This systematic literature search was conducted to identify research on AI-driven chatbot interventions for the treatment of depression. The objective was to evaluate the influence of chatbot-based digital tools on clinical outcomes, including symptom improvement, treatment effectiveness, and user adherence. Studies were included only if chatbots were used as therapeutic tools, while those focusing exclusively on diagnosis, screening, or prediction were excluded.

The search was conducted across 6 major academic databases, including Scopus, Web of Science, PubMed, IEEE Xplore, APA PsycINFO (Ovid), and Embase (Ovid). The strategy incorporated 2 primary concept domains, namely depression and chatbot or conversational systems. Only peer-reviewed journal articles and conference proceedings published in English, with coverage up to May 30, 2025, were included, as detailed in Textbox 1.

The search strategy was developed in accordance with guidance from the Cochrane Handbook. Consistent conceptual blocks were applied across all databases, with syntax adapted for each platform. Full database-specific search strategies are provided in Multimedia Appendix 1. Reporting of the search strategy and process adhered to PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) guidelines to ensure transparency and reproducibility [33]. The PRISMA-S checklist is presented in Checklist 1.

All records identified from the 6 databases, including Scopus, Web of Science, PubMed, IEEE Xplore, APA PsycINFO (Ovid), and Embase (Ovid), were imported into EndNote (version 21; Clarivate) for management and initial filtering based on 11 predefined bibliographic fields, including author, year, abstract, and keywords. A total of 3372 records were retrieved. After removing 4 erroneous records, 3368 records were retained for deduplication, resulting in 2097 unique articles. Titles and abstracts were independently screened by 2 reviewers (TH and WL) against the eligibility criteria. Any discrepancies were resolved through discussion with a third reviewer. In total, 87 articles were selected for full-text assessment, of which 8 met the inclusion criteria. An additional 3 eligible studies were identified through snowballing. Overall, 11 studies were included in the final analysis and subjected to meta-analysis. The screening and selection process is summarized in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; Checklist 2) flow diagram.

In addition to extracting general study characteristics, 6 interaction features of the chatbots were assessed, including interaction frequency, emotional responsiveness, self-disclosure encouragement, dialogue depth, feedback strategy, and user agency level. Each feature was rated on a 5-point scale (1=very low-5=very high) by 2 independent reviewers (TH and WL). A third reviewer adjudicated when ratings differed by ≥1 point, and the final score for each feature was calculated as the mean of the available ratings. Interrater reliability across all ratings was good (overall intraclass correlation coefficient [ICC]=0.71); detailed results are presented in the Multimedia Appendix 2.

For clarity, the 6 features were defined as follows: interaction frequency (the intensity and regularity of user-chatbot exchanges), emotional responsiveness (the chatbot’s ability to adaptively provide empathetic responses), self-disclosure encouragement (prompts guiding users to share personal experiences or emotions), dialogue depth (the richness and reflectiveness of conversations), feedback strategy (the presence of timely and tailored prompts or evaluative responses), and user agency level (the degree of control and choice available to the user).

User adherence was operationalized as intervention completion, defined as the proportion of participants who completed the intervention protocol relative to the number initially enrolled. This completion-based measure was consistently reported across the included studies and was therefore adopted to enable quantitative synthesis of adherence outcomes.

The methodological quality of the included studies was assessed independently by 2 reviewers (TH and WL) using the Cochrane Risk of Bias 2 (RoB 2) tool for RCTs. The tool evaluates potential biases across five domains: (1) bias arising from the randomization process, (2) bias due to deviations from intended interventions, (3) bias due to missing outcome data, (4) bias in the measurement of the outcome, and (5) bias in the selection of the reported result. Each domain was rated as “low risk,” “some concerns,” or “high risk” according to the signaling questions provided by RoB 2 guidelines.

Any discrepancies in assessments between the 2 reviewers (TH and WL) were resolved through discussion; if disagreement persisted, a third reviewer was consulted. The final risk-of-bias assessments were summarized in tabular and graphical form.

