A SEM–ANN analysis to examine impact of AI overreliance through a social cognitive lens: the roles of dependency, FOMO, addiction, and anxiety

Abstract

The rapid integration of artificial intelligence (AI) has transformed how people learn, work, and make decisions, while raising growing concerns about AI overreliance. Drawing on Social Cognitive Theory (SCT), we investigated how AI dependency, fear of missing out (FOMO), addiction, and Anxiety influence AI overreliance, and examined the mediating role of cognitive overload and the moderating role of technostress. We employed a hybrid analysis method (SEM-ANN), which was used to validate the theoretical relationships and explore nonlinear interactions, using survey data from 504 working adults in Shanghai. Results revealed that AI dependency, FOMO, addiction, and Anxiety have significant positive effects on AI overreliance. Cognitive overload partially mediates these relationships, while technostress moderates the relationship between AI dependency and AI overreliance. The findings are important for both businesses and policymakers. Promoting AI literacy and digital mindfulness will help people use technology more wisely. This will help to better facilitate collaboration between humans and AI.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40359-026-04246-6.

Introduction

The rapid diffusion of artificial intelligence (AI) technologies has profoundly transformed contemporary society, reshaping how individuals learn, work, and interact [45]. Intelligent systems now assist in decision-making, problem-solving, and information processing across both professional and personal contexts. While these technologies enhance efficiency and convenience, growing concerns have emerged regarding excessive reliance on algorithmic assistance [68]. When individuals depend too heavily on intelligent systems, their cognitive engagement, independent judgment, and problem-solving capacities may decline [25].

Empirical evidence has shown that students who excessively rely on generative AI tools such as ChatGPT exhibit diminished self-reflection and critical thinking abilities [34], whereas employees who heavily depend on AI-assisted decision-making tend to experience reduced creativity and innovative problem-solving [8]. Over time, such behavioral patterns weaken self-efficacy and intrinsic motivation, reinforcing passive dependence on technological assistance [65]. As AI becomes increasingly integrated into decision-making, communication, and creative processes, individuals often struggle to maintain balance between technological assistance and autonomous thinking [34]. Thus, the key issue is not whether to adopt AI, but how to integrate it responsibly—maximizing its advantages while preserving human cognitive independence and adaptability [65].

A growing body of research has begun to examine the psychological mechanisms driving AI overreliance. Wu et al.,and Shalaby found technology dependency, fear of missing out (FOMO), addictive tendencies, and Anxiety have been identified as key antecedents that heighten cognitive load during interactions with intelligent systems [59, 71]. However, most existing studies investigate these factors in isolation—focusing either on cognitive mechanisms such as automation bias and reduced vigilance, or on emotional and behavioral antecedents like dependency and technology addiction [17, 33]. Few empirical studies have integrated these perspectives into a unified model. In particular, the mediating role of cognitive overload and the moderating influence of technostress remain underexplored, even though they are crucial for understanding how psychological and contextual variables jointly shape AI overreliance.

To address these gaps, we adopts a social cognitive perspective to investigate how AI dependency, FOMO, addiction, and Anxiety influence AI overreliance through the mediating mechanism of cognitive overload, and how technostress moderates the relationship between AI dependency and AI overreliance. The research employs a hybrid methodological approach combining structural equation modeling (SEM) and artificial neural networks (ANN), which allows both confirmatory testing of theoretical relationships and exploration of nonlinear interactions among constructs.

This study makes several theoretical and methodological contributions. This study extends existing multidimensional models by explicitly integrating cognitive overload and technostress within an SCT-based framework focused on AI overreliance in workplace contexts. By extending SCT to the AI domain, it reveals how reinforcement and emotional regulation mechanisms contribute to dependency formation. Moreover, by applying a hybrid SEM–ANN framework, it provides a more nuanced understanding of interaction effects beyond linear assumptions. Finally, by focusing on working adults in Shanghai—a leading digital hub—this study offers valuable contextual insights for responsible AI governance, ethical system design, and the promotion of sustainable digital well-being.

In summary, this study advances theoretical and practical understanding of the cognitive and psychological foundations of AI overreliance. By revealing how emotional states, cognitive processes, and technological pressures interact within the social-cognitive framework, it contributes to the development of responsible AI use and promotes balanced human–AI collaboration in the digital era.

Literature review and hypothesis

Social cognitive theory and AI overreliance

Social Cognitive Theory (SCT) provides a theoretical framework for understanding human–technology interaction by conceptualizing behavior as the outcome of reciprocal interactions among personal factors, environmental influences, and self-regulatory processes [4, 6]. Individuals are agentic actors who regulate their behavior through cognitive, motivational, and emotional mechanisms rather than passive recipients of technological influence. However, the introduction of powerful cognitive tools such as artificial intelligence (AI) may disrupt this self-regulatory balance by externalizing judgment and decision-making processes. As AI systems increasingly structure task execution and decision contexts, users may experience a shift in the locus of control away from personal agency toward algorithmic outputs, thereby encouraging reliance on AI [61].

Grounded in SCT, this study conceptualizes AI Dependency as a form of habitual cognitive reliance that emerges from reduced self-efficacy and weakened perceived agency reinforced through repeated AI-assisted task success [7]. In parallel, affective and social-cognitive variables, including Anxiety (Anx) and Fear of Missing Out (FOMO), shape individuals’ outcome expectations by heightening perceived risks associated with non-use or autonomous decision-making. Additionally, Cognitive Overload constrains reflective self-regulation by exceeding individuals’ information-processing capacity, thereby increasing reliance on externally provided cognitive resources.

Within this framework, AI Overreliance is conceptualized not as a psychological disposition but as a behavioral outcome reflecting uncritical acceptance of AI-generated outputs, including reduced verification, suppression of contradictory cues, and delegation of judgment to the system [3, 31]. Consistent with SCT, AI overreliance is understood as a learned behavioral pattern arising from the interaction of cognitive and affective antecedents with impaired self-regulatory capacity, particularly under conditions of cognitive overload.

