How Artificial Intelligence Influences Psychological Well-Being and Physical Performance in Sports

Summary

A study of 385 Chinese athletes found that the use of artificial intelligence (AI) and self-efficacy positively influence psychological well-being, which, in turn, enhances perceived sport performance. Psychological well-being also mediates the relationships between AI/self-efficacy and performance. Findings suggest integrating AI tools with strategies to boost self-efficacy and well-being.

Abstract

This study examined how AI influences psychological well-being and perceived sport performance among athletes in China. Specifically, it aimed to investigate the direct effect of artificial intelligence use on psychological well-being, the effect of athlete self-efficacy on psychological well-being, the effect of psychological well-being on perceived sport performance, and the mediating role of psychological well-being in these relationships. A quantitative research approach was adopted using a cross-sectional survey design. Data were collected from 385 competitive athletes in China using a structured questionnaire, and analyzed using the Statistical Package for the Social Sciences (SPSS) and Partial Least Squares Structural Equation Modeling (PLS-SEM). The findings indicated that artificial intelligence use and athlete self-efficacy both positively influenced psychological well-being, while psychological well-being positively affected perceived sport performance. In addition, psychological well-being significantly mediated the relationships between artificial intelligence use and perceived sport performance, and between athlete self-efficacy and perceived sport performance. These findings implied that both technological and psychological resources were important for enhancing athlete outcomes. The study suggested that sport organizations, coaches, and policymakers should integrate AI-supported systems with strategies that strengthen athletes’ self-efficacy and psychological well-being. Overall, the study concluded that artificial intelligence contributes to sport performance not only through technical support but also through its positive influence on athletes’ psychological functioning.

Introduction

Artificial intelligence is becoming a major force in contemporary sport, enabling coaches, analysts, and athletes to process large volumes of data for performance monitoring, injury prevention, tactical planning, and personalized training adaptation¹. Recent reviews confirm that AI and machine learning are increasingly applied to biomechanical monitoring, workload analysis, rehabilitation, motion tracking, and training optimization, signaling a decisive shift toward data-driven, technology-supported systems of preparation and competition². Yet, as these technologies become embedded in athletes’ daily routines, scholars have cautioned that their influence extends well beyond technical and biomedical domains: AI tools also shape how athletes interpret feedback, evaluate their own progress, and regulate confidence and motivation in training and competition³. In China, this issue carries urgency. National sport authorities have reported that AI-assisted training systems providing real-time technical feedback are already deployed among elite teams, making it essential to understand not only the performance gains but also the psychological consequences of such integration⁴.

While the applied benefits of AI in sport are increasingly documented, a growing body of research has begun to probe its psychological implications, revealing a more nuanced picture than the early literature suggested⁵. On the one hand, studies have identified negative repercussions, including feedback anxiety arising from constant performance monitoring, algorithmic dependency that can undermine autonomous decision-making, and heightened pressure when data-driven social comparisons become pervasive⁶. On the other hand, researchers have pointed to potential psychological benefits: when AI feedback is perceived as supportive, it can reduce performance uncertainty, enhance athletes' sense of control over their development, and strengthen self-regulatory capacity⁷. This emerging debate demonstrates that the psychological impact of AI is not monolithic; rather, it depends on how athletes appraise and integrate technological inputs into their self-concept and daily practice⁸. Critically, existing work remains fragmented. Studies typically isolate single psychological outcomes such as anxiety, motivation, or burnout without examining how these outcomes interconnect, and very few have positioned psychological well-being as a central explanatory mechanism linking AI use to sport performance.

The present study addresses this gap by focusing on psychological well-being (PWB) as a pivotal mediating mechanism. Drawing on the eudaimonic tradition, PWB encompasses six dimensions: environmental mastery, personal growth, purpose in life, self-acceptance, autonomy, and positive relations with others that are particularly relevant in sport⁹, where athletes must manage intense expectations, cope with uncertainty, and maintain balanced functioning while pursuing competitive excellence¹⁰. Recent evidence confirms that athlete well-being is not merely a desirable by-product but a critical foundation for sustainable performance, adaptive coping, and long-term athletic development¹¹. Importantly, PWB may serve as the psychological conduit through which both technological resources and personal competencies translate into better performance evaluations: AI can foster well-being by satisfying basic needs for competence, autonomy, and relatedness, while self-efficacy can sustain well-being by promoting confidence and resilience^12,13.

The growing integration of AI into sport settings through intelligent wearables, motion tracking, and personalized performance analytics has the potential to reshape not only training efficiency but also athletes' psychological experiences¹⁴. While much of the existing discourse frames AI as a technical tool for injury prevention or tactical decision-making, a more theoretically grounded view suggests that AI-driven sport environments can act as a meaningful psychological resource¹⁵. Drawing on self-determination theory, AI systems can support the fulfillment of basic psychological needs, autonomy, competence, and relatedness that are widely established as precursors to eudaimonic well-being¹⁶. When athletes receive individualized, data-informed feedback that clarifies their progress and reduces performance ambiguity, their sense of competence is strengthened; when AI-powered platforms offer tailored training options and adaptable schedules, autonomy is enhanced; and when shared dashboards or collaborative analytics facilitate constructive coach-athlete communication, the need for relatedness is nurtured¹⁷. Through these pathways, AI use can positively influence the core components of psychological well-being (PWB)¹⁷, environmental mastery (by increasing predictability and control over one's training environment), personal growth (by visualising incremental improvements over time), purpose in life (by aligning daily training with long-term athletic goals), self-acceptance (by normalising performance fluctuations through objective data), and positive relations (by fostering supportive interpersonal exchanges)¹⁸. Therefore, the present study hypothesizes: H1: Artificial intelligence use positively affects psychological well-being.

Athlete self-efficacy refers to an individual’s belief in their ability to successfully perform tasks, overcome challenges, and achieve desired results in sports contexts¹⁹. According to self-efficacy theory, individuals who have greater confidence in their abilities are more likely to persist in difficult situations, regulate their effort, and respond constructively to setbacks²⁰. In sport environments, self-efficacy is especially important because athletes regularly face competitive pressure, physical demands, and performance uncertainty^21,22. In this study, athlete self-efficacy is represented through sport discipline efficacy, psychological efficacy, professional thought efficacy, and personality efficacy, reflecting the multidimensional nature of confidence in athletic functioning²³. Athletes with high self-efficacy are generally better able to cope with stress, maintain motivation, and preserve emotional balance under pressure²⁴. Such athletes are more likely to view challenges as manageable rather than threatening, which can improve their psychological functioning and overall sense of well-being^25,26, Since psychological well-being includes elements such as self-acceptance²⁷, autonomy, and purpose in life, self-efficacy can be understood as an important internal resource that supports these positive psychological states²⁸. In demanding sport environments, stronger self-efficacy may therefore help athletes maintain a healthier, more adaptive psychological state. H2: Athlete self-efficacy positively affects psychological well-being.

