Trust in the Automatic Train Supervision System and Its Effects on Subway Dispatchers’ Operational Behavior

Abstract

The evolution of Fully Automatic Operation (FAO) subway control systems has increased the importance of dispatchers’ trust in automation. This study focuses on the dynamics of trust formation among Chinese subway dispatchers and its downstream effects on their reliance behavior. Through the structural equation modeling of questionnaire data, we identify key factors influencing trust and its operational implications. Interface transparency and system capability redundancy directly strengthen trust, while dispatchers’ system knowledge indirectly shapes trust through interface interpretation. Trust strongly correlates with reliance behavior during routine monitoring, which in turn significantly influences reliance behavior in emergencies. Two mediation pathways approaching significance suggest that both capability redundancy and trust may influence emergency reliance behavior through routine reliance behavior. This implies that frequent daily use of automation systems enhances dispatchers’ understanding of system reliability. Our findings underscore the need for transparent interface design, adaptive redundancy strategies, and training programs that bridge system knowledge with operational practices to enhance both trust and scenario-specific reliance. Practical recommendations include designing user-friendly interfaces, implementing adaptive redundancy, and developing targeted training programs to improve dispatchers’ system knowledge and operational efficiency.

1. Introduction

The emergence of autonomous systems, such as driverless cars [1,2] and Fully Automatic Operation (FAO) subway systems [3], has significantly enhanced the operational efficiency of transportation systems while reducing the physical burden on human operators. Automation technology not only improves the reliability of operating systems but also introduces profound changes in the human–machine dynamic. By monitoring and supporting human operations, automation systems can reduce risks. However, in large, complex socio-technical systems, human professionals remain the critical backup, serving as the last line of defense in the event of system failure [4]. This reliance on humans is particularly important because automation inherently places operators in scenarios of uncertainty and incomplete knowledge, making it a pressing challenge to ensure their effectiveness across varying levels of automation [5].

Despite its benefits, automation does not always simplify or enhance work efficiency. Research on human–automation interaction reveals that it can lead to significant issues, such as the suppression of operators’ ability to detect critical signals or alarms [6] and the degradation of operational skills [7]. These problems may leave personnel ill-prepared to protect the system during automation failures. Moreover, increased automation can introduce new failure modes and alter personnel workloads [8], further complicating the human role in automated systems.

In rail transit systems, for example, dispatchers are primarily responsible for ensuring that trains adhere to their schedules [9]. While advances in signal systems have enabled the automatic correction of minor timing deviations, dispatchers must still maintain a high level of vigilance to prevent catastrophic accidents. A notable case is the head-on collision of two trains in Bad Aibling, Bavaria, in February 2016, which was attributed to a dispatcher’s error and an over-reliance on the safety redundancy of the signaling system [10]. This incident underscores the risks of excessive trust in automation and highlights the need for a balanced approach to human–automation interaction.

The challenges posed by automation extend beyond operational failures. Both excessively high and low workloads can introduce new cognitive and attentional demands [11], necessitating changes in operator training. Furthermore, since automation systems are not infallible, humans remain a crucial component in ensuring system safety. However, humans are not immune to the influence of automation systems. The degree of human reliance on automation in different contexts has thus become a topic of significant interest [12].

With the rapid advancement of rail transportation technology, Automatic Train Operation (ATO) systems have been widely adopted in many countries [13]. Automatic Train Supervision (ATS), a foundational component of ATO and its next-generation evolution, the FAO system, provides essential monitoring, control, and decision-making capabilities that underpin the functionality and reliability of both ATO and FAO systems. Currently, ATS is primarily used for routine monitoring and emergency response by dispatchers. In the future, within FAO systems, ATS will remain the primary channel through which dispatchers obtain information [14]. The level of trust that dispatchers place in the ATS system will significantly influence the performance of FAO systems.

Given the critical role of ATS in rail automation, it is essential to investigate the factors that shape dispatchers’ trust in ATS and how this trust influences their reliance behaviors. Understanding these dynamics can contribute to refining the theory of automation trust in safety-critical systems and optimizing the performance of FAO systems. Therefore, this study aims to explore the influencing factors of ATS automation trust and its impact on dispatchers’ reliance behaviors, providing both theoretical insights and practical implications for the design and operation of future rail automation systems.