Meta-analyses were conducted using random-effects models. Standardized mean differences (SMDs) were calculated for continuous outcomes, and odds ratios (ORs) were calculated for dichotomous outcomes. Heterogeneity was quantified using the I² and statistics [34]. Prediction intervals (PIs) were calculated to estimate the range of true effects in future comparable settings [35]. To provide more robust CI estimates, particularly given the relatively small number of included studies, the Hartung-Knapp-Sidik-Jonkman (HKSJ) adjustment was applied [36]. The DerSimonian-Laird estimator was also used for comparison [37]. To assess potential small-study effects, funnel plots were visually inspected. Egger regression test [38], the Begg-Mazumdar rank correlation test [39], and the Duval and Tweedie trim-and-fill procedure were applied where appropriate (≥10). All analyses were conducted in R (version 4.5.1; R Foundation for Statistical Computing).

The study selection process is shown in Figure 1. After full-text screening, a total of 11 articles were included in the analysis. In total, 2220 participants (1091 in the intervention and 1129 in the control groups) were included in our analysis. A summary of the characteristics of the studies is presented in Table 1.

A total of 11 [40-50] RCTs involving 2220 participants (1091 in the experimental group and 1129 in the control group) were included in the meta-analysis. The pooled results indicated a small-to-moderate effect of AI-driven chatbots on depressive symptoms, compared with control conditions (SMD −0.46, 95% CI −1.02 to 0.10; P=.01; as shown in Figure 2). Negative SMD values indicate greater symptom reduction in the chatbot intervention groups.

To further quantify the real-world implications of heterogeneity, 95% PIs were calculated. The PI ranged from −1.50 to 0.58, indicating that the true effect in a future comparable setting could vary substantially and may include no effect. Heterogeneity among studies was substantial (I²=87%), suggesting considerable variability in intervention effects across trials.

Sensitivity analyses excluding Fitzpatrick et al [41] reduced heterogeneity from 87% to 60%, while the direction of the pooled effect remained unchanged. Detailed results are provided in Multimedia Appendix 3. To further explore potential sources of heterogeneity, subgroup analyses were conducted to examine differences across AI chatbot types.

Across all 11 [40-50] included studies, the pooled analysis showed no significant difference in adherence between intervention and control groups, as shown in Figure 3 (OR 1.22, 95% CI 0.57‐2.62; P=.57). To further interpret heterogeneity in real-world settings, the review calculated the 95% PI for the overall adherence outcome. The 95% PI ranged from 0.24 to 6.18, indicating substantial between-study variability. This suggests that the true adherence effect in a comparable future setting could range from lower to substantially higher engagement than in control conditions. Heterogeneity was considerable (I²=74.8%).

All studies meeting the inclusion criteria reported depressive symptom outcomes at the end of the intervention, assessed using the PHQ-9. Overall, 7 studies [40,43-46,48,49] were judged to have a low risk of bias, while 4 studies [41,42,47,50] were assessed as having some concerns. The most common problem was missing outcome data (Domain 3), with insufficient reporting of outcomes in some studies. In contrast, all studies were judged to be of low risk across domains related to randomization, deviations from intended interventions, outcome measurement, and selective reporting. No study was considered to be at high risk of bias in any domain. Detailed domain-level assessments are presented in Figure 4 and Multimedia Appendix 4.

Certainty of evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach. As all included studies were RCTs, the evidence started at high certainty. For clinical effectiveness, certainty was downgraded because of very serious inconsistency, serious imprecision, and suspected publication bias, resulting in very low-certainty evidence. For user adherence, certainty was downgraded because of serious inconsistency and serious imprecision, resulting in low-certainty evidence. A summary of the GRADE assessments for the main outcomes is presented in Table 2, and a more detailed GRADE evidence profile is provided in Multimedia Appendix 5.

Before presenting the subgroup analyses, this review clarifies the classification of AI-driven chatbots. Although LLMs are technically a subset of NLP, in this review we distinguish LLM-based chatbots as systems built on large pretrained generative models (eg, GPT, Gemini, and Claude) that directly generate responses in an open-ended way. In contrast, NLP-/ML-based chatbots are systems that use more traditional NLP or ML methods, such as intent classifiers, decision trees, or response selection within constrained conversational flows. Rule-based chatbots are systems that rely on fixed scripts or expert-crafted rules. This operational categorization aligns with how the primary studies describe their systems and is consistent with recent surveys that distinguish LLM paradigms from traditional NLP approaches [51,52].

