2.1 Theoretical background

As digital health and intelligent technologies have continued to be embedded in sport and health management, researchers adopted and developed a range of established theoretical frameworks to explain users’ adoption, use, and future-use intentions toward related technologies [26,27]. For example, Chiu et al. [28] employed the expectation-confirmation model (ECM) to investigate factors influencing users’ intentions to continue using health and fitness applications. Similarly, Palos-Sanchez et al. [29] applied the technology acceptance model (TAM) to analyze individuals’ behavioral intention to use mobile health technologies. More recently, Li et al. [30] integrated the Theory of Planned Behavior (TPB) with the Health Belief Model (HBM) to test how attitude, subjective norms, perceived behavioral control, and health-belief-related factors affected exercise intention when users engaged with fitness mobile applications. Other frameworks, such as the Unified Theory of Acceptance and Use of Technology (UTAUT) and the task-technology fit (TTF) model, were also commonly used in health management technology research to characterize the roles of social influence, facilitating conditions, and the alignment between technological functionalities and task requirements in shaping intention [31,32].

GAI technologies are beginning to reshape the landscape of fitness services and to encourage traditional fitness models to undergo digital transformation [33]. Fitness enthusiasts are among the early groups encountering emerging GAI technologies capable of generating personalized training plans, but the present study does not assume that respondents used a particular commercial system. Therefore, examining users’ attitudes and intention to use this technology category when they perceived it as a high-quality information system was important. DeLone and McLean [34] noted that analyses of information system success had to consider quality dimensions and their effects on user satisfaction and intention. Given that GAI fitness assistants were inherently complex information systems integrating content generation, interactive services, and risk-laden outcomes, a single adoption perspective was insufficient to fully explain users’ decision-making processes. Accordingly, combining the holistic quality evaluation emphasized by the Information Systems Success Model (IS Success Model) with the psychological decision mechanisms proposed by the Theory of Planned Behavior (TPB), and further incorporating a perceived-value perspective capturing benefit-cost trade-offs, enabled a more systematic explanation of how users formed intentions to use GAI fitness assistants.

The Information Systems Success Model (IS Success Model), proposed by DeLone and McLean [34], provided a well-established framework for explaining how information system quality is associated with system success through user satisfaction and usage intention. The model highlighted the importance of information quality, system quality, and service quality (i.e., second-order information system quality) as key antecedents.

Meanwhile, the Theory of Planned Behavior (TPB) (see Fig 1) was extended from the Theory of Reasoned Action (TRA) to address the limitations of the original model in explaining behaviors that were not entirely under volitional control [35–37]. Ajzen [37] argued that behavioral intention was the most proximal antecedent of behavior, reflecting the extent of effort an individual was willing to invest in performing a behavior; whether the behavior occurred also depended on actual behavioral control, which was shaped by resources and opportunities such as time, money, skills, and cooperation from others. TPB further posited that behavioral intention was jointly predicted by attitude, subjective norm, and perceived behavioral control. In the context of GAI fitness assistants, this implied that users were more likely to report stronger intention to use when they held a more favorable evaluation of using the assistant, perceived support from important others, and believed they possessed the necessary resources and capabilities [38].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. TPB theoretical model.

https://doi.org/10.1371/journal.pone.0353384.g001

In addition, perceived value (PV) provided a critical complement for explaining technology adoption in high-risk contexts characterized by strong trade-offs. Compared with focusing only on usefulness or attitudinal tendencies, perceived value emphasized that individuals made an overall judgment about whether a technology was worth using by weighing expected benefits against expected costs [39,40]. In the context of GAI fitness assistants, benefits could be reflected in more personalized and efficient training recommendations, emotional support derived from continuous companionship, and cost savings by partially substituting for personal training services; costs included learning and interaction effort, concerns about privacy and data security, and physical risks arising from incorrect recommendations [41]. Therefore, even when users held a favorable attitude toward the technology, their intention to use could still decrease due to cost–risk trade-offs. Conversely, when users perceived that overall benefits substantially outweighed the costs and risks, perceived value became a direct predictor of intention to use. Within the integrated framework of this study, information system quality and the three TPB components jointly shaped users’ evaluations of benefits and costs and further influenced behavioral intention through perceived value.

Although TPB and the IS Success Model had been widely applied in studies of mobile health and general apps, research that simultaneously employed second-order information system quality and TPB mechanisms to explain intention to use generative AI fitness assistants remained relatively limited. More importantly, in the high-variability and high-risk context of fitness recommendation generation, users often did not make decisions solely based on attitude or social influence; instead, they further weighed benefits against costs to form an overall judgment of whether the technology was worth using [42,43]. Therefore, it was necessary to reveal how second-order information system quality (a technological attribute) worked through attitude, subjective norm, and perceived behavioral control (psychological attributes), and further via perceived value as a trade-off-based evaluative mechanism, to jointly influence intention to use. Accordingly, this study adopted an integrative three-theory perspective to develop and test the formation mechanism underlying users’ intention to use GAI fitness assistants, thereby addressing this gap.