2. Theoretical Background, Related Work, and Hypotheses

2.1. Foundations of Automation Trust in ATS Systems

Trust is a fundamental concept across numerous domains, including social psychology, where the study of interpersonal trust provides valuable insights for defining automation trust [15]. Empirical research has revealed significant similarities between automation trust and interpersonal trust [16]. For instance, human reactions to technology often mirror reactions to other humans [17], and the dynamics of trust in complex environments show strong parallels between interpersonal and automation contexts [18]. Several studies have identified the core features of automation trust within human–automation partnerships [19,20,21]. First, trust requires a delegator (e.g., the operator) to grant trust, a trustee (e.g., the automation system) to receive trust, and a task with potential risks [22]. Second, the trustee must be motivated to perform the task, typically driven by the designer’s intended use of the system. Finally, the possibility of task failure introduces uncertainty and risk. In this study, we adopt the widely accepted definition of automation trust by Lee and See as follows: “the attitude that an agent will help achieve individual goals in situations characterized by uncertainty and vulnerability” [23]. Empirical research further highlights that trust in automation is influenced by the following three main dimensions: the automation system, the operator, and the environment [24].

The human–machine interface serves as the critical link between operators and automation systems. Transparency, a key aspect of interface design, determines the operator’s ability to comprehend, perceive, and anticipate the automation system’s behavior [25]. Higher transparency enables operators to maintain better performance and trust, especially when the automation system encounters abnormalities [26]. Empirical studies in human–machine collaboration have consistently shown that interface transparency directly influences operators’ trust in automation agents [27]. In the context of automated decision making, providing transparency about the logic behind system recommendations has been found to improve the accuracy of automation use without increasing decision time or subjective workload [28]. However, excessive transparency, such as detailed visualizations of decision outcomes, may lead to faster but less accurate decisions [29]. For ATS systems, interface transparency reflects the extent to which dispatchers can review historical records, understand current system status, and predict future trends [30,31,32]. As the primary tool for decision making and information access, the perceived transparency of the ATS system over time is likely to shape dispatchers’ trust in automation, forming the basis of Hypothesis 1 in this study.

Automation trust is inherently human-centered, with operator factors playing a pivotal role in its formation [33]. Research has consistently shown that operators’ familiarity with the capabilities and limitations of their systems significantly influences their trust levels [34,35,36]. This finding has been replicated in safety-critical domains, such as spacecraft rendezvous and docking missions [37]. Dispatchers, whose roles demand high operational skills, system knowledge, and experience [38], are particularly affected by their understanding of automation systems. For example, a study using observational methods found that railway signalers’ trust in automation was heavily influenced by their comprehension of the system [39]. In summary, a holistic understanding of the ATS system is essential for building trust in it (Hypothesis 2a of this study). Additionally, a transparent interface serves as a crucial bridge between dispatchers’ system knowledge and the current system status (Hypothesis 2b of this study).

In safety-critical fields such as subway dispatching, environmental risks are mitigated primarily through system safety redundancies [40]. Operators’ awareness of the system’s capabilities and safety redundancies plays a crucial role in building trust. For instance, a study on seafloor visual inspection found that explicitly informing operators of the system’s false alarm rate increased their trust in automation [41]. Similarly, in a virtual decision-making experiment, the reliability of automated agents strongly influenced trust, particularly under conditions involving personal safety and economic costs [22]. Long et al. further demonstrated that system dependability is the most reliable predictor of initial trust [42]. However, operators often underestimate system reliability [43]. While subway operation systems maintain high reliability through redundant automation designs, this very redundancy can also disrupt trust in automation, forming the basis of Hypothesis 3 in this study.

2.2. Trust-Driven Reliance Behavior in Operational Contexts

Automation reliance, the primary behavioral outcome of automation trust, refers to the level of dependence on automation, and it is measured through users’ interactions with automated systems [44]. The relationship between trust and reliance behavior is complex, as trust and its influencing factors can have varied impacts on reliance [45]. Research has shown that reliance patterns vary significantly depending on the type of automation error. For instance, a high frequency of false alarms results in users that under-rely on automation during alarm states while over-relying on it during non-alarm states. Conversely, frequent automation failures result in over-reliance on alarms and under-reliance on automation in non-alarm states [46]. This phenomenon is further supported by a simulated seafloor sonar detection study, where participants’ trust and reliance decreased as the false alarm rate increased [41]. A similar correlation was observed in the context of anti-phishing tools, where user-calibrated trust significantly and positively influenced both current and future reliance on detection systems [47]. Given the complexity of scheduling tasks [48], the effect of trust on reliance behavior may differ across operational contexts. This study proposes that (4a) dispatchers’ trust in the ATS system affects their reliance behavior under routine monitoring, (4b) their trust in the ATS system affects their reliance behavior during emergency response, and (4c) their reliance behavior under routine monitoring further influences their reliance behavior during emergency response.

