadings varied from 0.776 to 0.868 (>0.70). Average variance extracted (AVE) varied from 0.693 to 0.734 (>0.50), supporting adequate convergent validity.⁶¹

Table 3.

Reliability and convergent validity of the measurement model.

Construct	Item	M	SD	FL	CA (α)	CR	AVE
TTF	TTF 1	5.26	1.267	0.842	0.876	0.915	0.73
	TTF 2	5.49	1.182	0.858
	TTF 3	5.314	1.242	0.848
	TTF 4	5.175	1.302	0.868
PEOU	PEOU 1	5.502	1.119	0.776	0.852	0.9	0.693
	PEOU 2	5.24	1.267	0.846
	PEOU 3	5.057	1.327	0.842
	PEOU 4	5.211	1.283	0.864
PU	PU 1	5.356	1.22	0.858	0.879	0.917	0.734
	PU 2	5.309	1.211	0.853
	PU 3	5.341	1.191	0.861
	PU 4	5.383	1.193	0.855
WS	WS 1	5.269	1.273	0.834	0.857	0.903	0.7
	WS 2	5.223	1.251	0.862
	WS 3	5.195	1.265	0.826
	WS 4	5.221	1.273	0.824
SI	SI 1	5.224	1.219	0.839	0.782	0.873	0.697
	SI 2	5.166	1.211	0.83
	SI 3	5.301	1.192	0.835
TR	TR 1	5.253	1.15	0.814	0.854	0.902	0.696
	TR 2	5.17	1.186	0.831
	TR 3	4.856	1.289	0.85
	TR 4	4.869	1.282	0.842
BI	BI 1	5.404	1.146	0.855	0.808	0.887	0.723
	BI 2	5.064	1.282	0.835
	BI 3	5.312	1.176	0.86
UB	UB 1	5.049	1.229	0.844	0.799	0.881	0.712
	UB 2	4.808	1.24	0.835
	UB 3	4.907	1.243	0.853

Note. M = Mean; SD = Standard Deviation; FL = Factor Loading; CA (α) = Cronbach’s alpha; CR=Composite Reliability; AVE=Average Variance Extracted. Task–Technology Fit=TTF; Perceived Usefulness=PU; Perceived Ease of Use=PEOU; Trust=TR; Work Stress=WS; Social Influence=SI; Behavioral Intention=BI; Use behavior=UB.

In terms of discriminant validity, as shown in Table 4, the square root of the average variance extracted (AVE) for each construct exceeded its correlations with all other constructs, thereby satisfying the Fornell-Larcker criterion.⁶⁴

In addition, the HTMT ratios for all constructs were below the recommended cutoff value of 0.90, which provides further evidence of adequate discriminant validity (Table 5).⁶⁵

Table 4.

Discriminant validity (Fornell-Larcker criterion).

	TTF	TR	BI	PEOU	PU	SI	UB	WS
TTF	0.854
TR	0.515	0.834
BI	0.509	0.672	0.85
PEOU	0.672	0.498	0.455	0.833
PU	0.755	0.53	0.547	0.723	0.857
SI	0.506	0.54	0.55	0.491	0.523	0.835
UB	0.438	0.586	0.638	0.428	0.457	0.49	0.844
WS	0.203	0.237	0.306	0.188	0.21	0.295	0.347	0.836

Note. Task–Technology Fit=TTF; Perceived Usefulness=PU; Perceived Ease of Use=PEOU; Trust=TR; Work Stress=WS; Social Influence=SI; Behavioral Intention=BI; Use Behavior=UB.

Table 5.

Discriminant validity: Heterotrait-monotrait (HTMT) ratios.

	TTF	TR	BI	PEOU	PU	SI	UB	WS
TTF
TR	0.593
BI	0.604	0.808
PEOU	0.778	0.582	0.547
PU	0.859	0.61	0.648	0.834
SI	0.61	0.658	0.691	0.6	0.628
UB	0.519	0.705	0.79	0.514	0.538	0.614
WS	0.234	0.276	0.368	0.221	0.242	0.361	0.419
WS x BI	0.171	0.2	0.25	0.171	0.154	0.202	0.224	0.155

Note. Task–Technology Fit = TTF; Perceived Usefulness = PU; Perceived Ease of Use = PEOU; Trust = TR; Work Stress = WS; Social Influence = SI; Behavioral Intention = BI; Use Behavior = UB.