In summary, this study used the IS Success Model to characterize the holistic quality foundation of GAI fitness assistants, applied TPB to explain the psychological pathways through which intention was formed, and incorporated perceived value to capture the decision outcome of benefit-cost trade-offs, thereby constructing an integrated framework that more closely reflected category-level intention formation in this emerging technology context.

2.2 Hypotheses development

Prior research had demonstrated that the Theory of Planned Behavior (TPB) was an important framework for explaining health technology adoption and designing behavior-change interventions. However, in the context of GAI fitness assistants—which emphasized conversational interaction, training safety, and sustained engagement—relying solely on attitude, subjective norm, and perceived behavioral control was insufficient to capture users’ judgments of overall system performance and their benefit–cost trade-offs. Therefore, building on TPB, this study incorporated second-order information system quality from the Information Systems Success Model and introduced the construct of perceived value, thereby forming an extended integrative model and proposing corresponding hypotheses (see Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Theoretical model.

https://doi.org/10.1371/journal.pone.0353384.g002

3.1 Ethics statement

Ethical approval for this study was obtained from the School of Physical Education, Yan’an University. Before completing the questionnaire, all participants were informed of the purpose of the study, the anonymous nature of the survey, and their right to withdraw at any time without penalty. Electronic informed consent was obtained from all participants prior to participation. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki.

3.2 Survey design

Given the lack of a generic scale specifically designed for GAI fitness assistants, this study adopted validated measurement items from related domains and contextualized them to ensure both comparability and theoretical traceability.

To reduce differences in respondents’ understanding of generative AI fitness assistants, this study presented a standardized visual scenario description at the beginning of the questionnaire, introducing the concept, core functions, and typical usage process of the target technology. The material described a generative AI fitness assistant as an intelligent digital tool that can understand users’ fitness goals, physical conditions, available time, and training feedback through text, voice, or multimodal interaction, and can generate, explain, and adjust personalized training plans. Respondents were instructed to answer the subsequent items based on this standardized hypothetical scenario rather than evaluating any specific commercial product, thereby improving the contextual consistency and interpretability of the questionnaire measurements.

Grounded in the Information Systems Success Model, the Theory of Planned Behavior (TPB), and perceived value, this study developed the questionnaire by combining the adoption of established scales with contextual adaptation. The overall procedure included scale selection, contextual rewriting, pilot testing, and item refinement, ultimately producing a formal measurement instrument for structural equation modeling (SEM).

Regarding measurement sources, the three first-order dimensions (information quality, system quality, and service quality) were measured using items primarily adapted from well-established and empirically validated scales widely used in the e-service quality and information systems literature, particularly those developed by Parasuraman et al. [59] and Nelson et al. [60]. The TPB constructs (attitude, subjective norm, and perceived behavioral control) were mainly adapted from the classic measures proposed by Ajzen [61], Fishbein and Ajzen [62], Van Lange et al. [63], and Compeau and Higgins [48]. Perceived value was primarily measured using items drawn from Kim et al. [39]. Behavioral intention was adapted from the scale developed by Taylor and Todd [64]. Based on these instruments, this study further added context descriptions related to generative AI (GAI) fitness assistants to enhance the coverage of the target technology and usage scenarios, ensuring that the meaning of each construct was adequately represented in the new technological context.

To enhance contextual consistency and participants’ comprehension, this study contextualized the wording of selected items by drawing on typical interaction patterns between users and generative AI fitness assistants in fitness coaching scenarios. Specifically, within the information system quality section, the items highlighted features related to generative interaction, including the stability and response speed of the conversational function, ease of use, and the extent to which generated recommendations were logical, clear, and actionable, thereby enabling participants to make judgments based on their category-level perceptions of such assistants. In the subjective norm section, the items focused on the influence of key referent groups relevant to participants’ daily study and exercise routines (e.g., parents, significant others, peers, and physical education teachers) on their usage decisions, thereby improving the fit of normative sources. The contextual adaptation process preserved the original theoretical meaning of each construct while replacing generic service-use wording with fitness-assistant scenarios. All revised items were reviewed during the pilot test for clarity, semantic overlap, and factor-loading performance. In the perceived value and behavioral intention sections, the items adopted a benefit-cost trade-off framing, emphasizing the relative gains of using the assistant in terms of time, effort, and input costs, and further assessing individuals’ future intention to use the technology category.

