AI typically refers to the simulation of human intelligence by computers (Gillath et al., 2021). This simulation process encompasses learning (acquiring information and using it to acquire rules), reasoning (using rules to reach conclusions), and self-correction. In essence, AI represents a sophisticated form of automation, enhancing its domain and efficacy. In this paper, we refer automation as traditional automated technologies that are distinct from AI. The distinction lies in automation being systems that perform repetitive tasks based on static rules or human commands, while AI involves systems skilled in dealing with uncertainties and making decisions in novel situations (Cugurullo, 2020; Kaplan et al., 2021).

Interestingly, in the initial research on human-automation interaction, AI was considered a technology difficult to implement (Parasuraman and Riley, 1997). However, in the 21st century and especially after 2010, AI technology has progressed significantly. Nowadays, the impact of AI technology and its applications pervade daily life and professional environments, encompassing speech and image recognition, autonomous driving, smart homes, among others. AI can autonomously deliver personalized services by analyzing historical human data (Lu et al., 2019). Particularly, the recent advancements in generative AI have provided a glimpse of the potential for achieving general AI. That said, understanding how current AI arrives at specific decisions or outcomes can be complex due to its reliance on vast amounts of data and intricate algorithms.

With the advancement and widespread applications of AI, trust in AI has indeed become a new focal point in the study of human-automation interaction. This transition redefines relationships between humans and automation, moving from reliance on technologies for repetitive, accuracy-driven tasks to expecting AI to demonstrate capabilities in learning, adapting, and collaborating (Kaplan et al., 2021). This evolution in trust requires AI systems to demonstrate not only technical proficiency but also adherence to ethical standards, legal compliance, and socially responsible behavior. As AI becomes more integrated into daily life and crucial decision making tasks, establishing trust in AI is essential. This necessitates a focus on enhancing transparency, explainability, fairness, and robustness within AI systems, which is also central to trustworthy AI.

While AI represents a new generation of automation technology with unique characteristics, research on trust in earlier forms of automation remains relevant. This historical perspective can inform the development of trust in AI by highlighting the important trust factors. Incorporating lessons from the past, the transition to AI additionally demands a renewed focus on ethical considerations, transparency, and user engagement to foster a deeper and more comprehensive trust.

Early theoretical models of trust in automation focused on attributes related to automation’s performance or competence, such as reliability, robustness, capability, and predictability (Sheridan, 1988; Malle and Ullman, 2021). In these models, trust was primarily grounded in the systems’ technical performance and their ability to meet user expectations reliably. Trust varied directly with the system’s demonstrated competence in executing tasks (Malle and Ullman, 2021). However, the progress of AI has broadened the scope of research on trust determinants to include considerations of automation’s inferred intentions and the ethical implications of its actions (Malle and Ullman, 2021).

Individual differences are critical to human-automation trust. Some individuals apply frequently the machine heuristic, which is similar to trust propensity and represents the tendency to perceive automation as safer and more trustworthy than humans (Sundar and Kim, 2019). Moreover, an individual’s self-efficacy in using automation—the confidence in their ability to effectively utilize automation technologies—also plays a crucial role in shaping trust (Kraus et al., 2020); a higher sense of self-efficacy correlates with greater trust and willingness to use automated systems (Latikka et al., 2019). Furthermore, the degrees of familiarity and understanding of automated systems contribute to a more accurate evaluation of these systems’ competence, promoting a well-calibrated trust (Lee and See, 2004). Trust in automation, therefore, is a dynamic process, continuously recalibrated with the input of new information, knowledge, and experience (Muir, 1987).

Initial research on trust in automation adopted frameworks from interpersonal trust studies, positing a similar psychological structure in them (Muir, 1987). In practical research, factors affecting human-automation trust can also be categorized into trustor (human factors), trustee (automation factors), and the interaction context. This tripartite framework has been validated across various studies, affirming its applicability in understanding trust dynamics (Hancock et al., 2011; Hoff and Bashir, 2015; Drnec et al., 2016). Automation technology has evolved into the era of AI, inheriting characteristics of traditional automation while also exhibiting new features such as learning capabilities and adaptability. Factors influencing trust in AI based on this three-dimension framework are analyzed in detail in the next section from a sociocognitive perspective.

