This study used a mixed methods exploratory sequential design [41]. The initial qualitative phase explored Alpha Mothers’ interactions with AI to understand maternal emotional dynamics and to refine and contextualize key constructs. These findings directly informed the subsequent quantitative phase by guiding item development, refining definitions, and contextualizing scale wording to reflect actual feeding experiences. Finally, the second phase empirically tested the derived structural model and research hypotheses within a larger population sample. The resulting conceptual framework, which maps the 14 proposed hypotheses and structural pathways, is illustrated in Figure 1.

Consistent with the exploratory sequential design, the 2 phases were connected at an explicit point of interface: the 5 literature-informed domains were contextually refined in phase 1, and the resulting subthemes and indicators informed the constructs measured in phase 2. Each of the 5 AI engagement dimensions—algorithmic trust, health information quality, personalization fit perception, information-seeking intensity, and social proof sensitivity—was operationalized into a measurable construct with a corresponding conceptual definition and a dedicated set of survey items.

The quantitative instrument was developed in several stages. Items were developed from the context-specific subthemes and indicators identified in phase 1 and refined with reference to previously validated instruments measuring related constructs. Before data collection, all items underwent face validity review by 2 subject-matter experts to assess clarity and relevance. The instrument was then pilot-tested with 30 respondents meeting the study eligibility criteria to evaluate item readability and preliminary internal consistency (Cronbach α). Based on the pilot results, several items were reworded for clarity and 1 item was removed before finalizing the instrument. These 30 pilot respondents were not included in the main analytic sample (N=442). All measurement items used a 5-point Likert response format.

The target population, referred to using the descriptive label “Alpha Mothers” as defined in the Introduction section, comprised urban mothers actively engaging with generative AI for child-related information. Eligibility was determined through a 2-stage process: a brief prescreening confirmed current parenting of a child aged 6‐59 months and prior AI use, while remaining criteria (duration and frequency of AI use, urban residence, and exclusion of health care professionals) were verified through the main survey and manual confirmation by the first author during recruitment. The complete screening and verification items are presented in Multimedia Appendix 1. The geographic scope covered Greater Jakarta, Bandung, Surabaya, and Medan.

For the qualitative phase, recruitment was conducted using purposive and snowball sampling techniques across diverse urban domiciles. The targeted sample size was established at 10-20 key informants, with interviews continuing until theoretical data saturation was achieved and no new themes emerged. In the quantitative phase, participants were recruited via a nonprobability purposive sampling approach.

Outreach efforts strategically used digital parenting communities, pediatric clinic networks, and school communities, channels deliberately selected for their alignment with the digitally literate profile of the target population with implications of this approach for sample composition discussed in the Limitations section. Sample size was determined a priori using G*Power (version 3.1.9.7; Heinrich Heine University Düsseldorf) [42]. Following a recommendation by Hair et al [43], the minimum sample size for PLS-SEM (partial least squares–structural equation modeling) was estimated based on the maximum number of predictors converging on a single endogenous construct in the structural model. The most complex regression equation in the model corresponded to the cognitive empowerment and parental self-compassion constructs, each receiving 5 direct paths from exogenous variables (AI Algorithmic Trust, AI Health Information Quality, AI Personalization Fit, AI Information-Seeking Intensity, and AI Social Proof Sensitivity). With f²=0.15 (medium effect size), α=.05, and statistical power=0.90, the minimum required sample size was 116 respondents. To ensure proportional geographic representation across the designated urban centers, stratified purposive sampling with regional quotas was applied.

Inclusion criteria stipulated that participants must (1) be mothers residing in the designated urban areas; (2) have at least 1 child aged 6-59 months; (3) actively use generative AI platforms (eg, ChatGPT, Gemini, Grok, etc) for at least 6 months, with a minimum frequency of 3 times per week, to seek childcare and nutritional information; and (4) provide voluntary informed consent to participate in the study. Mothers employed as health care professionals (eg, physicians, nurses, or nutritionists) were excluded from the study. This was a deliberate purposive sampling decision rather than a convenience-based one. Because the study examines how nonexpert mothers perceive, trust, and appraise AI-generated nutrition information, the inclusion of clinically trained mothers would have introduced a systematic confound: such participants evaluate health information through specialized professional knowledge and established clinical standards that differ fundamentally from the lay appraisal processes this study sought to capture [44,45]. Excluding them preserved the conceptual homogeneity of the target population and ensured that the measured constructs reflected lay maternal perception rather than professional clinical judgment. This exclusion was verified manually by the first author (LNA) during recruitment through direct confirmation of occupational status with each potential participant, in addition to the occupation question captured in the main survey. The complete eligibility screening and verification procedure is presented in Multimedia Appendix 1.