Regarding scale format, Brown [65] and Dawes [66] suggested that, compared with scales using fewer response categories, seven-point Likert scales typically demonstrated superior reliability and validity [65,66]. Therefore, this study used a seven-point Likert scale to measure all observed variables, with anchors ranging from 1 (strongly disagree) to 7 (strongly agree). The initial questionnaire contained 37 measurement items covering eight first-order latent variables, among which information quality, system quality, and service quality jointly formed the second-order latent construct of information system quality.

Before the formal survey administration, the research team conducted a pilot test and obtained 55 valid responses to assess item comprehensibility and preliminary measurement quality. Based on the results of reliability and validity analyses, two items (IQ3 and SER4) were removed due to low factor loadings and partial semantic redundancy. The remaining items were slightly refined in wording according to participant feedback. The final questionnaire comprised 35 measurement items (see S1 Appendix) and additionally included three demographic variables (gender, education level, and exercise frequency) for sample description and control analyses.

3.3 Data collection

The questionnaire was administered mainly from early January 2026 to late January 2026 through a professional online survey platform (Wenjuanxing, https://www.wjx.cn/). To ensure data integrity during distribution, the platform’s response-time tracking and IP restriction functions were enabled to prevent duplicate submissions. In total, 433 responses were collected from individuals with established fitness habits or an interest in digital health. Participants responded to a standardized category-level scenario and were not required to be verified current users of a named GAI fitness application.

Strict procedures were implemented to ensure sample quality, consistent with the methodological recommendations summarized by Hair et al. [67] and Wang et al. [68]. The screening process involved three steps. First, based on insights from the pilot test, completing the questionnaire typically required 2–6 minutes. Therefore, participants who finished the survey in an unusually short time (i.e., less than 120 seconds) were considered to have approached the task carelessly, and their responses were classified as invalid. Second, the questionnaire included an attention-check item (e.g., “Please select ‘5’ for this question”). Responses from participants who failed to answer this item correctly were treated as invalid to eliminate careless responding. Finally, response patterns were examined, and questionnaires in which the same option was selected for all items (e.g., choosing “7” for every question) were removed.

After screening, 91 invalid questionnaires were excluded. Specifically, 36 responses were discarded due to insufficient completion time, 36 were removed because of incorrect answers to the attention-check item, and 19 exhibited obvious straight-lining patterns. Consequently, 342 valid questionnaires were retained for formal data analysis, yielding an effective response rate of 79.0%. Given the 35 observed measurement items in the final questionnaire, the sample-to-item ratio was approximately 9.77:1, which is close to the commonly cited 10:1 guideline. Accordingly, the findings should be generalized cautiously beyond exercise-engaged and digitally interested potential users.

3.4 Data analysis

This study used SPSS 27.0 and AMOS 28.0 to conduct statistical analyses of the survey data. First, SPSS was used for data cleaning and descriptive statistics, as well as for preliminary reliability testing of the measurement items and profiling of sample characteristics. Subsequently, the proposed theoretical model was validated and the hypotheses were tested using covariance-based structural equation modeling (CB-SEM) in AMOS. The higher-order information system quality construct was specified reflectively because information quality, system quality, and service quality were treated as correlated manifestations of an overall perceived quality evaluation rather than as formative components that independently constitute the construct. Compared with variance-based approaches such as PLS-SEM, CB-SEM was chosen because it enabled the simultaneous estimation of relationships among latent variables and provided overall model-fit information, thereby better supporting statistical tests of the hypothesized model [67].

In terms of analytical procedure, a two-stage approach was followed [69]. In the first stage, the measurement model was evaluated by systematically assessing the reliability and validity of each latent construct to ensure stable and sound measurement quality. In the second stage, after the measurement model met the required criteria, the structural model fit was assessed and standardized path coefficients were estimated to test the hypothesized relationships. Once the reliability and validity tests are satisfied, model fit is then examined to ensure that the estimated paths accurately reflect the underlying relationships. Subsequently, the interrelationships among the different constructs are analyzed in depth [67].

4.1 Common method bias

All questionnaire data in this study were obtained from the same source via respondents’ self-reports and were measured using a fixed response format. Therefore, the data might have been subject to common method bias (CMB). CMB was typically induced by factors such as a shared measurement context, item framing, and item characteristics, which could result in spurious covariance between predictor and outcome variables [52,70]. To mitigate this risk and reduce systematic error attributable to common methods, multiple procedural remedies were implemented during questionnaire design and administration. First, measurement items for different constructs were placed on separate pages to reduce consistency motives and same-source priming effects caused by item adjacency, while also providing brief buffering between pages to limit the accumulation of common method variance [71,72].