Although AI Dependency, Addiction, FOMO, and Anxiety are interrelated psychological states, they represent theoretically distinct mechanisms within the social cognitive process of technology use. SCT posits that behavior emerges from the interaction of cognitive beliefs, affective states, and self-regulatory capacity [6], and prior research demonstrates that excessive technology use is driven by multiple coexisting psychological pathways rather than a single underlying factor [40]. Specifically, AI dependency reflects habitual cognitive reliance rooted in diminished self-efficacy [7, 41], whereas Addiction captures compulsive AI use characterized by impaired self-regulatory control and persistence despite negative consequences [40]. FOMO reflects socially learned outcome expectations driven by social comparison processes [54], while Anxiety represents an affective response to uncertainty and performance pressure associated with AI use [62]. Taken together, these constructs capture complementary cognitive, affective, social, and self-regulatory pathways through which individuals become vulnerable to AI overreliance. Modeling them jointly therefore provides a more comprehensive explanation of AI overreliance than any single construct alone.

AI dependency and AI overreliance

AI Dependency describes an individual’s perceived psychological and practical necessity to rely on AI systems for task performance and decision-making [46]. AI dependency represents an internal cognitive orientation toward AI as an indispensable support tool [36]. Recent studies further conceptualize AI dependency as a psychological state in which users habitually depend on AI tools for cognitive tasks and perceive AI assistance as necessary for effective problem-solving, while remaining consciously engaged in the decision process [47]. In such cases, users do not disengage from tasks but increasingly perceive their unaided capabilities as insufficient or inefficient [73, 74].

From the perspective of Social Cognitive Theory (SCT), AI dependency reflects diminished self-efficacy and a maladaptive shift in perceived agency from the individual to the technological artifact [4]. Repeated AI-assisted task success reinforces beliefs that effective performance is contingent upon system support, thereby reducing motivation for independent cognitive effort [13]. Importantly, AI dependency is conceptually distinct from AI overreliance. Dependency reflects a psychological sense of necessity and attachment to AI, whereas overreliance denotes a behavioral failure to critically evaluate AI-generated outputs. AI overreliance occurs when individuals uncritically accept AI recommendations despite contradictory information or available personal expertise [9, 24]. Accordingly, dependency represents a relatively stable cognitive orientation, whereas overreliance is behavioral outcome [15].

Within the SCT framework, AI dependency increases the likelihood of AI overreliance by weakening self-regulatory vigilance. When individuals habitually attribute superior competence to AI systems, they perceive the cost of verification as unnecessary, thereby increasing susceptibility to automation bias and overtrust [49] (Anser et al., 2024). Prior research further suggests that frequent interaction with AI systems can gradually diminish self-efficacy and externalize locus of control, reinforcing habitual reliance on algorithmic outputs [27, 69]. Consequently, AI dependency functions as a key antecedent that predisposes individuals to behavioral overreliance.

Hence, this study proposes the following hypothesis:

• H1: AI Dependency has a significant effect on AI Overreliance.

Fear of Missing Out (FOMO) and AI overreliance

Fear of Missing Out (FOMO) refers to the anxiety that others are gaining experiences, opportunities, or information from which one is excluded [64]. In AI-enabled workplaces, employees may perceive continuous AI use as a prerequisite for maintaining professional relevance and competitiveness [72].

Within the SCT framework, FOMO represents a form of observational learning. Individuals observe peers’ effective AI usage and internalize socially reinforced expectations that non-use entails disadvantage or obsolescence [12]. These socially learned outcome expectations motivate individuals to engage with AI beyond instrumental necessity, increasing reliance even when critical evaluation would be appropriate. Unlike generalized anxiety, which reflects internal emotional tension, FOMO is inherently social and comparative in nature. Over time, this socially driven motivation may reduce users’ verification efforts and increase their tendency to accept AI outputs uncritically, thereby contributing to AI overreliance [11, 23].

Hence, this study proposes the following hypothesis:

• H2: FOMO has a significant effect on AI Overreliance.

Addiction and AI overreliance

Addiction is characterized by compulsive engagement with technology despite awareness of its negative consequences [20, 67]. In AI contexts, users may repeatedly engage with generative systems to seek feedback, reassurance, or validation, resulting in excessive and habitual use [55].

From an SCT perspective, addiction reflects a failure of self-regulation driven by reinforcement learning. Immediate rewards such as rapid responses or task completion strengthen usage behavior, gradually overriding reflective control mechanisms [2, 35]. As self-regulatory capacity deteriorates, individuals increasingly default to AI-generated outputs, heightening the likelihood of overreliance.

Hence, this study proposes the following hypothesis:

• H3: Addiction has a significant effect on AI overreliance.

Anxiety and AI overreliance

Anxiety (Anx) is a psychological state characterized by fear, worry, and heightened arousal, often motivating avoidance-oriented coping behaviors [18]. In high-pressure AI-mediated environments, anxious individuals may perceive algorithmic recommendations as safer and more reliable than their own judgment [53].

Within SCT, anxiety shapes negative outcome expectations, whereby individuals anticipate adverse consequences from autonomous decision-making. Unlike FOMO, which is socially comparative, anxiety reflects internal affective discomfort under uncertainty [30]. Reliance on AI thus serves as an externalized coping strategy that temporarily alleviates stress [30] but reinforces behavioral dependence over time.

Hence, this study proposes the following hypothesis:

• H4: Anxiety has a significant effect on AI overreliance.

The mediating role of cognitive overload

From a Social Cognitive Theory (SCT) perspective, cognitive overload represents a situational constraint on self-regulation that limits individuals’ ability to monitor, evaluate, and adjust their own cognitive processes [4, 37]. When cognitive resources are strained, self-regulatory control weakens, making individuals more dependent on external support for task completion. Prior research has shown that high cognitive load reduces reflective thinking and increases reliance on automated or decision-support systems [49]. 

Although AI Dependency, Addiction, Fear of Missing Out (FOMO), and Anxiety (Anx) originate from different cognitive or emotional sources, they converge in their tendency to increase mental demands. AI dependency promotes cognitive offloading, addiction reinforces repetitive use, FOMO intensifies information-seeking behavior, and anxiety increases worry and mental tension [16, 67]. These processes collectively raise cognitive load, making cognitive overload a common psychological pathway through which diverse internal predispositions are translated into observable behavioral overreliance on AI systems.

