Empowerment or Depletion? Unpacking the Asymmetrical Pathways from Perceived Autonomy to Human–AI Trust

Abstract

As intelligent systems become decision-support tools in the architecture, engineering, and construction (AEC) industry, establishing human–AI trust is critical. However, in engineering consulting, the psychological mechanisms underlying trust formation remain unclear. Grounded in Self-Determination Theory and the Stereotype Content Model, this study utilized multi-wave survey data from Chinese engineering consulting employees to investigate these mechanisms. We examined how perceived autonomy influences human–AI trust through the competitive dual-mediation of warmth perception and competence perception, alongside the asymmetric moderating role of critical thinking. Results reveal that perceived autonomy directly enhances trust. However, social cognition acts as a competitive mechanism: autonomy positively impacts trust via warmth perception but generates a negative indirect effect via competence perception. Furthermore, critical thinking exerts an asymmetric boundary effect; it does not interfere with the intuitive warmth pathway but significantly intensifies the negative indirect effect through the competence pathway. Ultimately, these findings highlight that perceived autonomy exerts a double-edged sword effect in the context of human–AI collaboration. To mitigate professional defensive rejection and calibrate trust, AEC firms should prioritize human-in-the-loop deployment strategies, objective interface designs, and the cultivation of AI collaborative literacy.

1. Introduction

As digital and intelligent technologies increasingly penetrate the construction industry, intelligent systems are evolving from passive information-processing tools into decision-support entities capable of participating in judgment, recommendation, and coordination [1,2,3,4,5], as evidenced by their growing applications in AI-based defect inspection, building operation optimization, and energy consumption prediction [4,5]. In engineering consulting, intelligent systems have been widely embedded in key business processes such as scheme comparison, risk identification, cost analysis, and collaborative management, thereby reshaping the organization of professional tasks and the logic of decision-making [1,6,7,8]. These technologies are regarded as important enablers of industry development. On the one hand, they help address long-standing challenges in the construction industry, including labor shortages, knowledge gaps, and stagnant productivity growth. On the other hand, they provide an important foundation for the digital transformation of engineering consulting services and the upgrading of organizational capabilities. However, whether intelligent systems can be translated into stable organizational performance does not depend solely on algorithmic performance, data scale, or system functionality. For engineering consulting work that relies heavily on professional judgment and field experience, the realization of technological value largely depends on whether frontline employees are willing to accept, rely on, and continuously use these systems in real, high-risk tasks with clear accountability constraints [9,10].

In the context of human–AI collaboration in the architecture, engineering, and construction (AEC) industry, trust has become a critical psychological condition for the effective implementation of intelligent systems [11,12]. Unlike general office assistance or low-risk information retrieval, engineering consulting is characterized by strong professional judgment, high contextual uncertainty, cross-actor collaboration, and strict accountability [13,14]. When intelligent systems are involved in core activities such as scheme evaluation, cost estimation, and risk warning, employees are not merely assessing whether a tool is useful; rather, they are making complex trade-offs among efficiency gains, decision risks, professional responsibility, and accountability pressure [3,15,16]. Thus, whether employees allow intelligent systems to enter their core decision-making processes essentially reflects their willingness to form positive expectations of and develop reliance on these systems when facing uncertainty and potential risk exposure [17,18,19]. In this sense, Human–AI Trust is not a natural outcome of technology adoption but a prerequisite for intelligent systems to become integrated into professional decision-making.

Although the importance of Human–AI Trust has received increasing attention, existing research offers limited explanations of how such trust is formed in engineering consulting contexts. Traditional technology acceptance models explain adoption intention mainly through perceived usefulness and perceived ease of use [17,20]. This perspective is useful for understanding the functional basis of technology adoption, but it is less able to explain how employees form trust judgments toward intelligent systems in high-risk professional tasks. Meanwhile, although Human–AI Trust research has expanded from a focus on single technological attributes to broader frameworks incorporating human, machine, and environmental factors, meta-analytic evidence suggests that prior research has continued to emphasize machine characteristics such as transparency, reliability, accuracy, and explainability, while paying insufficient attention to users’ deeper psychological needs and cognitive heterogeneity [21]. Therefore, when intelligent systems are no longer merely passive tools but act as advisors, collaborators, or even quasi-decision-making agents in professional tasks, existing research still lacks a sufficient explanation of the internal psychological mechanisms through which employees form Human–AI Trust.

Addressing this gap requires shifting from a purely functional evaluation perspective to a social-cognitive perspective on human–AI collaboration. Engineering consultants’ evaluations of intelligent systems depend not only on whether the systems are accurate, efficient, or easy to use, but also on whether, during interaction, the systems are perceived as supporting professional autonomy, respecting human decision authority, and aligning with human goals. Drawing on self-determination theory [22], when employees feel that they retain task authority and decision control while using intelligent systems—that is, when they experience higher perceived autonomy—intelligent systems are more likely to be viewed as empowering resources rather than threats to professional identity and judgment authority. Conversely, when employees perceive that systems undermine their sense of control or replace their professional judgment, defensive reactions and psychological resistance may arise. Thus, perceived autonomy constitutes an important psychological foundation for the formation of Human–AI Trust.