While trust is a critical factor in automation reliance, Lee and See emphasize that it is not the sole determinant; reliance is also influenced by a combination of other factors [23]. For example, even when operators distrust an automation system, they may still rely on it under conditions of high workload and time pressure to complete tasks. Conversely, in situations where trust is high, operators might reduce their reliance on automation and opt for manual control due to concerns about economic costs or safety risks [49]. In such cases, trust acts as an intermediate variable influencing behavior. Additionally, a disconnect can occur between subjective trust evaluations and objective reliance behaviors. Operators may perceive their trust levels as unchanged, while their actual reliance behaviors shift significantly [50]. For subway dispatchers who rely on the ATS system as their primary information channel, reliance on the system is further influenced by its capability redundancy status. This study suggests that (5a) the ATS’s capability redundancy affects dispatchers’ reliance behavior under routine monitoring, and (5b) the ATS’s capability redundancy also affects their reliance behavior during emergency response.

3. Research Methodology and Empirical Analysis

3.1. Summary of Core Constructs and Hypotheses

By summarizing the established literature and sorting out the current state of practical work, nine research hypotheses are proposed as follows, and these can be summarized as shown in Figure 1.

Hypothesis 1.

The Transparency of the ATS Interface (TAI) affects the Dispatcher’s Trust in ATS system (DTA) (H1: TAI ⇒ DTA).

Hypothesis 2a.

The dispatcher’s ATS System Knowledge (ASK) directly affects the Dispatcher’s Trust in ATS system (DTA) (H2a: ASK ⇒ DTA).

Hypothesis 2b.

The dispatcher’s ATS System Knowledge (ASK) affects their understanding of the Transparency of the ATS Interface (TAI) (H2b: ASK ⇒ TAI).

Hypothesis 3.

The ATS’s Capability Redundancy (ACR) directly affects the Dispatcher’s Trust in ATS system (DTA) (H3: ACR ⇒ DTA).

Hypothesis 4a.

The Dispatchers’ Trust in the ATS system (DTA) affects their Reliance Behavior on ATS under Routine monitoring (RBR) (H4a: DTA ⇒ RBR).

Hypothesis 4b.

The Dispatchers’ Trust in the ATS system (DTA) affects their Reliance Behavior on ATS under Emergency response (RBE) (H4b: DTA ⇒ RBE).

Hypothesis 4c.

The dispatchers’ Reliance Behavior on ATS under Routine monitoring (RBN) affects their Reliance Behavior on ATS under Emergency response (RBE) (H4c: DTR ⇒ RBE).

Hypothesis 5a.

The ATS’s Capability Redundancy (ACR) affects dispatchers’ Reliance Behavior on ATS under Routine monitoring (RBR) (H5a: ACR ⇒ RBR).

Hypothesis 5b.

The ATS’s Capability Redundancy (ACR) affects dispatchers’ Reliance Behavior on ATS under Emergency response (RBE) (H5b: ACR ⇒ RBE).

Figure 1 presents the hypothesized model, illustrating the influence paths among the six key constructs in this study. To further clarify the theoretical foundation of these constructs,

Table 1. Conceptual definitions and sources of research constructs.

Construct	Factor Code	Conceptual Definitions and Sources
Dispatcher’s Trust in the ATS System	DTA	How much the dispatcher subjectively believes in the ATS system [39,51].
ATS’s Capability Redundancy	ACR	Redundancy in the ATS system’s capability to ensure safety that is perceptible to dispatchers [31,52,53].
Dispatcher’s ATS System Knowledge	ASK	The dispatcher’s familiarity with the ATS system, associated hardware- and software-related information, and operational dispatching rules and regulations [3,54].
Transparency of the ATS Interface	TAI	The degree to which the user interface allows the dispatcher to review past history records, understand the current system status, and assist in predicting future trends [30,31,32].
Reliance Behavior on ATS under Routine Monitoring	RBR	The extent to which the dispatcher relies on and benefits from the ATS system during routine monitoring conditions [55,56].
Reliance Behavior on ATS under Emergency Response	RBE	The extent to which the dispatcher relies on and benefits from the ATS system during emergency response conditions [55,56].

provides their detailed definitions, factor codes, and supporting sources from the literature.

Table 2. Hypotheses and variable relationships.

Hypothesis Code	H1	H2a	H2b	H3	H4a	H4b	H4c	H5a	H5b
Predictor Code	TAI	ASK	ASK	ACR	DTA	DTA	RBR	ACR	ACR
Dependent Code	DTA	DTA	TAI	DTA	RBR	RBE	RBE	RBR	RBE

summarizes the hypothesized relationships between predictors and dependent variables.

3.2. Research Design: Pre-Tests and Final Questionnaire Development

All methods adhered to the ethical principles and code of conduct of the American Psychological Association (2017), a widely accepted standard for questionnaire and human factors research. The experimental protocols were approved by the State Key Laboratory of Advanced Rail Autonomous Operation at Beijing Jiaotong University, which was upgraded from the former State Key Lab of Rail Traffic Control and Safety. The laboratory maintains a comprehensive management organization and ethics committee. All participants provided informed consent before the study and were free to withdraw at any time, with the option to delete their data. Data security was ensured throughout the research process.