Accordingly, cognitive overload is conceptualized as a mediating mechanism that transforms weakened self-regulatory capacity into increased reliance on external cognitive support. Existing studies indicate that prolonged exposure to complex digital environments increases perceived mental burden and encourages reliance on technological assistance for problem-solving and decision-making [25, 60]. Individuals with high levels of AI Dependency, FOMO, Addiction, or Anxiety are therefore more susceptible to cognitive overload due to frequent engagement with AI tools for validation, information seeking, or emotional regulation. As cognitive overload intensifies, users are more likely to adopt AI as a compensatory strategy to cope with mental complexity and decision fatigue, which in turn reinforces AI overreliance.

Hence, this study proposes the following hypothesis:

• H5: Cognitive Overload has a significant effect on AI Overreliance.

• H5a: Cognitive Overload mediates the relationship between AI Dependency and AI Overreliance.

• H5b: Cognitive Overload mediates the relationship between FOMO and AI Overreliance.

• H5c: Cognitive Overload mediates the relationship between Addiction and AI Overreliance.

• H5d: Cognitive Overload mediates the relationship between Anxiety and AI Overreliance.

The moderating role of technostress

Technostress refers to the psychological tension and discomfort individuals experience when using or adapting to new technologies [42]. This stress is often manifested as Anx, fatigue, reduced efficiency, and increased cognitive load resulting from excessive technology use or continuous connectivity. According to Social Cognitive Theory (SCT), technostress functions as an environmental determinant that interacts with individuals’ cognitive and emotional processes, thereby influencing their behavior [5]. Technostress, reflects an external situational pressure that intensifies habitual reliance under high system demands, thereby strengthening the impact of AI Dependency on AI Overreliance. Empirical evidence suggests that individuals in high technostress environments are more likely to develop maladaptive coping behaviors [42], such as excessive reliance on automated systems. Based on this reasoning, the following hypothesis is proposed:

• H6: Technostress moderates the relationship between AI Dependency and AI Overreliance.

Grounded in the above hypotheses, the resulting research framework is depicted in Fig. 1.

Fig. 1.

Research framework based on social cognitive theory

Methodology

Research design

This study adopted a non-probability sampling method, specifically purposive sampling, to collect data. This method allows researchers to deliberately select participants who are most relevant to the research objectives and possess the characteristics necessary to provide meaningful insights (Etikan, Musa, & Alkassim, 2016). Given the study’s focus on understanding AI overreliance among Shanghai residents, purposive sampling was deemed appropriate because it enables the researcher to target respondents with sufficient exposure to AI-related technologies and digital environments—factors critical for obtaining relevant and valid responses [66].

This study employed a questionnaire survey method for data collection. The questionnaire items was adapted from previously published and validated English questionnaires, with detailed sources provided in Appendix 1. Specifically, AI Overreliance was adapted from Maral et al. [44], comprising 11 items. AI Dependency was measured using five items adapted from Morales-García et al. [46]. Fear of Missing Out (FOMO) was assessed with eight items derived from Budnick et al. [10]. Addiction was adapted from Pawlikowski et al. [50], containing 12 items. Anx was measured with eight items adapted from Oei et al. [48]. Cognitive Overload was adapted from Malo et al. [43] with eight items, and Technostress was measured using 10 items adapted from Porcari et al. [52].

To ensure linguistic accuracy and contextual suitability for this study, the questionnaire underwent two stages before its official distribution. Experts in management and behavioral research evaluated its content validity. A small-scale pilot test was then conducted among a subset of respondents sharing similar demographic characteristics with the target sample. Based on feedback, several technical and linguistic revisions were made to enhance clarity and reliability.

The final questionnaire consisted of two main sections. Sect. 1: Collected demographic information, including participants’ name, age, gender, and occupation. Sect. 2: Assessed AI Overreliance, with measurement items derived from established theoretical constructs and previous studies (see Appendix 1).

All measurement items were rated using a seven-point Likert scale, ranging from 1 = strongly disagree to 7 = strongly agree. The seven-point format was chosen over the five-point scale because it provides greater response sensitivity, improves discriminant validity, and enhances reliability and precision in capturing participants’ attitudes [19].

Sample size determination

The minimum required sample size was determined using G*Power 3.1 software, with a medium effect size (f² = 0.15), an alpha value of 0.05, and a statistical power of 0.80, resulting in a minimum of 146 respondents. According to Hair et al. [29], in Partial Least Squares Structural Equation Modeling (PLS-SEM), the appropriate sample size should be at least ten times the maximum number of structural paths pointing to a construct in the model. Considering the complexity of the current model and to enhance statistical validity and generalizability, a total of 504 valid responses were ultimately collected and used for analysis. This sample size exceeds both the G*Power threshold and Hair et al.’s [29] recommendation, ensuring sufficient statistical power for robust hypothesis testing.

Data collection procedure and ethical considerations

The study targeted working adults of Shanghai, China, as the research population. Shanghai was selected as the research site based on theoretical and practical considerations rather than convenience sampling. Shanghai's AI industry output exceeded RMB 450 billion in 2024, ranked first nationally with 68 industry-specific generative AI applications registered. This concentration of AI-integrated workplaces provides a suitable setting to test SCT propositions about overreliance in technologically saturated environments. Methodologically, focusing on a single region controls for macro-level socio-cultural and policy variations that could confound cross-regional comparisons.

This study adopted a purposive sampling strategy, targeting individuals who met two predefined criteria: (1) being employed full-time in Shanghai and (2) regularly using AI tools in their daily work tasks. This sampling approach ensured that respondents possessed direct and sustained experience with AI-supported work environments, which is essential for examining AI overreliance.

Data were collected online between February 12 and May 2025 through the Wenjuanxing online survey platform, ensuring accessibility, anonymity, and voluntary participation. The online administration enabled broad reach across multiple industries. The questionnaire was self-administered and required approximately 10–15 min to complete.

To ensure data quality, two screening questions were presented at the beginning of the questionnaire. Participants were asked whether they were currently working in Shanghai and whether AI tools were frequently used in their daily work. Only respondents who answered “yes” to both questions were allowed to proceed with the survey, while others were automatically exited from the questionnaire.

After data collection, responses with missing values were removed. In addition, responses with the same ID were treated as duplicate entries and retained only once. Following these data screening and cleaning procedures, a total of 504 valid responses were retained for subsequent analysis.

