Antecedents of Trust and Adoption Intention toward Artificially Intelligent Recommendation Systems in Travel Planning: A Heuristic–Systematic Model

Adoption of AI Technology

AI refers to “the simulation of human intelligence processes that allows computer systems to automatically learn from experience and perform humanlike tasks to improve efficiency of daily tasks” (Li, Bonn, and Ye 2019). AI devices perform tasks based on stipulated rules and algorithms. Due to continuous technological developments, AI-based technologies nowadays can perform various tasks that are used to be performed by humans, such as driving vehicles, processing human language, recognizing faces in photos, analyzing big data, or conducting online searches (Anthes 2017). As noted by Wilson and Daugherty (2018), AI is transforming business models, customer behaviors, and even the whole world.

The adoption of AI technologies has attracted attention from scholars in various disciplines such as marketing (e.g., Goudey and Bonnin 2016), information systems (e.g., Gaudiello et al. 2016; Gessl, Schlögl, and Mevenkamp 2019), medical fields (e.g., Roham, Gabrielyan, and Archer 2012), and travel, tourism, and hospitality (e.g., Lu, Cai, and Gursoy 2019; Sunny, Patrick, and Rob 2019). Previous studies have mainly relied on traditional technology adoption theories such as theory of reasoned action (TRA), technology acceptance model (TAM), and the unified theory of acceptance and use of technology (UTAUT) to understand users’ adoption intention toward AI technologies (Huang et al. 2019; T. Zhang, Seo, and Ahn 2019). However, it is worth noting that some studies argue that traditional IT adoption models (e.g., TAM) are less suited for examining the adoption of AI technologies, which have unique features (e.g., humanlike mindset) that are significantly different from traditional technologies (Gursoy et al. 2019; Lu, Cai, and Gursoy 2019). In the travel-related service context, studies also point out that users may have mixed feelings toward the adoption of AI technologies. While users may expect higher service quality and performance because of the use of advanced technologies (Zalama et al. 2014), they may also feel isolated and uncomfortable because of the anthropomorphic features of AI technologies (Murphy, Gretzel, and Pesonen 2019). In addition, while the adoption of traditional recommendation agents relies heavily on the output quality, the adoption of artificially intelligent recommendation systems depends on agent characteristics, which allow users to assess the performance and effort they need to devote to using the smart system (Lin, Chi, and Gursoy 2020). Given that travel products are usually heterogeneous and intangible and AI devices are new to customers in the travel industry, both agent characteristics and environmental clues should be considered when examining the adoption intention (Lu, Cai, and Gursoy 2019).

While some studies in the travel and tourism field have made initial attempts to investigate travelers’ attitudes and behaviors toward AI technologies, most of them only assessed the initial adoption intention (e.g., willingness to use) (e.g., Choi, Liu, and Mattila 2019; Gursoy et al. 2019). Since the value of many AI-based technologies depends heavily on whether travelers can truly benefit from the use of AI technologies and rely on them for travel-related tasks (Choi, Liu, and Mattila 2019), studies need to further understand travelers’ postadoption attitudes and behaviors toward AI-based technologies. Among few studies that have investigated the postadoption attitudes of AI technologies, Duan, Edwards and Dwivedi (2019) suggested that cultural differences and personal values may affect the acceptance of AI support in decision making. Schneider and Leyer (2019) found that choice complexity has no impact on the willingness to delegate a strategic decision to an AI technology, and that people with low levels of situational awareness are more likely to delegate decision making to AI devices. These findings are different from previous findings in consumer research, which showed that consumers grant decision rights to other people when they perceive the choice to be difficult (Steffel and Williams 2018). These findings suggest that, when seeking help in decision making, customers hold different attitudes toward humans and AI devices. Previous studies show that people may hesitate to adopt intelligent devices in decision making because of a number of reasons, such as overconfidence in their own abilities, fear of being replaced and job losses, concerns about privacy infringements, and group pressure (Marler, Fisher, and Ke 2009; Li, Bonn, and Ye 2019).

The Heuristic–Systematic Model

The dual-process theory provides a theoretical lens for understanding how individuals process information, establish validity assessment, and subsequently perform decision making (Eagly and Chaiken 1993). One of the most prevalent models rooted in the dual-process theory is the heuristic–systematic model (HSM) (Chaiken 1980), which predicts individuals’ social judgments such as attitudes, impressions, and self-beliefs. The HSM posits two parallel routes of processing social information, which includes “the objective properties of the stimuli we think about and the properties that we bring into the perceptual experience” (Bohner, Moskowitz, and Chaiken 1995). One route is heuristic processing, in which individuals consider a few informational cues—or even a single informational cue—and form irrational judgments based on limited cues (Tam and Ho 2005). The other route is systematic processing, in which individuals consider all relevant pieces of information, elaborate on those pieces of information, and form rational judgments based on these elaborations (K. Zhang et al. 2014). Compared with heuristic information processing, systematic information processing requires the information recipient to spend more cognitive effort in evaluating pieces of information (Chaiken 1980). The HSM assumes that heuristic processing and systematic processing can occur simultaneously and even interact with each other during one’s decision-making process (K. Zhang et al. 2014).