4.4 Structural model and hypothesis testing

4.4.1 Collinearity diagnostics

Before testing the structural relationships, we examined the data for potential multicollinearity among the constructs. The VIF values for all structural paths ranged between 1.062 and 2.364, which are well below the commonly recommended cutoff of 3.0.^61,65 These results suggest that multicollinearity was unlikely to bias the structural estimates.

4.4.2 Structural model results and hypothesis testing

As recommended in prior research, we used 5,000 bootstrap samples to evaluate the statistical significance of the structural paths.⁶¹ Figure 3 and Table 6 report standardized path coefficients (β), t statistics, p values, and 95% confidence intervals (CIs). Nine hypotheses were supported (H1, H2, H4–H10), whereas H3 and H11 were rejected.

Figure 3.

Structural model results.

Table 6.

The results of hypothesis testing.

Path	β	t	p	95% CI (LL)	95% CI (UL)	Results
H1: TTF → PU	0.454	14.432	<0.001	0.389	0.515	Supported
H2: PEOU → PU	0.361	12.724	<0.001	0.306	0.416	Supported
H3: PEOU → BI	-0.035	1.06	0.289	-0.098	0.03	Not supported
H4: PU → BI	0.221	6.56	<0.001	0.156	0.286	Supported
H5:BI→ UB	0.398	12.957	<0.001	0.335	0.456	Supported
H6: SI → PU	0.116	5.205	<0.001	0.072	0.16	Supported
H7:SI → BI	0.201	7.637	<0.001	0.148	0.251	Supported
H8: TR → BI	0.464	16.597	<0.001	0.406	0.516	Supported
H9: TR → UB	0.274	8.056	<0.001	0.209	0.341	Supported
H10: WS → UB	0.155	6.252	<0.001	0.107	0.204	Supported
H11:WS×BI→ UB	-0.032	1.727	0.084	-0.068	0.006	Not supported

TTF showed the strongest association with PU (β = 0.454, p < 0.001), supporting H1. PEOU positively predicted PU (β = 0.361, p < 0.001), supporting H2. The direct effect of PEOU on BI was not significant (β = −0.035, p = 0.289), thus H3 was not supported. PU positively predicted BI (β = 0.221, p < 0.001), supporting H4. BI predicted UB (β = 0.398, p < 0.001), supporting H5. SI positively predicted PU (β = 0.116, p < 0.001) and BI (β = 0.201, p < 0.001), supporting H6 and H7. TR was the strongest predictor of BI (β = 0.464, p < 0.001) and also positively predicted UB (β = 0.274, p < 0.001), supporting H8 and H9. WS positively predicted UB (β = 0.155, p < 0.001), supporting H10. The interaction term (WS × BI → UB) was not significant (β =−0.032, p = 0.084), thus H11 was not supported.

4.4.3 Supplementary analysis of indirect effects

To better understand how the proposed constructs operate through intermediary paths in the structural model, supplementary indirect-effect estimates were examined (Table 7). The results showed that PEOU was indirectly associated with BI through PU (β = 0.080, p < 0.001), despite the absence of a direct effect. SI was also indirectly associated with BI through PU (β = 0.026, p < 0.001). Trust was also indirectly related to UB through BI (β= 0.185, p < 0.001). In contrast, the path PEOU → BI → UB was not supported by the data (β= −0.014, p = 0.288), indicating that PEOU may influence actual use primarily through increasing perceived usefulness rather than directly translating into use behavior.

Table 7.

Supplementary analysis of indirect effects.

Path	β	t	p	95% CI (LL)	95% CI (UL)
PEOU → PU → BI	0.08	6.018	<0.001	0.055	0.106
SI → PU → BI	0.026	4.069	<0.001	0.014	0.039
TR → BI → UB	0.185	10.777	<0.001	0.151	0.219
SI → PU → BI → UB	0.01	3.9	<0.001	0.006	0.016
PEOU → BI → UB	-0.014	1.064	0.288	-0.039	0.012
TTF → PU → BI → UB	0.04	5.161	<0.001	0.026	0.056
PU → BI → UB	0.088	5.703	<0.001	0.059	0.12
SI→ BI → UB	0.08	6.422	<0.001	0.056	0.105
PEOU → PU → BI → UB	0.032	5.306	<0.001	0.021	0.044
TTF → PU → BI	0.101	5.811	<0.001	0.068	0.135