The impacts of demographic variables on trust in AI are complicated. In a recent worldwide survey study conducted in 17 countries, Gillespie et al. (2023) found that gender differences in trust toward AI were generally minimal, with notable exceptions in a few countries (i.e., the United States, Singapore, Israel, and South Korea). Regarding age, while a trend of greater trust among the younger generation was prevalent, this pattern was reversed in China and South Korea, where the older population demonstrated higher levels of trust in AI. Additionally, the data indicated that individuals possessing university-level education or occupying managerial roles tended to exhibit higher trust in AI, pointing to the significant roles of educational background and professional status in shaping trust dynamics. The study further showed pronounced cross-country variations, identifying a tendency for placing more trust in AI in economically developing nations, notably China and India. Previous research has also found that culture and social groups can influence trust in AI (Kaplan et al., 2021; Lee and Rich, 2021). Therefore, the impact of demographic variables on trust in AI may be profoundly influenced by socio-cultural factors.

An individual’s familiarity with AI, rooted in their knowledge and prior interactive experience, plays a pivotal role in trust formation (Gefen et al., 2003). Such familiarity not only reduces cognitive complexity by supplying essential background information and a cognitive framework, but also enables the formation of concrete expectations regarding the AI’s future behavior (Gefen, 2000). Additionally, a deeper understanding of AI can reduce the perceived risks and uncertainties associated with its use (Lu et al., 2019). Relatedly, AI use self-efficacy, or individuals’ confidence in their ability to effectively use AI, significantly influences acceptance and trust (REF). Familiarity and self-efficacy are both related to past interactive experience with AI, and both are positively correlated with a precise grasp of AI, thereby facilitating appropriate trust in AI.

Hedonic motivation, an intrinsic form of motivation, can lead people to use AI in pursuit of enjoyment. Venkatesh et al. (2012) recognized this by incorporating hedonic motivation into the expanded Unified Theory of Acceptance and Use of Technology (UTAUT) model as a key determinant of technology acceptance. This form of motivation is instrumental in increasing user satisfaction and enjoyment derived from AI, thereby positively influencing their attitudes toward technology and enhancing their intention to use it (Gursoy et al., 2019).

Emotional experience in the context of AI refers to the sense of security and comfort users feel when relying on AI, often described as emotional trust. It can reduce people’s perception of uncertainties and risks, and thus increase trust in AI. The acceptance and utilization of AI are guided by both cognitive judgments and affective responses. As such, it is crucial for trust research in AI to address both the cognitive and the emotional components (Gursoy et al., 2019). Specifically, emotional experience has been shown to directly impact the willingness to adopt AI-based recommendation systems, as seen in the context of travel planning (Shi et al., 2020).

Sense of control represents individuals’ perception of their ability to monitor and influence AI decision-making processes. Dietvorst et al. (2018) found that algorithm aversion decreased when participants were allowed to adjust the outcomes of an imperfect algorithm, even if the adjustments were minimal. This finding underscores the importance of a sense of control in enhancing user satisfaction and trust in algorithms that are fundamental components of AI. Aoki (2021) found that AI-assisted nursing care plans that explicitly informed individuals that humans retained control over the decision-making processes significantly boosted trust in AI, compared to those who were not provided with this information. This highlights the importance of communicating human oversight in AI applications to enhance public trust. Similarly, Jutzi et al. (2020) found a favorable attitude toward AI in medical diagnosis when AI acted in a supportive capacity, reinforcing the value of positioning AI as an adjunct to human expertise.