The theoretical extensions of HSM have systematically illustrated the interactions between the two information processing routes and identified three types of effects including the additivity effect, attenuation effect, and bias effect (Chen and Chaiken 1999). The additivity effect illustrates that the heuristic processing and systematic processing can independently impact the judgmental decision making (Chaiken, Liberman, and Eagly 1989). The attenuation effect indicates that an elevated level of systematic processing may weaken the effects of heuristic processing on one’s judgmental decision making (Chaiken, Liberman, and Eagly 1989). Finally, the bias effect highlights that the heuristic processing may alter individuals’ evaluations indirectly through biasing systematic processing because heuristic cues may hinder individuals’ capability of validating systematic cues (Chaiken and Maheswaran 1994).

The HSM is widely applied to examine users’ information processing and decision making in the context of information technology use. According to Chaiken (1980), both message characteristics (e.g., validity of information) and source characteristics (e.g., popularity of a person) could impact individuals’ judgments. For instance, Ferran and Watts (2008) found that people are more influenced by the likeability of a speaker (i.e., source characteristic) than the quality of arguments (i.e., message characteristic) in videoconferences. In travel and tourism studies, the HSM is also widely applied to understand travelers’ decision-making processes such as their continuous usage of travel-related social media (Hur et al. 2017), attitudes and purchase intentions toward eco-resorts (Sparks, Perkins, and Buckley 2013), travel-related information processing (Jun and Vogt 2013), and so on. Specifically, Hlee, Lee, and Koo (2018) examined travelers’ processing of online travel information and found that travelers rely on both heuristic cues (e.g., identity, reputation, and expertise) and systematic cues (e.g., linguistic and lexical cues) in their travel-related decision making. In our study context, we adopt HSM to investigate travelers’ processing of the objective properties and the perceptual experience properties they bring into regarding the AI-based recommendation system. The outcome of such information processing can determine their judgments of whether they can trust and rely on the system for travel planning.

Perceived Risk

Perceived risk has long been recognized as central to consumers’ evaluations, choices, and behaviors in decision making (Dowling 1999). In consumer research, perceived risk is often defined in terms of uncertainty and consequences, which suggests that risk perception derives from uncertainty involved in decision making and potential negative consequences associated with decision making (Featherman and Pavlou 2003). Following this doctrine, risk perception is highly relevant to this study context because, when using AI-based recommendation systems for travel planning, travelers may not be able to determine how well the AI technology could function (i.e., uncertainty) and may be worried about faulty decisions (i.e., consequences). Risk perception is also closely related to users’ trust and adoption intention toward new and emerging technologies (Luo et al. 2010; Leung, Shi, and Chow 2020). In travel-related decision making, risk perception is also considered important because many travel-related products and services cannot be verified before actually experiencing them, rendering travelers vulnerable to faulty decisions (Sharifpour et al. 2014). Accordingly, risk perception is viewed as an important determinant of the level of cognitive effort consumers spent in the decision-making process (Dholakia 2001).

Consumers’ perception of risk may come from multiple sources (e.g., functional, physical, financial, social, and psychological) or from only one source (Campbell and Goodstein 2001). The importance of a particular type of risk depends on the specific context of decision making. For example, Adam (2015) found that perceived sociopsychological risk (e.g., impacting social standing) is least relevant for backpackers who travel to another country. In another study context, Dholakia (2001) revealed that social and psychological risk have the greatest impacts on consumers’ involvement, information seeking and dissemination in purchase decision making. In our study, we consider social risk to be the most important because travelers’ level of involvement and strategies of decision making would vary according to the social context (e.g., traveling with family and friends or traveling alone) associated with the travel planning (Bruwer and Cohen 2019). Studies have also found that social factors (e.g., opinions of others, social relations) significantly impact individuals’ travel planning process (Fotis, Buhalis, and Rossides 2011). Therefore, this study focuses on travelers’ perceived social risk in travel planning.

Hypotheses Development

Moderating Role of Perceived Risk

Dholakia (2001) reported that social risk perception derives from consumers’ involvement in decision making. According to Cheung, Xiao, and Liu (2014), under high levels of involvement, consumers are more likely to rely on systematic cues instead of heuristic cues in their decision making. In travel and tourism research, studies also found that when tourists perceive high social risks, they tend to spend more cognitive efforts in various risk-handling activities (e.g., systematic information search) in order to reduce potential negative consequences (Adam 2015). In addition, when perceived social risk is high, consumers become more wary and more conservative, which makes them more cautious about their evaluations and decisions (Dowling 1999). Under such circumstances, consumers will prefer the norm to the novel and their decisions will, thus, be less influenced by heuristic factors such as emotions and enjoyment (Campbell and Goodstein 2001).

Based on the above discussion, high levels of perceived social risk will motivate travelers to utilize more systematic cues to determine adoption intention through building cognitive trust. In contrast, low levels of social risk perceptions will lead travelers to depend more on heuristic cues to determine adoption intention through emotional trust. Thus we propose:

Research Design

The research model and proposed hypotheses were tested employing a hypothetical scenario method (Weber 1992). This approach uses vignettes that “present subjects with written descriptions of realistic situations and then request responses on a number of rating scales that measure the dependent variables of interest” (Trevino 1992). Scenario-based methods are widely used in previous studies to examine decision-making behaviors (e.g., Jun and Vogt 2013; J. Y. Park and Jang 2013). This approach is advantageous for this study because scenarios can incorporate situational details that might be important for decision making (Klepper and Nagin 1989). General survey questions that do not refer to a particular context can induce measurement errors if participants’ imagined circumstances do not match with those in real settings (Hong and Desai 2020). Therefore, scenarios provide an approach to enhance the realism of decision-making situations by providing more contextual details and ensuring that these details are consistent across respondents (Siponen and Vance 2010).