4.4.4 Structural model performance: Explanatory power, predictive relevance, and model fit

The overall performance of the structural model was assessed in terms of explanatory capability, predictive relevance as well as model fit, using the R², Q², f², and SRMR statistics, respectively.⁶¹

First, the R² values for all endogenous variables exceeded the suggested 0.25 benchmark: PU (R²=0.664), BI (R²= 0.527), and UB (R²= 0.476), indicating moderate to substantial explanatory capability. Second, the Q²values were also positive and above 0.25 for PU (0.483), BI (0.379), and UB (0.333), which implies that the model demonstrates strong predictive capability.^61,65 We further reported the effect sizes (f²) for key structural paths. According to Hair et al. (2022), f² values of 0.02, 0.15, and 0.35 indicate small, medium, and large effects, respectively. The results showed medium effects for the paths TTF→PU (f²= 0.312), TR→BI (f²= 0.278), PEOU→PU (f²= 0.201), and BI→UB (f²= 0.157), while the remaining effects were small.

Finally, the SRMR value was 0.045, which is below the 0.08, indicating an acceptable level of model fit.⁶⁶ The goodness-of-fit (GoF) index was 0.629, which is higher than 0.36. This adds to the evidence that the model fits well overall.⁶⁷

5. Discussion

Building on the TAM, this study incorporated TTF, TR, SI, and WS to develop an integrated framework and systematically examine the mechanisms underlying LLM adoption among hospital administrative staff. Overall, the findings support the core logic of TAM in the LLM context, while also suggesting that TAM alone is insufficient to fully explain LLM adoption in hospital administrative settings. The inclusion of TTF, TR, SI, and WS provides additional explanatory power by capturing task-related, uncertainty-related, and organisational mechanisms shaping adoption.

With respect to the core mechanism of TAM, our results are consistent with its basic pathway: PEOU influences PU, which in turn shapes BI, and BI subsequently leads to UB.¹⁴ This suggests that TAM remains informative for understanding LLM adoption among hospital administrative staff. However, inconsistent with the original TAM assumption, the direct effect of PEOU on BI was not significant in this study. A plausible explanation is that most LLM tools employ natural-language conversational interfaces that are relatively intuitive, allowing users to start using them without a complex learning process.⁶⁸ As a result, PEOU may be viewed by administrative staff as a basic system attribute rather than a decisive factor in their adoption of LLMs.

Furthermore, our findings indicate that TTF, TR, SI, and WS are important correlates of LLM adoption among hospital administrative staff. In the sections that follow, we elaborate on the potential mechanisms through which these constructs may shape LLM adoption, with reference to both the technical attributes of LLMs and the characteristics of hospital administrative work.

First, TTF showed a particularly prominent role in our model. This suggests that administrative staff’s value judgements regarding LLMs may follow a “fit-first logic”. This pattern is consistent with the core premise of TTF theory,¹⁵and aligns with prior findings that when a technology directly supports users’ core tasks, users are more likely to develop higher perceived usefulness.^34,69 Supplementary analysis of indirect effects (TTF→PU→BI; TTF→PU→BI→UB) further suggests a relatively coherent cognitive pathway: only when LLMs are perceived to fit concrete administrative tasks well do staff form a judgement that LLMs have practical value; this judgement then translates into intention and, ultimately, observable use behavior.

This mechanism is highly plausible in hospital administrative workflows. Administrative staff routinely deal with standardised text-intensive tasks, such as document drafting, meeting-minute preparation, and compilation of routine operational information. In such tasks, the value of LLMs lies less in replacing human judgement and more in whether they can reduce drafting effort, accelerate information integration, and improve work efficiency. If LLM-generated content requires substantial rework, even strong technical capability may not be perceived as consistently valuable in practice. Therefore, improving LLM adoption requires managerial efforts to design task-targeted use scenarios for high-frequency administrative tasks and to provide supporting templates, prompts, and formatting conventions, thereby enabling administrative staff to directly experience the practical usability of LLMs in concrete tasks.