For statistical assessment, Harman’s single-factor test was conducted to evaluate potential CMB. The results showed that the first unrotated factor accounted for 33.786% of the total variance, which was below the commonly used 50% threshold [73,74]. This finding suggested that no single factor dominated the variance structure and that common method bias was unlikely to materially affect the findings.

4.2 Validity and reliability testing

Before conducting the structural equation modeling (SEM) analyses, this study first performed a systematic assessment of the reliability and validity of the measurement scales (see Table 1). For reliability, Cronbach’s alpha coefficients were calculated using SPSS and were cross-validated with composite reliability (CR). The results showed that Cronbach’s alpha values for all latent constructs ranged from 0.794 to 0.921, exceeding the recommended threshold of 0.70 [73], which indicated good internal consistency. Meanwhile, CR values ranged from 0.795 to 0.922, all above the 0.70 benchmark [75], further supporting the stability and reliability of the measures.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. The results of the reliability and convergent validity of the measurement model.

https://doi.org/10.1371/journal.pone.0353384.t001

Because relying solely on Cronbach’s alpha was insufficient to fully demonstrate measurement quality, this study further examined convergent and discriminant validity. Convergent validity was primarily assessed using average variance extracted (AVE). The AVE values for all constructs ranged from 0.565 to 0.743, exceeding the 0.50 threshold [76], indicating that the latent variables explained a substantial proportion of variance in their indicators and that the measurement model demonstrated adequate convergent validity.

Regarding discriminant validity, the Fornell–Larcker criterion was applied by comparing the square roots of AVE with the inter-construct correlation coefficients [76]. As shown in Table 2, the bolded diagonal values represented the square roots of AVE for each construct (ranging from 0.752 to 0.862), whereas the off-diagonal values represented Pearson correlation coefficients among constructs. The results indicated that the square root of AVE for each construct was greater than its correlations with all other constructs, suggesting that each construct explained more variance in its own indicators than it shared with other constructs, thereby supporting adequate distinctiveness among the constructs. The largest inter-construct correlation was 0.697, which was below the commonly used threshold for severe construct overlap, further supporting discriminant validity. Although several moderate correlations were observed between constructs, the square roots of AVE for all constructs still exceeded the corresponding inter-construct correlations, indicating that discriminant validity remained at an acceptable level. Overall, the measurement model met commonly used criteria for discriminant validity in SEM research, providing a reliable measurement foundation for subsequent structural model analyses.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Discriminant validity.

https://doi.org/10.1371/journal.pone.0353384.t002

4.3 SEM fit indices

To avoid spurious associations among hypothesized paths resulting from model misspecification or sampling error, the overall SEM model fit was evaluated before testing the structural relationships (see Table 3). Following Hair et al. [67], threshold values suggested by Hayduk [77], Bagozzi and Yi [78], Hair et al. [79], and Hu and Bentler [80] were adopted as the evaluation criteria. Model fit indices were computed using Amos 28.0. The results showed that CMIN/DF = 1.292, GFI = 0.893, CFI = 0.978, TLI = 0.976, RMSEA = 0.029, and SRMR = 0.0495, all of which met or exceeded the recommended standards. Given that the chi-square statistic and degrees of freedom were sensitive to sample size, the combined evidence from multiple indices indicated that the model fit was satisfactory. These results suggested that the proposed model adequately fit the data and provided a robust basis for subsequent analyses of variable relationships and hypothesis testing.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Standard model fit indices.

https://doi.org/10.1371/journal.pone.0353384.t003

4.4 SEM validation

This study estimated the structural model using AMOS (v28.0) to obtain standardized path coefficients and to further assess the explanatory power of the endogenous variables (see Fig 3). Given that information system quality (Information System Quality, ISQ) was modeled as a second-order construct, it was necessary to further examine the contribution weights of its first-order dimensions to the higher-order construct (see Table 4). The results showed that the loadings of system quality (SQ), service quality (SER), and information quality (IQ) on ISQ all reached statistical significance (all paths p < 0.001), indicating that these three dimensions jointly constituted a valid representation of overall information system quality.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Weight distribution in second-order models.

https://doi.org/10.1371/journal.pone.0353384.t004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Structural model diagram.

https://doi.org/10.1371/journal.pone.0353384.g003

In terms of contribution weights, the path ISQ → SQ exhibited the highest weight (0.805), followed by ISQ → IQ (0.710), whereas ISQ → SER was relatively lower (0.639). This pattern suggested that, in the formation of users’ perceptions of the overall quality of GAI fitness assistants, system-level performance and stability served as the most central source of quality, followed by the quality of informational content, while service support factors played a significant yet comparatively secondary role. Overall, the distribution of weights in the second-order model supported the specification of ISQ as a higher-order quality construct and provided a reliable measurement basis for subsequent structural path analyses.