Scenario Design

A well-established AI-based travel planning application was chosen as the target application in scenarios. The scenario described a situation where the participant is going for a short trip next week and there is an AI-based recommendation system that could help him or her for travel planning. Afterwards, the main functions of the AI application were introduced through a video and pictures were used to show the process of using the AI application for making a travel plan (see Appendix A for details of the scenario). The introduction to the AI-based recommendation system shows the whole process of how travelers interact with the system and the characteristics and functions of the system, which allows travelers to gain a comprehensive understanding of the features (e.g., anthropomorphism) of the AI system.

Measurement

The measures were adopted from previous studies and the items were slightly modified to fit the study context (see Appendix B). All of the constructs were measured using a 5-point Likert scale, with 1 representing strongly disagree and 5 representing strongly agree. Disposition to trust, expertise knowledge, age, and gender were included as control variables because previous studies have demonstrated their impacts on trust and technology adoption. For example, disposition to trust, as an individual trait, has been shown to have an effect on trust development (Gefen 2000). Individuals’ expert knowledge of a technology has also been proven to impact their risk perception, which further impacts trust and adoption intention (Sharifpour et al. 2014). Sundar, Waddell, and Jung (2016) found that age will impact people’s attitudes and behavioral intentions toward AI technologies. Finally, gender was included because earlier studies have shown differences in male and female’s decision-making processes and acceptance of emerging technologies (Venkatesh, Morris, and Ackerman 2000).

Pretest

Before conducting the pretest, 5 professors and 10 doctoral students were invited to review the items and scenarios in our study. After making modifications based on their feedback, all professors and doctoral students reached a consensus that the items and scenarios were appropriate. To further validate the measurement items (Straub 1989), a pretest was conducted with 80 university students. The reliability and validity of the measurement items were tested by examining the Cronbach’s α, convergent validity, and discriminant validity. The assessments yielded satisfactory results, showing good validity and reliability of the measurement items.

Sample and Data Collection

Data were collected with the help of an online survey company in China, namely, Sojump. Sojump has millions of registered users from a wide range of demographic backgrounds and it is thus particularly useful for gathering self-reported data. The survey company helped recruiting participants with diverse backgrounds and geographical locations in China. To ensure the validity and reliability of responses, a number of validation and attention check questions were inserted into the online survey. We also included a screening question to make sure that participants had some prior experience with AI technologies so that they can have adequate understanding of the scenario. In total, 413 responses were collected and 364 were identified as valid ones. The respondents’ demographic characteristics are summarized in Table 1.

OPEN IN VIEWER

Table 1.

Demographic Profile of the Sample.

Items	Category	Total Number	%
Gender	Male	106	29.1
	Female	258	70.9
Age, years	18−28	6	1.6
	28−38	290	79.7
	38−48	40	11.0
	>48	23	6.3
Education	High school or below	15	4.1
	Diploma	21	5.8
	Bachelor	258	70.9
	Postgraduate or above	70	19.2
Travel frequency	Once a month	22	6.0
	Once every three months	92	25.3
	Once half a year	130	35.7
	Once a year	90	24.7
	Once in two years	5	1.4
	Once in more than two years	25	6.9

Data Analysis

Data were analyzed using a component-based structural equation modeling technique—partial least squares (PLS) (Chin and Marcoulides 1998). Specifically, the software package of Smart PLS 2.0 was used. PLS allows researchers to analyze the measurement and structural models simultaneously (Bollen 1989). PLS is advantageous for our study for the following reasons: (1) it makes minimal demands on sample size and the distribution of data and works with nominal, ordinal, and interval scaled variables (Chin and Marcoulides 1998); (2) PLS path models can be very complex (i.e., consist of many latent and manifest variables) without leading to estimation problems (Henseler, Ringle, and Sinkovics 2009); and (3) PLS can be used for analyzing the second-order formative construct, which is included in our research model.

Results

Assessment of the Measurement Model

First, the measurement model was tested by assessing both the convergent validity and discriminant validity. Three criteria were used to determine convergent validity (Fornell and Bookstein 1982): (1) composite reliability (CR) should be above 0.70; (2) the average variance extracted (AVE) should be at least 0.50; and all factor loadings should be greater than 0.60 and be statistically significant (i.e., more than twice the standard error). Based on previous studies (Dowell, Morrison, and Heffernan 2015; Shi and Chow 2015), cognitive trust (COT) was conceptualized as a formative second-order construct composed of multiple reflective, first-order dimensions. As shown in Table 2, all measurement items demonstrated good convergent validity.

OPEN IN VIEWER

Table 2.

Results of the Convergent Validity Analysis.