However, higher perceived autonomy does not necessarily translate directly into higher Human–AI Trust. Its influence may operate through two distinct forms of social-cognitive judgment. According to the stereotype content model [23], individuals evaluate social targets along two fundamental dimensions: warmth and competence. Although this model was originally developed to explain interpersonal and intergroup evaluations, employees in human–AI collaboration may similarly use these social-cognitive dimensions to understand intelligent systems. On the one hand, when employees perceive that a system respects their judgment, supports their work, and preserves their decision authority, the system is more likely to be perceived as benevolent, supportive, and aligned with human goals, thereby generating stronger warmth perception. On the other hand, higher employee autonomy may also produce a more complex competence attribution effect. Specifically, when employees occupy a stronger leading role in collaboration, they may be more inclined to attribute task success to their own professional competence rather than to the system, thereby weakening their evaluation of the system’s competence. Accordingly, perceived autonomy may strengthen the warmth perception of intelligent systems without necessarily enhancing their competence perception. This asymmetric social-cognitive mechanism provides an important yet underexplored theoretical lens for explaining the formation of Human–AI Trust.

Furthermore, the formation of Human–AI Trust depends not only on how employees perceive intelligent systems but also on how they examine and interpret these perceptual cues. Prior research suggests that in human–AI collaboration, overtrust may lead to blind compliance, responsibility shifting, and automation bias, whereas undertrust may result in technology disuse and inefficient collaboration [24,25,26]. Therefore, employees’ ability to evaluate system outputs in a careful and reflective manner is a key boundary condition for calibrated trust [26,27]. Critical thinking, as a cognitive trait emphasizing logical reasoning, evidence evaluation, and reflective judgment [28], helps individuals avoid being guided by superficial cues and enables more prudent assessments of intelligent system recommendations. However, in collaborative contexts where human–AI roles are being reconfigured, the effect of critical thinking is not necessarily unidirectional. On the one hand, individuals with higher critical thinking may be better able to recognize the system’s support for their autonomy, thereby strengthening their perception of the system’s benevolence and collaborative intent. On the other hand, they may also scrutinize the logic, evidence, and boundary conditions of system outputs more strictly, leading to more cautious or even more demanding evaluations of system competence [16,29,30]. Thus, critical thinking may not simply enhance or weaken Human–AI Trust; rather, it may asymmetrically shape the pathways from perceived autonomy to warmth perception and competence perception.

Based on abovementioned, this study focuses on employees in China’s engineering consulting industry and integrates self-determination theory with the stereotype content model to develop and empirically test a moderated dual-mediation model. Specifically, this study examines how employees’ perceived autonomy in collaboration with intelligent systems influences Human–AI Trust through two parallel social-cognitive pathways—warmth perception and competence perception—and further investigates how critical thinking moderates this process. By revealing the asymmetric psychological mechanisms through which perceived autonomy affects Human–AI Trust, this study aims to clarify the psychological process and boundary conditions under which employees form trust in intelligent systems in engineering consulting contexts characterized by high professionalism and strong accountability constraints.

This study makes three main contributions. First, it extends research on Human–AI Trust. Unlike prior studies that explain trust primarily through technological attributes such as system transparency, accuracy, and reliability, this study highlights the roles of employees’ autonomy needs and social-cognitive judgments in trust formation, thereby revealing the psychological mechanisms through which intelligent systems become trusted. Second, this study enriches the application of self-determination theory in human–AI collaboration contexts. By introducing perceived autonomy into the interaction between engineering consultants and intelligent systems, this study shows that autonomy not only affects employees’ acceptance of technology but also shapes their social evaluations of system warmth and competence. Third, this study advances the application of the stereotype content model to evaluations of nonhuman intelligent agents. It shows that employees’ trust in intelligent systems does not arise from competence judgments alone, but is jointly influenced by warmth perception and competence perception. More importantly, these two pathways may be asymmetrically moderated by critical thinking, thereby offering a new explanatory framework for understanding Human–AI Trust calibration in professional service settings.

2. Theoretical Background and Hypotheses Development

2.1. Human–AI Trust

Trust is a psychological state formed in uncertain and vulnerable environments, based on positive expectations of the other party’s behavior [19]. Trust in the human–AI interaction context follows the above definition of trust [17]. In engineering consulting scenarios, such as BIM collaborative platforms, intelligent drawing review systems, and generative AI, a user’s choice to rely on system outputs implies accepting a certain degree of risk exposure. Human–AI trust is precisely the psychological prerequisite for users’ willingness to bear this vulnerability and subsequently accept the system’s participation in decision-making [31]. Existing research indicates that the formation of technology acceptance variables, such as perceived usefulness and perceived ease of use, essentially depends on users’ trust in system reliability [32]. Therefore, trust, as a core antecedent mechanism running through the entire human–AI interaction process, directly determines practitioners’ acceptance intentions and collaborative behaviors toward emerging technologies. Prior explorations of the antecedents of human–AI trust have mostly focused on machine characteristics (e.g., system reliability, transparency) or environmental factors (e.g., task complexity), while the excavation of human (user)-side psychological and social cognitive factors remains insufficient [21]. In light of this, this study introduces users’ perceived autonomy, dual-dimensional social cognition (warmth and competence inferences), and critical thinking. It aims to systematically unpack the internal black box of human–machine trust formation from an integrated perspective of psychological motivation and cognitive traits.

2.2. Self-Determination Theory

Self-Determination Theory (SDT) posits that autonomy is the most core basic psychological need of human beings. It refers to an individual’s subjective experience that their actions stem from internal will and value choices, rather than external coercion or control [22]. When the external environment, such as the intervention mode of an intelligent system, supports an individual’s autonomous choices, it activates their intrinsic motivation, prompting positive psychological responses and evaluations. Conversely, if the situation is coercive or restrictive, it induces psychological reactance and defense [33]. In human–AI collaboration scenarios within engineering consulting, intelligent systems are gradually evolving from passive tools to active decision-makers. This inevitably reconstructs employees’ experiences of decision-making dominance. SDT provides a highly fitting theoretical lens to understand this phenomenon [34]. When a system is perceived as supporting rather than replacing employees’ judgments, their need for autonomy is satisfied, thereby triggering a positive attitude toward the system. Conversely, trust is weakened by a sense of autonomy deprivation. Thus, perceived autonomy constitutes a critical starting point influencing human–AI trust.