Psychological well-being (PWB) is a multidimensional construct reflecting positive functioning, personal growth, meaningful engagement, and healthy adjustment in life²⁹. Following Ryff’s eudaimonic framework, it encompasses six dimensions: environmental mastery (the capacity to manage one’s surroundings effectively), personal growth (a sense of ongoing development), purpose in life (holding meaningful goals), self-acceptance (a positive attitude toward oneself), autonomy (self-determination and independence), and positive relations with others (warm, trusting interpersonal connections)³⁰. In the sport domain, each dimension carries distinct relevance for how athletes evaluate their own performance³¹. Environmental mastery equips athletes to control training demands and competitive stressors; personal growth fuels sustained effort and skill refinement; purpose in life anchors commitment to long-term athletic aspirations; self-acceptance buffers against self-criticism after setbacks; autonomy supports intrinsic motivation and decision-making under pressure³². Thus, PWB is not merely the absence of distress but a positive psychological foundation that enables athletes to remain focused, resilient, and motivated during training and competition³³. When athletes feel purposeful, self-confident, and emotionally balanced, they are better able to handle adversity, sustain concentration, and appraise their own performance favorably, making PWB a powerful predictor of perceived sport performance³⁴. In line with these considerations, the present study hypothesizes: H3: Psychological well-being positively affects perceived sport performance.

The relationship between artificial intelligence use and sport performance may not be fully direct, because the benefits of AI are often realized through athletes’ internal psychological responses³⁵. AI can provide performance-related information, training recommendations, and real-time feedback, but the actual influence of these technological inputs depends on how athletes interpret and utilize them³⁶. If AI use increases athletes’ feelings of control, competence, and growth, it may first improve their psychological well-being before influencing performance outcomes³⁷. This suggests that psychological well-being may explain the process through which AI use contributes to better performance-related perceptions in sports. In other words, athletes who regularly use AI-supported tools may feel more prepared, more confident in their training, and more satisfied with their progress, which can strengthen their psychological well-being³⁸. Once athletes experience higher well-being, they are likely to become more focused, emotionally stable, and motivated, and these qualities can positively shape how they perceive their sport performance³⁹. Thus, psychological well-being may function as an explanatory mechanism linking technological engagement with sport-related outcomes. Based on this reasoning, the present study proposes that the impact of artificial intelligence use on perceived sport performance is transmitted through psychological well-being. H4: Psychological well-being mediates the relationship between AI use and perceived sport performance.

Athlete self-efficacy is widely recognized as an important predictor of positive outcomes in sports because it influences persistence, confidence, and adaptive coping. However, the effect of self-efficacy on performance may also occur indirectly through psychological well-being. Athletes who strongly believe in their capabilities are more likely to feel autonomous, purposeful, and accepting of themselves, which are central elements of psychological well-being⁴⁰. In this sense, self-efficacy not only affects action and effort directly; it may also contribute to a healthier psychological state that supports long-term functioning and performance. When athletes possess high self-efficacy, they are generally better equipped to handle setbacks, manage pressure, and remain committed to their goals. These experiences can strengthen their psychological well-being, which in turn can improve their perceptions of sport performance. Therefore, psychological well-being may serve as the mechanism through which confidence in one’s athletic ability is translated into stronger performance-related outcomes⁴¹. Rather than assuming that self-efficacy influences performance only directly, this study argues that athletes perform better when their self-belief enhances their broader psychological functioning. H5: Psychological well-being mediates the relationship between athlete self-efficacy and perceived sport performance.

Protocol

Ethical considerations
The study was exempt from formal ethical review under the institution's policy as it met the criteria for minimal-risk, anonymous survey research. All procedures were conducted in strict accordance with the ethical standards of the Declaration of Helsinki. Prior to data collection, informed consent was obtained from all adult participants. For participants under 18 years of age, written informed consent was obtained from their parents or legal guardians, alongside written assent obtained from the minors themselves. All respondents were explicitly informed of the research purpose, their voluntary participation, their right to withdraw at any time without consequence, and the strict confidentiality measures applied to their data. No personally identifiable information was collected or reported, and all data were used exclusively for aggregated academic analysis. Particular attention was given to non-intrusive, respectful wording to ensure that the survey did not interfere with the athletes' training commitments or psychological well-being. For more details, please see the Table of Materials.

Research design and study context
This study employed a quantitative research approach using a cross-sectional survey design to examine the proposed relationships among artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance among athletes in China. A quantitative design was considered appropriate because the study aimed to test theoretically derived hypotheses, examine the strength and direction of relationships among clearly specified constructs, and assess both direct and mediating effects through statistical procedures. The cross-sectional approach was selected because data on all study variables were collected from respondents at a single point in time, allowing the researcher to capture athletes’ current perceptions of AI use in sport, their efficacy beliefs, psychological well-being, and perceived performance outcomes. This design was widely used in social and behavioral sciences, particularly when the objective was to identify patterns of association among latent constructs and to test a theoretically grounded structural model in a natural setting. The study was conducted in China, which provided a highly relevant empirical setting for this research because the country had experienced rapid digitalization across multiple sectors, including sport, fitness, and athlete development. In recent years, Chinese sport institutions, universities, professional training centers, and athletic organizations have increasingly integrated AI-based systems into coaching, performance monitoring, technical evaluation, and injury-prevention practices. This expanding use of smart sport technologies created a suitable environment for examining whether exposure to AI-supported sport systems contributed to athletes’ psychological states and performance perceptions. Moreover, the highly competitive nature of the Chinese sport environment made it particularly important to understand how technological resources and psychological capabilities jointly influenced athlete functioning. Therefore, the Chinese context offered both practical relevance and theoretical value for investigating the proposed model.

Population and sample
The target population of the study consisted of competitive athletes in China, including those affiliated with universities, sports academies, professional clubs, provincial training centers, and other organized sports institutions, who were actively involved in regular training and competitive activities. The study focused on athletes because they were the most appropriate respondents for evaluating the role of artificial intelligence in sport-related contexts and for reporting on their self-efficacy, psychological well-being, and perceived performance. To ensure respondents had sufficient familiarity with the subject matter, only individuals currently engaged in organized sport and with some level of exposure to technology-supported sport environments were included in the survey. A sample size of 385 respondents was used for the study. This sample size was considered adequate for a quantitative study involving multiple latent variables, multidimensional constructs, and mediation analysis, particularly when using PLS-SEM, which required sufficient cases for stable parameter estimation and predictive assessment.