Constructs	Items	Factor Loading	Mean	Standard Deviation	CR	AVE	Cronbach’s Alpha
Performance Efficacy (PEE)	PEE1	0.755	3.418	0.827	0.92	0.622	0.897
	PEE2	0.818	3.566	0.829
	PEE3	0.838	3.527	0.828
	PEE4	0.839	3.629	0.804
	PEE5	0.774	3.464	0.758
	PEE6	0.823	3.442	0.83
	PEE7	0.655	3.734	0.781
Perceived Personalization (PEP)	PEP1	0.884	3.538	0.847	0.868	0.688	0.771
	PEP2	0.855	3.308	0.89
	PEP3	0.743	3.799	0.735
Anthropomorphism (ANT)	ANT1	0.833	2.566	1.036	0.939	0.794	0.913
	ANT2	0.919	2.236	0.956
	ANT3	0.926	2.08	0.934
	ANT4	0.884	2.047	0.971
Social influence (SOI)	SOI1	0.668	3.404	0.855	0.861	0.609	0.782
	SOI2	0.831	3.794	0.711
	SOI3	0.841	3.687	0.709
	SOI4	0.77	3.769	0.669
Cognitive trust of benevolence (CTB)	CTB1	0.894	3.382	0.869	0.902	0.755	0.838
	CTB2	0.859	3.621	0.771
	CTB3	0.853	3.5	0.901
Cognitive trust of competence (CTC)	CTC1	0.818	3.437	0.859	0.909	0.714	0.866
	CTC2	0.86	3.643	0.715
	CTC3	0.887	3.599	0.802
	CTC4	0.813	3.596	0.802
Cognitive trust of integrity (CTI)	CTI1	0.887	3.668	0.762
	CTI2	0.894	3.514	0.886	0.939	0.795	0.914
	CTI3	0.923	3.61	0.844
	CTI4	0.86	3.613	0.827
Emotional trust (EMT)	EMT1	0.911	3.376	0.774	0.924	0.803	0.877
	EMT2	0.864	3.382	0.789
	EMT3	0.912	3.349	0.797
Adopt as a decision aid (AID)	AID1	0.904	3.794	0.675	0.896	0.743	0.826
	AID2	0.839	3.654	0.724
	AID3	0.841	3.909	0.716
Adopt as a delegated agent (ADA)	ADA1	0.951	2.739	0.927	0.948	0.902	0.891
ADA2	0.948	2.717	0.962

Discriminant validity was assessed by analyzing whether the square root of the AVE values of each construct (i.e., the diagonal entry of each column) is larger than its correlations with other constructs (Chin and Marcoulides 1998). As shown in Table 3, the results indicated good discriminant validity. The potential multicollinearity issues were assessed using the variance inflation factor (VIF) test. A VIF value above 10 indicates multicollinearity problem (Harter, Schmidt, and Hayes 2002). Results showed that the VIF values for all independent variables ranged from 1.282 to 2.198, thus eliminating the threat of multicollinearity.

OPEN IN VIEWER

Table 3.

Correlations of Latent Variables for the First-Order Constructs.

	PEE	PEP	ANT	SOI	CTB	CTC	CTI	EMT	AID	ADA
PEE	0.788
PEP	0.481	0.830
ANT	0.287	0.272	0.891
SOI	0.335	0.300	0.227	0.780
CTB	0.467	0.449	0.285	0.311	0.869
CTC	0.423	0.453	0.156	0.430	0.514	0.845
CTI	0.471	0.348	0.198	0.265	0.491	0.500	0.891
EMT	0.345	0.495	0.286	0.431	0.535	0.466	0.470	0.896
AID	0.497	0.396	0.133	0.430	0.480	0.450	0.519	0.470	0.862
ADA	0.353	0.292	0.364	0.260	0.291	0.340	0.278	0.455	0.286	0.950

Note: Diagonal elements are the square root of AVE for each construct and the off-diagonal elements are the correlations between constructs.

Common method bias

Harman’s one-factor analysis was used to statistically test the severity of common method variance (CMV). All of the measures used in this study were subjected to a principal components analysis. Results indicated that the first factor accounted for 31.147% of the total variance, which suggested no single factor dominates the variance explained (see appendix C). By adding a common method factor in the research model, the CMV was further tested and the variance of each indicator was estimated using its principal construct and method factor (Liang et al. 2007). As shown in Table 4, the substantive factor loading (R1) of each item was significant at p <.001 and was much larger than the corresponding method factor loading (R2). Most of the method factor loadings were not significant at p <.05. Therefore, it was concluded that common method bias was not a serious concern in this study.

OPEN IN VIEWER

Table 4.

Common Method Bias Analysis.

Constructs	Item	Substantive Factor Loading (R1)	R12	Method Factor Loading (R2)	R22
Performance Efficacy (PEE)	PEE1	0.769***	0.592	−0.044	0.002
	PEE2	0.822***	0.676	0.095	0.009
	PEE3	0.841***	0.707	−0.100*	0.010
	PEE4	0.838***	0.703	−0.434***	0.188
	PEE5	0.773***	0.598	−0.283***	0.080
	PEE6	0.818***	0.670	0.021	0.000
	PEE7	0.640***	0.409	−0.021	0.000
Perceived personalization (PEP)	PEP1	0.880***	0.774	−0.402***	0.161
	PEP2	0.869***	0.755	0.000	0.000
	PEP3	0.733***	0.537	−0.131	0.017
Anthropomorphism (ANT)	ANT1	0.843***	0.711	0.001	0.000
	ANT2	0.920***	0.847	0.437***	0.191
	ANT3	0.924***	0.854	0.009	0.000
	ANT4	0.875***	0.766	0.428***	0.184
Social influence (SOI)	SOI1	0.661***	0.437	−0.115	0.013
	SOI2	0.839***	0.703	0.004	0.000
	SOI3	0.851***	0.724	−0.049	0.002
	SOI4	0.757***	0.573	0.082	0.007
Cognitive trust of benevolence (CTB)	CTB1	0.738***	0.544	−0.013	0.000
	CTB2	0.728***	0.529	0.039	0.001
	CTB3	0.743***	0.552	−0.06	0.004
Cognitive trust of competence (CTC)	CTC1	0.681***	0.464	−0.013	0.000
	CTC2	0.690***	0.476	0.364**	0.133
	CTC3	0.678***	0.459	0.154	0.024
	CTC4	0.636***	0.404	0.024	0.001
Cognitive trust of integrity (CTI)	CTI1	0.800***	0.640	−0.177**	0.031
	CTI2	0.754***	0.568	0.400***	0.160
	CTI3	0.801***	0.641	0.074	0.005
	CTI4	0.800***	0.640	−0.200**	0.040
Emotional trust (EMT)	EMT1	0.908***	0.824	0.111*	0.012
	EMT2	0.870***	0.757	−0.360***	0.130
	EMT3	0.909***	0.827	0.005	0.000
Adopt as a decision aid (AID)	AID1	0.906***	0.821	0.043	0.002
	AID2	0.828***	0.685	0.002	0.000
	AID3	0.851***	0.724	−0.011	0.000
Adopt as a delegated agent (ADA)	ADA1	0.950***	0.902	0.042	0.002
	ADA2	0.949***	0.901	0.049	0.002
Average (AVE)		0.807	0.695	−0.001	0.038