Data collection was conducted sequentially across 2 distinct phases. During the qualitative phase, the first author (LNA) conducted all semistructured in-depth interviews face-to-face over a 3-week period in December 2025. All study materials were administered in Bahasa Indonesia. Interview quotations were translated into English using AI-based tools and verified for meaning equivalence by 2 bilingual team members (FA and SA) against the original transcripts. AI tools were used only for translation and language editing; all coding and thematic analyses were conducted manually. As the quantitative instrument was administered in Bahasa Indonesia, the English wording of the measurement items is for reporting only and did not affect measurement validity. A semistructured guide comprising 16 open-ended questions across 5 core domains (AI trust formation, information quality evaluation, personalization expectations, information-seeking behaviors, and social validation) directed each session, with flexible probing questions to elicit deeper reflection. With participant consent, all sessions were audio-recorded and transcribed verbatim, each lasting 30-45 minutes. To enhance credibility, synthesized member checking following the approach by Birt et al [46] was conducted, returning summarized thematic findings to participants for confirmation.

The quantitative phase data were collected using a self-administered online questionnaire. This survey commenced with a brief screening section to verify that respondents met the designated “Alpha Mother” criteria. Several procedural remedies were implemented to mitigate common method bias (CMB), including randomization of item order within construct blocks, use of psychological separation between predictor and criterion measures, and ensuring respondent anonymity to reduce social desirability concerns [47,48]. Finally, rigorous data-cleaning protocols, including IP address tracking to identify and remove duplicate entries, were implemented prior to statistical analysis. An overview of the data collection process can be seen in Figure 2.

Qualitative data from the phase 1 interviews were analyzed using Braun and Clarke’s [49] 6-phase thematic analysis framework [50]. Because the interview guide was informed by 5 predefined conceptual domains from prior literature, the analysis followed a hybrid deductive-inductive approach. The predefined domains provided an initial structure for organizing the data, while open coding was used to identify context-specific meanings, indicators, and subthemes within and across those domains. Initial coding was conducted manually by the third author (SA), generating 29 initial codes that were collaboratively grouped into 13 subthemes and consolidated into 5 overarching themes. The coding framework and thematic structure were subsequently reviewed and refined through consensus discussion with LNA and FA to ensure conceptual coherence and that the themes remained grounded in participants’ accounts. A complete audit trail linking the predefined domains, interview questions, initial codes, subthemes, final themes, construct definitions, and survey items is provided in Multimedia Appendix 2. Data saturation was assessed iteratively rather than assumed. After each interview, the research team evaluated whether new codes, themes, or subthemes were emerging. No substantive new themes or subthemes were identified across the final 3 consecutive interviews (interviews 8‐10), indicating that thematic saturation had been reached by the 10th interview. This sample size is consistent with empirical guidance indicating that thematic saturation is typically achieved within 9-17 interviews in studies involving relatively homogeneous populations and narrowly defined research objectives [51], conditions that characterize this study’s focus on urban Alpha Mothers.

The first author (LNA), who conducted all interviews, is a mother with personal experience using AI for parenting decisions. This shared identity not only facilitated rapport and nuanced understanding of participants’ experiences but also introduced potential for overidentification. To mitigate this risk, initial coding was conducted by the third author (SA), who was not involved in conducting the interviews. The coding framework, subthemes, and final themes were subsequently reviewed and refined through consensus discussions with LNA and FA, helping ensure that interpretations remained grounded in participants’ accounts.

