A formal review protocol was not prospectively registered. Before screening began, the research team developed the review objectives, eligibility criteria, search strategy, screening procedures, and standardized data-extraction form. The complete database search strategies are provided in Multimedia Appendix 1, and the completed PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist is provided in Checklist 1 [15]. The review aimed to map and synthesize the existing literature on AI and chatbot-supported interventions targeting pediatric PA and obesity-related lifestyle behaviors, with particular attention to the extent and nature of family involvement.

Eligibility criteria were guided by the population, intervention, comparator, outcomes, and study design (PICOS) framework. Studies were included if they: (P) involved children or adolescents aged 0‐18 years or parents, caregivers, or families supporting the health behaviors of a child or adolescent within this age range; (I) evaluated an AI-enabled chatbot, conversational agent, or related digital system designed to support PA or obesity-related lifestyle behaviors; (C) included any comparator, such as usual care or an alternative intervention, or no comparator; (O) reported outcomes related to PA or exercise, sedentary behavior or screen time, obesity, overweight, or weight management; and (S) reported empirical quantitative, qualitative, or mixed methods findings.

Studies were eligible if they involved children or adolescents or parents, caregivers, or families supporting the health behaviors of children or adolescents. Direct parent or caregiver involvement was not required because family involvement was examined as an analytic dimension rather than an eligibility criterion.

Studies were excluded if they described general digital health tools without an automated conversational, chatbot, AI-enabled, adaptive, or personalized support component, were purely technical development papers without behavioral outcomes, or were nonempirical publications such as editorials, commentaries, protocols, or review articles. Only studies published in English were included.

All retrieved records were imported into Rayyan (Qatar Computing Research Institute), a web-based systematic review management platform [16]. Rayyan was used to manage records, identify potential duplicates, and support blinded title and abstract screening. Potential duplicates identified through bibliographic metadata were manually reviewed and confirmed before removal. Final eligibility decisions were made by the reviewers rather than by the software. Title and abstract screening were conducted independently by 2 reviewers (DP and SA), who were blinded to each other’s decisions during the initial screening phase. At this stage, records were excluded if they were clearly unrelated to pediatric, adolescent, or family populations; did not involve AI-enabled chatbots, conversational agents, or related automated digital systems; did not address PA, obesity, sedentary behavior, weight management, or related lifestyle behaviors; or were clearly nonempirical publications. Records were retained for full-text review when eligibility could not be determined from the title and abstract alone. Full-text screening was then conducted independently by 2 reviewers (DP and SA) using the complete eligibility criteria. At this stage, specific reasons for exclusion were recorded. Articles were excluded during full-text review if they did not report PA or related lifestyle outcomes, did not include children, adolescents, parents, caregivers, or families, did not include an AI-enabled or chatbot-supported intervention component, or provided insufficient information to determine eligibility. Disagreements were resolved through discussion, and unresolved disagreements were adjudicated by a third reviewer (QJ). Interrater reliability was assessed using Cohen kappa (κ=0.81).

Data were extracted using a standardized form developed for this review. Extracted variables included publication and study characteristics (author, year, country, design, and sample size); participant characteristics (target population, age, sex, and family involvement); intervention characteristics (intervention type, delivery platform, duration, comparator, wearable or sensor use, and behavior change components); AI-related features (AI or automation role, intervention system or platform, reported AI technique or automation logic, input data type, and AI tasks or outputs); theoretical foundations; and reported outcomes, including feasibility, acceptability, usability, engagement, PA, and related lifestyle outcomes, secondary outcomes, and key findings. Extracted data were synthesized narratively to characterize the scope of the evidence and identify patterns across study design, intervention delivery, AI functionality, family engagement, and outcomes.

A formal critical appraisal or methodological quality assessment was not conducted. The purpose of this scoping review was to map the characteristics, applications, and gaps within an emerging and heterogeneous evidence base rather than to assess the certainty of evidence or estimate pooled intervention effects.