*

p < .05; ^**p < .01; ^***p < .001.

Assessment of the structural model

Figure 2 presents the results of the structural model analysis, indicating that most of the hypotheses were supported except for hypothesis 2a. The two systematic cues of performance efficacy (β = 0.350, p < .001) and perceived personalization (β = 0.140, p < .01) were both positively related to cognitive trust, supporting hypotheses 4a and 4b. The two heuristic cues of anthropomorphism (β = 0.198, p < .001) and social influence (β = 0.386, p < .001) were both positively related to emotional trust, indicating support for hypotheses 5a and 5b. Cognitive trust (β = 0.476, p < .001) and emotional trust (β = 0.241, p < .001) were found to be positively associated with the intention to adopt AI-based recommendation systems as a decision aid, validating hypotheses 1a and 1b. Results also indicated that, compared with emotional trust, cognitive trust exerts stronger impact on the intention to adopt AI-based recommendation systems as a decision aid, supporting hypothesis 1c. Moreover, while emotional trust significantly influenced intention to adopt AI-based recommendation systems as a delegated agent (β = 0.300, p < .001), the impact of cognitive trust was not significant (β = 0.124, p > .10), supporting hypothesis 2b and rejecting hypothesis 2a. Results also indicated that emotional trust has a stronger impact on intention to adopt as a delegated agent than cognitive trust, supporting hypothesis 2c. Finally, emotional trust was positively associated with cognitive trust (β = 0.365, p < .001), supporting hypothesis 3. For control variables, age (β = 0.125, p < .01), gender (β = 0.105, p < .05), and disposition to trust (β = 0.185, p < .001) had significant impacts on intention to adopt as a delegated agent. The R² values showed that the model explained a sufficient variance in cognitive trust (53.77%), emotional trust (22.28%), intention to adopt AI-based recommendation systems as a decision aid (47.18%), and intention to adopt AI-based recommendation systems as a delegated agent (29.36%).

OPEN IN VIEWER

Figure 2.

Results of structural equation modeling.

Phase 1: Qualitative Assessment

Before we quantitatively examine the moderating role of perceived risk, a focus group was conducted to gain a deeper understanding of the relevance of perceived risk regarding travelers’ adoption of AI-based recommendation systems for travel planning.

Phase 2: Quantitative Assessment

Results

First, an independent samples t-test was conducted to assess the manipulation of perceived social risk. Results showed that participants perceived lower social risk in the low-risk scenario (M_low = 3.216, SD_low = 0.922) and perceived higher social risk in the high-risk scenario (M_high = 4.022, SD_high = 0.560). The levels of perceived social risk was significantly different across the two scenarios (t = 6.404, p < .001), suggesting successful manipulation. The measurement model assessment showed satisfactory convergent validity and discriminant validity (as shown in Table 5 and 6).

OPEN IN VIEWER

Table 5.

Results of the Convergent Validity Analysis in Study 2.

Constructs	Items	Factor Loading	Mean	Standard Deviation	CR	AVE	Cronbach’s Alpha
PEE	PEE1	0.760	3.418	0.827	0.891	0.540	0.858
	PEE2	0.711	3.566	0.829
	PEE3	0.780	3.527	0.828
	PEE4	0.696	3.629	0.804
	PEE5	0.749	3.464	0.758
	PEE6	0.777	3.442	0.830
	PEE7	0.664	3.734	0.781
PEP	PEP1	0.871	3.538	0.847	0.857	0.668	0.750
	PEP2	0.818	3.308	0.890
	PEP3	0.759	3.799	0.735
ANT	ANT1	0.917	2.566	1.036	0.918	0.737	0.902
	ANT2	0.898	2.236	0.956
	ANT3	0.838	2.080	0.934
	ANT4	0.774	2.047	0.971
SOI	SOI1	0.781	3.404	0.855	0.887	0.662	0.830
	SOI2	0.810	3.794	0.711
	SOI3	0.851	3.687	0.709
	SOI4	0.812	3.769	0.669
CTB	CTB1	0.835	3.382	0.869	0.874	0.699	0.785
	CTB2	0.844	3.621	0.771
	CTB3	0.829	3.500	0.901
CTC	CTC1	0.824	3.437	0.859	0.922	0.748	0.888
	CTC2	0.874	3.643	0.715
	CTC3	0.904	3.599	0.802
	CTC4	0.857	3.596	0.802
CTI	CTI1	0.887	3.668	0.762
	CTI2	0.880	3.514	0.886	0.940	0.798	0.916
	CTI3	0.910	3.610	0.844
	CTI4	0.896	3.613	0.827
EMT	EMT1	0.912	3.376	0.774	0.927	0.810	0.882
	EMT2	0.860	3.382	0.789
	EMT3	0.926	3.349	0.797
AID	AID1	0.897	3.794	0.675	0.911	0.773	0.853
	AID2	0.853	3.654	0.724
	AID3	0.887	3.909	0.716
ADA	ADA1	0.972	2.739	0.927	0.966	0.935	0.931
	ADA2	0.962	2.717	0.962
PSR	PSR1	0.685	3.432	1.004	0.900	0.753	0.866
	PSR2	0.986	3.757	0.973
	PSR3	0.904	3.730	0.966