To further explore heterogeneity in adherence, subgroup analyses were conducted by AI type (Figure 5; Figure 6). The pooled results indicated no significant overall difference in adherence between experimental and control groups (OR 1.22, 95% CI 0.57‐2.62; P=.57). To further interpret heterogeneity in real-world settings, we calculated the 95% PI for the overall adherence outcome. The 95% PI ranged from 0.24 to 6.18, indicating substantial between-study variability. This suggests that the true adherence effect in a comparable future setting could range from lower to substantially higher engagement than in control conditions.

At the subgroup level, no statistically significant effects were observed. Rule-based chatbots showed a nonsignificant pooled effect (OR 0.95, 95% CI 0.36‐2.51), while NLP/ML-based chatbots also did not reach statistical significance (OR 1.87, 95% CI 0.22‐16.28). The test for subgroup differences was not statistically significant (P=.47). The LLM-based subgroup was excluded from the adherence subgroup figure because one included study reported complete adherence in the intervention group, resulting in an extreme and clinically uninterpretable pooled OR estimate.

These findings suggest that chatbot type alone does not consistently predict user adherence across studies. Given the substantial heterogeneity, the results should be interpreted with caution. The findings further indicate that factors beyond chatbot type, such as specific interaction features, may play a more important role in sustaining user engagement. In the following section, we therefore turn to an analysis of 6 key interaction features, including interaction frequency, emotional responsiveness, self-disclosure encouragement, dialogue depth, feedback strategy, and user agency to clarify their contribution to adherence.

This section examines 6 interaction features of AI-driven chatbots: interaction frequency, emotional responsiveness, self-disclosure encouragement, dialogue depth, feedback strategy, and user agency level. Studies were classified into high (≥3.75) and low (≤3.5) groups based on expert ratings. The detailed scoring table is provided in Multimedia Appendix 2. These ratings were derived from structured coding of published study descriptions and were used as operational proxies of interaction characteristics, rather than as direct measurements of chatbot behavior. The analyses in this section were therefore conducted in an exploratory, hypothesis-generating manner. Unlike the chatbot-type subgroup analysis presented earlier, the analyses in this section represent stratified meta-analyses based on study-level coding of interaction features. For each feature, studies were categorized as high or low according to predefined criteria, and pooled estimates were calculated separately within these strata. These analyses examine whether variation in interaction design characteristics is associated with differences in clinical effectiveness and user adherence, rather than comparing participant-level subgroups within individual trials. Section “Interaction Features and Clinical Effectiveness” reports the relationship between interaction features and clinical effectiveness, while section “Interaction Features and User Adherence” analyzes their association with user adherence.

As shown in Figure 7 and summarized in Table 3, differences were observed between high and low-scoring subgroups across several interaction features. In general, high-scoring subgroups tended to show more consistent patterns of treatment effects, whereas low-scoring subgroups showed more heterogeneous and less stable estimates. These findings should be interpreted as exploratory contrasts rather than confirmatory evidence. Full model outputs and sensitivity analyses are provided in Multimedia Appendix 6 [40-50].

For dialogue depth, the high group was significantly associated with better outcomes (SMD=−0.35, 95% CI −0.61 to −0.10; P=.007; I²=0%), while the low group failed to reach significance and showed substantial heterogeneity (SMD=−0.54, 95% CI −1.20 to 0.11; P=.10; I²=97.0%). This pattern was largely influenced by the trial of Fitzpatrick et al [41], which used the Woebot platform and reported an exceptionally large effect size (SMD=−3.30, 95% CI −4.12 to −2.48). The authors noted frequent misunderstandings and repetitive dialogues as limitations, which may have shaped participants’ engagement and contributed substantially to the heterogeneity in this subgroup.