Ethical approval for this study was obtained from the Ethics Committee of SEGi University, Malaysia, prior to data collection (Approval No: SEGiEC/SR/GSB/56/2025–2026). All participants were informed of the study’s purpose, assured that their participation was voluntary, and that their responses would remain anonymous and confidential. Before completing the survey, each participant provided digital informed consent. The study was conducted in accordance with the Declaration of Helsinki (2013), ensuring respect, integrity, and data protection.

Data analysis procedures

After data collection, all responses were carefully screened and cleaned to remove incomplete, inconsistent, or outlier entries. The data analysis was conducted in two stages: Stage 1: Hypothesis testing was performed using SPSS 27.0 and SmartPLS 4.0, including common method bias, reliability analysis, validity assessment, and PLS-SEM structural model testing. Stage 2: To further validate the relationships among key variables, Artificial Neural Network (ANN) analysis was conducted using SPSS 27.0, following the dual-stage SEM–ANN approach recommended by prior research [75]. This hybrid method enhances predictive accuracy and provides nonlinear validation of structural relationships.

Demographic information

A total of 504 valid responses were collected from working adults in Shanghai. Table 1 summarizes the demographic characteristics of the sample. Males accounted for 56.15% of respondents, females for 41.27%, and 2.58% preferred not to disclose their gender. The sample was predominantly young, with 76.59% aged under 35. Most participants held a bachelor’s degree (57.54%) or higher. Respondents were distributed across major industries—including technology (24.01%), finance (23.21%), manufacturing (26.98%), and education (19.05%)—reflecting Shanghai’s diversified economy. The majority occupied entry-level roles (87.70%), with 40.48% having 1–3 years of work experience. Overall, the demographic composition aligns with the characteristics of Shanghai’s rapidly expanding digital workforce and supports the relevance of the study’s focus on AI use in dynamic work environments.

Table 1.

Demographic information

Variable	Category	Number	Percentage (%)
Gender	Male	283	56.15
	Female	208	41.27
	Prefer not to say	13	2.58
Age	Under 25	182	36.11
	25–34	204	40.48
	35–44	73	14.48
	45–54	36	7.14
	55 +	9	1.79
Education Level	High school or below	114	22.62
	Bachelor’s degree	290	57.54
	Master’s degree	76	15.08
	Doctorate	15	2.98
	Other	9	1.79
Industry	Education	96	19.05
	Finance	117	23.21
	Technology	121	24.01
	Healthcare	21	4.17
	Manufacturing	136	26.98
	Other	13	2.58
Position Level	Entry-level	442	87.7
	Middle management	55	10.91
	Senior management	4	0.79
	Other	3	0.6
Work Experience	< 1 year	167	33.13
	1–3 years	204	40.48
	4–6 years	78	15.48
	7–10 years	42	8.33
	> 10 years	13	2.58

Exploratory factor analysis

To further substantiate the conceptual distinction between AI Dependency and AI Overreliance, an Exploratory Factor Analysis (EFA) was conducted prior to the confirmatory procedures. The analysis yielded a clean two-factor solution. Factor 1 (Dependency) explained 38% of the variance, with items AID1–AID4 loading between 0.760 and 0.815. Factor 2 (Overreliance) explained 31% of the variance, with items AIO1–AIO11 loading between 0.722 and 0.766. Factor loadings: Correlation between item and factor (should be > 0.4) Cross-loadings: Item loads on multiple factors (should be < 0.3 for clean structure): No item cross-loaded, indicating that the psychological feeling of dependency is empirically distinct from the behavioral act of overreliance.

The analysis demonstrated that the items representing the two constructs load cleanly onto separate factors, supporting their empirical separability and addressing potential concerns regarding construct overlap. The results of the EFA, including factor loadings are presented in Table 2.

Table 2.

Rotated component matrixa

	Component
	1	2	3	4	5	6	7
AI_OL1		0.744
AI_OL2		0.728
AI_OL3		0.723
AI_OL4		0.722
AI_OL5		0.745
AI_OL6		0.739
AI_OL7		0.760
AI_OL8		0.731
AI_OL9		0.766
AI_OL10		0.740
AI_OL11		0.737
AI_Den1							0.811
AI_Den2							0.815
AI_Den3							0.803
AI_Den4							0.800
AI_Den5							0.760
FOMO1						0.806
FOMO2						0.816
FOMO3						0.772
FOMO4						0.793
FOMO5						0.801
FOMO6						0.804
FOMO7						0.787
FOMO8						0.788
AI_ADD1	0.809
AI_ADD2	0.846
AI_ADD3	0.858
AI_ADD4	0.826
AI_ADD5	0.824
AI_ADD6	0.852
AI_ADD7	0.833
AI_ADD8	0.850
AI_ADD9	0.829
AI_ADD10	0.834
AI_ADD11	0.834
AI_ADD12	0.837
Anx1					0.795
Anx2					0.807
Anx3					0.811
Anx4					0.803
Anx5					0.790
Anx6					0.822
Anx7					0.795
Anx8					0.805
CO1				0.835
CO2				0.820
CO3				0.831
CO4				0.833
CO5				0.820
CO6				0.833
CO7				0.860
CO8				0.831
TS1			0.781
TS2			0.779
TS3			0.797
TS4			0.767
TS5			0.792
TS6			0.805
TS7			0.774
TS8			0.767
TS9			0.794
TS10			0.789

Data analysis

PLS-SEM analysis

Common method bias

To address potential Common Method Bias (CMB) concerns, Harman’s single-factor test was conducted. The first unrotated factor accounted for 45.21% of the total variance, which is below the critical threshold of 50% [51]. This result indicates that CMB is not a serious threat to the validity of this study. Additionally, the variance inflation factor (VIF) values for all potential constructs were examined, revealing that all VIF values were below 3.3 [38], further confirming the absence of multicollinearity and common method variance issues. Together, these results provide reasonable assurance that the relationships observed among the study variables were not substantially inflated by measurement artifacts.

Confirmatory factor analysis

To assess the reliability and validity of the measurement model, a Confirmatory Factor Analysis (CFA) was conducted. As shown in Table 3, all constructs demonstrated strong internal consistency, with Cronbach’s alpha values ranging from 0.953 to 0.981 and composite reliability values between 0.961 and 0.983, exceeding the recommended threshold of 0.70 [29].

Table 3.