OPEN IN VIEWER

Table 6.

Correlations of Latent Variables for the First-Order Constructs in Study 2.

	PEE	PEP	ANT	SOI	CTB	CTC	CTI	EMT	AID	ADA	PSR
PEE	0.735
PEP	0.523	0.817
ANT	0.016	−0.091	0.859
SOI	0.403	0.357	0.092	0.814
CTB	0.448	0.422	−0.010	0.451	0.836
CTC	0.426	0.451	−0.093	0.502	0.540	0.865
CTI	0.431	0.314	0.120	0.456	0.434	0.555	0.893
EMT	0.491	0.502	0.126	0.466	0.454	0.509	0.380	0.900
AID	0.452	0.380	0.003	0.565	0.412	0.520	0.418	0.464	0.879
ADA	0.251	0.127	0.144	0.226	0.259	0.350	0.301	0.379	0.174	0.967
PSR	0.032	−0.019	0.019	0.031	0.020	−0.025	0.456	0.466	0.069	0.004	0.868

Next, moderated mediation analyses were conducted using PROCESS with a bootstrap method (Luna-Cortés, López-Bonilla, and López-Bonilla 2019) to examine the hypotheses, with systematic cues and heuristic cues as independent variables, cognitive trust and emotional trust as mediators, adoption as a decision aid, and adoption as a delegated agent as dependent variables, and perceived social risk (1 = high, 2 = low) as the moderator. In order to provide meaningful interpretations of predictors’ simple effects, variables were mean centered.

Results of the moderation effects are presented in Table 7. The upper part of the table shows the moderated indirect effects when adoption as a decision aid serves as the dependent variable. For systematic paths, the conditional indirect effect of performance efficacy mediated by cognitive trust were significant at both high and low level of perceived risk, with a decreasing magnitude of effects (t = 10.04, p < .001) as the level of perceived risk decreases. Meanwhile, the effect of perceived personalization through cognitive trust was significant at high level of perceived risk (95% CI: 0.157, 0.492) but was not significant at the low level (95% CI: –0.025, 0.171). For heuristic paths, the conditional indirect effects of anthropomorphism and social influence mediated by emotional trust were both significant at low level of perceived risk (95% CI: 0.018, 0.218; 95% CI: 0.040, 0.228), but were not significant at the high level (95% CI: –0.036, 0.040; 95% CI: –0.008, 0.119). Thus, Hypotheses 6a, 6b, 7a, and 7b were supported.

OPEN IN VIEWER

Table 7.

Indirect Effects of Systematic Cues and Heuristic Cues on Adoption Intentions at High and Low Levels of Perceived Risk Conditions.

DVs	Mediators	IVs	Level of Perceived Risk	Conditional Indirect Effects
						Boot 95% CI
				Effect	Boot SE	Boot LL	Boot UL
Adopt as a decision aid	Cognitive trust	Performance efficacy	High	0.252	0.082	0.117	0.441
			Low	0.133	0.057	0.027	0.248
		Perceived personalization	High	0.307	0.085	0.157	0.492
			Low	0.067	0.049	−0.025	0.171
	Emotional trust	Anthropomorphism	High	−0.006	0.018	−0.036	0.040
			Low	0.115	0.051	0.018	0.218
		Social influence	High	0.037	0.033	−0.008	0.119
			Low	0.125	0.049	0.040	0.228
Adopt as a delegated agent	Cognitive trust	Performance efficacy	High	0.363	0.14	0.109	0.653
			Low	0.182	0.119	−0.016	0.454
		Perceived personalization	High	0.161	0.122	−0.050	0.430
			Low	0.226	0.125	−0.004	0.489
	Emotional trust	Anthropomorphism	High	−0.012	0.04	−0.095	0.073
			Low	0.153	0.071	0.025	0.307
		Social influence	High	0.127	0.085	0.002	0.323
			Low	0.263	0.090	0.095	0.445

Note: DV, dependent variable; IV, independent variable; CI, confidence interval; LL, lower limit; UL, upper limit.

The lower part of Table 7 shows the moderated indirect effects when adoption as a delegated agent serves as the dependent variable. For systematic paths, the conditional indirect effect of performance efficacy mediated by cognitive trust was significant at high level of perceived risk (95% CI: 0.109, 0.653) but was not significant at the low level (95% CI: –0.016, 0.454). The conditional indirect effect of perceived personalization mediated by cognitive trust was not significant at both levels of perceived risk. For heuristic paths, the conditional indirect effect of anthropomorphism mediated by emotional trust was significant at low level of perceived risk (95% CI: 0.025, 0.307) but was not significant at the high level (95% CI: –0.095, 0.073). The conditional indirect effects of social influence through emotional trust were significant at both levels of perceived risk, with a decreasing magnitude of effects (t = 9.38, p < .001) as the level of perceived risk increases. These findings provided support for hypotheses 6c, 7c, and 7d, while hypothesis 6d was rejected.