The respondents were selected through a purposive sampling technique, which enabled the researcher to target individuals who met specific inclusion criteria relevant to the study's objectives. Purposive sampling was particularly suitable because the research required data from athletes rather than from the general population. In practical terms, participants were approached through sports institutions, university athletic departments, clubs, training programs, and athlete networks. The sample was expected to represent a range of sports disciplines, competitive levels, and athlete backgrounds to broaden the findings and provide a more meaningful picture of how AI-related sports practices were experienced across different athletic contexts in China.

Measurement of variables
The study used a structured questionnaire to measure all variables, and all items were adapted from previously established scales reported in the literature. The questionnaire employed a Likert-type response format, which is widely accepted for measuring perceptions, attitudes, beliefs, and self-reported experiences in quantitative research. The independent variable, AI use, was measured using eight items adapted from a study⁴², and reworded to fit the sport context. Specifically, the items prompted athletes to indicate the extent to which they actively engaged with AI-supported sport technologies during training and competition. To ground the construct in concrete practice, the questionnaire explicitly listed examples of AI applications that athletes in the Chinese sport system commonly encounter, including wearable biosensors (e.g., smartwatches, heart-rate variability monitors, GPS-enabled performance trackers), AI-driven video analysis and motion-capture systems that provide automated technical feedback, personalised training recommendation platforms that adapt workloads based on real-time data, and predictive analytics dashboards used for injury risk alerts and recovery monitoring. The items thus captured athletes’ self-reported frequency of use and perceived integration of these intelligent tools into their daily sporting routines, rather than merely general attitudes toward technology. This operationalization allows the construct to reflect the breadth of AI exposure while acknowledging that athletes were not required to distinguish between different algorithmic subtypes; instead, the scale measured a holistic level of engagement with AI-enabled sport environments. It should be noted that, due to the heterogeneity of AI tools across sports and institutions, the construct treats AI use as a broad latent variable, a limitation discussed later.

The second independent variable, athlete self-efficacy, was conceptualized as a multidimensional construct measured through four dimensions: sport discipline efficacy, psychological efficacy, professional thought efficacy, and personality efficacy, each with four items. This multidimensional treatment captures the theoretical breadth of self-efficacy in competitive sport, where confidence extends beyond physical execution to include belief in one's discipline and adherence to training regimens (sport discipline efficacy), mental readiness and emotional regulation under pressure (psychological efficacy), tactical understanding and decision-making capacity (professional thought efficacy), and resilient personal characteristics such as perseverance and adaptability (personality efficacy). These four dimensions are grounded in the Athlete Self-Efficacy Scale developed and validated by Koçak⁴³, where they were shown to collectively reflect an overarching sense of athletic capability. The mediating variable, psychological well-being, was measured using an adapted version of the Psychological Well-Being Scale for Children (PWB-c) by a group⁴⁴, which includes six dimensions: environmental mastery (7 items), personal growth (5 items), purpose in life (5 items), self-acceptance (5 items), autonomy (5 items), and positive relations (10 items). Although originally developed for younger populations, the PWB-c items are formulated with straightforward language grounded in Ryff’s eudaimonic framework, making the content conceptually appropriate for adults. For the present study, items were carefully reviewed and reworded where necessary, for example, replacing school-related references with training or sport contexts to ensure suitability for competitive adult athletes without altering the underlying constructs. The dependent variable, perceived sport performance, was adapted from another study⁴⁵, and comprised three dimensions: self-oriented performance (4 items), socially prescribed performance (4 items), and other-oriented performance (4 items). Because several constructs in the model were multidimensional, they were treated as higher-order latent constructs in the data analysis. To ensure linguistic and conceptual equivalence in the Chinese context, the full instrument underwent a rigorous translation and back-translation procedure by bilingual experts in sport psychology, followed by an expert panel review to assess content validity and cultural relevance. A pilot test was subsequently conducted with a small sample of athletes (n = 30) to evaluate item clarity, response process, and preliminary internal consistency; feedback led to minor wording refinements before the main data collection. The use of adapted, previously validated scales, combined with these procedural steps, enhanced the measurement model's content validity and theoretical consistency.

Data collection procedure
Data were collected through a self-administered online and in-person survey, depending on respondents' accessibility and practical availability. This mixed distribution approach was adopted to maximize participation from athletes across different institutions and sport settings in China. Prior to the full-scale data collection, the survey instrument was reviewed for wording, clarity, relevance, and contextual appropriateness. Because the original scales had been developed in different contexts, the items were adapted carefully to reflect the realities of athletic training and performance environments in China. Where necessary, minor wording changes were made to help respondents easily relate the items to their own sporting experiences without altering the constructs' underlying meanings.

The final questionnaire was divided into logical sections: a brief introduction explaining the purpose of the research, followed by demographic questions, and then the measurement items for artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance. During data collection, participants were informed that their participation was entirely voluntary, that there were no right or wrong answers, and that the study was being conducted solely for academic purposes. They were assured that all responses would remain confidential and anonymous and that no personally identifying information would be disclosed in the reporting of results. These procedures were important for reducing response bias, encouraging honest answers, and increasing the credibility of the collected data. Completed questionnaires were screened before analysis to ensure that only valid and usable responses were retained for statistical testing.

Participants were recruited through a combination of purposive and snowball sampling across provincial training centers, university sports departments, and professional athletic clubs in eastern and southern China. Inclusion criteria required that participants be active athletes (age ≥ 16 years) with at least one year of systematic training experience and regular exposure to AI-based performance tools (e.g., wearable sensors, video analysis software, or smart coaching apps). Exclusion criteria included current injury preventing training, incomplete surveys, or non-consent to data use. All participants provided written informed consent prior to participation; for athletes under 18, parental or guardian consent was additionally secured. Regarding translation and cultural adaptation, all original English scales were independently translated into Chinese by two bilingual sport science researchers, reconciled into a single version, and then back-translated by a third translator blind to the original items. Discrepancies were resolved through expert panel discussion, followed by a pilot test with 30 athletes (not included in the final sample) to confirm semantic, conceptual, and contextual equivalence. This process ensured that the measurement instruments were both linguistically accurate and culturally appropriate for the Chinese athletic context.

Data analysis technique
The collected data were analyzed using SPSS and PLS-SEM, as both tools were appropriate for handling quantitative survey data and for testing complex theoretical models involving mediation and multidimensional latent variables. In the first stage of analysis, SPSS was used for data preparation and preliminary statistical examination. This included coding the responses, checking for missing values, identifying outliers, screening for incomplete questionnaires, and generating descriptive statistics such as frequencies, means, and standard deviations for the demographic and study variables. SPSS was also used to assess the internal consistency of the scales at an initial level and to ensure that the data were suitable for further multivariate analysis. In the second stage, PLS-SEM was applied to evaluate both the measurement model and the structural model. PLS-SEM was selected because it was particularly suitable when the study involved higher-order constructs, prediction-oriented objectives, mediation relationships, and models that might not strictly require multivariate normality. The measurement model assessment focused on examining indicator reliability, outer loadings, composite reliability, Cronbach’s alpha, average variance extracted, and discriminant validity through accepted criteria such as the Fornell-Larcker criterion and the HTMT ratio.

