Chatting vs clicking: how does AI interaction style affect travel decisions? a comparative study of generative AI and GUI-based platform

The impact of Artificial Intelligence (hereinafter AI) has become increasingly widespread in various industries (Han et al., 2025). In particular, the tourism sector benefits from AI-based services, which offer several advantages for travelers (World Economic Forum, 2023). First, they use extensive destination information to facilitate decision-making. Second, they mitigate uncertainty by offering real-time and predictive data (Dwivedi et al., 2024). Third, based on personalized recommendations, they analyze user preferences to create customized travel plans (Hlee et al., 2025). Consequently, the AI travel planning assistant (hereinafter AI-TPA) is presently considered an essential tool for travel plans. According to the GuideGeek (2024) survey, more than one in five (i.e. over 22.8%) respondents had used these platforms to plan or assist with their travel.

Despite these numerous advantages, AI-based services still present certain limits for people seeking to use them (Statista, 2025). Previous studies have indicated that AI-based services exhibit weaknesses, such as providing inaccurate information and lacking human ambiance (Han et al., 2025). Furthermore, Lalicic and Weismayer (2021) explained that users are less willing to use AI-based services because of the stress associated with learning to use them. Further empirical studies are necessary to enhance the understanding of how AI-based services/technologies influence user acceptance.

With the rapid advancement of AI-TPA, previous studies have explored key factors influencing tourist behavior (Gu et al., 2019; Kanhed et al., 2024; Li et al., 2025). These recent studies have commonly identified the following factors as catalysts for AI-TPA adoption, including usefulness, convenience, personalization capabilities, information quality and efficacy. Prior studies on AI-based technology have widely applied validated frameworks such as TPB, task-technology fit (TTF), Technology Acceptance Model (TAM) and Innovation Diffusion Theory (IDT) to predict the relationship between causal and outcome variables (Foroughi et al., 2024; Hwang et al., 2024; Teng et al., 2024; Wang and Lin, 2019). Researchers argue that more than one theoretical model is required to explain complex technology adoption behavior, particularly for innovative services (Koenig, 2025). Specifically, TTF theory, proposed by Goodhue and Thompson (1995), suggests that task performance improves when technological functionalities closely align with task requirements. However, TTF alone provided limited insight into why users initially perceived a system as user-friendly or worth adopting – questions that were central to understanding AI-TPA acceptance, particularly during its early stages (Dhiman and Jamwal, 2023). The TAM, proposed by Davis (1989), tackled this issue by emphasizing users’ cognitive evaluations to explain their attitudes and behavioral intentions toward new technologies (Yen et al., 2010). Likewise, Dishaw and Strong (1999) demonstrated that integrating TTF and TAM constructs provides greater explanatory power than either model alone, as they encompass complementary aspects of technology acceptance, including rational task-performance fit and attitudinal factors. Additionally, they suggested that this integration addresses the gap in which user perceptions of usefulness may be disconnected from actual performance impact. Although TAM is an effective framework for investigating how consumers adopt new technologies, it may be inadequate for explaining consumers’ behavioral intentions across all contexts (Liu et al., 2024a, 2024b). Several studies have emphasized that promoting consumer adoption of emerging technologies requires a comprehensive understanding of the diffusion of innovative products and services within society (Lu et al., 2025). The IDT model effectively analyzes the adoption process and highlights the elements that contribute to one’s decision to embrace or discard technological innovations (Sujood et al., 2025). In this regard, the integration of TTF, TAM and IDT offers a robust framework for comprehending tourist attitudes and behavioral intentions toward AI-TPA, thereby enhancing the explanatory power of technology adoption within tourism contexts (Cai et al., 2023; Gu et al., 2019; Ho et al., 2021). However, investigations into AI-based travel planning applications from the integrated viewpoint of these three theoretical frameworks are scarce.