To explore the conceptual framework of the hypotheses and their influence paths shown in Figure 1 we initially developed 33 items based on the constructs outlined in Table 1. These items were evaluated by 20 manager-level dispatchers from the Beijing subway, who assisted in refining the wording and ensuring the clarity, validity, and non-redundancy of the questions. This process ensured that each question was simple, unambiguous, and distinct from others.

To verify the reliability and validity of the scale, a preliminary version of the questionnaire was distributed to 79 senior dispatchers (mean length of service: 8.6 years, SD = 3.6, min = 5, max = 38) who had participated in the China Dispatcher Skills Competition. After the reliability testing, cluster analysis, and post-response interviews, the questionnaire was further streamlined. The final version of the questionnaire is provided in Appendix A. The six latent variables and their corresponding questionnaire items (observed variables) are detailed in the left section of

Table A1. Correspondence between questionnaire items and research dimensions.

No.	Research Dimension (Automation Trust in ATS)	No.	Research Dimension (Task Complexity)
1	ASK.Q2	21	Emergency TC-1-Size
2	TAI.Q1 (Scale Reversed)	22	Emergency TC-2-Variety
3	DTA.Q2	23	Emergency TC-3-Ambiguity
4	RBR.Q2 (Scale Reversed)	24	Emergency TC-4-Relationship
5	ACR.Q2	25	Emergency TC-5-Variability
6	RBE.Q2	26	Emergency TC-6-Unreliability
7	RBR.Q1	27	Emergency TC-7-Novelty
8	TAI.Q3	28	Emergency TC-8-Incongruity
9	RBE.Q3 (Scale Reversed)	29	Emergency TC-9-Action Complexity
10	TAI.Q2	30	Emergency TC-10-Temporal Demand
11	ACR.Q3 (Scale Reversed)	31	Routine Monitoring TC-1-Size
12	ASK.Q1	32	Routine Monitoring TC-2-Variety
13	DTA.Q3 (Scale Reversed)	33	Routine Monitoring TC-3-Ambiguity
14	RBR.Q3	34	Routine Monitoring TC-4-Relationship
15	RBE.Q1 (Scale Reversed)	35	Routine Monitoring TC-5-Variability
16	ASK.Q3 (Scale Reversed)	36	Routine Monitoring TC-6-Unreliability
17	DTA.Q1 (Scale Reversed)	37	Routine Monitoring TC-7-Novelty
18	ACR.Q1 (Scale Reversed)	38	Routine Monitoring TC-8-Incongruity
19	Open-Ended Questions on Enhancing Trust in Automation	39	Routine Monitoring TC-9-Action Complexity
20	Open-Ended Questions on Reducing Trust in Automation	40	Routine Monitoring TC-10-Temporal Demand

Notes: TAI, Transparency of the ATS Interface; ASK, dispatcher’s ATS System Knowledge; ACR, ATS’s Capability Redundancy; DTA, Dispatcher’s Trust in the ATS system; RBR, Reliance Behavior on ATS under Routine monitoring; RBE, Reliance Behavior on ATS under Emergency response; TC, Task Complexity.

in Appendix B (items 1–18). The first half of the questionnaire comprised 18 questions assessed using a 7-point Likert scale, with at least one reverse question per latent variable to evaluate response quality.

The second part of the questionnaire focused on task complexity under two working conditions as follows: routine monitoring and emergency response. Based on the literature and the dispatchers’ actual work experience [57], task complexity was measured across 10 dimensions [58]. Pre-experiment interviews confirmed significant variability in task complexity between these conditions. The second part of the questionnaire measures task complexity across 10 dimensions for both routine monitoring and emergency response. The specific items for task complexity are included in Appendix A (items 21–40), with their alignment to the Task Complexity dimensions shown in the right section of Table A1 in Appendix B.

3.3. Data Collection and Sample Screening

During the formal questionnaire phase, 214 senior front-line dispatchers (average length of service = 13.3 years, SD = 6.8, min = 5, max = 25) from Beijing, Tianjin, Guangzhou, Shijiazhuang, and Suzhou participated. To ensure data quality, participants completed the questionnaire in a quiet, distraction-free meeting room under the supervision of duty managers or higher-level supervisors. The demographic characteristics of the survey participants, including age, length of service, and educational background, are summarized in

Table 3. Survey participant demographics.