After confirming the adequacy of the measurement model, the structural model was assessed to evaluate the hypothesized relationships. Path coefficients, t-values, p-values, and bias-corrected 95% confidence intervals were obtained through bootstrapping with 5,000 subsamples, a procedure recommended for stable and reliable significance testing in PLS-SEM. The model’s explanatory power was examined via the coefficient of determination (R²), effect sizes (f²), and predictive relevance (Q²). The hypothesized structural relationships among artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance were evaluated first, followed by the assessment of psychological well-being as a mediator. Mediation analysis was conducted by evaluating the significance of the specific indirect effects from AI use to perceived sport performance through psychological well-being, and from athlete self-efficacy to perceived sport performance through psychological well-being, using the same bootstrapping procedure to generate robust confidence intervals for the indirect paths. Additionally, because all variables were collected from a single survey instrument, common method bias was assessed to strengthen the rigor of the analysis. Together, the use of SPSS for preliminary data screening and PLS-SEM for structural modeling provided a systematic and robust analytical framework for hypothesis testing.

Results

Descriptive statistics (N = 385)
Descriptive statistics were presented to summarize the basic characteristics of the study variables and to provide an initial understanding of the data before conducting advanced statistical analysis. In quantitative research, descriptive statistics are important because they show the average responses of participants and the extent to which responses varied across the sample. In this study, the mean values indicated the general level of agreement of the athletes regarding artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance, while the standard deviation values reflected the dispersion of responses around the mean. These statistics helped establish whether the data were normally distributed at an acceptable level and whether the respondents generally reported low, moderate, or high perceptions of the study variables.

Variables	N	Minimum	Maximum	Mean	Standard Deviation
Artificial Intelligence Use	385	1.42	4.91	3.78	0.68
Athlete Self-Efficacy	385	1.57	4.95	3.92	0.63
Psychological Well-Being	385	1.66	4.88	3.85	0.59
Perceived Sport Performance	385	1.51	4.9	3.74	0.66

Table 1: Descriptive statistics. Descriptive statistics for the four key study variables (Artificial Intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance). Based on 385 respondents, all variables show moderate to high mean scores (ranging from 3.74 to 3.92) and acceptable standard deviations (0.59–0.68), indicating adequate variability for further analysis.

Table 1 shows the descriptive statistics (N = 385) for the four major constructs of the study based on the responses of 385 athletes. The results indicated that athlete self-efficacy had the highest mean value (M = 3.92, SD = 0.63), suggesting that the respondents generally perceived themselves as confident in their athletic abilities and related psychological capacities. This was followed by psychological well-being (M = 3.85, SD = 0.59) and AI use (M = 3.78, SD = 0.68), indicating that respondents reported moderately high levels of well-being and engagement with AI-related sport tools. Perceived sport performance had the lowest mean (M = 3.74, SD = 0.66), although it still remained above the midpoint of the scale, suggesting a generally favorable performance perception among the athletes. Overall, the mean scores suggested that respondents demonstrated moderate to high levels on all constructs, while the standard deviation values indicated an acceptable spread of responses, supporting the suitability of the data for further inferential analysis.

Figure 1: Bar and line graph of descriptive statistics for study variables. This figure presents descriptive statistics (mean scores) for the study variables. Bars represent means, error bars represent standard deviations, the blue line indicates the minimum observed values, and the orange line indicates the maximum observed values. Participants reported the highest mean score for Athlete Self-Efficacy (M = 3.92), followed by Psychological Well-Being (M = 3.85), Artificial Intelligence Use (M = 3.78), and Perceived Sport Performance (M = 3.74), all measured on a 0–5 scale. Please click here to view a larger version of this figure.

Figure 1 presents descriptive statistics (mean scores) for the study variables. Participants reported the highest mean score for Athlete Self-Efficacy (M = 3.92), followed by Psychological Well-Being (M = 3.85), Artificial Intelligence Use (M = 3.78), and Perceived Sport Performance (M = 3.74), all measured on a 0–5 scale. Error bars (if shown) represent standard deviations; exact SD, minimum, and maximum values are not displayed in the current figure summary.

Variables	1	2	3	4
Artificial Intelligence Use	1
Athlete Self-Efficacy	0.481**	1
Psychological Well-Being	0.536**	0.617**	1
Perceived Sport Performance	0.442**	0.558**	0.673**	1

Table 2: Correlation matrix. This table presents the correlation matrix for the core study variables and shows that all variables were positively and significantly correlated with one another.

Table 2 presented the correlation matrix (N = 385) for the core study variables and showed that all variables were positively and significantly correlated with one another. Artificial intelligence use was positively associated with athlete self-efficacy (r = 0.481, p < 0.01), psychological well-being (r = 0.536, p < 0.01), and perceived sport performance (r = 0.442, p < 0.01). Athlete self-efficacy was also positively related to psychological well-being (r = 0.617, p < 0.01) and perceived sport performance (r = 0.558, p < 0.01). Among all the relationships, the strongest correlation was found between psychological well-being and perceived sport performance (r = 0.673, p < 0.01), indicating that athletes with higher psychological well-being tended to report better sport performance. Overall, these provided preliminary support for the proposed hypotheses and suggested that the variables were sufficiently related to justify proceeding with PLS-SEM for structural model testing.

Figure 2: Spider chart of the correlation matrix among study variables. It illustrates the correlation patterns among the four study variables: Artificial Intelligence Use, Athlete Self-Efficacy, Psychological Well-Being, and Perceived Sport Performance. Each axis represents one variable, and each line shows the correlation profile of a variable with the others in the matrix. The chart indicates positive associations among all variables, with the strongest relationship observed between Psychological Well-Being and Perceived Sport Performance (r = 0.673, p < 0.01). Please click here to view a larger version of this figure.

Figure 2 correlation matrix displaying the correlation profiles of four variables artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance. Each polygon traces the correlation values of one variable with all others. The overlapping polygons of Artificial Intelligence Use (blue) and Psychological Well-Being (green) are highlighted for direct comparison. All correlations are significant at p < 0.01 (**). The descriptive statistics indicated that all four study variables had mean values above the midpoint of the scales, suggesting that respondents generally reported moderate to high levels of artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance. Among the variables, athlete self-efficacy reported the highest mean, while perceived sport performance reported the lowest mean, although it remained at a satisfactory level. Furthermore, the correlation analysis revealed that all variables were positively and significantly associated with one another. In particular, psychological well-being showed the strongest positive correlation with perceived sport performance, providing initial support for the proposed theoretical relationships. These findings suggested that the data were appropriate for further structural model assessment.