The propensity to trust refers to stable internal psychological factors affecting an individual’s inclination to trust, applicable to both people and technology. Research indicates that individuals with high trust propensity are more inclined to place trust in others, including strangers, and hold a general belief in the beneficial potential of technology (Brown et al., 2004). This tendency enables them to rely on technological systems without extensive evidence of their reliability. Attitudes toward new technologies vary significantly; some individuals readily adopt new technologies, while others exhibit skepticism or caution initially. This variation extends to AI, where trust propensity influences acceptance levels (Chi et al., 2021). Furthermore, trust propensity may intersect with personality traits. For instance, individuals experiencing loneliness may show lower trust in AI, whereas those with a penchant for innovation are more likely to trust AI (Kaplan et al., 2021).

Because of the complexity and potential wide-ranging impacts of AI, accountability is a key factor in establishing public trust in AI. Fears that AI cannot be held responsible hinder trust in AI. Therefore, people need assurance that clear processes exist to handle AI issues and that specific parties, like developers, providers, or regulators, are accountable.

When people think that AI cannot be held accountable, they are less willing to let AI make decisions and tend to blame it less. Research has found that in the service industry when service providers make mistakes that result in customer losses, participants believe that the robot responsible for the mistake bears less responsibility compared to a human service provider, and the service-providing company should bear more responsibility (Leo and Huh, 2020). This occurs probably because people perceive that robots have poorer controllability over tasks. People are reluctant to allow AI to make moral decisions because AI is perceived to lack mind perception (Bigman and Gray, 2018). Bigman et al. (2023) found that algorithmic discrimination elicits less anger, with people showing less moral outrage toward algorithmic (as opposed to human) discrimination and being less inclined to blame the organization, but it does lower the evaluation of the company. This might be because people perceive algorithms as lacking prejudicial motivation.

Anthropomorphism, the tendency to ascribe human-like qualities to non-human entities such as computers and robots, significantly affects individuals’ trust in these technologies (Bartneck et al., 2009). Cominelli et al. (2021) found that robots perceived as highly human-like are more likely to be trusted by individuals. Beyond physical appearance and vocal cues, emotional expression is a crucial aspect of anthropomorphism. Troshani et al. (2021) found that robots exhibiting positive emotions are more likely to receive increased trust and investment from people. Similarly, Li and Sung (2021) demonstrated through a network questionnaire that anthropomorphized AI correlates with more positive attitudes toward the technology. Experimental studies have corroborated these findings, suggesting that psychological distance plays a mediating role in how anthropomorphism influences perceptions of AI (Li and Sung, 2021).

The key to evaluating trustworthy AI is whether AI does what it claims to do (Schwartz et al., 2022). The claims of AI can be analyzed from two perspectives: one is whether it fulfills the functional requirements of its users, and the other is whether it demonstrates good intentions toward its users. This directly corresponds to the perceptions of competence and warmth of AI.

In both interpersonal trust and human-automation trust, competence and warmth are pivotal in shaping perceptions of trustworthiness (Kulms and Kopp, 2018). The stereotype content model (SCM) posits that stereotypes and interpersonal perceptions of a group are formed along two dimensions: warmth and competence (Fiske et al., 2002). Warmth reflects how one perceives the intentions (positive or negative) of others, while competence assesses the perceived ability of others to fulfill those intentions.

Trust and stereotype share a foundational link through the attitudes and beliefs individuals hold toward others (Kong, 2018). Moreover, in Mayer et al.’s (1995) trust model, the trustworthiness dimensions of ability and benevolence align closely with the competence and warmth dimensions in the SCM, respectively. Therefore, warmth and competence may be two core components of trustworthiness affecting interpersonal trust. The significance of these two dimensions is evident in their substantial influence on individuals’ evaluations and behaviors toward others (Mayer et al., 1995; Fiske, 2012).