Quantitative data from phase 2 were analyzed using PLS-SEM via SmartPLS 4 software (SmartPLS GmbH), following the evaluation guidelines by Hair et al [43]. The evaluation was conducted in 2 primary stages: the measurement model (outer model) and the structural model (inner model). For the measurement model, indicator reliability was assessed using outer loadings. Construct reliability was established using both Cronbach α and composite reliability (CR), while convergent validity was evaluated through the average variance extracted. Discriminant validity was assessed using the Heterotrait-Monotrait (HTMT) ratio of correlations, with thresholds and interpretation criteria following the guidelines of Hair et al [43,52].

Following a satisfactory measurement model, the structural model was evaluated. The overall quality and predictive relevance of the inner model were assessed by examining the inner variance inflation factor (VIF) for collinearity issues, the coefficient of determination (R²), Q²_predict values, and Cross-Validated Predictive Ability Test (CVPAT) [53,54]. Subsequently, a bootstrapping procedure was applied to test the significance of the 14 proposed research hypotheses, including the moderation effects. Following the a priori structural model, the 5 AI engagement constructs were linked to perceived nutrition-oriented feeding practices only via cognitive empowerment and parental self-compassion; direct AI engagement-to-outcome paths were not estimated. Consequently, the present analysis does not distinguish direct from indirect associations, and results are interpreted solely as indirect associations rather than as evidence of mediation. Indirect effects were tested using bias-corrected bootstrap confidence intervals, and an effect was considered supported when its 95% CI excluded zero. Finally, an Importance-Performance Map Analysis (IPMA) was conducted to generate actionable insights regarding priority areas for digital health interventions. To assess whether the modeled relationships were consistent across child developmental stages, a bootstrap-based multigroup analysis (MGA) was additionally conducted, comparing mothers of children aged 6‐24 months with those of children aged 25‐59 months (Multimedia Appendix 3) [43,55].

All research procedures were conducted in accordance with strict ethical principles for medical and social research. The study protocol was reviewed and approved by the Research Ethics Committee (KEP FEB), Faculty of Economics and Business, Universitas Pelita Harapan, under approval number 047/MARS/EC/XI/2025, issued on November 26, 2025. All recruitment and data collection commenced after this approval was granted. Prior to participation, all individuals were provided with a comprehensive participant information sheet detailing the study’s objectives, guarantees of data anonymity and confidentiality, and their right to withdraw at any time without penalty. Explicit informed consent obtained either in writing for the interviews or electronically for the survey was mandatory before any data collection occurred.

The second phase of this study used a quantitative survey to empirically validate the structural model derived from the qualitative findings. Specifically, this phase evaluated perceived nutrition-oriented feeding practices, operationalized as maternal self-report, not observed feeding behavior or objective child nutrition outcomes. The model then tested how nutrition-oriented feeding practices are associated with 5 AI-related predictors via 2 intervening constructs, cognitive empowerment (CE), and parental self-compassion (PSC), with eHealth self-efficacy (ESE) as a moderator.

This mixed methods study investigated how urban Indonesian Alpha Mothers, a digitally literate, highly educated, and evidence-driven caregiver profile, had an increasing engagement with generative AI as a decision support partner in toddler feeding. Beyond the structural model, the qualitative phase contextually refined 5 literature-informed AI engagement dimensions relevant to maternal AI-mediated toddler feeding—algorithmic trust, health information quality, personalization fit perception, information-seeking intensity, and social proof sensitivity. Although these dimensions align with existing constructs in the literature, their context-specific refinement through mothers’ lived experiences and their integration into a single framework for this specific context represent the study’s principal qualitative contribution. The quantitative outcome examined was maternal perception of nutrition-oriented feeding practices and not observed feeding behavior or child nutrition outcomes. The findings suggest that associations between AI engagement and perceived nutrition-oriented feeding practices may involve 2 distinct indirect associations: 1 cognitive and 1 emotional. Mothers who trusted AI and found its recommendations relevant to their child’s feeding context were better able to evaluate and adapt its suggestions on their own terms. Mothers who perceived AI as nonjudgmental and objective reported lower maternal guilt. In this study, it is captured as parental self-compassion. Both pathways were independently associated with better perceived nutrition-oriented feeding practices, although the cognitive pathway carried greater weight.