2.3. Stereotype Content Model

The Stereotype Content Model (SCM) is a classic framework in the field of social cognition. This model asserts that when individuals form impressions of others, including non-human entities, they primarily rely on two fundamental dimensions, Warmth and Competence [30]. The warmth dimension reflects inferences about the other’s intentions (e.g., whether they are friendly or pose a competitive threat), whereas the competence dimension evaluates their efficacy and professionalism in executing goals. SCM further reveals the structural antecedents of these two dimensions. Perceived non-competition positively predicts warmth judgments, while perceived status positively predicts competence judgments [35]. In recent years, the SCM framework has been effectively extended to the field of artificial intelligence. Studies have confirmed that people spontaneously engage in social anthropomorphic cognition of AI systems [36]. This study argues that employees’ trust regarding intelligent systems are not merely functional evaluations, but rather profound social cognitive processing. Introducing the SCM framework enables the precise capture of employees’ inferences about the system’s intentions and efficacy based on perceived competition and status under different autonomy experiences, thereby providing a dual-dimensional explanatory mechanism for the formation of trust.

2.4. Perceived Autonomy and Human–AI Trust

Perceived autonomy refers to an individual’s subjective experience of decision-making control, freedom of choice, and behavioral dominance when collaborating with intelligent systems [34]. As intelligent systems become deeply involved in reasoning and prediction processes, employees face the potential threat of compressed decision-making power. Previous research indicates that when a system exhibits supportive characteristics (such as providing clear recommendation rationales and allowing for final human decision), employees tend to view it as a collaborative tool that respects human judgment [1,37]. This reduces their defensive psychology against excessive system intervention and fosters a stable trusting attitude [38]. Conversely, when AI is perceived as replacing human decision-making, users experience a sense of compromised autonomy, which in turn triggers psychological reactance and diminishes trust [39]. Accordingly, the following hypothesis is proposed:

H1. 

Employees’ perceived autonomy positively influences their human–AI trust in intelligent systems.

2.5. The Mediating Roles of Warmth Perception and Competence Perception

According to the SCM, individuals’ cognitive evaluations of others are influenced by their role-interaction relationships. On the one hand, a non-competitive relationship is the core prerequisite for high warmth perception [23]. In collaborative contexts with high autonomy, the system assumes an empowering and supportive role. Employees are less likely to perceive it as a competitor threatening their occupational status, and are more inclined to interpret it as a collaborator with aligned interests [40,41]. This positive intention attribution can significantly enhance employees’ warmth perception of the system [42]. Furthermore, the positive perception of the system’s goodwill and friendliness effectively mitigates employees’ defensiveness, thereby facilitating the establishment of human–AI trust [17].

On the other hand, status discrepancy is a crucial cue determining competence perception [43]. High autonomy reinforces the employee’s positioning as the dominator in the task, allowing them to occupy a higher psychological status in the collaborative relationship, while relegating the system to a subordinate or auxiliary low-status role [15,44,45]. Based on the internal attribution tendency in Attribution Theory [46], employees possessing high autonomy are more likely to credit task success to their own professional judgment when evaluating performance, thereby relatively ignoring or even depreciating the system’s actual contribution [47]. Therefore, higher autonomy inversely induces employees’ underestimation of the system’s competence. However, competence perception, serving as a functional representation of whether the system can fulfill its promises, is the core foundation of trust [48]. A decrease in competence perception heightens employees’ subjective expectations of task failure risks, subsequently weakening trust. Synthesizing these two pathways, this study proposes the following hypotheses:

H2a. 

Employees will exhibit higher warmth perception when interacting with high perceived autonomy intelligent system compared to low perceived autonomy intelligent system.

H2b. 

Employees will exhibit lower competence perception when interacting with high perceived autonomy intelligent system compared to low perceived autonomy intelligent system.

H2c. 

Warmth perception mediates the relationship between perceived autonomy and human–AI trust.

H2d. 

Competence perception mediates the relationship between perceived autonomy and human–AI trust.

2.6. Moderation Role of Critical Thinking

Critical thinking is a stable cognitive trait where individuals actively apply analysis and logical reasoning to make prudent judgments when confronted with complex information [49]. This study posits that critical thinking moderates the first-stage paths from perceived autonomy to warmth and competence perceptions.

Regarding the warmth path, employees with lower critical thinking are more susceptible to surface cues. They tend to directly equate the non-competitive signals conveyed by high autonomy with the system’s goodwill, thus rapidly establishing warmth perception [50]. In contrast, employees with stronger critical thinking possess a stronger tendency toward skepticism and scrutiny. They do not easily equate superficial supportive design with the system’s genuine friendliness; instead, they deeply consider potential commercial motives or data hazards behind the system design [29]. Therefore, strong critical thinking attenuates the positive effect of high perceived autonomy on warmth perception.

Regarding the competence path, critical thinking similarly exerts a moderating function. As previously discussed, employees with lower critical thinking are more prone to falling into role-attribution biases. Their competence judgments are highly susceptible to the psychological suggestion of their own high autonomy (high dominant status), leading them to underestimate the system’s competence [51]. In comparison, employees with stronger critical thinking are more inclined to rationally evaluate system efficacy based on factual evidence, such as the system’s output quality and objective accuracy rates [52]. Consequently, they can effectively overcome the cognitive bias of “underestimating the system due to one’s own dominant status” [53]. Thus, strong critical thinking can mitigate the negative effect of high perceived autonomy on competence perception.