Age		Years of Service		Years of Experience in Dispatching		Educational Background
24–33:	99 (46.3%)	1–10:	84 (39.3%)	1–9:	150 (70.1%)	Bachelor’s Degree:	189 (88.3%)
34–43:	102 (47.7%)	11–20:	100 (46.7%)	10–19:	57 (26.6%)	Associate Degree:	17 (7.9%)
44–53:	11 (5.1%)	21–30:	27 (12.6%)	20–29:	4 (1.9%)	Graduate Degree:	8 (3.7%)
54–63:	2 (0.9%)	31–40:	3 (1.4%)	30–39:	3 (1.4%)

.

The formal questionnaire tested the initial research hypotheses, with questions presented in random order. Participants rated their agreement with descriptive statements on a 7-point Likert scale. Questionnaire items were fine-tuned to account for regional differences in terminology.

Questions 2, 4, 9, 11, 13, 15, 16, 17, and 18 (see Appendix B, Table A1) were reverse-scored. Samples were excluded if more than two dimensions of automation trust exhibited conflicting reverse scale scores or if response times were excessively long (indicating interruptions) or short (indicating careless responses). After screening, a final valid sample of 160 was obtained, with an average response time of 1408 s (SD = 529).

3.4. Structural Equation Modeling and Hypothesis Validation

Exploratory factor analysis (EFA) was conducted on the final valid data to assess construct validity. Using JAMOVI 2.3.21, principal components extraction with varimax rotation was performed. Components with eigenvalues greater than 1.0 were retained based on the scree test and Kaiser’s criterion. Item loadings exceeded the minimum threshold of 0.55, a level deemed suitable for principal component analysis [59]. Reliability was evaluated using Cronbach’s α.

The EFA extracted the following six factors: three trust impact factors, two reliance behavior outcomes, and one trust factor comprising marker variables. The initial scale data for SEM analysis excluded Automation Capability Redundancy Q1 due to its low factor loading (0.353). All other items exhibited ideal factor loadings, as shown in

Table 4. Final factor loadings, Cronbach’s α, and explained variance after excluding low-loading items (with pre-exclusion values for excluded items).

	Descriptives		Factor Loadings
Item	Mean	SD	F1	F2	F3	F4	F5	F6
ACR Q1 (excluded due to low factor loading)	3.701	1.758						0.353
ACR Q2	4.206	1.530						0.870
ACR Q3	3.775	1.890						0.853
ASK Q1	4.713	1.334					0.769
ASK Q2	5.112	1.534					0.863
ASK Q3	5.037	1.378					0.890
TAI Q1	3.938	1.482		0.785
TAI Q2	4.169	1.727		0.932
TAI Q3	3.681	1.531		0.851
DTA Q1	3.619	1.693			0.844
DTA Q2	3.737	1.523			0.897
DTA Q3	3.337	1.533			0.820
RBR Q1	4.875	1.444	0.868
RBR Q2	5.088	1.490	0.946
RBR Q3	4.938	1.512	0.842
RBE Q1	5.675	1.315				0.932
RBE Q2	5.094	1.268				0.805
RBE Q3	5.506	1.274				0.834
Explained Variance (%)			14.14	13.62	13.32	13.26	13.17	9.42
Cumulative Explained Variance (%)			14.14	27.76	41.08	54.34	67.51	76.93
Cronbach’s α			0.918	0.910	0.893	0.898	0.891	0.860

. The Kaiser–Meyer-Olkin Test for sampling adequacy was 0.80, Bartlett’s test of sphericity was significant (p < 0.001), and the six factors accounted for 76.9% of the total variance, supporting the continuation of structural equation modeling analysis.

The measurement model fitted the data well, as shown in

Table 5. Confirmatory factors of the SEM.

				95% Confidence Intervals
Latent	Observed	Estimate	SE	Lower	Upper	β	z	p	AVE
ACR	ACR.Q2	1	0	1	1	0.950			0.753
	ACR.Q3	1.052	0.122	0.813	1.292	0.809	8.612	<0.001
ASK	ASK.Q1	1	0	1	1	0.798			0.735
	ASK.Q2	1.209	0.102	1.009	1.408	0.838	11.891	<0.001
	ASK.Q3	1.207	0.093	1.025	1.389	0.932	13.015	<0.001
TAI	TAI.Q1	1	0	1	1	0.869			0.779
	TAI.Q2	1.204	0.081	1.046	1.362	0.898	14.938	<0.001
	TAI.Q3	1.043	0.072	0.902	1.184	0.877	14.482	<0.001
DTA	DTA.Q1	1	0	1	1	0.858			0.732
	DTA.Q2	0.927	0.070	0.789	1.064	0.884	13.229	<0.001
	DTA.Q3	0.870	0.071	0.732	1.009	0.824	12.336	<0.001
RBR	RBR.Q1	1	0	1	1	0.859			0.79
	RBR.Q2	1.125	0.071	0.985	1.265	0.938	15.769	<0.001
	RBR.Q3	1.056	0.074	0.911	1.200	0.866	14.339	<0.001
RBE	RBE.Q1	1	0	1	1	0.913			0.751
	RBE.Q2	0.868	0.065	0.740	0.996	0.822	13.272	<0.001
	RBE.Q3	0.911	0.065	0.784	1.037	0.859	14.107	<0.001

. The results of the measurement model and the reliability of each construct are presented in Table 5. All factoring loading values were significantly greater than 0.55, and all AVE were over 0.50, indicating that convergent validity was adequately achieved [60].