Four findings warrant closer attention. AI algorithmic trust was the strongest predictor of parental self-compassion, suggesting that lower emotional burden may be more closely associated with trust in the AI system rather than with its popularity. AI personalization fit perception was the strongest predictor of cognitive empowerment, more so than algorithmic trust, information quality, or information-seeking intensity. Alpha Mothers do not appear to value accurate AI content alone. The findings suggest that they also value advice tailored to their child, daily routines, and cultural feeding practices. Beyond personalization, AI social proof sensitivity predicted cognitive empowerment but had no bearing on parental self-compassion. Community support, whether in the form of ratings, peer recommendations, or popularity signals, may shape a mother’s opinion about AI. Yet, it will not affect her inner convictions as a parent. Self-compassion in this context comes from personal experience and not from external reassurance.

Finally, eHealth self-efficacy strengthened the pathway from cognitive empowerment to nutrition-oriented feeding practices. However, it played no moderating role in the self-compassion pathway. Digital confidence and emotional self-regulation appear to be separate capacities that call for separate intervention approaches. The findings suggest that AI may serve as a supportive tool that complements, rather than replaces, maternal judgment and professional nutrition counseling.

Algorithmic trust emerged as the strongest predictor of parental self-compassion in this study. In the digital health literature, trust in AI has largely been treated as a functional matter, assessed through criteria of technical performance, accuracy, and system credibility [70]. Less attention has been paid, however, to whether trust in AI also carries an emotional function. Recent empirical work suggests that patients’ experiences with AI health systems extend well beyond rational cost-benefit evaluation, with emotional responses such as anxiety, cognitive overload, and resistance playing a significant role in how users actually engage with these tools [71,72]. The concept of algorithmic authority provides a framework for interpreting this emotional dimension.

The present findings add a maternal caregiving dimension to this picture. For these urban Indonesian mothers, trusting an AI system was not simply a matter of judging whether its nutritional advice was clinically sound. It also determined whether that trust could absorb some of the guilt and cognitive load that accompanies daily toddler-feeding decisions. AI-driven pediatric tools may benefit from being designed with this emotional function in mind, not only to improve informational output but also to offer meaningful psychological relief to the caregivers who rely on them.

Trust in health AI has been shown to rest less on technical performance than on the degree to which a system reflects the user’s own values and ways of knowing [73]. For the Alpha Mothers in this study, factual accuracy was merely a baseline. Qualitative findings suggest that trust was closely tied to perceptions of commercial neutrality and value alignment.

The result indicated that AI personalization fit perception was the strongest predictor of cognitive empowerment compared with the other 4 variables. The dominance of AI personalization as a predictor of cognitive empowerment in these findings aligns with existing literature on parental digital behavior. This pattern is consistent with findings that parents of preschool children prioritize family-adapted technology over standardized content [57,74,75]. The findings suggest that mothers in this study preferred highly contextualized nutritional guidance. This finding is consistent with the interpretation that for this demographic, personalization is the strongest predictor associated with cognitive empowerment.

However, this study structural model also reveals that this technology-driven empowerment has distinct boundaries, particularly concerning the emotional complexities of parenting. Specifically, the rejection of the hypotheses regarding AI Social Proof Sensitivity (H10) and the moderating role of eHealth Self-Efficacy (H14) indicates that while AI can cognitively empower mothers, technical and social features were not associated with higher levels of Parental Self-Compassion in this study. Knowing that other mothers use AI (social proof) or possessing high technical skills to navigate health information (eHealth self-efficacy) does not organically translate into reduced maternal guilt or enhanced emotional grace when facing feeding challenges.