In summary, by moderating the first-stage paths, critical thinking further alters the strength of the overall mediation mechanisms. Accordingly, the following hypotheses are proposed:

H3a. 

Critical thinking negatively moderates the mediating role of warmth perception between perceived autonomy and human–AI trust. Specifically, the higher an employee’s critical thinking, the weaker the indirect effect of perceived autonomy on enhancing trust through elevated warmth perception.

H3b. 

Critical thinking negatively moderates the mediating role of competence perception between perceived autonomy and human–AI trust. Specifically, the higher an employee’s critical thinking, the weaker the indirect effect of perceived autonomy on diminishing trust through reduced competence perception.

In conclusion, the theoretical model is depicted in Figure 1.

3. Methodology

3.1. Measures

All measurement scales employed in this study were adapted from well-established literature and specifically contextualized to fit the research setting of the Chinese engineering consulting industry. These original scales are derived from articles published in leading academic journals [39,54,55,56], which focus on relevant phenomena in human–AI interaction scenarios. To ensure measurement reliability and contextual fit, we strictly followed Brislin’s (1980) translation and back-translation paradigm [57]. First, two bilingual researchers independently translated the English items into Chinese, discussed discrepancies in translation, and revised the wording to form the initial Chinese scale. Second, two independent bilingual researchers back-translated the Chinese version into English. Importantly, we ensured that these researchers were completely blind to the original English items prior to the task. Subsequently, the research team compared the back-translated version with the original English scale and fine-tuned the expressions to guarantee the semantic accuracy of Chinese items and their adaptability to the engineering consulting/AI context. Finally, prior to the formal survey, a pretest was conducted with 10 industry experts in engineering consulting to examine the clarity and readability of the questionnaire statements. Based on the experts’ pretest feedback, we further refined the questionnaire items to better suit the context of the engineering consulting industry, while strictly preserving their core conceptual meaning. To further enhance situational immersion, we designed specific contextual instruction (e.g., The following questions relate to your daily operations with intelligent systems; AI-aided design system, intelligent BIM platforms, Smart project management system, etc. Please recall the system you interact with most frequently and evaluate the extent to which you feel autonomous when using it) presented to participants immediately before assessing each scale item. (Further details of this process are provided in Supplementary Material S1.) The final scale used in this study is demonstrated in Appendix A. All items in this study were measured using a 7-point Likert scale.

Perceived Autonomy: At Time T1, perceived autonomy was measured using a 5-item scale developed by [39]. Respondents rated their level of agreement with each item based on their actual work performance during working hours on the survey day (1 = strongly disagree, 7 = strongly agree). A representative item is “The intelligent system provides choices based on my actual workflow needs.”

Critical Thinking: This construct was measured by a validated 3-item mature scale developed by [56]. A typical item is “I consider why the intelligent system generates specific analytical or design outputs.”

Warmth Perception: Drawing on the revised and validated 6-item scale proposed by [55], warmth perception was assessed in this study, with a representative item “I perceive this intelligent system as kind.”

Competence Perception: Competence perception was measured with reference to the mature scale developed by [55], consisting of six items. A representative item is “I perceive this intelligent system as logical.”

Human–AI Trust: An 11-item scale developed by [54] was employed to measure individuals’ trust in intelligent systems. A typical item is “I have confidence in using this intelligent system in my engineering tasks.”

In addition, based on previous of Human–AI interaction [56,58,59,60], age, gender, education, and working experience were selected as control variables.

3.2. Sample and Data

The participants of this study consisted of frontline professionals from engineering consulting firms located in Beijing, Tianjin, Shanghai, Hangzhou, Chongqing, Guangzhou, and Shenzhen, China. The online survey questionnaires were distributed with the assistance of the Human Resources departments within these companies. This specific industrial context was strategically selected because China’s engineering consulting sector is currently undergoing rapid digital transformation. There is a widespread adoption of AI-empowered tools across these enterprises, including AI-aided design systems, intelligent Building Information Modeling (BIM) collaborative platforms, and smart project management systems. Consequently, professionals in this industry are deeply immersed in human–AI collaboration within their daily workflows, providing an ideal empirical setting for this research.

To ensure high data quality and mitigate common method bias, this study employed a rigorously controlled, multi-wave longitudinal research design [61]. Prior to initiating the survey, a strict inclusion criterion for frequent human–AI collaboration was applied. Participants were required to answer a screening question, “In your current role, do you interact with AI-empowered systems (e.g., AI-aided design, intelligent BIM platforms, or smart project management tools) on a daily basis?” Only those who selected “Yes” were permitted to proceed. To address potential missing data issues, the online survey platform was configured with a forced-response mechanism, ensuring that questionnaires could only be submitted upon full completion, thereby eliminating item-level missing data. Furthermore, rigorous data screening rules were established. We embedded two attention check items (e.g., “Please select the number in the middle position” and “While watching TV, have you ever had a sudden fatal heart disease?”). Only participants who selected “4” for the first question and “No” for the second question passed the attention check. Surveys failing to correctly answer these checks or exhibiting straight-lining behaviors, where respondents provided identical answers across multiple items, were systematically excluded.

