Globally, health care systems face persistent challenges, including limited access to care, rising medical costs, and inefficiencies in service delivery [1]. These challenges are particularly pronounced in underresourced settings, where shortages of specialists and advanced medical equipment can lead to delayed diagnosis and adverse health outcomes. Against this backdrop, AI has emerged as a promising technological innovation with the potential to enhance health care accessibility [2,3], efficiency [4,5], and quality [6,7]. By reducing reliance on specialists and improving service delivery, AI health services may expand access to health information and support health management [2,8,9], particularly for populations facing barriers to health care [3]. Accordingly, AI health services are increasingly viewed not only as tools for improving health care delivery but also as technologies with the potential to support population health.

In recent years, health satisfaction has received increasing attention in public health research [10-13] because it captures individuals’ subjective evaluation of their own health and physical condition [14]. Unlike objective health status, which primarily reflects disease or functional limitations, subjective health evaluations integrate individuals’ perceptions of their health experiences and their perceived control over health [15]. These evaluations are also informed by bodily sensations that are directly accessible only to the individual and may reflect subtle physiological changes that are not readily captured by objective clinical indicators, thereby providing a unique source of information about health status [15]. Previous studies have shown that health satisfaction is associated with preventive health behaviors [16], health care utilization [17], and health outcomes [18]. Moreover, subjective health evaluations have been shown to be associated with mortality [19-21]. These findings indicate that health satisfaction captures meaningful dimensions of health that extend beyond objective clinical measures. It therefore offers a useful lens for examining whether AI health service use is related to how individuals perceive their own health.

South Korea provides a particularly informative setting for examining this question. As one of the world’s most digitally advanced societies, South Korea has achieved near-universal internet connectivity and widespread smartphone adoption, creating favorable conditions for the diffusion of AI health services. The Korean government has further promoted AI-Internet of Things–based health care services under its Fourth Industrial Revolution strategy [22], while the Fifth National Health Plan (HP2030) emphasizes reducing health disparities and strengthening health promotion [23]. These technological and policy developments make South Korea an informative setting for examining whether AI health services are associated with improved subjective health-related evaluation. Findings from this context may provide broader insights into the role of AI health services in improving population health in increasingly digitalized societies.

Despite these favorable conditions, empirical evidence examining whether AI health service use is associated with improvements in individual subjective health satisfaction remains limited, particularly in South Korea. Existing studies have primarily focused on efficiency gains [6,24], improvements in health care service delivery [25,26], users’ satisfaction [27], or have provided largely descriptive insights [1,8,28] rather than direct evidence on whether these services translate into meaningful improvements in individual subjective health satisfaction. Furthermore, little is known about whether the health benefits of AI health services are distributed evenly across socioeconomic groups. Addressing these gaps is important for understanding not only the effectiveness of AI health services but also whether their potential benefits are distributed similarly across different population groups.

Against this background, this study had 2 objectives: (1) to examine the association between AI health service use and health satisfaction, and (2) to investigate whether this association differs across socioeconomic groups. To achieve these objectives, nationally representative data from the 2024 Digital Divide Survey in South Korea were analyzed using a 2-stage least squares (2SLS) instrumental variable (IV) approach to address potential endogeneity arising from self-selection and reverse causality.

This study used data from the 2024 Digital Divide Survey commissioned by the Ministry of Science and Information and Communications Technology of South Korea and conducted by the National Information Society Agency (NIA). The survey covers both the general population and digitally vulnerable groups, including persons with disabilities, North Korean defectors, low-income individuals, marriage immigrants, and farmers and fishermen. The survey employed a stratified probability proportional sampling method organized by metropolitan local governments. Data were collected through face-to-face interviews between October and December 2024, yielding an initial sample of 15,000 respondents. This cross-sectional study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines (Checklist 1).

The data were obtained from the 2024 Digital Divide Survey (National Approved Statistics number 120017). The survey is an official government-administered statistical program conducted in accordance with national data protection regulations. All data were fully anonymized prior to public release. The dataset was obtained through the NIA, and the applicable terms of data use were followed. The present study involved a secondary analysis of pre-existing data; the authors had no direct contact with respondents and did not collect or record personally identifiable information. Under the relevant legislation of the Republic of Korea, Article 15(2) of the Bioethics and Safety Act and Article 13(1)(3) of its Enforcement Rule provide that research using publicly available information or preexisting data or documents without collecting or recording personally identifiable information may be exempt from institutional review [34]. On this basis, ethics review was not sought for the present analysis.

All analyses were conducted using R software (version 4.4.3; R Foundation for Statistical Computing). First, descriptive analyses were conducted to summarize the demographic and socioeconomic characteristics of the study sample. Categorical variables were reported as frequencies and percentages, while continuous variables were summarized using means and SDs.