Significantly, favorable attitudes toward technology enhance self-confident efficacy in using new technologies (Fakfare et al., 2023). Self-confident efficacy refers to individuals’ confidence in their capacity to effectively execute specific tasks using technology (Sun et al., 2024). In addition, self-confident efficacy has been recognized as a primary precursor of technology adoption across several contexts (Ahmed et al., 2025; Hlee et al., 2025; Montag et al., 2023). For instance, users with higher levels of self-confident efficacy perceive themselves as competent in their interactions with automated technology such as AI-TPA (Montag et al., 2023). Moreover, users are more inclined to manage AI-recommended travel schedules more efficiently (Hlee et al., 2025).

Finally, unlike previous studies, this study seeks to examine the moderating role of two distinct forms of AI-TPA. AI-TPA can be categorized into generative AI and graphical user interfaces (hereafter GUI). Generative AI enables users to directly pose inquiries or submit prompts in an input box (Kanhed et al., 2024). A GUI-based platform enables users to plan travel by clicking on graphic elements (such as buttons and maps) on an app or website (Cybulski and Horbiński, 2020). Thus, differences such as interface design may significantly impact users’ attitudes and intentions (Chenchu et al., 2025). For example, generative AI offers flexible conversational responses via natural language processing; however, it may occasionally generate inaccurate or inconsistent information (Kim et al., 2019). Conversely, GUI-based platforms offer structured navigation through predefined options, although it may exhibit limited flexibility for personalized travel plans (Liu et al., 2024a, 2024b).

In summary, this study uses a comprehensive framework that integrates TTF, TAM and IDT. Specifically, it aims to:

• predict user attitudes using the integrated framework;

• examine the effects of attitudes on self-confidence efficacy and intentions to use;

• investigate attitudes’ mediating role between theoretical antecedents and outcomes; and

• assess the moderating effect of AI-TPA types (generative AI vs GUI-based platforms).

AI refers to “computer systems capable of executing tasks that typically require human intelligence, such as learning, reasoning, and problem-solving” (NASA, 2024). AI-based services are extensively used to assist consumers in their daily lives, enhancing convenience in the tourism industry (Lalicic and Weismayer, 2021). For example, AI is used in various applications, including virtual assistants (e.g. Siri and Alexa), recommendation engines, self-driving cars and travel planning (McKinsey, 2024). Recent studies have demonstrated that AI-TPA enhances user’s travel experience through service optimization, including content-based filtering and personalization (Hlee et al., 2025). According to the 2025 Statista report, one-third of respondents regard AI-TPA as valuable tools for discovering new information, especially for their ability to streamline the decision-making process.

TTF, developed by Goodhue and Thompson (1995), emphasized the interaction between task, technology and individual. A key premise of TTF is that technology and tasks enhanced performance only when matched to the task requirements (Yen et al., 2010). TTF comprises three sub-variables: task characteristics, technology characteristics and task-technology fit. Task characteristics refers to “the actions conducted by individuals in transforming inputs into outputs” (Goodhue and Thompson, 1995, p. 216). Technology characteristics denote “tools utilized by individuals in performing their tasks” (Goodhue and Thompson, 1995, p. 216). Finally, task technology fit is described as “the extent to which a technology assists individuals in performing their portfolio of tasks” (Goodhue and Thompson, 1995, p. 216).

In tourism and hospitality, TTF has been effectively applied to explain technology adoption behaviors associated with hotel self-service kiosks, digital museums and service chatbots (Dhiman and Jamwal, 2023). In particular, TTF is crucial in shaping attitudes toward AI-based services, where personalization is a core feature. Prior research confirmed that task and technology characteristics significantly impact TTF and attitudes (Dhiman and Jamwal, 2023). Therefore, the following hypotheses were proposed:

The concept of TAM, as proposed by Davis (1989), elucidates the mechanisms via which users embrace new technologies. TAM comprises two key factors such as perceived ease of use and usefulness. Perceived ease of use is defined as “the extent to which a person believes that using a particular system is possible without effort,” whereas perceived usefulness refers to “the degree to which a person believes that using a particular system would improve his or her job performance” (Davis, 1989, p. 320). In travel and tourism contexts, TAM is extensively acknowledged for explaining how consumers’ perceptions of new technologies influence their attitudes and intentions to use them (e.g. smart hotels, travel apps, AI chatbots and travel assistant apps) (Du et al., 2025; Foroughi et al., 2024; Li et al., 2025; Liu et al., 2024a, 2024b). Research indicates that individuals are more willing to adopt innovative technologies when they believe these technologies effectively fulfill users’ needs and are user-friendly. For example, Foroughi et al. (2024) discovered that perceived ease of use and usefulness significantly influenced users’ attitudes toward travel apps. They demonstrated that when travelers could effortlessly learn and use the app with minimal effort, they were more inclined to perceive the app as useful for travel planning action, thereby fostering favorable attitudes toward continuous use.