Reliability and convergent validity

	Cronbach’s alpha	Composite reliability (rho_c)	Average variance extracted (AVE)
AI_ADD	0.981	0.983	0.826
AI_Den	0.953	0.964	0.842
AI_OL	0.969	0.973	0.765
Anx	0.970	0.974	0.824
CO	0.972	0.976	0.834
FOMO	0.966	0.971	0.807
TS	0.954	0.961	0.709

Convergent validity was also confirmed as all Average Variance Extracted (AVE) values were above 0.709, surpassing the 0.50 benchmark, indicating that each construct explained more than half of the variance of its indicators [22, 29].

This study employed the Fornell-Larcker criteria and the HTMT (Heterotrait-Monotrait ratio) method to examine the discriminant validity of latent variables.

First, as shown in Table 4, the diagonal elements represent the square roots of the average-variance-extracted (AVE) reliability coefficients for each latent variable (0.842–0.917), while the off-diagonal elements denote the correlation coefficients between latent variables (0.367–0.594). The square root of the AVE reliability coefficient for each latent variable exceeds its correlation coefficients with other latent variables, indicating that each latent variable possesses good discriminant validity [22].

Table 4.

Discriminant validity

	Fornell-Larcker criteria							HTMT value
Variables	AI_ADD	AI_Den	AI_OL	Anx	CO	FOMO	TS	AI_ADD	AI_Den	AI_OL	Anx	CO	FOMO
AI_ADD	0.909
AI_Den	0.458	0.917						0.473
AI_OL	0.578	0.539	0.875					0.592	0.559
Anx	0.486	0.48	0.594	0.908				0.498	0.498	0.612
CO	0.435	0.459	0.525	0.463	0.913			0.445	0.475	0.54	0.476
FOMO	0.453	0.474	0.57	0.542	0.507	0.899		0.465	0.493	0.589	0.56	0.523
TS	0.437	0.463	0.531	0.397	0.367	0.419	0.842	0.45	0.483	0.551	0.411	0.38	0.435

Additionally, discriminant validity was evaluated using the Heterotrait–Monotrait HTMT values. From Table 4 Heterotrait–Monotrait (HTMT) ratios among constructs were all below 0.90 [32], further supporting discriminant validity.

Regarding model fit, the SRMR value was 0.025 and NFI was 0.931, both indicating a good model fit according to Hair et al. [29]. The R² values of AI_OL (0.578) and CO (0.359) suggested that the model has a moderate explanatory power for endogenous variables (show in Table 5).

Table 5.

R² and model fit

R-square		Saturated model
AI_OL	0.578	SRMR	0.025
CO	0.359	NFI	0.931

Path Coefficient

The structural model was evaluated using path coefficients, t-values, and p-values derived from the bootstrapping procedure (5000 resamples) to assess the significance of hypothesized relationships. As presented in Table 6, all hypothesized paths were statistically significant at the 0.05 level, supporting all proposed hypotheses (H1–H7)(show in Table 6).

Table 6.

Path coefficient

Hypothesis	Relationship	β	T	P	Results
H1	AI_Den—> AI_OL	0.120	3.032	0.002	Supported
H2	FOMO—> AI_OL	0.155	3.665	0.000	Supported
H3	AI_ADD—> AI_OL	0.211	4.685	0.000	Supported
H4	Anx—> AI_OL	0.215	5.078	0.000	Supported
H5	CO—> AI_OL	0.135	3.640	0.000	Supported
H6	TS x AI_Den—> AI_OL	0.101	2.582	0.010	Supported

AI Dependency (β = 0.120, t = 3.032, p = 0.002), FOMO (β = 0.155, t = 3.665, p < 0.001), AI Addiction (β = 0.211, t = 4.685, p < 0.001), and Anx (β = 0.215, t = 5.078, p < 0.001) exerted significant positive effects on AI Overreliance, indicating that individuals who feel more dependent on AI, experience stronger FOMO, have higher addictive tendencies, and suffer from greater Anx are more likely to over-rely on AI technologies. The path from AI Dependency to AI Overreliance was positive and significant (β = 0.145, p < 0.001). Although this coefficient represents a small effect according to conventional PLS-SEM benchmarks [28], such magnitudes are typical in multivariate behavioral models where several cognitive, emotional, and contextual predictors simultaneously influence the outcome. Thus, the effect size should be interpreted as an incremental yet theoretically meaningful contribution to users’ behavioral tendency toward AI Overreliance.

Moreover, Cognitive Overload had a significant positive effect on AI Overreliance (β = 0.135, t = 3.640, p < 0.001), indicating that users who experience greater mental strain from AI use tend to depend even more heavily on it. Finally, Technostress was found to significantly moderate the relationship between AI Dependency and AI Overreliance (β = 0.101, t = 2.582, p = 0.010), implying that when individuals experience higher levels of technostress, the positive effect of AI dependency on overreliance becomes stronger.

Mediation analysis

Regarding the direct effects all four relationships remained significant (show in Table 7), namely AI_Den → AI_OL (β = 0.120, t = 3.032, p = 0.002), FOMO → AI_OL (β = 0.155, t = 3.665, p < 0.001), AI_ADD → AI_OL (β = 0.211, t = 4.685, p < 0.001), and Anx → AI_OL (β = 0.215, t = 5.078, p < 0.001). This indicates that both behavioral (dependency and addiction) and emotional (Anx and FOMO) factors directly contribute to users’ overreliant behaviors toward AI systems.

Table 7.

Mediation analysis

Direct effects
Hypothesis	Relationship	β	T	P	Results
H1	AI_Den—> AI_OL	0.120	3.032	0.002	Supported
H2	FOMO—> AI_OL	0.155	3.665	0.000	Supported
H3	AI_ADD—> AI_OL	0.211	4.685	0.000	Supported
H4	Anx—> AI_OL	0.215	5.078	0.000	Supported
Specific indirect effects
Hypothesis	Relationship	β	T	P	Results
H5a	AI_Den—> CO—> AI_OL	0.025	2.685	0.007	Supported
H5b	FOMO—> CO—> AI_OL	0.036	2.996	0.003	Supported
H5c	AI_ADD—> CO—> AI_OL	0.021	2.445	0.015	Supported
H5d	Anx—> CO—> AI_OL	0.021	2.118	0.034	Supported
Total effects
	Relationship	β	T	P
	AI_Den—> AI_OL	0.145	3.544	0.000
	FOMO—> AI_OL	0.191	4.592	0.000
	AI_ADD—> AI_OL	0.232	5.197	0.000
	Anx—> AI_OL	0.236	5.242	0.000

Further, the specific indirect effects through Cognitive Overload (CO) were also significant, confirming CO as a crucial mediator. Specifically, AI_Den → CO → AI_OL (β = 0.025, t = 2.685, p = 0.007), FOMO → CO → AI_OL (β = 0.036, t = 2.996, p = 0.003), AI_ADD → CO → AI_OL (β = 0.021, t = 2.445, p = 0.015), and Anx → CO → AI_OL (β = 0.021, t = 2.118, p = 0.034) were all supported. Among these, the mediation effect of FOMO was the strongest, suggesting that users with higher fear of missing out experience greater cognitive strain during AI use, which subsequently enhances their overreliance.