The nonsignificant relationship between AI social proof sensitivity and parental self-compassion may reflect a dynamic specific to digital parenting environments. When mothers evaluate AI guidance through the lens of social proof, they are simultaneously exposed to a wide and often contradictory range of parenting standards. Far from offering reassurance, this collective validation tends to fuel upward social comparison and maternal competition. Recent empirical research indicates that exposure to heterogeneous parenting norms in digital spaces intensifies the internalization of an unrealistic “good mother” ideology [18,76]. True self-compassion requires self-kindness and the acceptance that all parents struggle. However, this mindset is incredibly hard to maintain in digital spaces that constantly push idealized standards, leaving mothers feeling like they fall short. High sensitivity to social proof, in this reading, does not appear to soften self-judgment and may instead be associated with greater self-criticism.

eHealth self-efficacy moderated the cognitive empowerment pathway but not the pathway from parental self-compassion to nutrition-oriented feeding practices. Digital self-efficacy has been shown to associate positively with health self-management behaviors through chain mediation in general population samples [77]. The divergence in these findings reflects the nature of self-compassion itself. Cognitive empowerment involves active information processing and responds to greater digital confidence. Self-compassion, by contrast, is an affective response rooted in self-acceptance. A mother’s ability to navigate health platforms does not make her more forgiving of herself when feeding goes wrong. These 2 observed patterns may reflect differences in the underlying psychological constructs and require different intervention approaches. Ultimately, these different paths lead to the question: what can a model focused on AI engagement realistically explain?

Taken together, these findings can be interpreted through the algorithmic authority lens introduced earlier [32]. This interpretation is offered as a theoretical reading of the cross-sectional associations observed in this study rather than as a directly tested causal account. The strongest pathway, from AI algorithmic trust to parental self-compassion, may be consistent with early-stage algorithmic authority formation, in which higher trust is associated with reliance on AI not only for information but also for emotional reassurance. Within this theoretical framing, such reliance could mark the point at which AI begins to substitute for, rather than support, maternal judgment, although the present data cannot establish whether this substitution actually occurs. The nonsignificant association for H10 refines this picture: in maternal caregiving, awareness that other mothers use AI was not associated with parental self-compassion, a pattern interpreted here as peer signals inviting social comparison rather than reassurance [18,76]. The moderation finding adds further insight: eHealth self-efficacy strengthened only the cognitive pathway, suggesting that digital literacy helps mothers think critically about AI advice but does not protect them from becoming emotionally dependent on it.

Together, these patterns may suggest, at a theoretical level, that conditions for algorithmic authority are most pronounced when high trust in AI coexists with low emotional resilience, the pattern the AI parenting nutrition alignment approach is intended to address. This raises a conceptual concern about potential algorithmic dependency: if high trust in AI and emotional relief through its use were to develop together, mothers might progressively reduce consultations with health care professionals, perceiving clinical guidance as less necessary [32,42]. Consequently, AI-generated inaccuracies may remain undetected, potentially delaying appropriate referral when a child presents with genuine nutritional concerns. In the toddler-feeding domain specifically, these risks take concrete forms: AI-generated advice may recommend age-inappropriate foods, overlook individual allergies, or suggest incorrect nutrient dosages [28], concerns corroborated by qualitative participants who reported receiving recommendations containing potentially harmful ingredients for their child’s condition. Because detecting such errors depended on the mother’s own vigilance, the risk was unevenly distributed—least for mothers with stronger nutritional knowledge and critical appraisal skills, and greatest for those most reliant on AI. This remains a hypothesized risk that the cross-sectional design cannot confirm and that warrants longitudinal investigation.

The effect size pattern warrants careful interpretation. AI Algorithmic Trust showed the strongest contribution to Parental Self-Compassion, while the other supported pathways showed smaller contributions. Some pathways, although statistically significant, contributed only modestly to the outcome. Two factors help explain this pattern. First, nutrition-oriented feeding practices are shaped by many caregiving factors, such as economic, cultural, situational, and child-specific factors that fall outside any AI engagement model. Second, moderation effects are typically smaller than main effects in behavioral models [69]. The findings should therefore be read as statistically robust but practically modest contributions of AI engagement to perceived nutrition-oriented feeding practices, rather than as evidence that AI engagement is the dominant driver of nutrition-oriented feeding practices. The same reality is reflected in the modest overall explanatory power for nutrition-oriented feeding practices. This suggests that AI engagement, together with the 2 intervening constructs, captures only part of what shapes feeding practices, with substantial remaining variance attributable to factors outside the model, such as socioeconomic conditions, family food culture, child temperament, maternal mental health, and situational demands. Furthermore, perceived measures of caregiving behavior—as opposed to directly observed practices—are known to be shaped by momentary contextual influences and subjective interpretation, which further constrains the variance explainable by any structural predictor set. Structural models targeting behavioral outcomes rarely explain more than a small proportion of variance, even when theory-driven and methodologically robust. The identification of algorithmic trust and parental self-compassion as significant psychosocial pathways therefore represents a meaningful contribution within this explanatory context. The perceived nutrition-oriented feeding practices construct in this model captures the nutritional and structural dimensions of infant and young child feeding guidance, namely, schedule consistency, nutritional appraisal, and limitation of low-nutrient foods, together with attentiveness to satiety signals. It does not operationalize the reciprocal responsiveness components central to the WHO or UNICEF definition of responsive feeding, which additionally require assessment of coercive feeding pressure and child autonomy support [5,6]. While the findings suggest that AI engagement may reinforce these nutrition-oriented practices via its association with cognitive empowerment, current AI platforms remain unbenchmarked against WHO or UNICEF responsive feeding frameworks or the structured competencies delivered through the UNICEF C-IYCF Counselling Package [25]. AI-assisted feeding support should therefore complement, not replace, established C-IYCF counseling models.