IDT, as proposed by Rogers (1995), is conceptualized as a framework that describes why individuals adopt novel ideas or technologies. IDT encompasses five categories: relative advantage, compatibility, complexity, trialability and observability, and examines whether an individual or an organization influences the adoption or non-adoption of those sub-factors of IDT (Rogers, 1995). Relative advantage is denoted as the degree to which an innovation is perceived as superior to the idea it supersedes (Rogers, 1995). Compatibility is defined as the degree to which innovation is perceived as being consistent with existing values, past experiences and needs of potential adopters (Rogers, 1995). Complexity refers to the degree to which an innovation is perceived as relatively complex to understand and to use (Rogers, 1995). For instance, consumers adopt innovations, such as new concepts, novel applications and advanced products, as they diffuse over time (Han et al., 2025). This process transpired when users evaluate advantages in comparison to conventional alternatives (Sujood et al., 2025). Thus, innovative technology, such as AI-TPA, is less likely to be adopted if it is perceived as overly complex or incompatible. Prior researchers have applied IDT sub-factors to examine adoption of innovative technologies in tourism contexts, including ChatGPT, VR tourism, travel apps and smart tourism platforms (Han et al., 2025; Lu et al., 2025; Lim et al., 2022). Lim et al. (2022) affirmed that relative advantages, compatibility and complexity were significant determinants that influenced attitudes toward mobile travel applications.

More importantly, previous studies focused on consumers’ attitudes through the integrated TAM and IDT framework in various fields associated with technology-based service (Gu et al., 2019; Sujood et al., 2025; Terrah et al., 2024). TAM and IDT both rely on the assumption that ease of use and perceived complexity are fundamental determinants of technology acceptance (Venkatesh et al., 2003). In other words, users are less inclined to adopt even innovative technology when they anticipate excessive effort or difficulty. By contrast, users are more inclined to accept technology that provides distinct performance benefits while being user-friendly (Dishaw and Strong, 1999). In this context, several empirical studies have shown that integrating TAM and IDT enhances the prediction of technology-related attitudes and behavioral intentions. For example, Gu et al. (2019) indicated that the integrated TAM and IDT framework demonstrated higher explanatory power than single-theory models in the context of mobile tourism shopping. Recently, Terrah et al. (2024) validated the effectiveness of integrating TAM and IDT in predicting hotel guests’ adoption of AI-enabled technologies. Based on the above discussion, we proposed the following hypothesis:

Attitude is denoted as “the degree to which a person possesses a favorable or unfavorable evaluation or appraisal of the behavior” (Ajzen, 1991, p. 188). In other words, attitude is a fundamental concept in the consumer decision-making literature, as it reflects the positive or negative evaluation of a service or product. Existing studies have demonstrated the significance of attitude and its significant correlation with diverse variables and intentions in technology-based service-related tourism settings (Ahmed et al., 2025; Fakfare et al., 2023; Sun et al., 2024). In addition, Lim et al. (2022) revealed that attitude significantly mediates the nexus between multi-dimensional factors of IDT and behavioral intentions within the context of travel apps.