To present the mechanisms among variables more clearly, this study further tested the hypothesized paths and reported the standardized path coefficients (β) and significance levels of the structural model (see Table 5). Regarding the formation of attitude (ATT), both subjective norm (SN) and information system quality (ISQ) exerted significant positive effects on attitude (SN → ATT, β = 0.260, p < 0.001; ISQ → ATT, β = 0.623, p < 0.001), with the effect of ISQ being stronger than that of SN. Taking ISQ → ATT as an example, the result could be interpreted as follows: when ISQ increased by one standard deviation, users’ attitude increased by approximately 0.623 standard deviations, indicating that overall system quality served as a key driver of positive attitude formation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Structural model path analysis results.

https://doi.org/10.1371/journal.pone.0353384.t005

Regarding perceived value (VAL), both attitude and subjective norm exerted significant positive effects (ATT → VAL, β = 0.608, p < 0.001; SN → VAL, β = 0.174, p = 0.004). The effect of attitude was more pronounced, indicating that individuals’ favorable evaluations of GAI fitness assistants served as an important source of value judgments. For perceived behavioral control (PBC), subjective norm and information system quality also had significant positive effects on PBC (SN → PBC, β = 0.270, p < 0.001; ISQ → PBC, β = 0.485, p < 0.001), suggesting that subjective norm and overall system quality jointly enhanced users’ perceptions of whether they possessed sufficient capability and necessary conditions to use the assistant, thereby strengthening perceived behavioral control.

With respect to behavioral intention (BI), attitude, perceived behavioral control, and perceived value were all significant predictors (ATT → BI, β = 0.275, p < 0.001; PBC → BI, β = 0.163, p = 0.009; VAL → BI, β = 0.447, p < 0.001), with perceived value showing the largest effect. In contrast, the direct effect of subjective norm on behavioral intention was not significant (SN → BI, β = −0.080, p = 0.192). Accordingly, H1, H2, H3, H4, H5, H7, H8, H9, and H10 were supported, whereas H6 was not supported. The negative sign of the SN → BI coefficient was therefore not interpreted substantively, because the effect was statistically negligible.

In terms of explanatory power, the model explained 58.4% of the variance in attitude (R² = 0.584), 50.7% in perceived value (R² = 0.507), 41.2% in perceived behavioral control (R² = 0.412), and 50.9% in behavioral intention (R² = 0.509), indicating that the structural model exhibited meaningful but incomplete explanatory capability for intention to use GAI fitness assistants.

4.5 Indirect effects testing

To formally examine the mediating mechanisms implied by RQ2 and the structural model, standardized specific indirect effects were tested using bootstrapped confidence intervals (see Table 6). Overall, all indirect paths reported in Table 6 were statistically significant, with 95% confidence intervals excluding zero. This pattern indicated that the non-significant direct path from subjective norm to behavioral intention did not imply the absence of social influence; rather, subjective norm was associated with intention through attitude, perceived behavioral control, and perceived value.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Results of standardized specific indirect effects.

https://doi.org/10.1371/journal.pone.0353384.t006

For subjective norm, the significant paths included SN → PBC → BI (standardized effect = 0.044, 95% CI [0.009, 0.093], p = 0.044), SN → ATT → BI (standardized effect = 0.072, 95% CI [0.020, 0.142], p = 0.020), SN → VAL → BI (standardized effect = 0.078, 95% CI [0.026, 0.144], p = 0.010), and the serial path SN → ATT → VAL → BI (standardized effect = 0.071, 95% CI [0.030, 0.118], p = 0.001). Thus, subjective norm operated through multiple mediated routes rather than as a direct normative-pressure pathway.

Indirect effects involving information system quality and attitude were also significant. ISQ was indirectly associated with behavioral intention through ATT (standardized effect = 0.171, 95% CI [0.072, 0.274], p = 0.001), through PBC (standardized effect = 0.079, 95% CI [0.021, 0.139], p = 0.008), and through the serial path ISQ → ATT → VAL → BI (standardized effect = 0.169, 95% CI [0.105, 0.247], p < 0.001). ATT also showed a significant indirect association with BI through VAL (standardized effect = 0.272, 95% CI [0.181, 0.373], p < 0.001), while the upstream indirect paths ISQ → ATT → VAL (standardized effect = 0.379, 95% CI [0.278, 0.489], p < 0.001) and SN → ATT → VAL (standardized effect = 0.158, 95% CI [0.071, 0.246], p < 0.001) further indicated that attitude contributed to the formation of perceived value.