Data collection was conducted at three separate time points with a one-month interval between each wave to establish a clear temporal sequence and provide chronological evidence for verifying the causal chain. To minimize sample bias risks caused by participant attrition across the waves, a tiered monetary incentive structure was implemented, complemented by follow-up reminders. Sample matching across the three waves was realized using the last four digits of respondents’ mobile phone numbers. Specifically, at Time 1 (T1), 600 questionnaires measuring demographic characteristics, perceived autonomy, and critical thinking were distributed. After applying our screening rules, 526 valid responses were retained. Participants who provided valid responses received an initial reward of 5 RMB. One month later, at Time 2 (T2), targeted questionnaires measuring warmth perception and competence perception were delivered to the 526 valid T1 respondents. Following the exclusion of those who failed attention checks or showed straight-lining behavior, 431 valid questionnaires were collected, and these respondents were rewarded with an additional 10 RMB. After another one-month interval, at Time 3 (T3), questionnaires measuring human–AI trust were sent to the T2 respondents. Ultimately, after strict data screening, 368 valid matched samples were obtained, yielding an overall effective response rate of 61.33%. Participants completing the final valid wave received a 15 RMB reward.

Moreover, we conducted attrition analysis in accordance with well-established methodological procedures in management research [62,63]. Specifically, we categorized the initial target sample at T1 (N = 526) into two distinct groups, comprising stayers who successfully completed all three waves (N = 368) and leavers who dropped out at T2 or T3 (N = 158). To compare the stayers and leavers groups, we conducted Chi-square tests for demographic variables (age, gender, education, and work experience) and independent-samples t-tests for focal constructs (perceived autonomy and critical thinking). The results revealed no significant differences between stayers and leavers group (p > 0.05).

The demographic characteristics of the samples are presented in

Table 1. Sample demographic information (N = 368).

Variable	Item	Number	Percentage
Gender	Male	189	51.4%
	Female	179	48.6%
Age	18–30	85	23.1%
	31–40	88	23.9%
	41–50	89	24.2%
	>51	106	28.8%
Education	Undergraduate degree or below	87	23.6%
	Bachelor	93	25.3%
	Master	108	29.3%
	Doctor	80	21.7%
Work experience	<3 years	89	24.2%
	3–10 years	95	25.8%
	11–20 years	93	25.3%
	>20 years	91	24.7%

. In terms of gender distribution, males accounted for 51.4% and females accounted for 48.6%. Regarding age structure, participants aged 18–30 years accounted for 23.1%, 31–40 years for 23.9%, 41–50 years for 24.2%, and 51 years and above for 28.8%. In terms of educational background, participants with a bachelor’s degree or below accounted for 23.6%, bachelor’s degree holders for 25.3%, master’s degree holders for 29.3%, and doctoral degree holders for 21.7%. In terms of work experience, employees with less than 3 years of work experience accounted for 24.2%, 3–10 years for 25.8%, 11–20 years for 25.3%, and more than 20 years for 24.7%.

4. Results

4.1. Common Method Bias

Based on the recommendations of [61], this study adopts a three-wave longitudinal research design to control for common method bias (CMB). Meanwhile, Harman’s single-factor test was conducted in this study. The results indicate that the cumulative variance explained by the 5 extracted common factors reaches 73.703%, and the variance explained by the first single factor accounts for 35.242%, which is below the critical threshold of 40%. For a more robust assessment of CMB, the unmeasured latent method construct (ULMC) approach was employed [64]. The results indicated no significant changes in model fit indices (Before: χ²/df = 1.122, RMSEA = 0.018, TLI = 0.993, CFI = 0.994, SRMR = 0.028; After: χ²/df = 1.130, RMSEA = 0.019, TLI = 0.993, CFI = 0.994, SRMR = 0.027). Based on the approach of [65], the aforementioned results indicate no substantial difference between the two models. Therefore, these findings alleviate concerns regarding common method bias and confirm that it does not pose a threat to this study.

The descriptive statistics and correlation coefficients of all variables are presented in

Table 2. Descriptive Statistics and Correlation Analysis.

Variables	1	2	3	4	5	6	7	8	9
1 Gender	-
2 Age	−0.057	-
3 Education	−0.01	0.015	-
4 Experience	0.012	0.017	−0.081	-
5 PA	0.017	0.084	0.016	−0.008	-
6 CT	0.000	0.062	0.01	−0.056	0.054	-
7 WP	−0.04	0.102 *	0.048	0.009	0.542 **	0.185 **	-
8 CP	−0.012	0.021	0.026	0.051	−0.297 **	0.053	−0.105 *	-
9 TRU	−0.042	0.126 *	0.013	0.037	0.424 **	0.100	0.501 **	0.161 **	-
Mean	0.510	2.590	2.490	2.510	4.067	3.947	4.015	4.012	3.992
SD	0.500	1.133	1.077	1.110	1.116	1.153	1.134	1.116	1.102

Note: N = 368. Correlations are Pearson correlation coefficients. * indicates significance at the p < 0.05 level; ** indicates significance at the p < 0.01 level.

. The data analysis results reveal that perceived autonomy (PA) is significantly positively correlated with warmth perception (WP) (β = 0.542, p < 0.01), significantly negatively correlated with competence perception (CP) (β = −0.297, p < 0.05), and significantly positively correlated with human–AI trust (TRU) (β = 0.424, p < 0.01). Critical thinking (CT) has a significant positive association with warmth perception (β = 0.185, p < 0.01). Warmth perception (WP) is significantly positively correlated with human–AI trust (TRU) (β = 0.501, p < 0.01). Additionally, competence perception (CP) exhibits a significant positive correlation with human–AI trust (TRU) (β = 0.161, p < 0.01).

4.2. Confirmatory Factor Analysis

To examine the reliability and validity of the measurement scales, confirmatory factor analysis (CFA) was performed using Mplus 7.4, with the results summarized in

Table 3. Results of Composite Reliability, Convergent Validity and Discriminant Validity.