Furthermore, researchers contended that the relationship between attitude and self-confident efficacy determines behavioral intentions in tourism contexts (Hlee et al., 2025). Self-confident efficacy can be described as “the respondent’s belief in their ability to accurately assess the attributes of a destination” (Valencia and Crouch, 2008, p. 27). Self-confident efficacy is considered an essential factor, as it significantly influences tourists’ travel-related decision-making (Fakfare et al., 2023). For instance, Sun et al. (2024) discovered that tourists’ attitudes toward travel positively influence self-confident efficacy. They also explained that individuals are more likely to be favorable toward international travel when they feel confident in their ability to plan their trips. Currently, Ahmed et al. (2025) have shown that a positive association between attitudes and self-confident efficacy significantly affects intentions to use in the context of AI voice-based assistant services. Hence, the following hypothesis was proposed:

AI-TPA is subdivided into two types based on the operation method (i.e. generative AI and GUI-based platforms). Generative AI can be implemented in the form of ChatGPT or a chatbot grounded in natural language-based interaction (Dwivedi et al., 2024). The effectiveness of generative AI in influencing users’ positive perceptions has been verified in the hospitality and tourism field. Current studies have highlighted the practical benefits of generative AI, with Li et al. (2025) reporting that hotel customers appreciated the usefulness of ChatGPT in facilitating their service needs. According to the Expedia (2025) report, chatbot-based AI service agents enabled over 50% of travelers to independently fulfill their service requirements without human intervention, resulting in significantly higher customer satisfaction ratings, more than double those of users who used call centers for assistance. In contrast, a GUI-based platform enables users to navigate and book travel options via visual elements such as buttons, menus and icons, rather than relying solely on text-based commands (Cybulski and Horbiński, 2020). Recently, Chenchu et al. (2025) demonstrated that the interactive visual interface of the platform is essential for promoting users’ engagement and facilitating more intuitive decision-making in travel planning. Liu et al. (2024a, 2024b) discovered that the graphical features in travel platforms enhance users’ cognitive information processing, thereby positively influencing behavioral intentions in the trip planning phase.

Previous studies investigated the impact of various operational methodologies of AI technologies on user perceptions. For example, Kim et al. (2019) compared participant response characteristics between chatbot- and Web-based survey modes. They verified that chatbot users generated more varied responses but expressed diminished satisfaction with the survey experience than Web respondents. Similarly, Nguyen et al. (2022) compared chatbots and menu-based interfaces in the travel context, showing that users perceived autonomy and cognitive effort differently across the two platforms. In this regard, users with various interaction experiences and capabilities may respond differently to each platform type, indicating that generative AI or GUI-based platforms may be more effective depending on individual user characteristics (Nguyen et al., 2022). Based on the above discussion, we proposed the following hypothesis:

In this study, the questionnaire comprised 42 validated measurement items on a five-point Likert scale and demographic questions. Measurement items for task characteristics under the TTF model were adopted from Schrier et al. (2010) and Dhiman and Jamwal (2023), comprising four items. Technology characteristics were measured using five items sourced from Dhiman and Jamwal (2023). Four items for TTF were derived from Cai et al. (2023), as well as Dhiman and Jamwal (2023). Measurement items for user motivation factors based on the TAM were adopted from Davis (1989), Kim et al. (2007) for ease of use (three items) and from Kim et al. (2019) and Ha and Stoel (2009) for perceived usefulness (three items). As for the IDT factors, three items for relative advantage were drawn from Hosseinikhah Choshaly (2019), four items for compatibility from Lu et al. (2015) and three items for complexity from Min et al. (2021). Attitude was measured using four items from Rivera et al. (2015), while self-confident efficacy was assessed based on four items from Fakfare et al. (2023). Finally, behavioral intentions to use the technology were measured with four items adapted from Venkatesh et al. (2003).

Data were collected via an online panel survey through SK Telecom, which provides access to approximately 16 million demographically diverse panel members in South Korea (SK Telecom, 2024). Screening questions ensured respondents had AI-TPA experience and classified them by usage type: generative AI (e.g. ChatGPT 3.5 or 4.0) or GUI-based platform (e.g. Myrealtrip, TRIPLE). Of 652 responses collected between June 2 and 6, 623 valid responses were retained for final analysis.

Data were analyzed using partial least squares structural equation modeling (PLS-SEM) via SmartPLS 4.1.1. PLS-SEM is appropriate for complex integrated models with high predictive power (Hair et al., 2010). A bootstrapping method with 5,000 samples assessed path coefficient significance (Hair et al., 2017). The proposed research model is presented in Figure 1.