Second, trust plays a central role in LLM adoption among hospital administrative staff. The findings indicate that trust not only affects whether staff are willing to try LLMs, but also whether they incorporate LLMs into formal work processes; the supplementary analysis of indirect effects (TR→BI→UB) is consistent with this mechanism. This pattern aligns with broader evidence in AI adoption research: in contexts involving algorithmic decision-making or automatically generated content, trust often facilitates the transition from intention to final use.^70,71 Compared with educational or general settings, hospital administrative settings places greater emphasis on accuracy, traceability, and clearly defined accountability. Accordingly, administrative staff may attend not only to efficiency gains but also to whether LLM outputs are reliable and verifiable. For example, when LLMs are used for operational data interpretation or information retrieval, output accuracy can directly shape trust; once staff observe discrepancies between generated results and actual information, they may be reluctant to use LLMs for related tasks thereafter. In this sense, trust may function as a “gatekeeping mechanism”: only when LLM outputs are perceived as controllable and reliable are staff likely to integrate them into routine practice.

Third, SI was associated with both PU and BI, and the supplementary analysis of indirect effects further suggests that SI may relate to UB through the PU and BI pathway. This indicates that, in hospital administrative settings, adoption is shaped not only by individual judgements about LLM functionality but also by the organisational environment. This is consistent with prior studies on generative AI adoption showing that leadership advocacy, peer modelling, and organisational climate can influence employees’ value appraisals of generative AI and subsequently shape their intention to use it.^19,72 In public hospitals characterised by clearer hierarchies and stricter institutional norms, such social influence may be particularly salient. When deciding whether to use LLMs, administrative staff may rely not only on personal experience but also on leaders’ attitudes and colleagues’ practices. When leaders openly support LLM use and peers accumulate experience and share it within the organisation, staff may be more likely to view LLMs as “permitted and useful”, thereby strengthening adoption intention. This mechanism may be especially influential while administrative uses of LLMs remain exploratory. For instance, in tasks such as first-draft generation and revision, meeting-minute preparation, and operational report drafting, some staff may be unfamiliar with potential use cases or uncertain about risks; in such circumstances, peer demonstration and experience sharing may become important drivers of adoption.

Finally, we found that WS was positively associated with UB, but it did not significantly strengthen the translation from BI to UB. This suggests that, in hospital administration, WS may prompt staff to try LLMs, but may not necessarily promote sustained routine use. This pattern can be interpreted through the “resource compensation” logic of the Job Demands-Resources (JD-R) framework: under high workload conditions, employees may actively seek additional resources or tools to reduce stress.^21,53 LLMs may thus be treated as a readily accessible auxiliary resource—for example, to generate initial drafts or integrate information—thereby alleviating time pressure and documentation burden to some extent.

However, WS does not automatically strengthen the conversion of BI into UB. In hospital administrative practice, LLM-generated content often still requires verification and revision. If verification costs remain high, or accountability boundaries are unclear, staff may not continue using the tool even when they express intention. Therefore, if hospital managers expect LLMs to genuinely reduce administrative workload, it is necessary to establish clear use policies, review mechanisms, and standardised templates to reduce the verification costs of LLM-generated content.

From a theoretical perspective, this study makes two main contributions. First, it extends LLM adoption research to hospital administration, thereby extending prior research beyond educational, research, and clinical contexts. Second, by integrating TTF, TR, SI, and WS with TAM, this study provides a more context-sensitive account of LLM adoption among hospital administrative staff, highlighting the roles of task alignment, uncertainty management, organisational norms, and workload pressure. Taken together, these contributions offer a useful foundation for future LLM adoption research in similarly regulated, high-responsibility settings.

From a practical perspective, the findings suggest several implications. First, when deploying LLMs, hospitals should adopt a task-oriented approach by prioritising high-frequency and standardised administrative tasks and providing standardised templates, prompts, and usage guidelines to improve task-technology fit, strengthen administrative staff’s perceived usefulness, and increase actual use. Second, hospitals should implement secure and controllable LLM-use environments and clarify scenario boundaries, thereby enhancing perceived controllability and trust in LLM outputs; training and operational guidance can also further help staff understand relevant norms and risks and reduce concerns. Third, hospitals should foster a supportive climate for LLM use through leadership endorsement and experience sharing among administrative staff, thereby strengthening PU and PEOU and facilitating integration into routine workflows. Finally, to ensure that LLMs meaningfully reduce administrative workload, hospital managers should invest in staff training to improve LLM-related competencies and human-AI interaction, thereby enhancing LLM output accuracy and usability and reducing post-use verification burden.