Figure 1 presents the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram. The initial database search identified 2730 records. After removing 363 duplicate records, 2367 unique articles remained for title and abstract screening. Of these, 2231 were excluded because they were clearly outside the review scope. A total of 136 full-text articles were assessed for eligibility. During full-text review, 122 articles were excluded for the following primary reasons: no PA or related lifestyle outcome, not child-, adolescent-, parent-, caregiver-, or family-focused, no AI-enabled or chatbot-supported component, or insufficient information to determine eligibility. Ultimately, 14 studies met the inclusion criteria and were included in the final sample (Figure 1).

A total of 14 original studies were included in this scoping review [17-30] (Table 1). The studies were published between 2021 and 2026, with 12 of the 14 published in 2023 or later. The 14 studies were conducted across 10 countries. The United States and South Korea were the most frequently represented settings (n=3 each), followed by one study each from Turkey, Spain, Italy, the United Kingdom, Belgium, Switzerland, Greece, and Canada. Although the included studies were geographically diverse, all were conducted in middle- or high-income contexts.

Among included studies, 3 studies were randomized controlled trials [19,24,26], one was quasi-randomized [25], one was a pre-post feasibility study with a comparison group [17], and 3 were single-arm or uncontrolled pilot or pre-post studies [18,22,27]. The remaining 6 were noneffectiveness-focused designs, including one qualitative adaptation study [30], one co-design mixed methods study [23], 2 development or usability studies [20,28], one cross-sectional survey [21], and one autoethnographic feasibility or development paper [29]. Sample sizes ranged from 9 to 900 participants, with most studies enrolling relatively small samples; only one large, randomized trial included more than 500 participants [19].

The 14 studies targeted a broad range of pediatric populations, including infants, school-aged children, and adolescents (Table 2). Most studies focused on school-aged children or adolescents, whereas one large trial focused on infants from birth to 24 months [19]. Several studies targeted specific clinical or high-need populations, including youth with autism spectrum disorder (n=2) [17,30], children or adolescents with obesity or overweight (n=5) [18,22,24,26,29], childhood or adolescent cancer survivors (n=1) [21], and adolescents facing barriers to PA or living in under-resourced settings (n=2) [23,27].

Ten [17-21,25,26,28-30] of the 14 studies included a parent, caregiver, or family component in the target population or intervention design, such as parent-child dyads, family-based obesity management, or studies collecting both child and caregiver perspectives. The remaining 4 [22-24,27] were primarily youth-focused, although family influence was acknowledged as an important contextual factor. The degree of family involvement varied considerably, ranging from minimal parental support or consent to parent-mediated delivery, co-participation in activities, shared monitoring, and parent-facing intervention interfaces.

Reporting of demographic and contextual characteristics was inconsistent. Child age was available in all 14 studies. While parents’ age was reported in only 4 studies [17,19-21]. Female participation or sex distribution was reported in 11 studies [17-24,26,27,29]. Among studies reporting child sex distribution, the proportion of female participants ranged from 16.7% to 65.0% [20,24]. Language or linguistic context could be identified in all 14 studies. However, cultural tailoring was reported in only one study [19], and 4 additional studies [20,21,23,30] described some form of contextual, linguistic, developmental, or population-specific tailoring.

Intervention design and delivery characteristics were presented in Table 3. Mobile app– or smartphone-based delivery was the most common format, appearing in 9 studies [17-19,22,24-27,29]. Six studies [19,20,23,28-30] incorporated a clearly web-based component, such as a dashboard, web platform, online interface, or browser-based system. Other delivery modes included chatbot-supported platforms, WhatsApp (Meta)-supported programming, SMS text messaging or responsive text messaging, serious games, metaverse-supported systems, and hybrid platforms integrating child-, parent-, and clinician-facing tools.