Constructs	Items	Factor Loadings	Composite Reliability	Convergent Validity	Discriminant Validity
			CR	AVE	PA	WP	CP	CT	TRU
Perceived Autonomy	PA1–PA5	0.803–0.842	0.913	0.677	0.823
Warmth Perception	WP1–WP6	0.798–0.840	0.933	0.700	0.584	0.837
Competence Perception	CP1–CP6	0.808–0.874	0.929	0.686	−0.324	−0.107	0.828
Critical Thinking	CT1–CT3	0.809–0.852	0.864	0.679	0.063	0.210	0.058	0.824
Human–AI Trust	TRU1–TRU11	0.732–0.856	0.958	0.675	0.454	0.532	0.174	0.110	0.822

Note: The bold values on the diagonal are the square roots of the AVE, and the lower triangle presents the Pearson correlations among the constructs.

. The factor loadings of all variables exceeded 0.7, satisfying the evaluation criteria proposed by [66]. The composite reliability (CR) values of all constructs were higher than 0.8, and the average variance extracted (AVE) exceeded 0.6, which conforms to the methodological norms established by [66,67]. These findings demonstrate that the measurement scales in this study possess good composite reliability and convergent validity.

Additionally, the Pearson correlation coefficient between any two variables was lower than the square root of the corresponding AVE value. This outcome meets the discriminant validity criterion recommended by [67], indicating satisfactory discriminant validity of the research scales.

In addition, this study compared the model fit between the baseline five-factor model and other competing models (see

Table 4. Model Fit Comparison.

Model	Model Structure	χ2	df	χ2/df	CFI	TLI	RMSEA	SRMR
5 Factors	PA, WP, CP, CT, TRU	474.917	424	1.1	0.994	0.994	0.018	0.028
4 Factors	PA + CT, WP, CP, TRU	1009.274	428	2.4	0.934	0.928	0.061	0.063
3 Factors	PA + CT, WP + CP, TRU	2653.514	431	6.2	0.746	0.726	0.118	0.148
2 Factors	PA + CT + TRU, WP + CP	3661.687	433	8.5	0.631	0.604	0.142	0.171
1 Factor	PA + CT + TRU + WP + CP	4833.788	434	11.1	0.497	0.462	0.166	0.182

). The results revealed that the five-factor model exhibited the best fit performance (χ²/df = 1.1, CFI > 0.9, TLI > 0.9, RMSEA < 0.08, SRMR < 0.05). It is thereby verified that the hypothesized five-factor structure is significantly superior to other alternative models, demonstrating adequate discriminant validity among the five core variables.

4.3. Hypothesis Test

First, Mplus 7.4 was employed to examine the main effect of perceived autonomy on human–AI trust, as well as the mediating roles of warmth perception and competence perception. The path models were evaluated using the Maximum Likelihood (ML) estimator. Furthermore, the significance of all direct and indirect effects was tested using a bootstrapping procedure with 5000 resamples to generate 95% percentile confidence intervals. All reported path coefficients are unstandardized estimates (b). The empirical results are presented in

Table 5. Results of Direct and Indirect Effect.

Path	Estimate	S.E.	Z Value	95% Percentile Confidence Interval
				Lower	Upper
Direct Effect
PA → TRU	0.343	0.062	5.571	0.216	0.470
Indirect Effect
PA → WP → TRU (Path WP)	0.204	0.039	5.294	0.132	0.289
PA → CP → TRU (Path CP)	−0.107	0.023	−4.606	−0.154	−0.066
Total Indirect Effect	0.097	0.048	2.041	0.004	0.198
Contrast
Path WP vs. Path CP	0.311	0.042	7.394	0.233	0.399

. The findings indicated that perceived autonomy had a significant positive effect on human–AI trust (b = 0.343, 95% percentile CI [0.216, 0.470]), thus supporting H1. Furthermore, perceived autonomy exerted a positive indirect effect on human–AI trust via warmth perception (b = 0.204, 95% percentile CI [0.132, 0.289]) and a negative indirect effect via competence perception (b = −0.107, 95% percentile CI [−0.154, −0.066]). Since none of the 95% percentile CIs encompassed zero, hypotheses H2c and H2d were supported.

Second, this study adopted a full path model via Mplus 7.4 to examine the proposed moderated mediation effect. The model was estimated using the Maximum Likelihood (ML) estimator with 1000 Bootstrap samples. The results are presented in Figure 2, which displays the unstandardized path coefficients (b). The results indicated that perceived autonomy had a significant positive effect on warmth perception (b = 0.849, SE = 0.206, p < 0.001), whereas its effect on competence perception was non-significant (b = 0.337, SE = 0.187, p > 0.05). Therefore, H2a was supported, whereas H2b was not supported.

In addition, the moderating effect of critical thinking on the relationship between perceived autonomy and warmth perception was non-significant (b = −0.073, SE = 0.051, p > 0.1). By contrast, critical thinking exerted a significant negative moderating effect on the relationship between perceived autonomy and competence perception (b = −0.167, SE = 0.047, p < 0.001). To visually demonstrate such moderating trends, simple slope plots were drawn based on the mean value plus and minus one standard deviation, as illustrated in Figure 3.

Third, to verify H3a and H3b, this study adopted the Bootstrap method (Bootstrap = 1000; Estimator = ML) within the overall path model to examine the indirect effects of perceived autonomy on human–AI trust via warmth perception and competence perception at high and low levels of critical thinking. The relevant results are presented in

Table 6. Results of Moderated Mediation.

Mediator	Critical Thinking	Indirect Effect	S.E.	95% Percentile Confidence Interval
				Lower	Upper
Warmth Perception	Low critical thinking (−1 sd)	0.231	0.047	0.140	0.323
	High critical thinking (+1 sd)	0.175	0.041	0.094	0.256
	Between difference	−0.027	0.019	−0.064	0.011
	Index of Moderated Mediation	−0.027	0.019	−0.064	0.011
Competence Perception	Low critical thinking (−1 sd)	−0.050	0.026	−0.100	0.001
	High critical thinking (+1 sd)	−0.166	0.032	−0.228	−0.104
	Between difference	−0.055	0.016	−0.086	−0.025
	Index of Moderated Mediation	−0.055	0.016	−0.086	−0.025

.