Structural equation modeling (SEM) was conducted to examine the influence of factors on trust and its outcomes, thereby testing the hypotheses. The analysis was performed using JAMOVI 2.3.21 with the semlj 1.1.6 module. The initial hypothetical structural model revealed two insignificant or weak path coefficients as follows: between automation trust and reliance behavior in emergencies (β = −0.049, p = 0.583, Hypothesis 4b) and between personal knowledge and automation trust (β = 0.090, p = 0.431, Hypothesis 2a), as shown in

Table 6. Initial hypothetical model: standardized path coefficients between latent variables and their significance levels.

				95% Confidence Intervals					Hypothesis Code	Validation Status
Dep	Pred	Estimate	SE	Lower	Upper	β	z	p
DTA	TAI	0.204	0.115	−0.021	0.429	0.182	1.778	0.075	H1
DTA	ASK	0.122	0.155	−0.181	0.425	0.090	0.788	0.431	H2a	Rejected
TAI	ASK	0.662	0.102	0.461	0.862	0.548	6.463	<0.001	H2b
DTA	ACR	0.206	0.098	0.015	0.398	0.208	2.113	0.035	H3
RBR	DTA	0.172	0.076	0.024	0.321	0.201	2.269	0.023	H4a
RBE	DTA	−0.040	0.074	−0.185	0.104	−0.049	−0.549	0.583	H4b	Rejected
RBE	RBR	0.277	0.084	0.112	0.442	0.286	3.290	0.001	H4c
RBR	ACR	0.168	0.076	0.020	0.316	0.197	2.228	0.026	H5a
RBE	ACR	0.187	0.074	0.042	0.332	0.227	2.522	0.012	H5b

. Contrary to the prior hypotheses, personal knowledge did not significantly affect trust, leading to the rejection of Hypothesis 2a as follows: the dispatcher’s ATS system knowledge directly affects the dispatcher’s trust in the ATS system. Similarly, Hypothesis 4b—dispatchers’ trust in the ATS system affects their reliance behavior on ATS under emergency response—was also rejected.

These insignificant parameters were subsequently removed from the model. The goodness-of-fit (GOF) of the measurement and structural models was evaluated using recommended indices for sample sizes under 250 and observed variables between 12 and 30 [60]. The fine-tuned model demonstrated good overall fit, as follows: χ²(111, N = 160) = 220 (p < 0.001), RMSEA = 0.078, TLI = 0.930, CFI = 0.943, and NFI = 0.893. The standardized path estimates for the model are presented in Figure 2, with detailed coefficients between latent variables provided in

Table 7. Fine-tuned model: standardized path coefficients between latent variables and their significance levels.

				95% Confidence Intervals					Hypothesis Code	Validation Status
Dep	Pred	Estimate	SE	Lower	Upper	β	z	p
DTA	TAI	0.251	0.097	0.060	0.442	0.225	2.575	0.010	H1	Supported
TAI	ASK	0.665	0.103	0.464	0.866	0.549	6.482	<0.001	H2b	Supported
DTA	ACR	0.224	0.088	0.053	0.396	0.227	2.560	0.010	H3	Supported
RBR	DTA	0.170	0.076	0.021	0.319	0.197	2.238	0.025	H4a	Supported
RBE	RBR	0.267	0.082	0.106	0.429	0.276	3.245	0.001	H4c	Supported
RBR	ACR	0.169	0.075	0.022	0.317	0.199	2.253	0.024	H5a	Supported
RBE	ACR	0.174	0.071	0.035	0.314	0.211	2.451	0.014	H5b	Supported

.

The structural equation modeling analysis supported seven hypotheses, as illustrated in Figure 2 and detailed in Table 7. Key findings include the significant influence of interface transparency on dispatchers’ trust in the ATS system (Hypothesis 1) and the role of dispatchers’ knowledge mastery in shaping their understanding of interface transparency (Hypothesis 2b). Additionally, the capability redundancy of the ATS system was found to directly enhance dispatchers’ trust (Hypothesis 3). Trust, in turn, primarily influenced reliance behavior during routine monitoring operations (Hypothesis 4a), which further affected reliance behavior during emergency responses (Hypothesis 4c). Finally, the capability redundancy of the ATS system was confirmed to impact reliance behavior in both routine monitoring and emergency scenarios (Hypotheses 5a,b). These results highlight the interconnected roles of interface design, system knowledge, and capability redundancy in fostering trust and shaping reliance behavior in automation systems.