Table 1 presents the demographic characteristics of the respondents. Among the participants, 61.5% were females and 38.5% were males. Respondents were aged 20 years 13.0%, 27.1% were aged 30 years and 59.9% were aged 40 years. In addition, 58.4% of the participants had an undergraduate education. Moreover, 41.3% of the respondents were married. Finally, the respondents reported an average monthly income of US$2,475.

Descriptive analysis showed means between 2.61 and 3.68 with standard deviations ranging from 0.791 to 0.834. Normality assumptions were validated using Shapiro–Wilk tests (p > 0.05) alongside acceptable skewness (±1.0) and kurtosis (±1.5) indices, confirming appropriate data distribution for parametric procedures (George and Mallery, 1994).

Additionally, common method bias (CMB) can inflate relationships between variables because of self-report measures. Therefore, Harman’s single-factor test was used to assess CMB. Results showed that a single factor accounted for 45.236% of the variance, below the 50% threshold (Podsakoff et al., 2003), indicating CMB is not a significant concern.

 Appendix 1 presents the constructions used in this study with their standardized factor loadings of the two data sets as follows:

1. Generative AI data; and

2. GUI platform data sets.

In addition, no factor loading values fell below 0.736.  Appendix I shows that all the composite reliability (CR) values exceeded 0.700 (between 0.870 and 0.946), providing evidence for the internal consistency of the measurement models (Hair et al., 2006). Moreover, all the average variance extracted (AVE) values exceeded 0.500 (between 0.655 and 0.814), which reveals a high level of convergent validity (Fornell and Larcker, 1981).

As shown in  Appendix 2, discriminant validity between constructs was evaluated using two established methods: the Fornell–Larcker criterion (Fornell and Larcker, 1981) and the heterotrait-monotrait (HTMT) technique (Henseler et al., 2016). The diagonal elements, shown in bold, indicate the average variance extracted (AVE), while the off-diagonal values represent the squared correlations between the latent constructions. All values of the AVE exceeded the squared correlations values between pairs of constructs, which supports a high level of discriminant validity (Bagozzi and Yi, 1988). Furthermore,  Appendix 2 presents that all the Cronbach’s α values exceeded 0.700 (between 0.824 and 0.902), justifying the reliability of constructs (Henseler et al., 2016). The HTMT correlation ratios presented in  Appendix 2 ranged from 0.091 to 0.862, all remaining below the recommended threshold of 0.900 (Henseler et al., 2016), thus indicating satisfactory discriminant validity among the constructs.

Model fit adequacy was assessed using Standardized Root Mean Square Residual (SRMR) and Normed Fit Index (NFI) as important indicators of goodness of fit. The results demonstrated satisfactory fit, with SRMR achieving 0.050 (below the 0.08 threshold) and NFI reaching 0.817 (exceeding the 0.8 criterion) (Hu and Bentler, 1999).

The structural model was examined using R² and path coefficient significance (β). The model explained substantial variance: attitudes (R² = 76.4%), self-confident efficacy (R² = 59.5%) and intentions to use (R² = 76.9%). Based on established thresholds including strong (R² ≥ 0.75), moderate (R² ≥ 0.50) or weak (R² ≥ 0.25) (Hair et al., 2019), R² values demonstrated strong performance (≥0.75) for attitudes and intentions to use and moderate performance (≥0.50) for self-confident efficacy. Effect sizes were assessed using f² criteria: small (f² ≥ 0.02), medium (f² ≥ 0.15) or large (f² ≥ 0.35) (Cohen, 1988) and exceeded zero for all endogenous constructs in Table 2, supporting predictive adequacy (Hair et al., 2019).

Path coefficient significance was determined through bootstrapping with 5,000 resamples. Bootstrapping is an analytical technique showing the significance level of the paths between each construct (Henseler et al., 2016). Multicollinearity was assessed using VIF values, which ranged from 1.000 to 3.602 – all below the acceptable threshold of 3–5 (Hair et al., 2019), indicating no collinearity issues in the model. The hypothesis testing results showed that 10 of 12 hypotheses were supported. These results are summarized in Figure 2.