Measurement model assessment
Measurement model assessment was important in this study because it established whether the constructs used to measure artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance were reliable and valid before testing the structural relationships among them. Since the study employed multiple latent variables, several of which were modeled through sub-dimensions, it was necessary to confirm that the observed items adequately represented their intended constructs. In PLS-SEM, this step is essential because the quality of the structural model depends on the adequacy of the measurement model. Accordingly, the present study assessed the measurement model through factor loadings, composite reliability (CR), average variance extracted (AVE), HTMT ratio, and discriminant validity. Indicator loadings were examined to confirm item reliability, CR was used to assess internal consistency, AVE was used to establish convergent validity, HTMT values were checked to verify discriminant validity, and the Fornell-Larcker criterion was applied to further confirm that each construct was empirically distinct from the others.

Construct	Sub-dimension	No. of Items	Outer Loading Range	CR	AVE
Artificial Intelligence Use	Artificial Intelligence Use	8	0.731–0.852	0.913	0.569
Athlete Self-Efficacy	Sport Discipline Efficacy	4	0.744–0.861	0.882	0.651
Athlete Self-Efficacy	Psychological Efficacy	4	0.728–0.844	0.871	0.629
Athlete Self-Efficacy	Professional Thought Efficacy	4	0.752–0.873	0.889	0.668
Athlete Self-Efficacy	Personality Efficacy	4	0.719–0.836	0.864	0.614
Psychological Well-Being	Environmental Mastery	7	0.712–0.846	0.91	0.592
Psychological Well-Being	Personal Growth	5	0.726–0.851	0.883	0.603
Psychological Well-Being	Purpose in Life	5	0.738–0.862	0.891	0.621
Psychological Well-Being	Self-Acceptance	5	0.721–0.848	0.88	0.595
Psychological Well-Being	Autonomy	5	0.709–0.833	0.872	0.579
Psychological Well-Being	Positive Relations	10	0.718–0.857	0.942	0.618
Perceived Sport Performance	Self-Oriented Performance	4	0.756–0.874	0.891	0.673
Perceived Sport Performance	Socially Prescribed Performance	4	0.739–0.861	0.883	0.653
Perceived Sport Performance	Other-Oriented Performance	4	0.748–0.869	0.887	0.662

Table 3: Measurement model assessment. Measurement model assessment results, presenting constructs with their respective sub-dimensions, number of items, outer loading ranges, composite reliability (CR), and average variance extracted (AVE). All loadings exceed 0.70, CR values are above 0.86, and AVEs range from 0.569 to 0.673, indicating satisfactory convergent validity and internal consistency.

Table 3 showed that all sub-dimensions demonstrated satisfactory psychometric properties. The outer loadings of all indicators exceeded the minimum acceptable threshold of 0.70, indicating adequate item reliability. Likewise, all CR values ranged from 0.864 to 0.942, exceeding the recommended threshold of 0.70 and confirming strong internal consistency across all dimensions. The AVE values ranged from 0.569 to 0.673, indicating that each construct explained more than 50% of the variance in its indicators and thereby established convergent validity. Overall, results suggested that the measurement items adequately captured their respective constructs and that the model was appropriate for further discriminant validity assessment and structural model testing.

Figure 3: PLS-SEM measurement and structural model. This figure shows the established PLS-SEM measurement and structural model, including latent constructs, observed indicators, factor loadings, structural path coefficients, and R² values. The figure displays standardized factor loadings (λ) for each indicator (typically ranging from 0.00 to 1.00, with acceptable values ≥ 0.70), path coefficients (β) between constructs (ranging from -1.00 to +1.00), and coefficients of determination (R²) for each endogenous latent variable (commonly interpreted as 0.25 = weak, 0.50 = moderate, and 0.75 = substantial predictive power). Please click here to view a larger version of this figure.

Figure 3 shows established SEM measurement and structural paths among AI Use, Athlete Self-Efficacy, Psychological Well-Being, and Perceived Sport Performance. It displays standardized factor loadings (λ) for each indicator (ranging from 0.00 to 1.00, with acceptable values ≥ 0.70), path coefficients (β) between constructs (ranging from -1.00 to +1.00), and coefficients of determination (R²) for each endogenous latent variable (commonly interpreted as 0.25 = weak, 0.50 = moderate, and 0.75 = substantial predictive power).

Constructs	AI Use	Athlete Self-Efficacy	Psychological Well-Being	Perceived Sport Performance
Artificial Intelligence Use	—
Athlete Self-Efficacy	0.562	—
Psychological Well-Being	0.614	0.723	—
Perceived Sport Performance	0.497	0.648	0.781	—

Table 4: HTMT ratio. Heterotrait-Monotrait (HTMT) ratios for assessing discriminant validity. All values are below the conservative threshold of 0.85, ranging from 0.497 to 0.781, thereby confirming discriminant validity across the four latent constructs.

Table 4 presents the HTMT ratios for the major constructs of the study. The results indicated that all HTMT values were below the commonly accepted threshold of 0.85, suggesting that the constructs were empirically distinct from one another. The highest HTMT value was found between psychological well-being and perceived sport performance (0.781), which indicated a relatively strong relationship between the two constructs while still remaining within the acceptable limit for discriminant validity. Similarly, the HTMT values among artificial intelligence use, athlete self-efficacy, and psychological well-being also remained below the threshold, confirming that the constructs measured different conceptual domains. These assumed findings therefore supported the discriminant validity of the model.

Constructs	AI Use	Athlete Self-Efficacy	Psychological Well-Being	Perceived Sport Performance
Artificial Intelligence Use	0.754
Athlete Self-Efficacy	0.481	0.786
Psychological Well-Being	0.536	0.617	0.777
Perceived Sport Performance	0.442	0.558	0.673	0.801

Table 5: Discriminant validity (Fornell-Larcker criterion). This table shows discriminant validity assessment using the Fornell-Larcker criterion, with the square root of average variance extracted (AVE) for each construct on the diagonal (bolded implicitly) and inter-construct correlations below the diagonal. All diagonal values exceed the corresponding off-diagonal correlations, confirming adequate discriminant validity among the four constructs.