Second, the primary association between AI health service use and health satisfaction was estimated using a 2SLS IV approach. In the first stage, individual AI health service use was regressed on the IV. In the second stage, the predicted value of AI health service use from the first stage was included in a linear regression model of health satisfaction. All models adjusted for age, gender, educational attainment, household income, residential location, disability status, and living arrangement. Robustness analyses were conducted using an alternative sample restriction, an alternative estimation method, and alternative IV specifications. In both the primary and robustness analyses, survey sampling weights provided by NIA were incorporated to account for the complex survey design, including the oversampling of digitally vulnerable groups, and to improve the national representativeness of the estimates. In both sets of analyses, cluster-robust standard errors were also estimated based on the peer groups used for IV construction.

Finally, subgroup analyses were conducted to examine potential socioeconomic heterogeneity in the association between AI health service use and health satisfaction. Analyses were stratified by gender, age group (<19, 19‐45, 46‐64, and ≥65 y), residential location (capital area vs noncapital area), living arrangement (living alone vs not living alone), education level (below high school vs high school or above), disability status (with disability vs without disability), and household economic level (low income vs high income). Within each subgroup, the 2SLS model was re-estimated using the same covariate specification as the primary model, with survey sampling weights applied. To account for multiple subgroup comparisons, Benjamini-Hochberg adjusted P values were calculated. Interaction terms between AI health service use and each subgroup variable were additionally tested in the full sample to evaluate effect modification. In the IV interaction models, the interaction between AI health service use and each subgroup variable was also instrumented. Specifically, AI health service use was instrumented using the peer adoption instrument, and the interaction term was instrumented using the interaction between the peer adoption instrument and the corresponding subgroup variable.

All reported regression coefficients are unstandardized coefficients with corresponding 95% CIs. All statistical tests were 2-sided, and statistical significance was defined as P<.05.

Table 1 presents the descriptive characteristics of the study population (N=15,000). The mean age of respondents was 49 (SD 18.19) years. Slightly more than half of the participants were men (n=7583, 50.55%), while 7417 (49.45%) were women. Regarding socioeconomic characteristics, 10,676 (71.17%) respondents had completed high school education or above, and 2364 (15.76%) were classified as high-income households. Approximately 2481 (16.54%) participants were living alone. In terms of residential location, 4568 (30.45%) resided in the capital area, whereas 10,432 (69.55%) lived in noncapital regions. Additionally, 2367 (15.78%) respondents reported having a disability. With respect to AI health services, 2541 (16.94%) participants reported use, while 12,459 (83.06%) reported no use. The mean level of health satisfaction was 2.73 (SD 0.74) on a 4-point scale.

Table 2 presents the 2SLS regression results with cluster-robust SEs. Because the leave-one-out peer adoption rate could not be calculated for peer groups containing only a single individual, 15 respondents were excluded, resulting in an analytical sample of 14,985 individuals. In the first-stage regression, the IV was strongly associated with AI health service use (β=0.625, SE 0.043, 95% CI 0.541‐0.710; P<.001), supporting the relevance of the instrument. The first-stage weak instrument F statistic was 340.10, substantially exceeding the critical value of 104.7 [37], indicating that the instrument was sufficiently strong. In the second-stage regression, AI health service use was positively associated with health satisfaction (β=0.739, SE 0.178, 95% CI 0.391‐1.088; P<.001). Overall, these findings support Hypothesis 1.

Figure 1 summarizes subgroup-specific estimates and formal interaction tests for the association between AI health service use and health satisfaction. Overall, AI health service use was positively associated with health satisfaction across most subgroups, and most subgroup-specific associations remained statistically significant after Benjamini-Hochberg correction. However, the subgroup-specific estimates were not statistically significant among respondents younger than 19 years or among those with below-high-school education.

Formal interaction tests showed significant subgroup differences by gender, household economic status, residential location, living arrangement, and disability status. The positive association was stronger among men than women (β=0.892 vs 0.478; P for interaction=.02), low-income than high-income respondents (β=0.896 vs 0.670; P for interaction=.001), respondents living outside the capital area than those living in the capital area (β=0.960 vs 0.418; P for interaction=.02), individuals living alone than those not living alone (β=1.259 vs 0.717; P for interaction=.02), and respondents with disabilities than those without disabilities (β=0.908 vs 0.695; P for interaction <.001).

No statistically significant interactions were observed for age (P for interaction=.48) or educational level (P for interaction=.85), indicating that the association between AI health service use and health satisfaction did not appear to differ significantly across these subgroups. Overall, these findings provided partial support for Hypothesis 2. The interactions for household economic status, residential location, living arrangement, and disability status were consistent with the Resource Substitution perspective. A significant gender interaction was also observed, although the stronger association among men was not directly predicted by Hypothesis 2, while age and educational attainment showed no evidence of effect modification.

This study examined whether AI health service use was associated with health satisfaction and whether this association differed across socioeconomic groups using nationally representative data from South Korea and an IV approach to address potential endogeneity. The findings demonstrated that AI health service use was positively associated with health satisfaction. This finding extends previous research that has primarily focused on the technological capabilities of AI health care, including improvements in health care efficiency, diagnostic performance, and service delivery [24], by showing that AI health service use was also positively associated with individuals’ health satisfaction.