Although this study provides both theoretical and practical insights, several limitations should be acknowledged. First, the data were collected at a single time point, so we were unable to capture how attitudes or behavior may change over time. Future research could therefore consider longitudinal designs to provide stronger evidence on causal relationships. Second, as this study relies on a survey-based design, it does not directly capture real-time deployment or workflow-level use of LLMs in hospital settings. Future research could therefore complement these findings through case studies and field experiments to provide more fine-grained and practice-oriented evidence. Third, our sample was drawn mainly from public hospitals in Guangxi. This regional focus means that the findings may not fully reflect conditions in other settings. Replicating the study in different regions or cultural contexts would help to assess the robustness and wider relevance of the proposed model.

6. Conclusion

This study examined the factors influencing LLM adoption among hospital administrative staff by integrating TAM with TTF, TR, SI, and WS. The findings suggest that LLM adoption is shaped not only by the core elements of TAM, but also by TTF, TR, SI, and WS, reflecting the joint influence of technological, psychological, and organisational settings.

These results provide a more comprehensive understanding of LLM adoption in hospital administrative settings and offer practical insights for public hospitals seeking to support the effective and responsible use of LLMs in administrative work.

Footnotes

Acknowledgements

We thank the administrative staff from the 40 participating public hospitals for their time and support in completing the survey.

ORCID iD

Xuepeng Du

Author contributions

Xuepeng Du: conceptualisation and study design, data analysis and interpretation, manuscript drafting.

Chunqing Ling: data analysis.

Lan Yang: supervision.

Liusi Wei: manuscript drafting and revision.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.*

AI disclosure statement

The authors used generative AI tools (e.g., ChatGPT) solely for language editing (grammar and sentence rephrasing), and all outputs were reviewed and edited by the authors who take full responsibility for the manuscript’s content.

Appendix

Table A1.

The survey questionnaire items

Construct	Items
Task-Technology Fit (TTF)	TTF1: The functions of LLMs closely match the tasks involved in my daily administrative work. TTF2: The functions provided by LLMs effectively support the completion of my administrative tasks. TTF3: LLMs are well suited for handling complex or demanding administrative tasks. TTF4: Overall, the functions of LLMs align well with the requirements of my administrative tasks.
Perceived Ease of Use (PEOU)	PEOU1: The use of LLMs is straightforward and easy to learn. PEOU2: I can learn how to use the main functions of LLMs within a short period of time. PEOU3: I can easily integrate LLMs into my existing administrative work processes. PEOU4: Overall, LLMs are easy for me to use.
Perceived Usefulness (PU)	PU1: Using LLMs can improve the efficiency of my administrative work. PU2: Using LLMs enables me to complete more administrative tasks within the same amount of time. PU3: Using LLMs enhances the quality and outcomes of my administrative work. PU4: Overall, LLMs make an important contribution to my administrative work performance.
Social Influence (SI)	SI1: Many of my colleagues use LLMs in their administrative work. SI2: Hospital leadership encourages the appropriate use of LLMs in administrative work. SI3: LLMs are generally portrayed positively in the media and public discussions.
Trust (TR)	TR1: I am confident that LLMs are reliable in administrative work. TR2: The information generated by LLMs is generally dependable. TR3: When handling administrative tasks, I am willing to rely on the content or suggestions provided by LLMs. TR4: Overall, I consider LLMs to be trustworthy tools.
Work Stress (WS)	WS1: I often handle many administrative tasks. WS2: I often feel pressed for time in my administrative work. WS3: I carry a high level of responsibility in my administrative work. WS4: My administrative work covers many areas and is often complex.
Behavioral Intention (BI)	BI1: I intend to use LLMs in my administrative work. BI2: I plan to use LLMs as an important supporting tool in my administrative work. BI3: When administrative tasks arise, I will give priority to using LLMs.
Use Behavior (UB)	UB1: Currently, I use LLMs frequently in my daily administrative work. UB2: Currently, I spend a certain amount of time using LLMs to complete administrative tasks. UB3: Currently, some of my administrative tasks rely on the use of LLMs.