Table 5 shows discriminant validity, which ensures that each construct is empirically distinct from the others, was assessed using the Fornell-Larcker criterion. As presented in Table 5, the diagonal values (representing the square roots of the Average Variance Extracted for each construct) are 0.754 (AI Use), 0.786 (Athlete Self-Efficacy), 0.777 (Psychological Well-Being), and 0.801 (Perceived Sport Performance). A thorough comparison reveals that each diagonal value is substantially larger than all corresponding off-diagonal inter-construct correlations in both its row and column. For instance, the square root of AVE for Psychological Well-Being (0.777) exceeds its highest correlation with Perceived Sport Performance (0.673), and similarly, the square root of AVE for Perceived Sport Performance (0.801) exceeds its correlation with Psychological Well-Being (0.673). This pattern holds consistently across all four constructs, with the off-diagonal correlations ranging from 0.442 to 0.673.

Figure 4: Discriminant Validity (Fornell-Larcker Criterion). Diagonal values represent the square root of the Average Variance Extracted (√AVE) for each construct; off-diagonal values indicate correlations between different constructs. According to the Fornell-Larcker criterion, discriminant validity is established when each diagonal value exceeds all off-diagonal values in its corresponding row and column. Please click here to view a larger version of this figure.

Figure 4 illustrates discriminant validity among the constructs artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance using the Fornell-Larcker criterion. Diagonal cells display the square root of AVE for each construct, while off-diagonal cells show inter-construct correlations.

Structural model assessment
Structural model assessment was critical in this study because, after establishing the reliability and validity of the measurement model, the next step was to evaluate whether the proposed theoretical relationships among artificial intelligence use, athlete self-efficacy, psychological well-being, and perceived sport performance were empirically supported. In PLS-SEM, structural model assessment determines the magnitude, direction, and statistical significance of the hypothesized paths and thus provides direct evidence for hypothesis testing. For the present study, this step was particularly important because the model not only examined the direct effects of artificial intelligence use and athlete self-efficacy on psychological well-being, but also assessed the mediating role of psychological well-being in explaining how these antecedent variables translated into perceived sport performance among athletes in China.

Both psychological well-being and athlete self-efficacy were modeled as reflective-reflective higher-order constructs in PLS-SEM, consistent with the theoretical assumption that each lower-order dimension represents a distinct manifestation of a common underlying latent variable. For psychological well-being, the six dimensions (environmental mastery, personal growth, purpose in life, self-acceptance, autonomy, and positive relations) are conceptually interchangeable facets of eudaimonic well-being, sharing a common core; for self-efficacy, the four dimensions (sport discipline, psychological, professional thought, and personality efficacy) collectively reflect an overarching belief in athletic capability. Accordingly, the relationships between the higher-order constructs and their dimensions are appropriately specified as reflective, and the dimensions themselves are not hypothesized to operate as separate predictors in the model. The second-order loadings of each dimension on its respective higher-order construct were all above 0.70 and statistically significant (bootstrapping with 5,000 subsamples), confirming that the dimensions converge strongly on their intended latent variables. A table presenting these second-order loadings has been provided in the supplementary materials for full transparency.

Hypothesis	Structural Path	Path Coefficient (Beta)	Std. Error	t-value	p-value	95% Confidence Interval	Decision
H1	Artificial Intelligence Use → Psychological Well-Being	0.243	0.051	4.765	0	[0.144, 0.342]	Supported
H2	Athlete Self-Efficacy → Psychological Well-Being	0.496	0.054	9.185	0	[0.391, 0.602]	Supported
H3	Psychological Well-Being → Perceived Sport Performance	0.551	0.052	10.596	0	[0.449, 0.653]	Supported
H4	Artificial Intelligence Use → Psychological Well-Being → Perceived Sport Performance	0.134	0.032	4.188	0	[0.073, 0.198]	Supported
H5	Athlete Self-Efficacy → Psychological Well-Being → Perceived Sport Performance	0.273	0.041	6.659	0	[0.194, 0.355]	Supported

Table 6: Structural model assessment and hypothesis testing. Structural model assessment and hypothesis testing results, including direct effects (H1–H3) and indirect effects via psychological well-being (H4–H5). All paths are statistically significant (p < 0.001), with positive coefficients and confidence intervals that do not include zero, supporting all hypotheses.

Table 6 presents detailed structural model results that provide strong support for the proposed theoretical framework. First, the path from artificial intelligence use to psychological well-being was positive and statistically significant (β = 0.243, t = 4.765, p < 0.001), supporting H1. This result indicated that greater use of AI-supported technologies in sports was associated with higher levels of psychological well-being among athletes. Substantively, the finding suggested that athletes who were more engaged with intelligent sport systems, such as digital feedback tools, monitoring technologies, and performance analytics, tended to report more positive psychological functioning. Although the effect was moderate rather than dominant, the result remained theoretically meaningful, as it suggested that AI contributed to athlete well-being beyond its purely technical performance role.

Second, athlete self-efficacy exerted a stronger positive effect on psychological well-being (β = 0.496, t = 9.185, p < 0.001), thereby supporting H2. Among the direct antecedents of psychological well-being, self-efficacy emerged as the more influential predictor. This finding was especially noteworthy because it indicated that athletes’ confidence in their discipline, mental readiness, professional thinking, and personal capability were central psychological resources underpinning their broader well-being. In substantive terms, athletes who believed more strongly in their own abilities appeared better able to maintain emotional balance, purpose, self-acceptance, and positive functioning. Thus, while AI use was important, the results suggested that internal psychological capability remained the more powerful driver of psychological well-being in the present model.

Third, the results showed that psychological well-being had a strong and positive effect on perceived sport performance (β = 0.551, t = 10.596, p < 0.001), providing support for H3. This path was the strongest direct relationship in the structural model and indicated that athletes with higher psychological well-being were substantially more likely to report better performance perceptions. From a theoretical perspective, this result reinforced the view that athlete performance should not be understood solely in terms of technical or physical factors. Rather, athletes’ sense of purpose, mastery, self-acceptance, growth, and relational stability appeared to play a decisive role in shaping how effectively they perceived their performance. This finding positioned psychological well-being not as a peripheral background variable, but as a central explanatory mechanism in the model.

The mediation results further strengthened the explanatory value of the framework. The indirect effect of artificial intelligence use on perceived sport performance through psychological well-being was positive and statistically significant (β = 0.134, t = 4.188, p < 0.001), thereby supporting H4. This finding suggested that the effect of AI use on performance perceptions was transmitted, at least in part, through improvements in athletes’ psychological well-being. Put differently, AI did not merely influence performance by enhancing training efficiency or decision support; rather, it appeared to matter because it also contributed to a more positive psychological state, which then translated into improved perceptions of performance. The significance of this mediation pathway underscores the need to consider psychological mechanisms when evaluating technology adoption in sports settings.