As shown in Table 3, the mediation analysis revealed that attitude serves as a significant mediating factor in the relationship between various independent variables and the dependent variables (i.e. self-confident efficacy and intention to use). Specifically, the mediating effect of attitude was strongest in the relationship between perceived usefulness and self-confident efficacy (β = 0.055), followed by compatibility (β = 0.054) and task-technology fit (β = 0.035). However, attitude did not significantly mediate the relationship between perceived ease of use and self-confident efficacy (β = 0.006) and complexity and self-confident efficacy (β = −0.002). Notably, attitude fully mediated the relationship between relative advantage and self-confident efficacy (β = 0.067). Likewise, the mediating effect of attitude was strongest in the relationship between relative advantage and intention to use (β = 0.148), followed by perceived usefulness (β = 0.123) and compatibility (β = 0.119). However, attitude did not significantly mediate the relationship between perceived ease of use and intention to use (β = 0.013) and between complexity and intention to use (β = −0.003).

As Henseler et al. (2016) suggested, measurement invariance should be tested with MICOM before conducting multigroup analysis (MGA). MICOM involves three-steps:

1. configural invariance;

2. compositional invariance; and

3. equal means and variances assessment.

The first step established configural invariance by maintaining identical model specifications across groups (Hair et al., 2024). The second step confirmed compositional invariance when original correlation met or exceeded the 5.00% quantile (Henseler et al., 2016). As presented in Table 4, adequate compositional invariance was achieved. The third step assessed whether mean and variance difference fell within permutation-based confidence intervals (Hair et al., 2024; Matthews, 2017). As shown in Table 4, full measurement invariance was established as all three conditions were satisfied. According to Henseler et al. (2016), partial measurement invariance (Steps 1 and 2) provides sufficient grounds for MGA regardless of Step 3 results. Therefore, the analysis proceeded to test the group-specific differences.

MGA results indicated significant differences in path estimates between generative AI and GUI groups (Table 5). Perceived usefulness positively affected attitudes in both groups (generative AI: β = 0.309, p < 0.001 and GUI: β = 0.136, p = 0.008) with a significant difference between groups (Henseler’s MGAp-value = 0.047). Compatibility has a positive effect on attitudes in both groups (generative AI: β = 0.259, p < 0.001 and GUI: β = 0.186, p = 0.009), showing significant group difference (Henseler’s MGAp-value = 0.009). Moreover, attitudes significantly affected self-confident efficacy in both groups (generative AI: β = 0.811, p < 0.001 and GUI: β = 0.713 and p < 0.001) with a significant difference between groups (Henseler’s MGAp-value = 0.028). Finally, attitudes significantly affected intention to use in both groups (generative AI: β = 0.609, p < 0.001 and GUI: β = 0.712, p < 0.001), showing a significant difference between groups (Henseler’s MGAp-value = 0.014). Therefore, hypotheses H15f, H15h, H15j and H15k were supported, indicating that the generative AI and GUI groups exhibit significantly different behavioral patterns in technology adoption processes. These results are presented in Figure 3.

To assess model robustness, quadratic effects analysis was conducted (Hair et al., 2024). The analysis revealed curvilinear relationships in two paths: compatibility demonstrated a U-shaped relationship with attitude (β = 0.044 and p < 0.05) and attitude exhibited an inverted U-shaped effect on self-confident efficacy (β = −0.043 and p < 0.05). When quadratic terms were included, corresponding linear effects became non-significant, confirming the curvilinear specification. Despite the limited incremental variance explained (ΔR² ≤ 0.008), which falls below Cohen’s (1988) threshold for small effect sizes, the statistical significance of these quadratic terms supports their theoretical relevance. Following established robustness check protocols, we incorporated these quadratic terms in our final model to provide a more accurate representation of the underlying relationships while maintaining theoretical parsimony for the remaining linear paths.