Discussion of Results

Study 1 and 2 jointly provide strong evidence that, when determining adoption intention toward AI-based recommendation systems, systematic cues that require more cognitive effort to process enhance individuals’ rational cognition, while heuristic cues that require less cognitive effort facilitate individuals’ affective evaluations. Moreover, the strength of impacts of systematic and heuristic cues on adoption intention is dependent on the perceived risk associated with travel planning.

Findings of study 1 provide support for most of the hypotheses regarding the impacts of systematic and heuristic cues on adoption intentions through cognitive and emotional trust. Unlike the additivity effect, which is fully confirmed, the attenuation effect between systematic processing and heuristic processing is partially supported. Contrary to our prediction, when determining the intention to adopt an AI-based recommendation system as a delegated agent, only the impact of emotional trust is significant. This can be explained by the unique characteristics of AI-based technologies as compared with traditional technologies that does not involve intelligent features and personal interactions (Gursoy 2019). As suggested by the uncanny valley theory, the humanlike features of AI applications can generate fear and anxiety among individuals (Murphy, Gretzel, and Pesonen 2019). Therefore, compared with cognitive trust, which depends heavily on rational evaluations, emotional trust that derives from individuals’ positive inner feelings has a stronger predictive power toward the intention to completely rely on AI-based recommendation systems for travel planning.

Results also confirm the bias effect in the HSM by showing that emotional trust has a positive impact on cognitive trust, suggesting that heuristic factors can alter travelers’ adoption intentions indirectly through influencing their systematic processing (Chaiken and Maheswaran 1994). These results reveal the critical role of emotional trust in determining a traveler’s level of dependence on AI-based recommendation systems. Since travelers’ knowledge about AI-based recommendation systems (e.g., how it functions and on whose behalf) is usually low and ambiguous, emotional trust could serve as a trigger to reduce travelers’ anxiety and negative emotions toward the adoption of AI-based recommendation systems in travel planning. It is also worth noting that the variances explained in constructs of emotional trust and adoption as a delegated agent are relatively small as compared to other constructs. This suggests that, when determining full dependence on an AI-based recommendation system, the heuristic processing alone may not be sufficient. Other psychological, social, and moral aspects could be further considered as suggested by previous literature (Gessl, Schlögl, and Mevenkamp 2019; Schneider and Leyer 2019).

Results of study 2 further suggest that when travelers perceive high social risk, they tend to rely more on systematic cues to determine their adoption intentions toward AI-based recommendation systems through building cognitive trust. However, when the level of perceived social risk is low, travelers are more likely to adopt the approach that require the least amount of effort and depend more on heuristic cues to form adoption intention through establishing emotional trust. However, the indirect effects of perceived personalization on adoption as a delegated agent through cognitive trust are not significant at both levels of risk perception. This may be explained by the fact that personalization is regarded as a basic feature of AI technologies and the dependence on personalization to form cognitive trust and delegated adoption intention may be weakened by other factors under either low or high risk perceptions (Schneider and Leyer 2019).

Implications for Research

This research contributes to the extant literature with several important theoretical implications. First, while previous travel and tourism research mainly focuses on the initial acceptance of AI technologies by travelers (e.g., Gursoy et al. 2019; Sunny, Patrick and Rob 2019; Lin, Chi, and Gursoy 2020), this study extends the research line by examining travelers’ postadoption intention (i.e., AI-assisted decision making) of AI-based recommendation systems in travel planning. Compared with initial adoption, understanding postadoption experiences and behaviors are critical in facilitating individuals’ continuance usage (Hur et al. 2017; Li, Bonn, and Ye 2019). While previous travel and tourism research focuses heavily on using AI technology for forecasting (e.g., Yu and Schwartz 2006; Law et al. 2019; Höpken et al. 2020), this research is a response to the calls for providing new insights on travelers’ use of technology for travel planning due to rapid technological advancements (Xiang et al. 2015).

Second, based on the dual process theory and HSM, this study identifies two different paths that affect travelers’ cognitive trust and emotional trust toward the AI-based recommendation systems, respectively. Although the dual-process theory has been widely used to explicate travelers’ information-processing patterns (e.g., Jun and Vogt 2013; Sparks, Perkins, and Buckley 2013; M. J. Kim and Petrick 2020), little research has specifically examined the potential relevance of the HSM in understanding travelers’ adoption of AI technologies in decision making. Furthermore, while previous studies mainly focus on users’ cognitive evaluations toward AI technologies (e.g., Gessl, Schlögl, and Mevenkamp 2019; Lin, Chi, and Gursoy 2020) and travelers’ cognitive trust in travel settings (e.g., Gursoy et al. 2016; Agag and El-Masry 2017; S. Park and Tussyadiah 2019), this study expands our understanding by investigating the impacts of both travelers’ cognitive assessments and affective evaluations utilizing the trust-centered HSM. This study also investigates the joint effects of both systematic and heuristic evaluations on trust development and adoption intentions, and interactions (i.e., additivity, attenuation, and bias effects) between systematic and heuristic routes.