The following implications should be interpreted as preliminary directions derived from cross-sectional associations rather than as established prescriptions. The policy, clinical practice, and system development levels outlined in this section each warrant validation through longitudinal investigation, intervention trials, and contextual adaptation prior to large-scale implementation. At the policy level, these findings highlight the growing importance of governance frameworks for AI-assisted parenting and nutrition support. As generative AI becomes increasingly integrated into health care, international guidance has emphasized transparency, accountability, safety, and human oversight in AI-enabled health systems [78,79]. Indonesia has similarly established foundational principles for responsible AI development through Ministerial Circular No. 9/2023 [80]. However, no dedicated framework currently addresses AI-generated child nutrition advice or clearly defines accountability when such advice is associated with harm. Future policy efforts may therefore benefit from establishing standards for transparency, evidence attribution, and alignment with WHO and UNICEF responsive feeding guidance, while ensuring that AI-assisted nutrition support complements rather than replaces professional nutrition counseling.

At the clinical practice level, the findings suggest that pediatricians, nutritionists, and public health practitioners may increasingly serve as facilitators who help mothers critically evaluate AI-generated nutrition advice and integrate it with professional guidance and their child’s individual circumstances. Within the Indonesian context, future research may explore whether AI literacy and nutrition counseling can be incorporated into existing maternal and child health services, including Posyandu and routine well-child care. Such approaches should position AI as a complement to, and not a substitute for, professional nutrition counseling while strengthening maternal cognitive empowerment and critical appraisal skills.

At the system development level, the findings provide preliminary insights that may inform future AI-assisted feeding support tools. The AI Parenting Nutrition Alignment Approach highlights the potential value of contextualized guidance, transparent explanations, and support for critical evaluation of AI-generated recommendations. Future systems may also benefit from benchmarking against WHO and UNICEF responsive feeding guidance to improve consistency with evidence-based nutrition recommendations. Development and implementation of such tools would likely require collaboration among pediatricians, nutritionists, public health practitioners, AI developers, and policymakers to ensure appropriate clinical oversight, safety, and governance. However, these implications remain conceptual and require evaluation through longitudinal and intervention-based research before broader implementation.

Several limitations should be acknowledged. First, the sample was drawn from a specific demographic profile—urban, digitally active mothers who actively use generative AI—referred to in this study using the descriptive label “Alpha Mother.” This label functions as a recruitment descriptor rather than a validated theoretical construct, and findings should not be generalized beyond this defined demographic profile. Whether the patterns observed here apply to mothers from different socioeconomic, educational, or digital engagement profiles remains an open empirical question. Second, recruitment through digital parenting communities likely introduced self-selection bias toward mothers already favorably disposed toward generative AI. Findings should therefore be interpreted as reflecting the experiences of digitally active, AI-engaged urban mothers rather than urban Indonesian mothers more broadly.