First, this study advanced AI-based technology research by developing a comprehensive framework that incorporated TTF, TAM and IDT to enable an enhanced comprehension of the complex antecedents that underpin users’ attitudes and intentions in the context of AI-TPA. However, prior combined frameworks provided partial explanations TTF and IDT models neglected cognitive evaluations (Wang and Lin, 2019), while TAM and IDT approaches overlooked task-specific fit (Terrah et al., 2024). In this regard, this comprehensive framework addresses these theoretical gaps by incorporating complementary perspectives that collectively explain technology acceptance more effectively than previous dual-theory models, thereby establishing a more robust foundation for future AI tourism research.

Second, the analysis results confirmed a significant impact of the integrated variables TTF, TAM and IDT on attitude. From TTF perspective, task and technology characteristics emerged as crucial indicators of TTF, which significantly influenced attitudes toward AI-TPA. This result aligns with prior research (Dhiman and Jamwal, 2023), which emphasizes that TTF is crucial in shaping attitudes in the context of service chatbots. In the TAM scope, perceived ease of use positively influenced perceived usefulness, which subsequently significantly influenced attitudes, aligning with Foroughi et al.’s (2024) studies on travel apps. However, perceived ease of use insignificantly impacted attitudes, contradicting prior studies by Du et al. (2025). This may be because of users prioritizing functional value over ease when planning their travel itinerary (Ozturk et al., 2016). From an innovation perspective, relative advantages and compatibility were validating as positively influencing attitudes, supporting findings that technology acceptance is contingent upon performance benefits and aligns with user needs (Lim et al., 2022). Notably, complexity insignificantly impacted attitudes, suggesting that as AI-based technology becomes mainstream, researchers should reconsider complexity as a diminishing barrier to adoption in technology acceptance models.

Third, this study identified attitude as a crucial factor in influencing both self-confident efficacy and intentions to use AI-TPA. Importantly, self-confident efficacy demonstrated a stronger positive association with intentions to use AI-TPA than attitude alone, revealing that favorable technology evaluations are insufficient without users’ confidence in their ability to use the technology. These findings underscore the essential role of self-confident efficacy in shaping intentions to use AI-TPA. This result indicates that a positive evaluation of AI-TPA must be accompanied by confidence in its use to enhance the intention to use it.

Fourth, this study examined the mediating role of attitude within the integrated TTF, TAM and IDT model, linking antecedent variables to self-confident efficacy and intentions to use in the context of AI-TPA. The findings demonstrated that attitude is essential in converting users’ technology-related perceptions into intentions to use AI-TPA, aligning with Lim et al. (2022), who demonstrated that attitude significantly mediates among tourists using travel apps. Notably, the complete mediation effect of attitude in the correlation between relative advantage and self-confident efficacy was confirmed. This finding indicated that the objective superiority of technology alone cannot directly enhance users’ confidence. Moreover, self-confident efficacy functions as a mediating variable that transforms attitudes into intentions to use, supporting the findings of Sun et al. (2024). The current finding demonstrated that favorable attitudes require users’ confidence in their AI-TPA capabilities to generate intentions to use. This result reveals that enhancing self-confident efficacy can promote the adoption of AI-TPA more effectively in users’ decision-making. This finding enhances the existing AI technology literature by bridging the gap between the perception of technology and the actual intention to use it.

Finally, this study investigated the moderating effect of AI-TPA interaction types in influencing users’ attitudes and intentions. Existing studies focused on factors that can positively affect consumer attitudes toward AI-TPA; however, researchers have typically targeted users of only one type of AI-TPA, either generative AI or GUI-based platforms (Cybulski and Horbiński, 2020; Kanhed et al., 2024). That is, examining the impact on attitude across two different types of AI-TPA has been overlooked. To the best of our knowledge, this is the first empirical evidence comparing generative AI and GUI interaction regarding a moderating effect within an integrated model in the context of AI-TPA. This study advances theoretical understanding by corroborating the findings of Nguyen et al. (2022), who identified differing consumer responses to chatbot versus menu-based interface interactions.