Finally, regarding the total effects, all four antecedents exhibited significant positive influences on AI Overreliance (AI_OL). Specifically, AI Dependency (β = 0.145, t = 3.544, p < 0.001), FOMO (β = 0.191, t = 4.592, p < 0.001), AI Addiction (β = 0.232, t = 5.197, p < 0.001), and Anx (β = 0.236, t = 5.242, p < 0.001) all significantly predicted AI Overreliance. Among these, Anx and AI Addiction exerted the strongest total effects, suggesting that emotional and behavioral dependence serve as key precursors to users’ excessive reliance on AI technologies. Since both direct and indirect effects are significant, cognitive overload exhibits partial mediation effects across all pathways [29].

Moderation analysis

This study examined the moderating role of Technostress (TS) in the relationship between AI Dependency (AI_Den) and AI Overreliance (AI_OL). As show in Fig. 2 the three regression lines representing different levels of Technostress (low TS = –1 SD; mean TS = mean; high TS = + 1 SD) all show a positive trend, indicating that individuals with higher levels of AI dependency tend to exhibit stronger AI overreliance overall. However, the strength of this relationship varies across levels of technostress. Specifically, under high technostress conditions (+ 1 SD), the positive effect of AI dependency on AI overreliance is most pronounced, as reflected in the steepest slope. Under average technostress, the relationship remains positive but weaker, whereas under low technostress (–1 SD), the association becomes relatively weak or even nonsignificant.

Fig. 2.

Moderation analysis

Consistent with the statistical results (β = 0.101, t = 2.582, p = 0.010), these findings confirm that technostress significantly moderates the effect of AI dependency on AI overreliance. In other words, individuals experiencing higher levels of technostress are more likely to transform functional dependence on AI tools into irrational overreliance. This finding aligns with previous studies suggesting that technostress can exacerbate problematic technology use and compulsive digital behavior [42]. Overall, the results highlight technostress as not only a source of psychological strain but also a contextual factor that amplifies excessive reliance on AI systems.

All path coefficient results are shown in Fig. 3.

Fig. 3.

Results of path coefficient

ANN analysis

To further validate the robustness of the PLS-SEM results and capture potential non-linear relationships among the variables, an Artificial Neural Network (ANN) analysis was conducted. PLS-SEM was employed as the primary analytical technique to test theoretically derived hypotheses and causal mechanisms grounded in Social Cognitive Theory. ANN was subsequently applied to assess the robustness and relative importance of predictors under non-linear conditions, thereby complementing SEM results rather than replacing theory-driven explanations.

In this study, the neural network model was constructed with six input nodes corresponding to the core predictors identified as significant in the PLS-SEM analysis, namely AI Dependency, AI Addiction, Fear of Missing Out (FOMO), Anxiety, Cognitive Overload, and Technostress. The model architecture consisted of two hidden layers containing four and three neurons, respectively. Both hidden layers employed the sigmoid activation function, which is suitable for modeling nonlinear relationships in behavioral and psychological data. The output layer was designed to predict the dependent variable, AI Overreliance (AI_OL), and also adopted the sigmoid activation function to accommodate continuous outcome prediction. Figure 4 illustrates the overall structure of the ANN model, including the input layer, hidden layers, and output layer.

Fig. 4.

Artificial Neural Network (ANN) analysis

During model training, a typical holdout validation strategy was adopted to reduce the risk of overfitting and to evaluate the generalization capability of the model. Specifically, the full sample (N = 504) was randomly divided into a training set comprising 400 cases (79.4%) and a testing set comprising 104 cases (20.6%). The ANN model converged rapidly within 0.03 s, indicating satisfactory computational efficiency. For the training sample, the sum of squared errors (SSE) was 5.264, with a relative error of 0.297. For the testing sample, the SSE was 1.765, with a relative error of 0.519. Although the testing error was slightly higher than the training error, the difference remained within an acceptable range, suggesting that the model demonstrates adequate predictive stability and does not exhibit signs of overfitting.

Neural network models involve inherent stochasticity due to random parameter initialization and training–testing data partitioning, which may lead to variability in variable importance estimates across different runs. Therefore, sensitivity analysis based on a single training process may lack sufficient robustness.

To address this issue, prior studies recommend repeated model training and averaging importance values to reduce stochastic effects. Specifically, Gevrey et al. indicated that more than 10 repetitions already provide reliable reference values [26]. Following these recommendations, this study conducted 12 independent repetitions of the RBF neural network model. The averaged importance values across all runs were used as the final sensitivity analysis results (in Table 8), thereby improving the stability, interpretability, and robustness of the ANN findings.

Table 8.

Independent variable importance (average)

	Importance	Normalized importance
TS	0.471	94.32%
Anx	0.204	45.79%
CO	0.089	20.86%
AI_Den	0.095	20.38%
AI_ADD	0.077	16.58%
FOMO	0.064	14.03%

Overall, the ANN results align with the PLS-SEM findings, confirming that Technostress and Anxiety are the strongest predictors of AI Overreliance, while Dependency, FOMO, and Addiction serve as secondary but meaningful contributors. These findings are consistent with prior studies emphasizing that technostress and emotional strain can amplify users’ irrational trust and overdependence on intelligent systems [42]. Thus, the ANN model provides additional empirical evidence supporting the robustness and explanatory power of the structural model.

Discussion

This study examines the cognitive and psychological mechanisms underlying excessive reliance on artificial intelligence from the perspective of Social Cognitive Theory (SCT), based on data collected from 504 employed adults in Shanghai.