Nine studies [17-19,22,24-27,29] reported a defined intervention or exposure period, with durations ranging from 2 weeks to 24 months. Four studies [17,24,25,27] had relatively short intervention periods of 2-4 weeks, 3 [22,26,29] lasted approximately 8-12 weeks, and 2 [18,19] extended beyond 5 months, including one 24-month randomized trial and one 12-month obesity management program. The remaining 5 studies [20,21,23,28,30] were qualitative, developmental, survey-based, or case-based and therefore did not involve a conventional intervention period.

Most interventions were multicomponent and incorporated established behavior change strategies. Self-monitoring or tracking components were present in 9 studies [18-22,25-27,29], personalized feedback or adaptive content in 10 studies [17-20,23,25-27,29,30], and education or informational support in 9 studies [18,19,21-23,25-27,30]. Goal setting or challenge-based components were identified in 7 studies [19,22,25-28,30], gamification or reward-related elements in 7 studies [17,20,22,24-26,29], and social support, co-participation, or human-supported features in 8 studies [17-20,23,25,29,30]. Reminders or prompts were described as a distinct feature in 4 studies [17,24,26,27]. Although all studies were relevant to PA, only 5 [17,23,24,27,28] focused primarily on PA or exercise promotion; the remaining 9 [18-22,25,26,29,30] embedded PA within broader obesity-management, lifestyle, rehabilitation, or healthy living interventions.

Wearable devices, sensor-based inputs, or objective digital monitoring were reported in 6 studies [18,24,25,27-29]. These included Fitbit (Google LLC) devices, accelerometers, computer vision–based gesture recognition, smartphone- or app-linked monitoring, and planned smartwatch integration. However, most studies still relied partly or primarily on self-reported behavior, questionnaires, or user-entered information.

The included studies varied substantially in the sophistication and reporting of their AI and automation features (Table 4). To improve clarity, we distinguished the named intervention system or platform from the reported AI technique or automation logic. Three studies [17,27,28] used generative AI or large language model–based approaches directly in intervention delivery or content development. Several studies used scripted, predefined, or automated chatbots and tailoring functions based on structured pathways, branching logic, or automated rules [19,22,25,26,30]. Moore et al [23] described natural language understanding for intent recognition and dialogue routing. The study by de Arriba Muñoz et al [18] used a hybrid personalization engine, Zarkogianni et al [29] used a machine learning–based recommendation system, Oh et al [24] used computer vision and deep learning for gesture recognition, and Kim et al [20] described a broader integrated chatbot-supported digital health ecosystem.

Across studies, the most common role of AI was to support personalization and automated interaction rather than diagnosis or prediction. Tailored feedback, adaptive support, or personalized recommendations were described in 10 studies [17-20,22,23,25-27,29]. Chatbot-based or conversational coaching features were present in 8 studies [17,21-23,25-27,29], activity generation or exercise recommendation in 4 studies [17,24,28,29], and monitoring or self-management support in 6 studies [18,19,22,25,26,29].

Input data were most commonly participant- or caregiver-entered, including text prompts, questionnaires, goal selections, behavior logs, profile information, or interaction history. Six studies [18,24,25,27-29] incorporated wearable, sensor, image, or other objectively tracked data as part of the AI-supported system. Technical reporting of AI methods was often limited. Seven studies [20,21,23,27-30] provided only limited, conceptual, or partial descriptions of AI logic or architecture, whereas 6 [17-19,22,24,25] provided at least moderate detail about system workflow, personalization logic, or model type. One study [21] did not evaluate a live AI system but instead assessed attitudes toward a proposed future chatbot.

Theoretical frameworks were identified in most studies (Table 5), including Social Cognitive Theory [19]; the capability, opportunity, motivation, and behavior (COM-B) model; the Theoretical Domains Framework [23]; the Health Action Process Approach; the Elaboration Likelihood Model; Persuasive Systems Design [25]; Self-Determination Theory [27,29]; the Health Promotion Model; the Technology Acceptance Model [21], and the Multi-Process Action Control (M-PAC) model [28]. Other studies referenced broader approaches such as cognitive-behavioral principles, family-centered care, or user-centered design, whereas Aydemir [17] and Oh et al [24] did not report an explicit theoretical framework.