4.6 Robustness check with control variables

To account for differences in participants’ engagement levels and related experience, gender, education level, and exercise frequency were included as control variables. The purpose of incorporating these controls was to reduce potential confounding effects of demographic characteristics and exercise behavior differences on the relationships among the core constructs, thereby improving the accuracy and explanatory power of structural path estimates. The results (see Fig 4) showed that after controlling for gender, education level, and exercise frequency, the main path directions and significance levels among second-order information system quality (ISQ), subjective norm (SN), attitude (ATT), perceived behavioral control (PBC), and perceived value (VAL) remained consistent with those of the model without controls, and the overall fit indices did not exhibit any substantial changes. Further analyses indicated that the direct effects of gender, highest education level, and exercise frequency on behavioral intention (BI) were all non-significant. These findings suggested that the key structural relationships of the model did not change markedly across these control variables within the current sample.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Structural model with control variables.

https://doi.org/10.1371/journal.pone.0353384.g004

5.1 The role of information system quality

Second-order information system quality (ISQ) significantly promoted both attitude and perceived behavioral control (ISQ → ATT, β = 0.623; ISQ → PBC, β = 0.485), and system quality contributed the largest weight to the higher-order construct (0.805). This empirical result indicates that users first formed a broad judgment about whether GAI fitness assistants were stable, responsive, understandable, and supportive, and this judgment then shaped both favorable evaluations and perceived ability to use the assistant effectively.

This pattern is consistent with the IS Success Model and with prior digital health research showing that information, system, and service quality shape satisfaction, benefit perception, and intention [34,44,46]. It also aligns with research on conversational agents and large language models, where users are especially sensitive to fluency, response consistency, and the risk that plausible outputs may be inaccurate or unsupported [13,81,82].

In the GAI fitness context, quality is technology-specific because conversational interaction and iterative plan refinement are not merely interface conveniences; they shape exercise decisions with potential safety consequences. If an assistant hallucinates an exercise prescription, ignores injury history, misinterprets wearable data, or fails to explain intensity and recovery logic, users may perceive both lower value and higher injury risk. The practical implication is that developers should prioritize stable dialogue flows, wearable-data integration with clear consent, traceable explanations, risk warnings for unsafe exercise advice, and escalation pathways to human coaches or clinicians when the user’s condition exceeds the assistant’s safe boundary.

5.2 The role of subjective norm

Subjective norm had significant positive effects on attitude, perceived behavioral control, and perceived value (SN → ATT, β = 0.260; SN → PBC, β = 0.270; SN → VAL, β = 0.174), whereas its direct path to behavioral intention was not significant (SN → BI, p = 0.192). Empirically, social influence did not operate as a direct pressure to use the assistant; instead, it worked indirectly by shaping evaluations, perceived controllability, and value judgments. The formal indirect-effect tests further confirmed this interpretation, because the SN-related indirect paths through PBC, ATT, VAL, and the serial ATT → VAL route were all significant, whereas the direct SN → BI path remained non-significant.

This indirect pattern is broadly consistent with technology-adoption research showing that social influence often works through usefulness beliefs, attitude, and facilitating conditions rather than through a simple direct path [50,51]. Fitness-app studies have also shown that peer support, social features, and social comparison can influence exercise-related intentions and behavior by supporting self-regulation [26,28,83], while generative AI adoption studies in education similarly position subjective norm alongside attitude, perceived control, and AI literacy [38,52].

For GAI fitness assistants, the specific mechanism is informational rather than merely normative. Peer demonstrations, teacher or coach recommendations, and community discussions can provide prompt examples, safe-use cautions, and lessons about how to refine a plan after soreness, schedule changes, or wearable feedback. These forms of peer learning reduce trial-and-error costs and make perceived value more concrete. The implication is that promotion should emphasize support-oriented diffusion – such as prompt libraries, peer-shared cases, safety checklists, and coach-endorsed examples – rather than relying on social pressure or unsupported fear-of-missing-out narratives.

5.3 The role of attitude

Attitude exerted a significant direct effect on behavioral intention (ATT → BI, β = 0.275) and a stronger indirect pathway through perceived value (ATT → VAL, β = 0.608; VAL → BI, β = 0.447). This result shows that favorable evaluations of using GAI fitness assistants mattered in their own right, but they also became influential because they increased users’ belief that the assistant was worth the time, effort, cost, and risk involved. The bootstrapped indirect effect of ATT → VAL → BI was also significant, indicating that perceived value transmitted a substantial part of the association between favorable attitude and intention.