Cognitive and psychological factors of AI overreliance

PLS-SEM results suggest that AI Dependency contributes to AI overreliance through both direct and indirect paths, while ANN shows that AI Dependency has a stable but secondary predictive role. Together, these findings suggest that dependency reflects a baseline tendency toward reliance rather than a dominant behavioral driver. This interpretation is consistent with SCT, which emphasizes that repeated external reinforcement can shift control away from internal regulation [58]. In highly digitalized work settings, such as Shanghai, reliance on AI may therefore become a routine component of task execution [1, 57].

Structural results indicate Fear of Missing Out (FOMO) also shows a significant association with AI overreliance, while ANN results suggest relatively limited predictive importance. Prior studies show that individuals with higher FOMO rely on AI tools to maintain social or professional relevance [70, 72]. While such use may improve short-term efficiency, it may also increase mental fatigue and reliance on fast decision-making, consistent with evidence that cognitive overload encourages heuristic choices [21]. These results suggest that FOMO acts as a conditional motivational factor rather than a direct determinant of overreliance.

PLS-SEM results confirm AI Addiction shows both direct and indirect associations with AI overreliance, while ANN indicates a moderate predictive role. Existing research suggests that sustained habitual use may reduce vigilance and independent judgment over time, even without uncritical acceptance of AI outputs [73, 74]. This finding points to a tension in AI-supported work, where perceived performance gains may coincide with reduced cognitive autonomy.

While PLS-SEM shows significant direct and indirect effects, ANN ranks Anxiety as the second most important predictor overall. Anxiety emerges as one of the most influential individual-level factors related to AI overreliance. Prior research suggests that anxious individuals view AI as a stable and predictable aid that reduces uncertainty and emotional discomfort [14]. From an SCT perspective, this pattern reflects a coping-oriented shift from self-regulation to external support, which may reduce confidence in personal abilities over time [56].

Mediating effect of cognitive overload

The results indicate that Cognitive Overload serves as a key mediating mechanism linking multiple psychological antecedents to AI overreliance. Individuals with higher AI Dependency, FOMO, Addiction, and Anxiety report greater information burden and mental fatigue, which increases reliance on AI. This is consistent with evidence that excessive information exposure impairs cognitive control and critical evaluation [39, 63].

Within the SCT framework, cognitive overload reflects a condition in which personal motivations and emotions interact with demanding environments to constrain self-regulation. ANN results further show that the predictive importance of Cognitive Overload is comparable to that of AI Dependency, indicating that overload is not a by-product of technology use but a central pathway through which psychological factors influence overreliance.

Moderating effect of technostress

PLS-SEM results indicate that technostress strengthens the relationship between AI Dependency and AI Overreliance. This result aligns with SCT, which emphasizes that environmental conditions influence the strength of personal behavior [4, 6].

In contrast, the ANN analysis identifies technostress as the most influential predictor overall. This aligns with prior research showing that technostress diminishes perceived control [42]. By capturing cumulative and non-linear effects, ANN reveals that technostress amplifies AI reliance through multiple pathways, highlighting the critical role of managing technological stress to prevent overreliance. These findings suggest that technostress may shift AI use from a supportive tool toward a necessary coping resource, reinforcing dependency-driven reliance patterns in digital work settings.

Taken together, the results show that AI overreliance arises from the joint influence of cognitive tendencies, psychological states, and technostress. While dependency use provide a baseline orientation toward AI reliance, anxiety and technostress appear more influential in shaping actual reliance behavior in practice. Cognitive overload functions as a central process through which these factors translate into overreliance.

Significance and limitations

Theoretical significance

This study advances the literature on AI–human interaction by extending Social Cognitive Theory (SCT) to contexts characterized by intensive and continuous AI use. It demonstrates that AI dependency, FOMO, anxiety, and addictive tendencies undermine self-efficacy and promote reliance on external cognitive systems, clarifying how AI-related psychological states reshape self-regulatory processes. By identifying cognitive overload as a mediating mechanism, the study explains how information complexity and decision fatigue translate psychological and technological pressures into AI overreliance. Moreover, the moderating role of technostress highlights the importance of contextual technological demands, showing that environmental pressure amplifies the effects of individual cognition on reliance behavior. Together, these findings refine SCT by integrating AI-specific psychological constructs and technological stressors into a unified explanatory framework..

Practical significance

The findings offer actionable insights for organizations, system designers, and policymakers seeking to manage AI overreliance in digital work environments. Organizations may mitigate excessive reliance by fostering digital self-regulation and critical AI literacy, thereby sustaining cognitive engagement without compromising efficiency. For technology designers, reducing cognitive load through transparent reasoning and user-centered interfaces may encourage reflective rather than automatic AI use. At the policy level, initiatives that address technostress and protect digital well-being—such as guidelines for responsible AI use—may help maintain a balance between productivity and psychological sustainability, ensuring that AI functions as a cognitive augmentation tool rather than a substitute for human judgment.

Limitation and future research

Despite its contributions, this study has several limitations. First, the cross-sectional design limits causal inference. Second, the sample was drawn from Shanghai, China’s most AI-intensive metropolis with a young and highly educated workforce, which may limit the generalizability of the findings to regions with lower AI penetration (e.g., Western China). Third, reliance on self-reported data may introduce subjective or social desirability bias. Fourth, although validated scales were adapted to the research context, cultural interpretation differences may still affect measurement accuracy. Future research should employ longitudinal designs, multi-regional samples, and mixed-method approaches to strengthen causal inference and enhance generalizabili.