Outcome assessment approaches varied across studies (Table 5). Six studies [18,19,24,25,27,29] included an objective or performance-based PA-related measure, such as accelerometer-assessed activity, Fitbit-derived metrics, step counts, calorie expenditure, physical fitness testing, or device-recorded movement patterns. Four studies [17,21,23,26] relied primarily on self-report or parent-report PA-related measures, including exercise questionnaires, activity hours, or perceived confidence and motivation for PA. The remaining 4 studies [20,22,28,30] did not directly assess intervention-related PA outcomes.

The studies frequently examined outcomes beyond PA alone. Anthropometric or weight-related outcomes, including BMI, BMI z score, body fat, waist circumference, or obesity-related indicators, were reported in 6 studies [18,19,22,24,26,29]. Diet- or nutrition-related outcomes were reported in 4 studies [18,19,21,29], screen time or sedentary behavior in 3 studies [19,25,27], and sleep-related outcomes in 3 studies [19,21,25]. Psychosocial, motivational, quality-of-life, and usability-related outcomes were also commonly reported, suggesting that many AI-supported family-based interventions were conceptualized as broader healthy lifestyle or obesity-management programs rather than narrowly focused PA interventions.

Among the 8 studies that reported direct PA-, exercise-, movement-, or fitness-related outcomes, 4 [17,23-25] reported significant improvements in PA participation, exercise performance, movement-related behavior, or fitness-related indicators. Two studies [26,27] reported little or no statistically significant improvement in PA despite favorable findings for engagement or other health outcomes. One study [18] collected wearable-based PA data but reported stronger clinical obesity-related effects than PA-specific effects, and one additional study [19] described PA-related findings as part of a broader multibehavior framework without demonstrating a clearly significant PA effect.

All 14 original studies reported at least one feasibility-, acceptability-, usability-, or engagement-related outcome (Table 5). Five studies [20,21,23,28,30] were primarily qualitative, developmental, survey-based, or case-based in design and therefore emphasized user preferences, accessibility needs, perceived usefulness, desired features, implementation barriers, or design priorities rather than behavioral efficacy. These studies generally reported favorable perceptions of AI-supported tools, particularly when interventions were viewed as personalized, interactive, and supportive of family routines or youth autonomy.

Among the intervention-oriented studies, quantitative implementation indicators were commonly reported. Examples included high parent satisfaction and 92% participation in WhatsApp discussions in one parent-mediated ChatGPT study [17]; a System Usability Scale score of 72 in one obesity-management intervention [18]; 86.3% retention at 24 months in the large primary care trial; 85% friendliness, 100% perceived utility, and 69% acceptability in one pediatric obesity app pilot [22]; 61 total chatbot conversations and 73% definite intention to use in one conversational-agent co-design study [23]; average daily app usage of 71.5% in one obesity management trial [26]; and a mean of 8.9 (SD 7.6) behavior entries, 30 reflections, and 33.5 AI feedback interactions in one school-based chatbot study [27]. One quasi-randomized trial reported substantial nonuse attrition, with 18% of participants never using the app and 30% stopping use by week 2 [25].

This scoping review explored the emerging evidence on AI and chatbot-supported interventions targeting PA and obesity-related lifestyle behaviors among children and adolescents, with attention to family involvement. Across the 14 studies included, the evidence indicates that AI-supported interventions are being used in increasingly diverse ways to support PA, obesity management, and related health behaviors in pediatric populations, but the literature remains varied in terms of intervention design, AI capabilities, family involvement, and outcome measurement.