4. Discussion

4.1. Key Factors Shaping Dispatchers’ Trust in ATS Systems

The transparency of the ATS interface significantly enhances dispatchers’ trust in the system, supporting hypothesis 1. A transparent interface provides operators with clear and accessible information, which is critical for building trust, especially under high time pressure [29]. This aligns with traditional research showing that transparent interfaces improve operator trust and subjective acceptance of automation systems, provided operators are proficient in their use [61]. Additionally, the capability redundancy of the ATS system directly and positively affects dispatchers’ trust, supporting Hypothesis 3. This finding is consistent with studies highlighting the importance of system reliability redundancy in fostering automation trust [43,62]. Furthermore, dispatchers’ knowledge of the ATS system significantly influences their perception of interface transparency, supporting Hypothesis 2b. This suggests that a deeper understanding of the system enhances operators’ ability to interpret and utilize interface information effectively [27].

However, Hypothesis 2a, which posited that dispatchers’ knowledge directly affects their trust in the ATS system, was not supported. This contrasts with Balfe’s study, which identified understanding of automation as the strongest dimension for constructing trust [39]. The discrepancy may stem from differences in the methodology, as follows: while Balfe’s study used observational methods with a small sample, this study examined the impact of long-term accumulated knowledge on trust in a larger sample. Additionally, subway dispatchers’ work is highly complex and information-intensive, which may lead to a disconnect between knowledge and trust [63]. This phenomenon is also observed in other domains, such as self-driving vehicles, where operators may express difficulty in fully grasping automation-related knowledge but may still trust and rely on the system [64,65].

4.2. Impact of Trust on Dispatchers’ Behavioral Responses

Trust in the ATS system significantly influences dispatchers’ reliance behavior during routine monitoring, supporting Hypothesis 4a. This aligns with studies showing that operators’ trust in automation systems increases their reliance on the system, particularly when the interface effectively displays system status [41]. Furthermore, reliance behavior during routine monitoring positively affects reliance behavior during emergency responses, supporting Hypothesis 4c. This suggests that daily interactions with the ATS system shape operators’ confidence and behavior in high-stakes scenarios [57]. Additionally, the capability redundancy of the ATS system significantly impacts reliance behavior in both routine and emergency scenarios, supporting Hypotheses 5a,b. This finding corroborates traditional research emphasizing the role of system reliability in shaping operator dependence [66,67].

However, Hypothesis 4b, which proposed that trust directly affects reliance behavior during emergency responses, was not supported. This may be because, in emergency scenarios, dispatchers prioritize situational awareness and immediate action over their trust in the system. Interestingly, reliance behavior during emergencies is indirectly influenced by trust and capability redundancy through routine monitoring, as evidenced by the marginally significant mediation effects of ACR ⇒ RBR ⇒ RBE (p = 0.059) and DTA ⇒ RBR ⇒ RBE (p = 0.064). This indicates that daily reliance behavior serves as a bridge between trust, capability redundancy, and emergency responses. The lack of a direct effect of trust on emergency reliance may also reflect the unique demands of subway dispatching, where operators must balance trust in automation with their own expertise and situational judgment [68,69].

4.3. Task Complexity in Emergency vs. Routine Monitoring Situations: A Comparative Discussion

A non-parametric repeated measures ANOVA (Friedman test) was conducted on task complexity data across conditions in Part 2 of the study. The results revealed that all dimensions of contingency tasks were significantly more challenging than those in routine conditions, except for the third dimension, ambiguity, as shown in

Table 8. Variability in task complexity under different operating conditions.

		Descriptive Statistics						Significance Statistics
		Emergency Condition			Normal Condition			Repeated Measures ANOVA (Non-Parametric)
Task Complexity Factors	N	Mean	Std. Error Mean	Standard Deviation	Mean	Std. Error Mean	Standard Deviation	$\mathbf{\text{Friedman}} {\mathit{\chi }}^2$	p	Mean Difference (I-J)
1—Size	160	5.319	0.112	1.416	4.856	0.127	1.609	11.636	<0.001	0.463
2—Variety	160	5.650	0.099	1.255	5.069	0.114	1.446	26.042	<0.001	0.581
3—Ambiguity	160	2.594	0.118	1.493	2.650	0.118	1.489	0.821	0.365	−0.056
4—Relationship	160	5.369	0.104	1.316	4.669	0.131	1.651	15.042	<0.001	0.700
5—Variability	160	5.963	0.076	0.964	4.200	0.148	1.866	70.738	<0.001	1.763
6—Unreliability	160	3.931	0.138	1.742	3.288	0.137	1.732	15.696	<0.001	0.644
7—Novelty	160	3.856	0.131	1.663	3.331	0.134	1.692	19.463	<0.001	0.525
8—Incongruity	160	4.819	0.119	1.500	4.231	0.139	1.753	7.353	0.007	0.588
9—Action Complexity	160	4.894	0.111	1.399	3.881	0.139	1.757	37.586	<0.001	1.013
10—Temporal Demand	160	5.756	0.098	1.238	4.094	0.145	1.832	67.930	<0.001	1.663