Among them, the moderated mediation index with warmth perception as the mediating path was non-significant (95% percentile CI = [−0.064, 0.011], containing zero), which invalidates H3a. In contrast, the moderated mediation index with competence perception as the mediating path was significant (95% percentile CI = [−0.086, −0.025], excluding zero), thereby supporting H3b. In the context of low critical thinking, the indirect effect of perceived autonomy on human–AI trust through competence perception was non-significant (95% percentile CI = [−0.100, 0.001], containing zero). However, under high critical thinking, perceived autonomy exerted a significant negative indirect effect on human–AI trust via competence perception (b = −0.166, 95% percentile CI = [−0.228, −0.104], excluding zero).

Following best-practice recommendations for the treatment of control variables in organizational research [68,69], we re-estimated the mediation model and full path model without the inclusion of any demographic control variables (i.e., age, gender, education, and work experience). The results revealed that the significance levels and the direction of all hypothesized paths remained entirely consistent with the models that included the controls. This consistency confirms that our focal findings are robust and not statistical artifacts driven by the inclusion of covariates. The detailed results of these tests are presented in Supplementary Material S2.

5. Discussion

5.1. Research Findings

Drawing upon multi-wave survey data, this study investigates the underlying mechanisms through which employees’ perceived autonomy influences human–AI trust toward intelligent systems within the context of engineering consulting. Furthermore, it explores the complex interplay of social cognitive dimensions and individual cognitive traits in this process. The main conclusions are drawn as follows:

First, perceived autonomy serves as a core antecedent driving human–AI trust. The results demonstrate that perceived autonomy exerts a significant positive direct effect on human–AI trust. In human–AI collaboration, granting employees more decision-making autonomy directly enhances their psychological acceptance of intelligent systems. This conclusion corroborates the highly interdependent relationship between users’ sense of control and their trust levels in highly automated work environments [18,26].

Second, social cognition plays an asymmetric competitive dual-mediation role in the formation of human–AI trust. The results indicate that perceived autonomy has a significant positive impact on employees’ warmth perception, whereas its direct effect on competence perception is non-significant. However, the test of mediation effect reveals contrasting transmission mechanisms. Perceived autonomy exerts a significant positive indirect effect on Human–AI trust through the warmth perception, while simultaneously exerting a significant negative indirect effect through the competence perception. This suggests that granting employees high autonomy explicitly builds trust by signaling benevolence and support (warmth); yet, beneath the surface, there lies a hidden depleting mechanism that hinders trust-building by undermining instrumental expectations (competence). This finding empirically deepens the multidimensional competition hypothesis of social cognition in complex interactive environments [30], demonstrating that trust calibration is not only driven explicitly by affective paths but is also constrained by implicit negative mechanisms within instrumental paths.

Third, critical thinking plays an asymmetric moderating role in the dual-mediating pathways of social cognition. The results demonstrate that in the negative indirect pathway, where perceived autonomy affects human–AI trust via competence perception, high levels of critical thinking significantly exacerbate this negative indirect effect. Specifically, when employees with high critical thinking are granted high decision-making autonomy in collaboration, their strong analytical and evaluative capabilities enable them to more acutely identify the functional limitations of intelligent systems in complex tasks. This cognitive trait amplifies employees’ negative evaluations of the system’s capabilities, thereby leading to a further decline in human–AI trust. Another possible explanation is that individuals with high critical thinking establish a significantly higher internal baseline for competence. Consequently, once they perceive any shortcomings, it triggers a more profound sense of expectancy violation, thereby amplifying their negative evaluations [70]. This finding provides compelling evidence for how cognitive traits profoundly intervene in individuals’ trust calibration toward intelligent systems [50].

5.2. Theoretical Contribution

This study makes three primary theoretical contributions to the literature:

First, by integrating Self-Determination Theory (SDT) into the human–AI trust domain, this study empirically validates that psychological need satisfaction serves as the fundamental driver of trust formation. Prior research on the antecedents of human–AI trust has predominantly focused on objective system characteristics (e.g., transparency, reliability) or environmental constraints, largely overlooking the perspective of users’ intrinsic psychological needs within their work contexts [21,71]. This research demonstrates that perceived autonomy exerts a significant positive main effect on human–AI trust. When employees perceive that intelligent systems preserve their decision-making dominance, the fulfillment of basic psychological needs effectively mitigates defensive reactions associated with viewing AI as a substitutive threat, thereby fostering profound psychological acceptance. This finding not only resonates with [72]’s design philosophy of Human-Centered AI, but also provides robust empirical evidence for [73] assertion that respecting user autonomy is a crucial foundation for establishing human–AI trust, significantly broadening the psychological theoretical perspective on trust formation mechanisms.

Second, this study extends the dynamic application of the Stereotype Content Model (SCM) to non-human entities, uncovering the asymmetric dual-pathways of social cognitive judgments. While previous studies have largely conceptualized human–AI trust as the outcome of a single-logic driven process, this study establishes that trust formation is a complex mechanism jointly shaped by multidimensional social cognitions. Specifically, perceived autonomy yields a positive indirect effect through warmth perception, alongside a negative indirect effect through competence perception. The discovery of this double-edged sword effect aligns with SCM’s classical propositions regarding the correlation between status and competence [43]. This finding successfully extends [74] assertion regarding the primacy of the warmth dimension in social judgments from traditional interpersonal contexts to the highly technical domain of human–AI interaction in engineering consulting. Furthermore, this dual-pathway structure responds to [75] theoretical call that the warmth and competence dimensions are not homogenous and thus require differentiated processing mechanisms.