Abbreviations

AI_OL

AI Overreliance

AI_ADD

AI Addiction

AI_Den

AI Dependency

Anx

Anxiety

CO

Cognitive overload

FOMO

Fear of Missing Out

TS

Technostress

Appendix 1

AI Overreliance [44]

1. I constantly have thoughts related to AI tools lingering in my mind.

2. I often open AI applications even when I had no initial intention to use them.

3. I feel anxious or irritable when I cannot access AI tools.

4. I find myself spending progressively more time using AI.

5. I have tried to reduce my AI use but have been unsuccessful.

6. I have lost interest in activities I previously enjoyed due to AI use.

7. My use of AI leads me to procrastinate or delay completing necessary tasks.

8. I devote excessive time to AI tools despite negative consequences.

9. My sleep is negatively affected by excessive AI use.

10. I sometimes hide the extent of my AI usage from others.

11. I turn to AI tools to alleviate feelings of stress, helplessness, or Anx.

AI dependency [46]

1. I feel unprotected when I do not have access to AI.

2. I am concerned about being left behind at work if I do not use AI.

3. I do everything possible to stay updated with AI to remain competitive.

4. I need validation or feedback from AI systems to feel confident in my decisions.

5. I worry that AI might replace my current skills or abilities.

FOMO [10]

1. I worry that I might miss important work-related updates.

2. I worry that I might miss valuable information relevant to my job.

3. I fear missing important professional news or trends.

4. I am concerned about not knowing what is happening in my workplace or industry.

5. I get anxious that I will miss opportunities to strengthen professional connections.

6. I often think about potential business opportunities I might miss.

7. I worry that my colleagues may form professional networks that I miss out on.

8. I fear that others at work might advance because they have more AI-related information or access.

Addiction [50]

1. I stay online using AI tools longer than I intended.

2. I tell myself “just a few more minutes” when using AI.

3. I neglect personal tasks to spend more time using AI.

4. I have tried to reduce AI use but failed.

5. My work performance has suffered due to excessive AI use.

6. I lose sleep because of late-night AI usage.

7. I prefer using AI tools rather than socializing.

8. I try to hide how much time I spend using AI.

9. I get annoyed when someone interrupts my AI use.

10. I feel anxious or moody when I can’t use AI, but better once I return.

11. I often think about AI even when I’m not using it.

12. I become defensive if someone asks about my AI use.

Anx [48]

1. I am aware of my heartbeat even without physical exertion.

2. I experience difficulty breathing.

3. I feel scared for no apparent reason.

4. My mouth feels dry when I’m anxious.

5. I feel close to panic in stressful situations.

6. I worry about embarrassing myself in front of others.

7. I experience trembling or shaking.

8. I find it hard to relax or calm down.

Cognitive Overload [43]

1. I struggle to meet the cognitive demands of my work.

2. I feel uncertain about how to perform some of my work tasks.

3. I often have difficulty mastering new AI-related tasks.

4. I sometimes feel overwhelmed by the amount of information I must process.

5. I find it hard to decide which AI information or advice to trust.

6. I feel mentally exhausted after using AI tools for extended periods.

7. I often find it hard to focus because of excessive AI-related information.

8. I feel that using AI increases my mental workload.

Technostress [52]

1. Prolonged use of digital tools reduces my concentration.

2. Continuous use of AI negatively affects my job performance.

3. Frequent technological updates make me feel overworked.

4. Using AI tools makes me feel constantly available to others.

5. I feel pressured to keep up with new AI systems.

6. Technical problems with AI tools cause stress and interruptions.

7. Frequent use of digital tools causes me physical discomfort.

8. Using AI tools for long periods causes Anx and tension.

9. I feel that technology use invades my personal life.

10. When I receive AI-related notifications, I feel compelled to respond immediately.

Appendix 2

ANN1			ANN2
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.158	30.10%	AI_ADD	0.029	4.50%
AI_Den	0.039	7.50%	AI_Den	0.086	13.30%
Anx	0.105	19.80%	Anx	0.112	17.20%
CO	0.062	11.70%	CO	0.066	10.10%
FOMO	0.109	20.70%	FOMO	0.057	8.80%
TS	0.527	100.00%	TS	0.65	100.00%
ANN3			ANN4
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.042	7.40%	AI_ADD	0.162	40.70%
AI_Den	0.091	16.10%	AI_Den	0.081	20.30%
Anx	0.148	26.00%	Anx	0.16	40.00%
CO	0.076	13.40%	CO	0.103	25.90%
FOMO	0.074	13.10%	FOMO	0.095	23.80%
TS	0.568	100.00%	TS	0.399	100.00%
ANN5			ANN6
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.078	11.40%	AI_ADD	0.062	11.80%
AI_Den	0.051	7.30%	AI_Den	0.084	15.90%
Anx	0.116	16.80%	Anx	0.189	35.80%
CO	0.019	2.70%	CO	0.068	12.90%
FOMO	0.046	6.60%	FOMO	0.069	13.10%
TS	0.691	100.00%	TS	0.527	100.00%
ANN7			ANN8
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.127	32.40%	AI_ADD	0.076	16.20%
AI_Den	0.113	28.90%	AI_Den	0.114	24.30%
Anx	0.208	53.10%	Anx	0.467	100.00%
CO	0.084	21.50%	CO	0.125	26.90%
FOMO	0.075	19.20%	FOMO	0.029	6.20%
TS	0.392	100.00%	TS	0.19	40.60%
ANN9			ANN10
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.037	9.30%	AI_ADD	0.064	14.40%
AI_Den	0.121	30.60%	AI_Den	0.142	31.90%
Anx	0.288	73.10%	Anx	0.173	38.80%
CO	0.127	32.20%	CO	0.12	26.90%
FOMO	0.034	8.70%	FOMO	0.054	12.10%
TS	0.394	100.00%	TS	0.446	100.00%
ANN11			ANN12
	Importance	Normalized Importance		Importance	Normalized Importance
AI_ADD	0.038	6.40%	AI_ADD	0.045	14.40%
AI_Den	0.137	23.30%	AI_Den	0.078	25.10%
Anx	0.169	28.90%	Anx	0.311	100.00%
CO	0.037	6.30%	CO	0.186	59.80%
FOMO	0.032	5.40%	FOMO	0.095	30.60%
TS	0.587	100.00%	TS	0.284	91.20%

Authors’ contributions

Conceptualization: Z, Y, R.R, K Data curation: Z, K, G Methodology: Z, Y, Writing—original draft: Z, Y, G, K Writing—revised draft: Z,Y Software: Z, Y Supervision/Overall project guidance: Y, R.R

Funding

There is no funding supporting.

Data availability

The raw data supporting the conclusions of this article will be made available by the authors on request.

Declarations

Ethics approval and consent to participate

The study protocol was reviewed and approved by the Ethical Committee of SEGi Research Ethics Committee, with approval number SEGiEC/SR/GSB/56/2025–202. Participants provided written informed consent. The study adhered to ethical standards as defined in the 1975 Declaration of Helsinki and its subsequent amendments.

Competing interests

The authors declare no competing interests.