Another important finding of this study is that the association between AI health service use and health satisfaction was not uniform across population groups. The positive association was significantly stronger among individuals with lower household income, those living outside the capital area, people living alone, and individuals with disabilities. These findings are broadly consistent with the Resource Substitution perspective, which suggests that newly available resources may generate greater benefits among individuals with fewer existing resources [32]. Although previous discussions of digital health inequalities have emphasized that disadvantaged populations may be less likely to access and adopt emerging technologies [38-40], the present findings suggest a more nuanced picture. While inequalities in technology adoption may persist, those who do adopt these services may derive greater health-related benefits when alternative health care resources are limited. These findings add to existing discussions of digital health equity [2,3].

The interaction tests did not identify significant variation in the association by age or educational attainment. Although men are not conventionally characterized as socioeconomically disadvantaged, the stronger association observed among men may be interpreted through a broader application of the Resource Substitution perspective that considers differences in existing health-information resources. Previous research in South Korea has found that women are more likely than men to use the internet for health information [41] and report higher levels of health literacy in several domains [42]. Women may therefore already draw on a broader range of health-information resources, reducing the additional value provided by AI health services. In contrast, AI health services may offer men an additional, self-directed channel for obtaining health information and managing their health, resulting in a greater marginal association with health satisfaction.

The observed positive association in this study may be explained through several potential pathways. First, AI health services may promote health self-management by providing personalized exercise coaching [43,44], dietary guidance [45,46], personalized health reminders [47], and continuous monitoring [48]. Behavioral improvements facilitated by AI health services may increase health management self-efficacy, ultimately contributing to higher health satisfaction. In addition, AI health services provide continuous access to personalized health information and immediate feedback, which may alleviate health-related anxiety and improve individuals’ health satisfaction. These behavioral and psychological processes may help explain the positive association.

This study makes 3 contributions to the existing literature. First, by conceptualizing health satisfaction as a user-centered, self-reported appraisal of physical and mental health, this study broadens the evaluation of AI health services beyond conventional indicators such as technological performance [7,49], diagnostic accuracy [24,50], and operational efficiency [24,51]. Second, the heterogeneity findings further refine the applicability of the Resource Substitution [33] perspective in the context of digital health. The larger estimated associations observed among lower-income respondents, residents outside the capital area, people living alone, and people with disabilities are consistent with this perspective, suggesting that AI health services may have greater marginal value when conventional health information, social support, or health-management resources are relatively limited. Third, this study provides new empirical evidence from South Korea as a highly digitalized critical case. Even in a setting characterized by advanced digital infrastructure, widespread connectivity, and strong policy support for AI health services, the association between AI health service use and health satisfaction remained socially patterned. This suggests that individuals’ ability to translate technological access into meaningful health-management support continues to be shaped by complementary socioeconomic and social resources.

Several limitations should also be acknowledged. First, although the IV approach and multiple robustness analyses strengthened confidence in the findings, the cross-sectional design prevents definitive conclusions regarding long-term causal effects. Second, health satisfaction was measured using a self-reported indicator and may be influenced by reporting differences across individuals. Future research should incorporate objective health indicators to provide a more comprehensive assessment of health outcomes. Third, due to the limitations of the available data, the survey did not ask respondents to report the specific types of AI health services they used. Therefore, this study could not further examine whether different types, frequencies, or intensities of AI health service use were differently associated with health satisfaction. Finally, although positive associations were observed across the primary and alternative IV specifications, the exclusion restriction cannot be directly tested. Peer adoption may still capture unobserved group-level characteristics, such as health literacy or shared attitudes toward technology, which may also be related to health satisfaction. Therefore, the findings should be interpreted as reflecting associations rather than definitive causal relationships. Future research may consider using more rigorous research designs, such as randomized controlled trials, to validate the causal effects of AI health service use on health satisfaction.

These findings also have important implications for AI health policy. In promoting AI health services, policymakers may consider paying greater attention to socioeconomically disadvantaged groups, including individuals with lower income, residents outside the capital area, people living alone, and individuals with disabilities. Targeted strategies, such as improving affordability, accessibility, and digital support for these groups, may facilitate their effective use of AI health services and help them derive greater health-related value from these technologies.

Nevertheless, these findings should not be interpreted as suggesting that AI health services are universally beneficial. Previous studies have shown that large language models and AI health applications may generate incomplete or misleading health information, particularly when responding to complex clinical questions [52,53]. Therefore, AI health services should be viewed as complementary tools that support, rather than replace, professional health care.

AI health service use was positively associated with health satisfaction, with stronger associations observed among men, individuals with lower income, residents outside the capital area, people living alone, and individuals with disabilities. These findings suggest that AI health services may serve as complementary health resources for populations with relatively limited health and social resources. More broadly, these findings highlight the importance of evaluating AI health services not only in terms of technological performance but also in terms of their potential contributions to subjective health outcomes and the distribution of these benefits across socioeconomic groups. These insights may help inform the design and implementation of AI-enabled health policies that better reflect the diverse needs of different population groups.