The analysis revealed that the relationships between perceived usefulness and attitude and compatibility and attitude differ between interaction types. The moderating effect on perceived usefulness and attitude was significant only in the generative AI group, indicating that users recognize customization through natural language, resulting in more favorable attitudes. These results align with Dwivedi et al. (2024), who indicated that generative AI users may express nuanced preferences via conversational input. In contrast, GUI-based users are constrained by pre-structured menu choices (Cybulski and Horbiński, 2020), which may not foster favorable attitudes. This finding aligns with Kim et al. (2019), who emphasized the differential impact of AI interaction types. Compatibility significantly affected attitude formation in both groups, with a more pronounced effect among generative AI users. This suggests users exhibit positive attitudes when the system matches their preferences via conversational interaction. This finding supported Nass and Moon (2000), who proposed social response theory, positing that people instinctively regard human-like technologies as social partners. For GUI-based platforms, compatibility holds less significance as users are already familiar with point-and-click interfaces. These findings indicate that conversational AI aligns more closely with user preferences than conventional click-based systems. This study offers a framework for systematically comparing various AI-TPA types, addressing the heterogeneity that single-platform studies cannot capture.

First, this study suggests that perceived usefulness significantly affects attitudes toward AI-TPA and intentions to use. Marketers can create advertisements highlighting how AI-TPA enables faster access to travel information compared to manual searches. Additionally, software developers should incorporate features demonstrating time savings and user ratings (e.g. “1,000 travelers rated this 4.1 out of 5 stars”). This social proof approach is effective because individuals tend to follow others’ attitudes in uncertain situations (Amblee and Bui, 2011). Successful platforms such as Booking.com and TripAdvisor effectively use social proof by displaying real-time booking activities and user ratings.

Second, perceived compatibility significantly affects attitudes and intentions to use AI-TPA, with differential effects across AI-TPA types. Generative AI companies should emphasize how conversational interaction aligns with natural communication styles, as real-time engagement positively influences user satisfaction (Dwivedi et al., 2024). GUI-based platforms benefit from reduced compatibility barriers because of user familiarity with conventional interfaces; however, they should highlight visual clarity and intuitive navigation to compete against generative AI alternatives. Various consumer segments perceive compatibility differently, requiring tailored approaches. Providers should initially target consumers with strong digital literacy and innovativeness as early adopters (Teng et al., 2024). To broaden the user base, simple tutorials and hands-on demonstrations can assist less tech-savvy consumers in becoming comfortable with AI-TPA. Moreover, customized strategies based on age, education and lifestyle characteristics will further enhance adoption across various market segments.

Finally, the findings reveal that self-confident efficacy functions as a crucial mediating factor between attitudes toward AI-TPA and the intention to use it. As more travelers adopt AI-based travel planning tools and traditional travel agent roles diminish (Ho et al., 2021), this finding offers a strategic pathway for the tourism industry to navigate this transition. This presents both challenges and opportunities for the tourism industry. Travel companies should invest in upskilling their workforce to work alongside AI technology rather than competing with it. Specialized training programs can enable employees to become AI-TPA consultants that assist customers in optimizing the technology’s benefits. By repositioning staff as expert facilitators rather than traditional agents, companies can preserve human touchpoints while embracing technological advancement. This hybrid approach mitigates employment concerns while enhancing consumers’ self-confident efficacy and intentions about AI-TPA through the synergy of human expertise and AI technology.

Despite the significant theoretical and practical contributions of this research, certain limitations require thorough consideration. First, focusing solely on Korea may limit generalizability; future research should compare markets with varying digital adoption rates and literacy levels. Second, the cross-sectional design captured attitudes at one point; longitudinal studies are needed to examine how initial skeptical attitudes evolve, shifts in preference between generative AI and GUI and generational differences as users gain familiarity with AI technology. Third, while user preferences were quantitatively measured, mixed-methods research is essential to explore how users’ demand for control and travel styles influence AI tool selection and platform preferences. Fourth, future studies should address ethical and sustainability dimensions, including data privacy, algorithmic bias, employment displacement, digital equity and environmental impacts of AI recommendations on sustainable tourism practices.