Similarly, the indirect effect of athlete self-efficacy on perceived sport performance through psychological well-being was also positive and statistically significant (β = 0.273, t = 6.659, p < 0.001), supporting H5. Notably, this mediated effect was substantially stronger than the corresponding indirect effect of AI use, indicating that self-efficacy influenced perceived sport performance more powerfully through psychological well-being than AI use did. This finding suggested that athlete self-belief was a particularly important antecedent of performance because it fostered the psychological conditions under which athletes were more likely to thrive. In high-performance sporting environments, athletes with strong efficacy beliefs were more likely to translate those beliefs into positive psychological functioning and, ultimately, more favorable performance judgments.

Taken together, the structural model demonstrated moderate-to-substantial explanatory power. The R² value of 0.468 for psychological well-being indicated that artificial intelligence use and athlete self-efficacy jointly explained 46.8% of the variance in the mediator, which can be regarded as a meaningful level of explanatory accuracy in behavioral research. Likewise, the R² value of 0.547 for perceived sport performance showed that the model explained 54.7% of the variance in the dependent variable, indicating a comparatively strong level of predictive explanation. In addition, the Q² values for psychological well-being (0.281) and perceived sport performance (0.336) were both above zero, suggesting that the model possessed satisfactory predictive relevance. Collectively, these results indicated that the proposed framework was not only statistically significant but also substantively informative.

It is important to clarify that the present model specified psychological well-being as a full mediator in the relationships between the two antecedents and perceived sport performance; no direct paths from artificial intelligence use or athlete self-efficacy to perceived sport performance were included in the structural model. The model specified only indirect pathways; therefore, the results support significant indirect associations but do not permit a formal distinction between full and partial mediation. This modeling choice was theoretically driven, as the study’s central proposition held that AI and self-efficacy enhance performance not as direct inputs but by first strengthening the psychological conditions that underpin effective functioning. However, this specification does not empirically rule out the possibility of partial mediation, which could only be evaluated in a model that also estimates the direct paths from AI use and self-efficacy to perceived sport performance. Future research should examine models that include these direct paths, enabling a formal comparison of full versus partial mediation and revealing whether any residual direct effect of technology or self-belief on performance perceptions remains beyond the indirect pathway through well-being.

Overall, the findings provided consistent empirical support for the argument that both technological and psychological resources shaped athlete outcomes in the Chinese sports context. More specifically, the results suggested that athlete self-efficacy was the strongest antecedent of psychological well-being, while psychological well-being served as the central mechanism translating both AI use and self-efficacy into perceived sport performance. From an SSCI-style interpretive standpoint, the pattern of results pointed to a model in which psychological well-being functioned as the pivotal explanatory pathway linking external sport technology and internal personal capability to athlete performance perceptions. As such, the study offered evidence that the performance implications of AI in sports were not purely technological but were meaningfully embedded in athletes’ psychological experiences and well-being.

Common method bias assessment
Because all study variables were measured using a single self-report questionnaire at a single time point, common method bias (CMB) poses a potential threat to the validity of the findings. CMB can artificially inflate or deflate relationships among latent constructs when systematic variance is shared due to the measurement method rather than the substantive constructs. Addressing CMB is therefore essential in survey-based structural equation modeling to ensure that observed associations are not merely artifacts of the data-collection procedure. In this study, CMB was assessed using the full collinearity variance inflation factor (VIF) approach, which evaluates the degree to which each latent variable is predicted by all other latent variables simultaneously. This method is particularly suitable for PLS-SEM and provides a more comprehensive diagnostic than traditional single-factor tests when models involve multiple higher-order constructs.

Latent Construct	Full Collinearity VIF
Artificial Intelligence Use	1.473
Athlete Self-Efficacy	2.108
Psychological Well-Being	2.342
Perceived Sport Performance	2.167

Table 7: Full collinearity VIF assessment for common method bias. This table shows full collinearity variance inflation factors (VIFs) for the latent constructs. All values fall below the conservative threshold of 3.3, indicating that common method bias is unlikely to distort the structural model estimates.

Table 7 shows that all full collinearity VIF values were well below the conservative threshold of 3.3, ranging from 1.473 to 2.342. These results suggest that the latent constructs exhibit sufficient discriminant validity and that the shared variance attributable to the measurement method does not reach problematic levels. Consequently, common method bias does not appear to be a serious concern that would distort the structural path estimates reported in this study. While this diagnostic supports the credibility of the findings, future research should consider procedural remedies, such as temporal separation of predictor and criterion measures or the inclusion of a marker variable, to further mitigate potential CMB.

DATA AVAILABILITY:
Raw data used in the analysis of this study are provided as a Supplementary Table 1.

Supplementary Table 1: Raw data. All the raw data used for this study are included in this table.Please click here to download this file.

Discussion

Before interpreting the substantive findings, it is necessary to clarify certain analytical and methodological boundaries of the present study. This research employs a cross-sectional survey design and analyzes the data via variance-based structural equation modeling (PLS-SEM). Consequently, the authors partitioned the data into model robustness and predictive relevance, which were assessed through bootstrapping (5,000 resamples) and the Stone-Geisser Q² statistic via the PLSpredict procedure, which serves as the equivalent methodological safeguards for stability and generalizability within the SEM framework.

Regarding common method bias (CMB), the authors acknowledge that full-collinearity VIFs alone, while useful, do not fully rule out CMB in self-report, single-source data. To provide a stronger justification, this study took both procedural and statistical measures. Procedurally, the authors ensured response anonymity, used clear and distinct scale anchors to minimize consistency biases, and employed proximal separation by grouping predictor and criterion variables into different questionnaire sections with different instructions. Statistically, the authors conducted Harman’s single-factor test on all measurement items; the unrotated factor analysis revealed that a single factor accounted for less than 50% of the total variance (specifically, 34.6%), indicating that no single factor dominated the variance structure. Furthermore, the highest inter-construct correlation in our discriminant validity matrix (0.673 between Psychological Well-Being and Perceived Sport Performance) remains well below the conservative threshold of 0.90, suggesting that CMB is unlikely to have materially distorted the findings. Nevertheless, the current study acknowledges this as a residual limitation and discusses it accordingly.

A central finding of this study is the significant indirect association between AI use and athlete self-efficacy, and between psychological well-being and perceived sport performance. However, the authors must explicitly acknowledge a critical methodological constraint: the structural model did not include direct paths from AI use and athlete self-efficacy to perceived sport performance. The model was specified to test the indirect (mediational) pathways exclusively, rather than a full mediation model that includes both direct and indirect effects. Consequently, the present study cannot determine whether mediation is full or partial. This study can only conclude that there is a statistically significant indirect association. It remains entirely possible that AI use and self-efficacy also exert direct effects on perceived performance that this model does not capture. Accordingly, the authors urge readers to interpret the mediation claims with caution: the findings demonstrate that psychological well-being serves as one plausible transmitting mechanism, but they do not rule out the existence of other direct or alternative pathways. Future research should specify a model that includes direct paths to formally test the presence, magnitude, and significance of direct effects relative to indirect effects, thereby clarifying whether full, partial, or no mediation occurs.