References

Chen

Liu

, et al. Adoption of large language model AI tools in everyday tasks: Multisite cross-sectional qualitative study of Chinese hospital administrators. J Med Internet Res 2025; 27: e70789.

Han

Wang

Zhang

, et al. A Review of Large Language Models: Fundamental Architectures, Key Technological Evolutions, Interdisciplinary Technologies Integration, Optimization and Compression Techniques, Applications, and Challenges. Electronics 2024; 13(24): 5040.

Nazi

Peng

. Large language models in healthcare and medical domain: A review. Informatics 2024; 11(3): 57.

Khan

Umer

. ChatGPT in finance: Applications, challenges, and solutions. Heliyon 2024; 10: e24890.

Mai

DTT

Hanh

. The use of ChatGPT in teaching and learning: A systematic review through SWOT analysis approach. Front Educ 2024; 9: 1328769.

Wang

Zhang

. Large language models in medical and healthcare fields: Applications, advances, and challenges. Artif Intell Rev 2024; 57: 299.

Holmström

Carroll

. How organizations can innovate with generative AI. Business Horizons 2025; 68: 559–573.

, et al. Determinants of continuance intention to use GAI in academic research among graduate students:Findings from PLS-SEM and fsQCA. Educ Inf Technol 2025; 30: 20289–20315.

Zhang

Khan

Mahmud

, et al. Exploring the role of large language models in the scientific method: From hypothesis to discovery. npj Artif Intell 2025; 1: 14.

10.

Kunze

Nwachukwu

Cote

, et al. Large language models applied to health care tasks may improve clinical efficiency, value of care rendered, research, and medical education. Arthroscopy 2025; 41: 547–556.

11.

Dai

Yang

, et al. Attitudes, perceptions, and factors influencing the adoption of AI in health care among medical staff: Nationwide cross-sectional survey study. J Med Internet Res 2025; 27: e75343.

12.

Reis

Lenz

Gossen

, et al. Practical applications of large language models for health care professionals and scientists. Jmir Med Inf 2024; 12: e58478.

13.

Zhu

Wen

, et al. Identifying highly correlated determinants influencing student nurses’ behavioral intention of using generative artificial intelligence (generative AI): A network analysis. Nurs Educ Today 2025; 154: 106855.

14.

Davis

. Perceived usefulness, perceived ease of use and user acceptance of information technology. Mis Quart 1989; 13(3): 319–340.

15.

Goodhue

Thompson

. Task-technology fit and individual performance. Mis Quart 1995; 19(2): 213–236.

16.

Afroogh

Akbari

Malone

, et al. Trust in AI: Progress, challenges, and future directions. Hum Soc Sci Commun 2024; 11: 1–30.

17.

Huang

, et al. Developing trustworthy artificial intelligence: Insights from research on interpersonal, human-automation, and human-AI trust. Front Psychol 2024; 15: 1382693.

18.

Venkatesh

Morris

Davis

, et al. User acceptance of information technology: Toward a unified view. Mis Quart 2003: 425–478.

19.

Changalima

Amani

Ismail

. Social influence and information quality on generative AI use among business students. Int J Manag Educ-Oxf 2024; 22: 101063.

20.

Cavanaugh

Boswell

Roehling

, et al. An empirical examination of self-reported work stress among US managers. J Appl Psychol 2000; 85: 65.

21.

Bakker

Demerouti

. Job demands–resources theory: Taking stock and looking forward. J Occup Health Psych 2017; 22: 273.

22.

Venkatesh

Davis

. A theoretical extension of the technology acceptance model: Four longitudinal field studies. Manage Sci 2000; 46: 186–204.

23.

Balaskas

Tsiantos

Chatzifotiou

, et al. Determinants of ChatGPT adoption intention in higher education: Expanding on TAM with the mediating roles of trust and risk. Information 2025; 16: 82.

24.

Noor

Tong

Zainol

. ChatGPT and higher education student well-being: Role of subjective norm and anthropomorphism with TAM. J Inf Commun Ethics 2025: 1–20.

25.

Şimşek

Cengiz

GŞT

Bal

. Extending the TAM framework: Exploring learning motivation and agility in educational adoption of generative AI. Educ Inf Technol 2025; 30: 20913–20942.

26.