Across the included studies, AI features span a continuum of complexity, from rule-based scripted chatbots and automated tailoring systems (eg, structured coaching or tailored text messaging) to hybrid systems that incorporated limited natural language understanding or personalization. AI was most commonly used to support personalization, user interaction, and ongoing behavioral support, rather than for diagnostic, predictive, or decision-making purposes. Many studies provided only high-level or conceptual descriptions of system functionality, with limited detail on model architecture, training data, validation procedures, or performance across subgroups. This lack of transparency makes it difficult to assess reproducibility, evaluate potential bias, or determine whether AI-driven personalization meaningfully contributes to outcomes beyond standard digital tailoring approaches. These gaps are particularly important given the increasing integration of more complex AI systems, which introduce additional concerns related to accuracy, safety, and algorithmic bias. Future research should adopt AI-specific reporting standards and explicitly document how AI outputs are validated, monitored, and updated over time across diverse populations.

Regarding the delivery platform and intervention components, most interventions embed common behavior change functions, including self-monitoring, goal setting, education, feedback, and rewards, within various platforms. These platforms range from low-burden text messaging and web dashboards to multiapp family–clinician ecosystems, wearable-linked systems, higher-complexity immersive environments, and sensor-rich exercise apps. Engagement strategies also include gamification, relational dialogue, narrative content, and school- or family-embedded routines. Within this structure, AI primarily serves as a delivery and personalization mechanism that shaped how these components were implemented. AI enabled more dynamic forms of support, such as routine coaches, on-demand assistants, and interactive educational or motivational agents [23,25,28]. In this sense, AI modifies the timing, intensity, and responsiveness of intervention delivery rather than fundamentally changing its core behavioral content.

A central contribution to this review is the characterization of family involvement, which was important but inconsistently operationalized. In some studies, parents were critical to implementation, such as delivering activities at home, receiving caregiver education, or interacting with the digital system directly [17-19,29]. In others, youth were the primary users and family influence was acknowledged but not fully integrated into the intervention structure [23,24,27]. This variability suggests that the field has not yet converged on a clear definition of what constitutes a “family-based” AI intervention. Importantly, most studies do not measure family engagement systematically enough to test whether caregiver involvement mediates or moderates intervention effects. More precise conceptualization of whether caregivers are intended as co-users, facilitators, role models, or targets of behavior change may improve both intervention design and interpretation of outcomes.

With respect to cultural tailoring and bilingual delivery, evidence remains limited. One large randomized clinical trial embedded bilingual delivery (English/Spanish) and cultural appropriateness processes in a racially and ethnically diverse population [19]. Although all studies could be situated within identifiable linguistic or contextual settings, specific cultural adaptation was rare, and the evidence base was entirely in middle- and high-income countries. This limits the transferability of current findings to lower-resource contexts and to culturally diverse family populations whose routines, values, digital access, and barriers to health behavior change may differ substantially. The lack of consistent reporting on cultural tailoring, bilingual adaptation, and structural access issues also limits understanding of how well current interventions align with the realities of underserved families. These limitations introduce broader concerns in digital health that technologies may inadvertently reinforce inequities if they are developed primarily in well-resourced settings without sufficient adaptation for different social and cultural contexts [31].

Consistent with the aims of this review, PA emerged as the most prominent target behavior, either as a primary intervention focus or as part of broader multibehavior lifestyle and obesity-management programs. Several studies focused directly on PA promotion, home exercise, or movement-related confidence and performance [17,23,24,27], whereas others addressed PA within broader interventions targeting obesity prevention, rehabilitation, or healthy lifestyle routines [18,19,25,26,29]. This pattern likely reflects both the importance of PA in pediatric health promotion and the broader reality that children’s movement behaviors are closely associated with diet, sleep, sedentary behavior, and family routines. Rather than treating PA as an isolated outcome, many of the included studies positioned it within a more comprehensive family lifestyle framework.