Third, this research clarifies the asymmetric boundary mechanisms and individual heterogeneity of critical thinking in the evolution of Human–AI trust. The study reveals an asymmetric moderating effect of critical thinking on the dual mediating pathways. Drawing upon Dual-Process Theory [50], warmth perception—as an intuitive and affective judgment of system intentions—relies on the automated and rapid processing of the intuitive system. Consequently, it is marginally affected by critical thinking, which corroborates the perspectives of [76]. In contrast, competence perception, being a cognitive judgment that necessitates logical reasoning, is highly subject to the moderation of the analytical system. Employees with high critical thinking abilities are capable of overcoming role attribution biases, enabling them to objectively evaluate the actual performance of the intelligent systems.

5.3. Practical Implication

The findings of this study offer valuable practical implications for engineering consulting firms and intelligent system developers in advancing the management and design of human–AI symbiosis.

First, regarding intelligent system deployment strategies, organizations must adhere to the human-in-the-loop principle and prioritize safeguarding employee autonomy. When introducing generative AI or intelligent design-review platforms, firms often fall into the trap of exclusively pursuing technical automation rates. Psychological reactance theory [77] suggests that compromised autonomy triggers strong defensive reactions among individuals. Therefore, during process reengineering, enterprises should explicitly retain employees’ decision-making dominance and latitude for selecting alternatives. For core tasks involving professional judgment, system outputs must be positioned as high-value references rather than definitive verdicts, thereby preserving employees’ psychological primacy. This corroborates the managerial recommendation by [10] that, in the initial stages of human–AI collaboration, prioritizing users’ psychological acceptance is far more critical than merely emphasizing algorithmic efficacy. This implication applies not only to the engineering consulting sector but also to other knowledge-intensive industries—such as healthcare (e.g., AI-assisted diagnosis) and financial auditing—where professionals universally face threats to their professional identity when confronted with AI.

Second, regarding system interaction design, developers should facilitate the dual transmission of warm intentions and objective capabilities. This study finds that warmth perception is the dominant pathway fostering trust, unaffected by individual cognitive traits. This necessitates that system interfaces emit goodwill signals—framing the AI as a collaborative partner—through interaction designs such as anthropomorphic advisory language and transparent recommendation rationales. However, organizations must also remain vigilant about the potential risks associated with the competence perception pathway to avoid the cognitive paradox of the more deferential to the user, the more incompetent it is perceived. To counteract the competence devaluation bias induced by high autonomy, systems should simultaneously establish objective performance display modules (e.g., real-time historical accuracy, saved time costs) while demonstrating supportive intentions. Utilizing such empirical data will reinforce the system’s professional authority. Because the trade-off between warmth perception and competence perception is rooted in the universally applicable Stereotype Content Model (SCM), this interaction design paradox offers critical insights for global AI-empowered tools developers. When implementing anthropomorphic design [78] and Explainable AI (XAI) [79], developers across various sectors must carefully balance empathetic interfaces with objective performance metrics to avoid undermining the system’s authority.

Third, concerning talent cultivation and organizational support, firms should foster AI collaboration literacy to mitigate defensive rejection. The findings indicate that critical thinking is paramount in rectifying cognitive biases regarding system competence. Consequently, organizations should not limit their training to basic operational skills for intelligent systems; rather, they should integrate critical thinking and AI collaborative literacy into their core professional development frameworks. By regularly reviewing AI collaboration case studies and enhancing employees’ data interpretation skills, managers can assist staff of varying cognitive styles in learning to evaluate system outputs based on objective facts and logical reasoning, rather than their professional ego. This approach facilitates the transformation of employees’ instinctive resistance toward emerging technologies into rational acceptance and deep value co-creation. This implication transcends the Chinese cultural context—which traditionally emphasizes interpersonal harmony—and the specific AEC industry. As mitigating algorithm aversion and overcoming professional ego are universal challenges in the global Future of Work transition, all organizations aiming to enhance efficiency through human–AI collaboration must prioritize AI collaboration literacy.

6. Limitation and Further Research

Drawing upon Self-Determination Theory and the Stereotype Content Model, this study develops and tests a moderated mediation model to elucidate how perceived autonomy influences Human–AI trust. Our findings uncover the dual mediation pathways of warmth perception and competence perception, while highlighting the asymmetrical moderating effects of critical thinking in these processes.

However, this study offers valuable insights, certain limitations should be acknowledged. First, although a multi-wave survey design was employed to mitigate common method variance, the data fundamentally relies on self-reporting. Future research could incorporate experimental designs or extract objective interaction logs (log data) from system backends to more precisely capture users’ actual collaborative behaviors. Second, the formation of trust is a dynamic process that evolves over time. As this study primarily captures cross-sectional characteristics at a specific temporal stage, future endeavors could employ longitudinal designs to explore the dynamic fluctuations of warmth perception and competence perception throughout prolonged human–AI adaptation. Third, this study did not restrict the specific AI systems evaluated. While this mixed-system approach enhances external validity by capturing universal psychological mechanisms, we acknowledge that system heterogeneity may confound users’ judgments of autonomy, competence, warmth, and trust. Future research should isolate specific AI systems to explore how this heterogeneity shapes these dual-mediating pathways. Finally, this study focuses on the engineering consulting context, which is characterized by a high cognitive threshold. Whether these conclusions are fully generalizable to other domains, such as routine administrative tasks in other construction management scenario, requires further external validity testing through cross-contextual sampling.