Third, all data relied on self-report. Mothers may have overstated their trust in AI or their sense of self-compassion to align with perceived parenting ideals, particularly given the social pressures that surround toddler feeding in urban Indonesian contexts. Furthermore, perceived nutrition-oriented feeding practices capture maternal self-assessment, not observed feeding behavior or objective child nutrition outcomes. The instrument assesses nutrition-focused and structural aspects of feeding; coercive feeding pressure, caregiver responsiveness to the full range of hunger and satiety cues, and child autonomy support were not measured. Findings therefore speak to the nutritional and structural dimensions of feeding rather than to caregiver-child interactional quality and should not be read as evidence regarding responsive feeding as defined by WHO and UNICEF. Future work should triangulate this construct against behavioral instruments such as the Comprehensive Feeding Practices Questionnaire and against anthropometric indicators.

Fourth, the cross-sectional design and the modest variance explained limit interpretation. The associations identified through PLS-SEM are statistically robust but cannot establish whether AI engagement changes maternal behavior over time, and the low explained variance indicates that AI engagement is only one of several factors shaping feeding practices. The design also permits alternative explanations the present data cannot adjudicate: the trust-outcome relationship may be reversed or reciprocal, as mothers who already possess greater feeding confidence, knowledge, or emotional resilience may be more predisposed to trust AI and report favorable feeding experiences. An unmeasured third variable, such as general health literacy or socioeconomic advantage, could likewise elevate both, and such confounding is difficult to exclude given the self-selected sample. The reported associations should therefore be read as compatible with, but not confirmatory of, the mechanistic interpretation advanced in the Principal Findings section. Improvements in perceived nutrition-oriented feeding practices should therefore not be interpreted as evidence of improved child nutritional or developmental outcomes. In addition, although the sample spanned a broad age range (6‐59 months), an MGA confirmed that the structural estimates did not differ significantly between younger and older children (Multimedia Appendix 3); the younger stratum nonetheless follows the survey’s 6‐ to 24-month category, 1 month wider than the WHO 6‐ to 23-month cutoff, and its smaller size warrants caution in interpreting subgroup estimates.

Fifth, social desirability and CMB cannot be fully excluded, as all constructs were self-reported through a single instrument at one time point. Although procedural remedies and a full collinearity assessment (all inner VIF <3.3) indicate that CMB is unlikely to be severe, these safeguards reduce rather than eliminate the risk. In addition, the 5 AI engagement dimensions were newly developed here and have not been externally validated, and the AI parenting nutrition alignment approach that organizes them adapts an existing framework [44] not yet validated in this domain. Findings involving them should be regarded as preliminary.

Finally, this study did not independently audit the clinical accuracy of AI-generated nutritional advice received by participants; the constructs measured reflect mothers’ perceptions of AI quality rather than objective evaluation of AI output. Future research should systematically benchmark AI-generated feeding advice against WHO or UNICEF responsive feeding recommendations and the C-IYCF counseling competency framework, a critical step toward developing clinically validated AI-assisted feeding platforms.

This mixed methods study suggests that generative AI is not independently associated with perceived nutrition-oriented feeding practices. Associations between AI engagement and perceived nutrition-oriented feeding practices were strongest when mothers trusted the system and perceived its recommendations as fitting their child and household realities. Two distinct indirect associations may help describe these associations. Algorithmic trust was associated with parental self-compassion, which was linked to lower emotional burden during feeding challenges, whereas personalization fit was associated with cognitive empowerment, which in turn was associated with mothers’ self-reported capacity to critically evaluate and adapt AI advice.

These indirect associations may reflect potentially important constructs by which AI engagement is associated with perceived nutrition-oriented feeding practices. Within the data, the perceived value of AI in toddler feeding appeared to relate less to algorithmic sophistication than to perceived trust, contextual fit, and support for maternal judgment. Beyond these associations, and extending past what the present cross-sectional data can establish, the findings may hold tentative implications for clinicians, policymakers, and developers. They may inform future research on, and the eventual design and evaluation of, AI-assisted feeding support, provided such tools are positioned to complement rather than supplant professional nutrition counseling. These implications remain hypotheses for future longitudinal and intervention research rather than evidence-based recommendations. In addition to these quantitative pathways, this study contributes a qualitative typology of AI engagement constructs specific to the toddler-feeding context, providing a conceptual foundation for future research at the intersection of AI, maternal cognition, and child nutrition.