The reviewed studies also indicate that AI-supported interventions were generally more consistent in demonstrating feasibility, acceptability, and engagement than in behavioral efficacy. These findings align with previous youth-focused syntheses emphasizing that chatbots and AI-supported tools are often perceived as appealing and acceptable when they are interactive, personalized, and low burden [12]. However, the current review also found that favorable usability does not necessarily translate into sustained engagement or robust changes in PA participation, exercise performance, calorie expenditure, cardiopulmonary function, or confidence and motivation for PA [17,23-25], and some broader obesity-management or prevention interventions also showed favorable changes in clinical or anthropometric outcomes [18,19,29]. At the same time, findings were not uniform across studies. Tan et al [27] did not detect significant changes in PA, and Stasinaki et al [26] did not show clear superiority of the conversational-agent intervention over the comparison condition for BMI-SD score (BMI-SDS). Moreover, many included studies were feasibility, pilot, or developmental studies rather than adequately powered effectiveness trials. These findings suggest that while AI-supported interventions show promise, the current evidence remains stronger for feasibility and acceptability than for clinically meaningful and sustained improvements in PA and related outcomes.

This review highlights several implications for intervention design, implementation, and future research. In practice, AI-supported family interventions may be most useful when integrated into broader family-based, school-based, or clinical programs rather than positioned as isolated tools. Their strengths appear to lie in delivering tailored education, self-monitoring support, reminders, coaching, and feedback in ways that are flexible, scalable, and responsive to daily family routines. However, the findings also indicate that intervention success likely depends on more than the presence of AI alone. Additional attention to developmental appropriateness, caregiver roles, response quality, burden of use, and contextual relevance will be essential to sustaining engagement and producing meaningful behavior change.

Future research and program development should prioritize family-centered co-design, more rigorous comparative and longitudinal study designs, and greater consistency and transparency in reporting AI features, personalization logic, and behavior change techniques. More culturally and linguistically responsive adaptation is also needed, particularly for diverse and underserved family populations. Additional work should examine implementation outcomes such as reach, adoption, maintenance, scalability, and cost, alongside behavioral and clinical outcomes. Emerging approaches such as generative AI, wearable-integrated personalization, and adaptive feedback systems may offer added value, but their safety, accuracy, transparency, and practical benefit require more careful evaluation in pediatric settings. Continued interdisciplinary collaboration will play an important role in ensuring that AI-supported family interventions are not only innovative but also ethical, accessible, and equitable.

Several limitations should be acknowledged. First, the number of included studies was small, and the evidence base was highly heterogeneous. Further, the included studies varied widely in population, intervention modality, duration, AI sophistication, and measured outcomes, which limits comparability. In addition, many studies were pilots, feasibility studies, or usability-focused investigations rather than adequately powered trials. Further, reporting of demographic, contextual, and technical AI details was often incomplete, making it difficult to assess generalizability or identify active intervention ingredients. The search was limited to English-language publications; therefore, relevant studies published in other languages may not have been captured, which may have introduced selection bias. Further, the search was limited to peer-reviewed publications indexed in the selected databases and did not include a formal gray-literature search or supplementary citation tracking. Therefore, relevant evidence reported in theses, reports, conference materials, policy documents, unpublished sources, or publications not indexed in the selected databases may not have been captured. A formal critical appraisal or methodological quality assessment was not conducted; therefore, the methodological rigor of the included studies was not evaluated.

In conclusion, this scoping review found that AI and chatbot-supported interventions for pediatric PA and obesity-related lifestyle behaviors are a rapidly emerging but still early-stage field. The existing literature suggests that these interventions are generally feasible, acceptable, and promising as scalable tools for supporting healthy behavior change among children and families. However, evidence for effectiveness, particularly for PA outcomes, remains mixed and limited by methodological heterogeneity, short follow-up, and inconsistent outcomes. The review also highlights important gaps in family-based design, cultural tailoring, equity, and evaluation of advanced AI capabilities. Future research should prioritize rigorous trials, clearer conceptualization of family engagement, objective outcome measurement, and culturally responsive development to ensure that AI-supported interventions are not only innovative but also effective, equitable, and meaningful for diverse families.