Prior TPB and generative AI adoption studies similarly identify attitude as a central antecedent of intention [37,38,52,53]. However, attitude differs from perceived value: attitude captures whether users evaluate the behavior positively, whereas perceived value captures whether benefits exceed sacrifices [39,40]. Modeling attitude as an antecedent of perceived value is therefore appropriate because positive affect, perceived convenience, and perceived usefulness can increase the salience of benefits while lowering the subjective burden of interaction effort.

In GAI fitness, attitude is fragile because a smooth conversational tone may mask hallucinated or unsafe exercise advice. Positive attitude is more likely to convert into value when the assistant explains why a plan is appropriate, admits uncertainty, uses wearable data transparently, and encourages human-coach escalation for pain, injury history, or medical-risk cases. The implication is that platforms should not rely only on enjoyable dialogue; they should design explainable and correctable interactions that let users revise prompts, compare alternatives, and understand the rationale behind load, frequency, progression, and recovery recommendations.

5.4 The role of perceived behavioral control

Perceived behavioral control had a significant but relatively modest effect on behavioral intention (PBC → BI, β = 0.163), and it was strengthened by both ISQ and SN. The empirical result suggests that perceived controllability functions as a feasibility threshold: users may value the assistant, but intention weakens when they doubt their ability to communicate goals, judge the safety of outputs, or integrate the assistant into daily training routines.

This result is consistent with TPB and AI adoption research emphasizing perceived capability, effort expectancy, self-efficacy, and facilitating conditions [15,16,47,48]. In generative AI adoption research, AI literacy and self-efficacy are also repeatedly identified as conditions that shape intention [52,55].

For GAI fitness assistants, PBC is technology-specific because users need prompt literacy, basic training knowledge, privacy and permission awareness, and the ability to interpret wearable-data-driven feedback. Users also need to know when a recommendation is too generic, too intense, or unsuitable for pain and injury conditions. The implication is that systems should provide onboarding, example prompts, adjustable plan templates, plain-language safety explanations, confidence or uncertainty cues, and simple routes to human coaching support, thereby turning perceived control into safer future-use intention rather than blind reliance.

5.5 The role of perceived value

Perceived value exhibited the strongest direct effect on behavioral intention (VAL → BI, β = 0.447). Empirically, users’ intention toward GAI fitness assistants was predicted most strongly by whether they believed the overall benefits of using the category outweighed the sacrifices involved.

This finding is consistent with value-based adoption research, which argues that intention depends on a trade-off between perceived benefits and perceived sacrifices [39,40]. It is also compatible with mHealth and fitness-app studies showing that usefulness, service experience, social support, and investment can influence adoption or future-use intention [18,26,28]. In generative AI chatbot services, perceived competence and warmth have likewise been linked to functional, social, and emotional value [54].

In the GAI fitness setting, perceived value combines distinctive benefits – conversational interaction, on-demand personalization, iterative plan refinement, motivational companionship, and possible integration with wearable data – with distinctive costs and risks, including privacy concerns, subscription costs, prompt effort, hallucinated or unsafe exercise advice, and injury risk. The implication is that providers should communicate value through verifiable outcomes and risk controls: transparent data use, explainable recommendations, clear safety boundaries, opt-in wearable integration, human-coach escalation, and pricing that matches the reliability and personalization actually delivered.

5.6 Model explanatory power and robustness

At the model level, the proposed framework explained 50.9% of the variance in behavioral intention, while the R² values for attitude (0.584), perceived value (0.507), and perceived behavioral control (0.412) were also meaningful. This empirical result indicates that integrating second-order information system quality, TPB mechanisms, and perceived value captured a substantial portion of intention formation, but it also leaves approximately 49.1% of behavioral intention unexplained.

Prior adoption studies similarly show that intention is rarely explained by a single theoretical lens; technological quality, social influence, trust, risk, effort, habit, and context-specific skills often work together [15,16,41]. For GAI fitness assistants, the unexplained variance is especially important because the technology category combines generative AI uncertainty with exercise safety, personal health data, and wearable-device ecosystems.

The omitted factors most likely include trust, perceived risk, privacy concerns, AI literacy, prompt self-efficacy, subscription cost, habit, wearable-device integration, prior fitness-app experience, and product-specific familiarity. The practical implication is that future models should treat the present framework as a meaningful baseline rather than a complete explanation. Longitudinal designs, usage logs, prompt and dialogue records, wearable indicators, and churn analyses would be needed to test whether these perceptions translate into actual use, safe training behavior, or discontinuance over time. Because the present data are cross-sectional and self-reported, the reported paths should be interpreted as theory-guided associations rather than causal effects.