Dang

. Unveiling trust in AI: The interplay of antecedents, consequences, and cultural dynamics. Ai Soc 2026; 41(1): 669–692.

27.

Park

Young Yoon

. AI algorithm transparency, pipelines for trust not prisms: Mitigating general negative attitudes and enhancing trust toward AI. Hum Soc Sci Commun 2025; 12: 1–13.

28.

Schepman

Rodway

. Initial validation of the general attitudes towards artificial intelligence scale. Comput Hum Behav Rep 2020; 1: 100014.

29.

Uludağ

Kılıç

Çelik

. Artificial intelligence, social influence, and AI anxiety: Analyzing the intentions of science doctoral students to use ChatGPT with PLS-SEM. Hum Soc Sci Commun 2025; 12: 1–17.

30.

Alyoussef

. Acceptance of e-learning in higher education: The role of task-technology fit with the information systems success model. Heliyon 2023; 9(3): e13751.

31.

Dishaw

Strong

. Extending the technology acceptance model with task–technology fit constructs. Inform Manage-Amster 1999; 36: 9–21.

32.

Elçi

Abubakar

. The configurational effects of task-technology fit, technology-induced engagement and motivation on learning performance during covid-19 pandemic: An fsQCA approach. Educ Inf Technol 2021; 26: 7259–7277.

33.

Zheng

Jiang

Shen

, et al. Investigating clinicians’ intentions and influencing factors for using an intelligence-enabled diagnostic clinical decision support system in health care systems: Cross-sectional survey. J Med Internet Res 2025; 27: e62732.

34.

Tao

Chen

Qin

, et al. Modeling consumer acceptance and usage behaviors of m-health: An integrated model of self-determination theory, task–technology fit, and the technology acceptance model. Healthcare-Basel 2023; 11: 1550.

35.

Al-Mamary

Alfalah

Alshammari

, et al. Exploring factors influencing university students’ intentions to use ChatGPT: Analysing task-technology fit theory to enhance behavioural intentions in higher education. Futur Bus J 2024; 10: 119.

36.

Chakraborty

Troise

Bresciani

. Exploring consumer intentions to continue: Integrating task technology fit and social technology fit in generative AI based shopping platforms. Technovation 2025; 142: 103189.

37.

Wang

, et al. Exploring user adoption of ChatGPT: A technology acceptance model perspective. Int J Hum-Comput Int 2025; 41: 1431–1445.

38.

Mutammimah

Rejeki

Kustini

, et al. Understanding teachers’ perspective toward ChatGPT acceptance in English language teaching. Int J Technol Educ 2024; 7: 290–307.

39.

Zhang

Schießl

Plößl

, et al. Acceptance of artificial intelligence among pre-service teachers: A multigroup analysis. Int J Educ Technol High Educ 2023; 20: 49.

40.

Kaya

Adıgüzel

. Exploring the acceptance of ChatGPT in higher education: A comprehensive quantitative study of university students and faculty. Front Educ 2025; 10: 1652292.

41.

Kim

Blazquez

. Determinants of generative AI system adoption and usage behavior in korean companies: Applying the UTAUT model. Behav Sci-Basel 2024; 14: 1035.

42.

Owan

Mohammed

Bello

, et al. Higher education students’ ChatGPT use behavior: Structural equation modelling of contributing factors through a modified UTAUT model. Contemp Educ Technol 2025; 17: ep592.

43.

Ajzen

. The theory of planned behavior. Organ Behav Hum 1991; 50: 179–211.

44.

Bali

Suwandi

Chen

T-C

, et al. Social influence, personal views, and behavioral intention in ChatGPT adoption. J Comput Inform Syst 2024: 1–12.

45.

Pham

Klaus

Bateh

. Key factors influencing intention to use ChatGPT: An empirical study of US students. Acta Psychol 2025; 260: 105516.

46.

Wang

Zhu

. Why do users stay with generative artificial intelligence? J Comput Inform Syst 2025: 1–19.

47.

Dai

Shi

, et al. Structural equation modeling for influencing factors on behavioral intention to adopt medical AI among Chinese nurses: A nationwide cross-sectional study. BMC Nurs 2025; 24: 1084.

48.

Wang

Liu

, et al. Investigating the factors influencing users’ adoption of artificial intelligence health assistants based on an extended UTAUT model. Sci Rep-Uk 2025; 15: 18215.