Third, while some previous studies have considered the importance of trust in the acceptance of AI technologies (e.g., Gaudiello et al. 2016), this study is among the first to examine the roles of different types of trust in AI technology adoption. By doing so, we also respond to the calls for including both cognitive and affective trust in analyzing tourists’ behavioral intentions (M. Kim and Kim 2020). Findings reveal that including both cognitive trust and emotional trust to AI adoption models especially in the postadoption decision-making context is of great value. Moreover, this study offers new insights into the relationship between cognitive trust and emotional trust in determining travelers’ level of dependence on AI-based recommendation systems for travel planning. While cognitive trust may have a stronger impact on users’ adoption intention toward conventional technologies, this study reveals that emotional trust predicts a higher level of dependence in the context of AI technology adoption. As also suggested by Kuo and Wu (2012), individuals’ final decision to accept or reject the use of AI devices during service encounters are likely to be determined by their emotions.

Finally, this study has important implications for understanding travelers’ AI-assisted decision making under different levels of risk perceptions. In the travel and tourism context, perceived risk is considered vital due to the ambiguity and uncertainty involved in decision making (Sharifpour et al. 2014; S. Park and Tussyadiah 2017). Though some studies have suggested that specific context characteristics should be considered when investigating AI technology adoption (Schneider and Leyer 2019), little empirical research has examined perceived risks associated with using AI technologies. Findings suggest that travelers depend more on the systematic route rather than the heuristic route, such that they become more wary and conservative in their decision making, when travel planning using AI-based recommendation systems involves higher risk. While findings indicate that systematic processing and heuristic processing conjointly determine travelers’ adoption intentions, this study also uncovers the boundary condition under which travelers’ dependence on systematic and heuristic cues varies. By doing so, this study contributes to the current AI adoption literature by revealing the relative importance of different informational cues under different decision-making contexts.

Implications for Practice

Findings of this research also generate valuable implications for practitioners and managers. First, designers of AI-based recommendation systems may leverage the impacts of AI technology characteristics through the lens of our research model. According to our findings, practitioners should improve the functional design of AI-based recommendation systems through enhancing systematic cues such as including patented technology, collecting and analyzing personal preferences, and providing high-quality information in order to enhance travelers’ cognitive evaluations. In terms of heuristic cues, practitioners should pay attention to humanlike features of AI-based recommendation systems so that travelers feel a sense of human contact and generate personal feelings during their interactions with the system. They should also spend considerable effort in showcasing the acceptance and use of AI technologies for travel-related tasks in society to improve the effects of social influence. Practitioners should also design the AI-based recommendation systems in a way that it generates a sense of comfort and security in order to facilitate emotional trust.

Moreover, findings provide guidance for the implementation of AI-based recommendation systems in the travel industry. Given that nowadays many companies in the travel industry adopt AI-based recommendation systems to assist decision making, managers should be aware of the most important factors that determine the actual value of the system. For example, companies may need to provide educational materials and instructions in order to enhance customers’ knowledge about the personalized functions and performance of the system. Further, managers should pay close attention to not only the functional qualities of the system but also heuristic cues that drive positive emotions. For example, hotels, airlines, and the travel agencies could use robotic AI devices with humanlike features to provide recommendation services for travelers. In order to stimulate positive social influence, managers can establish a platform to encourage users of AI-based recommendation systems to share their experiences.

Last but not the least, when designing and implementing AI-based recommendation systems for travel planning, practitioners should consider travelers’ risk perceptions in decision making. Our findings clearly indicate that in situations where the perceived risk is high, travelers tend to depend more on systematic cues to form cognitive trust and adoption intentions. Therefore, it is important to identify high-risk travel decisions using data collected from travelers before generating recommendations from the system. For example, under high-risk conditions, the AI-based recommendation systems should provide more details of the recommendation-generation process and highlight the reliability and validity of recommendations in order to facilitate the cognitive evaluations by travelers. Moreover, more flexibility should be offered in system operations so that travelers can modify the recommended travel plans according to their own intuition and evaluation in order to reduce risk in decision making.

Limitations and Future Research

Although the contribution of this study is evident, it has certain limitations that could be addressed by future studies. First, given that AI-based recommendation systems are adopted for various tasks, the generalizability of our findings may be limited because this study only focuses on travel planning as the research context. Future research may consider extending the findings to other contexts of AI-assisted decision making. The generalizability of study 2 may also be limited because we invited university students for experimentation. Future studies can include different customer groups in order to further validate our findings. Second, although this study empirically examined the moderating effects of social risk perception, the moderating effects of other dimensions of risk (e.g., financial risk) and other possible factors (e.g., consumer involvement and expertise) were not tested. Future research is thus recommended to expand our understanding of the boundary conditions that impact travelers’ evaluations and decision-making processes with AI-based recommendation systems. Also, other factors (e.g., media coverage about AI technology) shaping public opinion and perceived risk could be further analyzed. Finally, findings show that the model explains 22.3% of variances in emotional trust and 29.4% of variances in adoption intention as a delegated agent. Although the R² value above 10% is considered as adequate by previous studies (Falk and Miller 1992), these two moderate R² values indicate that some other important predictors may be missing in the proposed research model. For example, according to the model of AI device use and acceptance (AIDUA) (Gursoy et al. 2019), travelers’ positive and negative emotions (e.g., pleasure and anxiety) toward the AI-based recommendation system can be considered when examining their delegation intention. Therefore, future studies are encouraged to enrich our understanding of the impacting factors of these two constructs. Moreover, future studies are encouraged to examine travelers’ actual usage of AI-based recommendation systems instead of measuring behavioral intentions.