Competence is the key factor influencing trust in automation (Drnec et al., 2016). When users observe errors in the automated system, their trust in it decreases, leading them to monitor the system more closely (Muir and Moray, 1996). However, an AI agent that is competent but not warm might not be trusted, because, in certain situations, intention is a crucial influencing factor of trust (Gilad et al., 2021). For instance, although users may recognize the technical proficiency of autonomous vehicles (AVs) in navigating complex environments, concerns that AVs may compromise safety for speed or prioritize self-preservation in emergencies can undermine trust (Xie et al., 2019). Thus, AI needs to demonstrate good intentions to build trust. This is exemplified by AI’s social intelligence, such as understanding and responding to user emotions, which significantly bolsters trust in conversational agents (Rheu et al., 2021). Moreover, trust in AI is generally lower in domains traditionally dominated by human expertise, potentially due to concerns about the intentions of AI (Lee, 2018).

The rapid development of AI, facilitated greatly by network technology, raises privacy concerns, especially when third parties access data through networks without user consent, risking privacy misuse (Featherman and Pavlou, 2003). Additionally, while AI’s ability to tailor services to individual needs can enhance user satisfaction, this often requires accessing personal information, thus creating a dilemma between personalization benefits and privacy risks (Guo et al., 2016). Network technologies have amplified privacy risks, resulting in individuals losing control over the flow of their private data. As a result, privacy concerns play a crucial role in establishing online trust, and internet users are highly concerned about websites’ privacy policies, actively seeking clues to ensure the protection of their personal information (Ang et al., 2001). Research has found that providing adequate privacy protection measures directly influences people’s trust in AI and their willingness to use it (Vimalkumar et al., 2021; Liu and Tao, 2022).

Sheridan (1988) argued that robustness should be an important determinant of trust. The robustness of AI refers to the reliability and consistency of its operations and results, including its performance under diverse and unexpected conditions (High-Level Expert Group on Artificial Intelligence (AI HLEG), 2019). The fairness of AI involves treating all users equitably, making unbiased decisions, and not discriminating against any group (Shin and Park, 2019). Robustness is a key factor influencing trust in AI. Rempel et al. (1985) identified three components of trust from a dynamic perspective, including predictability (the consistency of actions over time), dependability (reliability based on past experience), and faith (belief in future behavior). Based on the definitions, these components also correspond to the formation of the perception of robustness. Compared to trust in humans, building trust in AI takes more time; moreover, when AI encounters problems, the loss of trust in it happens more rapidly (Dzindolet et al., 2003). Furthermore, the simpler the task in which the error occurs, the greater the loss of trust (Madhavan et al., 2006).

Because robustness and fairness are vulnerable to data bias, from both theoretical and practical standpoints, these two factors are closely related. Robustness serves as a crucial foundation for fairness, with the presence of discrimination and bias often signaling a lack of robustness. For example, the training data used for developing large language models often contain biases, and research has found that ChatGPT replicates gender biases in reference letters written for hypothetical employees (Wan et al., 2023). Such disparities underscore the importance of aligning AI with human values, as perceived fairness significantly influences users’ trust in AI technologies (Angerschmid et al., 2022).

Users need to understand how and why AI makes specific decisions, which corresponds, respectively, to transparency and explainability, before trusting in it. However, this is not an easy task for AI practitioners and stakeholders. Because of the mechanisms of algorithms, particularly the opacity of neural networks, it is difficult for humans to fully comprehend their decision-making process (Shin and Park, 2019).

Transparent AI models with clear explanations of their decision-making processes help users gain confidence in the system’s capabilities and accuracy. Moreover, transparency in AI design and implementation helps identify potential sources of bias, allowing developers to address these issues and ensure the AI system treats all users fairly. The concepts of transparency and explainability are deeply interconnected; explainability, in particular, plays a crucial role in reducing users’ perceived risks associated with AI systems (Qin et al., 2020). Additionally, providing reasonable explanations after AI errors can restore people’s trust in AI (Angerschmid et al., 2022).

That said, the impact of transparency and explainability on trust in AI shows mixed results. Leichtmann et al. (2022) found that displaying AI’s decision-making process through graphical and textual information enhances users’ trust in the AI program. However, Wright et al. (2020) found no significant difference in trust levels attributed to varying degrees of transparency in simulated military tasks for target detection. Furthermore, in a task of using AI assistance to rate movies, Schmidt et al. (2020) observed that increased transparency in AI-assisted movie rating tasks paradoxically reduced user trust.