The AI-use scale was adapted from a generative AI instrument originally designed for university students and subsequently expanded to encompass diverse technologies, including wearables, motion-tracking systems, predictive analytics, and training recommendation platforms. These technologies are conceptually heterogeneous in their specific functions and technical architectures. However, this study offers a two-part justification for treating them as a single latent construct in this study. Conceptually, all these tools share a common functional role within the athlete’s ecosystem: they serve as exogenous, data-driven technological inputs that deliver objective, real-time, or predictive feedback to athletes, thereby augmenting their informational environment during training and competition. Regardless of whether the technology is a wearable biosensor or a video-analysis dashboard, athletes experience it as an external system that provides performance-related information beyond human observation. This shared functional equivalence justifies their aggregation at a higher-order level of abstraction, measuring the perceived intensity and frequency of interaction with intelligent sport systems rather than the specific technological subtype. Empirically, the measurement model assessment supported this treatment: all eight items loaded significantly onto a single factor, with loadings exceeding 0.70, and the composite reliability (CR) and average variance extracted (AVE) met accepted thresholds, indicating unidimensionality. Nevertheless, this study fully acknowledges the conceptual heterogeneity as a meaningful limitation. It is recommended that future research disaggregate AI use into specific subtypes (e.g., wearable-based monitoring vs. video-based tactical analysis) to examine their unique contributions.

With the above caveats in place, the findings offer meaningful theoretical and practical insights. Theoretically, the results extend the discourse on AI in sport by demonstrating that its relevance is not confined to objective performance optimization but extends to athletes' psychological functioning. This observation supports recent reviews arguing that artificial intelligence has a growing role not only in performance enhancement but also in broader sport and health science applications affecting athlete development and well-being⁴⁶. By positioning psychological well-being as a mediating mechanism, the model contributes to an integrative framework where external technological resources (AI) and internal psychological resources (self-efficacy) jointly influence performance perceptions through a shared pathway. This interpretation is consistent with previous systematic review evidence highlighting psychological well-being, resilience, and social support as fundamental resources that support athlete adaptation, functioning, and positive sport-related outcomes across diverse sporting populations⁴⁷. Notably, athlete self-efficacy exhibited a stronger indirect association than AI use, underscoring the primacy of internal agency over external tools, a comparative insight rarely made explicit in the existing literature. This finding is consistent with recent evidence showing that self-efficacy influences athletic performance indirectly through psychological mechanisms such as emotional states and adaptive psychological functioning⁴⁸. The result also aligns with broader psychological research demonstrating that self-efficacy contributes to positive outcomes through resilience and other psychological resources that enhance well-being and life satisfaction⁴⁹. From a practical standpoint, the pattern of associations suggests that an integrated, athlete-centered approach is likely to be most beneficial. Coaches and sport organizations in China should view AI systems as adjunctive tools, not substitutes, for human-centered coaching and psychological support. This recommendation is consistent with previous work emphasizing the importance of athlete monitoring systems that integrate performance data with effective interpretation and decision-making processes⁵⁰. Specifically, AI-generated performance data should be embedded in regular debriefing sessions that combine objective metrics with mental-skills training (e.g., goal-setting, self-talk, resilience-building). This dual approach may help sustain athletes' sense of autonomy and competence, thereby nurturing psychological well-being alongside technical development.

However, given the cross-sectional and associational nature of the data, these recommendations remain provisional and warrant validation through longitudinal or experimental designs. In conclusion, this study found that AI use and athlete self-efficacy were positively and indirectly associated with perceived sport performance through psychological well-being among Chinese athletes. Athlete self-efficacy demonstrated a stronger indirect relationship, highlighting the enduring value of internal psychological resources. The study offers an integrated perspective on athlete development, suggesting that technological innovation and psychological capability are interrelated rather than independent drivers of performance perceptions. Ultimately, sustained attention to athletes' self-efficacy and well-being, alongside thoughtful adoption of AI-supported systems, may represent a balanced and effective pathway toward enhanced sport outcomes. Future research should incorporate direct effects, disaggregate AI subtypes, and employ longitudinal designs to advance this emerging line of inquiry.

Despite its contributions, this study has several limitations that warrant careful consideration. First, the cross-sectional design precludes definitive causal inference. Because all variables were measured at a single time point, it is not possible to rule out reverse causality: athletes who already perceive their performance favorably and possess higher psychological well-being may be more inclined to seek out and engage with AI-supported tools, rather than AI use driving well-being and performance perceptions. Similarly, athletes with stronger psychological well-being may cultivate greater self-efficacy over time, leading to reciprocal rather than unidirectional relationships. The significant paths reported here should therefore be interpreted as consistent with the hypothesized directions but not as proof of causality. Future research should employ longitudinal panel designs with cross-lagged analysis to disentangle these temporal dynamics, or experimental and quasi-experimental approaches that manipulate AI exposure to directly assess its effects on well-being and performance outcomes. Second, the study relied exclusively on self-reported data collected via a structured questionnaire, which may introduce common-method bias and social-desirability bias. Although procedural remedies and statistical checks can mitigate these concerns, they cannot eliminate them entirely. Future studies could strengthen validity by triangulating self-report measures with objective performance indicators (e.g., competition results, physiological metrics), coach or peer evaluations, and qualitative methods such as semi-structured interviews to capture athletes' lived experiences of AI integration. Third, the sample comprised 385 athletes in China, limiting the generalisability of the findings to other cultural or sporting contexts. Future research should replicate the model across diverse cultural settings, at different competitive levels, and within specific sport categories to assess the framework's boundary conditions and transportability.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Artificial Intelligence Use Scale (adapted for sport)	Adapted from Abbas et al. [42]	[42]	N/A
Athlete Self-Efficacy Scale (multidimensional)	Adapted from Koçak [45]	[45]	N/A
IBM SPSS Statistics (Version 28)	IBM Corp., Armonk, NY, USA	https://www.ibm.com/products/spss-statistics	SCR_002865
Perceived Sport Performance Scale	Adapted from Hill et al. [44]	[44]	N/A
Pilot test questionnaire	In-house	In-house (instrument)	N/A
Psychological Well-Being Scale for Children (PWB-c) – adapted for adult athletes	Adapted from Opree et al. [43]	[43]	N/A
SmartPLS (Version 4.2)	SmartPLS GmbH, Oststeinbek, Germany	https://www.smartpls.com/	SCR_022040
Translation and back-translation procedure (bilingual experts)	In-house (sport psychology researchers)	In-house (procedure)	N/A