. This suggests that while most aspects of emergency response tasks are inherently more complex, the level of ambiguity remains consistent across both routine and emergency scenarios. This finding aligns with the nature of subway dispatching, where ambiguity in information interpretation is a constant challenge, regardless of the operational context.

The interviews further highlighted that emergency response tasks require greater flexibility and adaptability from dispatchers. The significant differences in task complexity between routine monitoring and emergency response conditions lead to substantial changes in task content and workflow [48]. In emergency situations, dispatchers tend to rely on multiple channels for the repeated confirmation of the same information, as depending solely on the ATS system is often impractical [70]. This multi-channel approach reflects the heightened need for accuracy and reliability during high-stake scenarios. Conversely, in routine monitoring environments, the security and completeness of the operation system are relatively robust, allowing dispatchers to place stronger reliance on the ATS system based on their established trust in its performance. However, the trust developed under routine conditions may not seamlessly translate to emergency response situations, where operational demands and task dynamics undergo drastic changes. This discrepancy underscores the importance of designing automation systems that can adapt to varying levels of task complexity and operational contexts.

4.4. Future Directions and Practical Implications

To address automation trust calibration, this study recommends designing transparent ATS interfaces that provide clear, real-time feedback, enabling dispatchers to maintain an appropriate level of trust. Integrating targeted training programs to enhance dispatchers’ system knowledge can further optimize trust calibration. For instance, our research team is currently developing a next-generation ATS human–machine interface (see Figure 3) that features transparent information channels, prototype completion, and ongoing interaction design and human factors validation. This interface aims to enhance dispatchers’ understanding of system operations and improve trust calibration. Additionally, we are designing an adaptive training system (see Figure 4) that incorporates adaptive learning modules and usability testing, ensuring dispatchers can effectively build and maintain system proficiency. Future research should explore the interaction between system transparency, feedback mechanisms, and trust calibration, aiming to identify the optimal level of transparency that balances trust with cognitive workload. Advancements in artificial intelligence and deep learning also offer promising avenues for automating trust assessment and calibration. While this study focused on Chinese subway dispatchers, extending the research to other regions and operational contexts will help establish a more comprehensive understanding of automation trust in safety-critical systems.

4.5. Limitations and Generalizability

This study has several limitations that should be acknowledged. First, the heterogeneity of ATS systems—ranging from differences in system logic and interface design to terminology and operational rules across cities—constrained the selection of subjects and limited the sample size. These variations, while reflective of real-world conditions, may affect the generalizability of the findings. Second, due to sample size constraints, the study focused on a limited set of factors derived from pre-experiment interviews and questionnaires. Future research with larger and more diverse samples could explore a three-dimensional automation trust model encompassing system factors (e.g., failure history and safety redundancy), dispatcher factors (e.g., individual tendencies and skill levels), and environmental factors (e.g., task complexity and multitasking demands). Finally, the study’s focus on Chinese subway dispatchers raises questions about cross-cultural applicability, necessitating further validation in other regions and operational contexts.

5. Conclusions

This study investigates the factors influencing the formation of automation trust in ATS systems among subway dispatchers and its behavioral consequences. Through structural equation modeling, we validated the direct positive impact of interface transparency on automation trust, highlighting its critical role in fostering dispatchers’ confidence in the system. While dispatchers’ knowledge of the operational system did not directly affect their trust in the ATS, it significantly enhanced their understanding of the interface information, underscoring the importance of system familiarity. Furthermore, the capability redundancy of the automation system not only directly influenced dispatchers’ trust but also shaped their reliance behavior across different operational conditions. Notably, reliance behavior during emergency responses was more closely tied to routine behavior and system safety redundancy than to automation trust itself. This divergence likely stems from the inherent differences in task complexity and operational demands between routine monitoring and emergency scenarios. These findings provide valuable insights for designing transparent interfaces, optimizing system redundancy, and developing targeted training programs to enhance automation trust and operational efficiency in safety-critical systems.