Therefore, the relationship between transparency and trust in AI is intricate. Appropriate levels of transparency and explainability can enhance people’s trust in AI, but excessive information might be confusing (Kizilcec, 2016), thereby reducing their trust in AI. The absence of clear operational definitions for AI’s transparency and explainability complicates the determination of the optimal transparency levels that effectively build trustworthy AI. In general, lack of transparency indeed hurts trust in AI, but high levels of transparency do not necessarily lead to good results.

AI is surrounded by various unknowns, including ethical and legal uncertainties, that are critical evaluations of the interactive environment. Lockey et al. (2020) emphasized uncertainties as a key factor influencing trust in AI. Similarly, Jing et al. (2020) conducted a literature review and discovered a negative correlation between perceived uncertainties and risks with the acceptance of autonomous driving. Furthermore, perceived uncertainties in the application of AI vary across different applications, particularly pronounced in medical expert systems and vehicles (Yang and Wibowo, 2022).

Perceived benefits, such as time savings, cost reductions, and increased convenience, highlight the recognized advantages of using AI (Kim et al., 2008). Liu and Tao (2022) found that perceived benefits, such as usefulness, could facilitate the use of smart healthcare services. Although perceived benefits can be viewed as characteristics of AI the trustee, they can also be highly socially dependent, mainly because the impacts of these benefits are not uniform across all users: For instance, while AI applications may enhance work efficiency for some, they could pose a risk of unemployment for others (Pereira et al., 2023). Therefore, perceived benefits are intricately linked to the social division of labor, underscoring their importance within the broader interactive context of AI usage.

Drawing from the concept of institution-based trust, safeguards are understood as the belief in existing institutional conditions that promote responsible and ethical AI usage (McKnight et al., 1998). Because AI is perceived as lacking agency and cannot be held accountable for its actions (Bigman and Gray, 2018), safeguards play a crucial role in ensuring human trust in AI.

Lockey et al. (2020), however, found that a mere 19–21% of Australians considered the current safety measures adequate for AI’s safe application, underscoring a significant trust gap. Their analysis further showed that perception of these safeguards was a strong predictor of trust in AI. In today’s AI landscape, establishing legal frameworks to protect human rights and interests is crucial for fostering trust. A prime example is the European Union’s General Data Protection Regulation (GDPR). Enacted in 2018, GDPR introduces stringent privacy protections and sets clear standards for algorithmic transparency and accountability (Felzmann et al., 2019).

Social influence is defined by the extent to which an individual perceives endorsement of specific behaviors by their social network, including encouragement from influential members to adopt new technologies (Gursoy et al., 2019). It is a crucial construct in the UTAUT (Venkatesh et al., 2003). Social influence theory posits that individuals tend to conform to the norms and beliefs of their social network (Shi et al., 2020). When individuals perceive that the use of AI is socially acceptable, they are more likely to experience positive emotions toward it, leading to an increase in their emotional trust in AI.

Facilitating condition is another critical variable in the UTAUT model, referring to the extent to which individuals perceive organizational, group, or infrastructural support in using AI (Venkatesh et al., 2003). Chi et al. (2021) found that facilitating robot-use conditions could improve users’ trust in social robots in service scenarios.

Cultural factors can significantly influence trust and acceptance of AI. For instance, cultures with high uncertainty avoidance are more inclined to trust and depend on AI (Kaplan et al., 2021), and the level of trust in AI also varies between individualistic and collectivistic cultures (Chi et al., 2023). Moreover, cultural influences may interact with economic factors to affect AI trust. Gillespie et al. (2023) found that individuals in the emerging economies, such as Brazil, India, China, and South Africa, exhibit higher levels of trust, in comparison to developed nations, such as the United Kingdom, Australia, Japan, and France. Furthermore, the impact of culture on AI trust can be mediated through social influence, highlighting the importance of social norms (Chi et al., 2023).