6.1 Theoretical implications

The Information Systems Success Model and the Theory of Planned Behavior explain user adoption from the perspectives of technological quality evaluation and psychological decision-making mechanisms, respectively. However, in the digital health context of generative AI fitness assistants—characterized by high interactivity, substantial individual heterogeneity, and safety-related consequences—any single theory is often insufficient to capture the full chain from technology perceptions to intention formation. This study structurally integrated the IS Success Model with TPB and introduced perceived value as a key mediator, thereby developing an explanatory framework in which technological attributes, social cognition, and value trade-offs jointly explain intention to use. In doing so, the study extended the applicability boundaries and clarified the complementary relationship of these classic theories in generative-AI-enabled health behavior settings.

More specifically, this study consolidated information quality, system quality, and service quality into a second-order information system quality construct to capture users’ holistic judgments of the overall performance of GAI fitness assistants. It further demonstrated the significant effects of this higher-order quality construct on attitude and perceived behavioral control, highlighting the upstream role of quality perceptions in the intention-formation process. This specification not only responded to the integrated demands for stability, content reliability, and support assurance in fitness recommendation generation, but also provided a testable modeling approach for adopting higher-order quality constructs in future research on emerging generative AI applications. Accordingly, this study contributes to explaining category-level intention formation toward GAI fitness assistants in a safety-sensitive service context, without extending its claims to product-specific adoption, actual use, or post-adoption retention.

In addition, the findings showed that perceived value exerted the strongest explanatory power for intention to use, and that attitude and subjective norm indirectly promoted intention formation by enhancing value judgments. This pattern suggested that, in high-risk and sustained-engagement technology use, users’ decisions relied more on an overall weighing of benefits versus costs than on simple evaluations of liking or perceived usability. Meanwhile, subjective norm did not show a significant direct effect on intention to use but had significant effects on attitude, perceived behavioral control, and perceived value, indicating that social influence in this context was more likely to function as an antecedent source of cognitive cues and resource support rather than as a mechanism of direct normative pressure. These findings offered new empirical evidence regarding the boundary conditions of TPB’s social influence pathway in low-visibility, highly individualized health management behaviors. By formally testing the indirect effects, this study further clarified that subjective norm contributed to intention formation mainly through mediated routes, strengthening the theoretical argument that social influence in this context functions as informational and value-shaping support.

6.2 Practical implications

This study provides clear implications for the product design, operational strategies, and governance practices of generative AI fitness assistants. First, developers should treat improvements in second-order information system quality as foundational work for increasing adoption intention. In particular, system quality should be prioritized as the core source of overall quality by enhancing response speed, interaction stability, error tolerance, and process guidance to ensure a seamless conversational experience. This can help prevent training workflows from being disrupted by lag, interruptions, or complex operations, which would otherwise undermine users’ attitudes and perceived controllability. Second, information quality should be strengthened systematically by establishing content generation and verification mechanisms centered on recommendation accuracy, relevance, actionability, and contextual fit. For example, integrating movement-risk alerts, progressive overload rules, identification of individual constraints, and explanatory feedback can reduce uncertainty and safety concerns and improve users’ judgments of recommendation credibility and usefulness. Third, although service quality carries a relatively lower weight, it remains an important lever for reducing learning costs in the early stage of diffusion. Providing onboarding tutorials, prompt/example libraries, self-service troubleshooting for common issues, and, when necessary, human-AI collaborative support can reduce frustration and further enhance perceived behavioral control.

Second, operators should build the core growth logic around perceived value. Given that perceived value is the strongest driver of intention, platforms need to simultaneously strengthen the benefit side while lowering the cost side. On the benefit side, platforms should highlight gains from personalized plan generation, efficient dynamic adjustment, time and money savings, and experiential benefits derived from companion-like interaction. On the cost side, efforts should focus on reducing learning and communication burdens, privacy and data security concerns, and risk expectations associated with incorrect recommendations. Concretely, practices such as data minimization and transparent consent, traceable explanations for training recommendations, clear risk-boundary prompts, and appropriate referral mechanisms can increase certainty in the perceived return on investment, making it easier for users to form an overall judgment that the assistant is worth using in the future.

Third, managing subjective norms should shift from pressure-based promotion to support-oriented diffusion. Given the limited direct effect of subjective norm on intention, a more effective approach is to leverage communities and word of mouth to provide experiential cues and learning resources. Examples include building peer-shared training cases and dialogue templates, and encouraging users to present improvement processes and key precautions. This can translate social endorsement into more positive attitudes, stronger perceived controllability, and higher value judgments, rather than relying on others’ expectations alone. Finally, at the regulatory and industry levels, quality and value can be used as governance levers to promote safety and disclosure standards for generative-AI-based exercise recommendations, clarifying requirements for risk warnings, data protection, and responsibility boundaries. Such measures can reduce users’ perceived costs and uncertainty at the institutional level and provide external safeguards for the sustainable use of these tools in digital health contexts.