49.

Constantinescu

Kaptein

. Responsibility gaps, LLMs & organisations: Many agents, many levels, and many interactions. Sci Eng Ethics 2025; 31: 36.

50.

Roustan

Bastardot

. The clinicians’ guide to large language models: A general perspective with a focus on hallucinations. Interact J Med Res 2025; 14: e59823.

51.

Zhang

, et al. Unpacking perceived risks and AI trust influences pre-service teachers’ AI acceptance: A structural equation modeling-based multi-group analysis. Educ Inf Technol 2025; 30: 2645–2672.

52.

Wang

Cheung

ACK

, et al. In GenAI we trust: An investigation of university students’ reliance on and resistance to generative AI in language learning. Int J Educ Technol High Educ 2025; 22: 59.

53.

Bakker

Demerouti

Sanz-Vergel

. Job demands–resources theory: Ten years later. Annu Rev Organ Psychol Organ Behav 2023; 10: 25–53.

54.

Loureiro

SMC

Bilro

Neto

. Working with AI: Can stress bring happiness? Serv Bus 2023; 17: 233–255.

55.

Samson

Luriaga

Ebardo

. Do Academic Stress and Risk Propensity Affect Behavioral Intention to Use ChatGPT among University Students? In: International Conference on Computers in Education, Manila, Philippines, 2024. https://doi.org/10.58459/icce.2024.4938

56.

Beaton

Bombardier

Guillemin

, et al. Guidelines for the process of cross-cultural adaptation of self-report measures. Spine 2000; 25: 3186–3191.

57.

Chen

Ning

, et al. Task/technology fit or technology attraction? The intentions of STEM teachers to use AI technologies for teaching innovation. Asia‐Pacific Edu Res 2025; 35: 219–233. https://doi.org/10.1007/s40299-025-01020-x

58.

Perrig

SAC

Scharowski

Brühlmann

. Trust issues with trust scales: Examining the psychometric quality of trust measures in the context of AI. In: Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, pp. 1–7.

59.

Đerić

Frank

Milković

. Trust in generative AI tools: A comparative study of higher education students, teachers, and researchers. Information 2025; 16: 622.

60.

Yang

Guo

, et al. Do challenge stress and hindrance stress affect quality of health care? Empirical evidence from China. Int J Env Res Pub He 2018; 15: 1628.

61.

Hair

Hult

GTM

Ringle

, et al. A primer on partial least squares structural equation modeling (PLS-SEM). Third edition : Sage Publications, Inc, 2022.

62.

Podsakoff

MacKenzie

Lee

J-Y

, et al. Common method biases in behavioral research: A critical review of the literature and recommended remedies. J Appl Psychol 2003; 88: 879.

63.

Kock

. Common method bias in PLS-SEM: A full collinearity assessment approach. Int J e-Collab 2015; 11: 1–10.

64.

Fornell

Larcker

. Evaluating Structural Equation Models with Unobservable Variables and Measurement Error. J Marketing Res 1981; 18: 39–50.

65.

Hair

Risher

Sarstedt

, et al. When to use and how to report the results of PLS-SEM. Eur Bus Rev 2019; 31: 2–24.

66.

Bentler

. Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Struct Equ Modeling 1999; 6: 1–55.

67.

Tenenhaus

Vinzi

Chatelin

Y-M

, et al. PLS path modeling. Comput Stat Data An 2005; 48: 159–205.

68.

Liu

. ChatGPT: Perspectives from human–computer interaction and psychology. Front Artif Intell 2024; 7: 1418869.

69.

Wang

Amini

. Associations of task technology fit and perceived usefulness with responsible generative AI use among university students. Sci Rep 2026; 16: 2759.

70.

Choudhury

Shamszare

. Investigating the impact of user trust on the adoption and use of ChatGPT: Survey analysis. J Med Internet Res 2023; 25: e47184.

71.

Yang

Xiao

Liu

, et al. Factors influencing adoption of large language models in health care: Multicenter cross-sectional mixed methods observational study. J Med Internet Res 2025; 27: e84918.

72.

Korchak

Getman

, et al. The role of social influence in generative artificial intelligence ChatGPT adoption intentions among undergraduate and graduate students. Innov Educ Teach Int 2025; 62: 1559–1573.