By contrast, He et al [42] was consistently classified into high groups across all 6 interaction features, with particularly high scores in emotional responsiveness (4.75). He et al [42] found that enhanced emotional awareness significantly predicted superior therapeutic outcomes (F_{2, 145}=3.636; P=.03), a finding corroborated by the meta-analysis (SMD=−0.55, 95% CI −1.01 to −0.08; P=.02; I²=80%). This provides indicative evidence that empathetic and adaptive chatbot responses may enhance clinical effectiveness. In addition, Liu et al [45] was classified into the low group for interaction frequency (score=3). The intervention demonstrated a significant negative effect (SMD −0.94, 95% CI −1.46 to −0.42), suggesting that insufficient interaction intensity may limit sustained therapeutic gains.

Taken together, these exploratory findings suggest that dialogue depth, emotional responsiveness, and interaction frequency may be associated with variation in clinical effectiveness across studies. However, given the limited number of included trials and the substantial heterogeneity observed in several subgroups, these patterns should be interpreted with caution and regarded as hypothesis-generating rather than definitive evidence.

Subgroup analyses were conducted to examine the association between 6 interaction features and user adherence, using the completion-based adherence definition described in the Methods (as shown in Figure 8 and Table 4). Overall, differences were observed between high- and low-scoring subgroups across several interaction features. In contrast to the clinical effectiveness outcomes reported in Figure 7, effect estimates for user adherence showed greater variability in magnitude and precision across studies.

For emotional responsiveness, the high-scoring subgroup showed a statistically significant association with adherence (OR 3.03, 95% CI 1.45‐6.36; P=.003; I²=14%), whereas the low-scoring subgroup did not reach statistical significance (OR 0.87, 95% CI 0.53‐1.44; P=.59; I²=70%). He et al [42], who were classified in the high group (score of 4.75), reported a statistically significant association between emotional awareness and adherence. He et al [42] illustrates how emotionally responsive chatbot interactions may be associated with adherence outcomes in specific contexts, rather than providing confirmatory evidence of a causal relationship.

In the domain of feedback strategy, high-scoring studies yielded robust and consistent associations (OR 4.62, 95% CI 1.96‐10.91; P<.001; I²=0%). Fitzpatrick et al [41] and He et al [42] both applied structured feedback mechanisms, and their findings largely drove the statistical significance of this subgroup. While these findings highlight a consistent pattern within the available data, they should be interpreted cautiously given the small number of contributing studies.

With regard to interaction frequency, the high group demonstrated a significant advantage (OR 4.18, 95% CI 1.10‐15.87; P=.04; I²=78%), whereas the low group did not (OR 0.75, 95% CI 0.49‐1.13; P=.17; I²=53%). Chen et al [40], and Yasukawa et al [50], which were included in the high-frequency subgroup, both reported patterns consistent with sustained adherence under more frequent interactions. Rather than implying a causal relationship, this contrast illustrates variability in adherence outcomes across different interaction intensity profiles.

Taken together, these exploratory results suggest that emotional responsiveness, feedback strategy, and interaction frequency may be associated with variation in adherence outcomes across studies. In contrast, self-disclosure encouragement, dialogue depth, and user agency level showed weaker or nonsignificant associations. Given the heterogeneity of adherence estimates and the limited number of included trials, these patterns should be interpreted as hypothesis-generating and contingent on contextual and design-specific factors.

Funnel plots were generated to assess potential small-study effects for the primary outcomes (Multimedia Appendix 7). For clinical effectiveness, visual inspection suggested possible funnel plot asymmetry. Egger regression test (P=.05) and the Begg-Mazumdar rank correlation test (P=.04) both indicated statistically significant asymmetry. The trim-and-fill procedure imputed 3 potentially missing studies, and the adjusted pooled estimate was attenuated (SMD=−0.13, 95% CI −0.47 to 0.21) and no longer statistically significant. However, substantial heterogeneity was present (I²=87%), and funnel plot asymmetry in this context may reflect genuine between-study variability rather than publication bias alone. For user adherence, statistical testing did not indicate significant funnel plot asymmetry (Begg test P=.07). Although the trim-and-fill method imputed 3 studies, the adjusted pooled estimate remained nonsignificant (OR 0.84, 95% CI 0.49‐1.45), suggesting that small-study effects did not materially alter the overall conclusion for adherence outcomes.

This systematic review evaluated the clinical effectiveness of AI-driven chatbots for depression and examined how interaction features relate to treatment outcomes and user adherence. AI-driven chatbots tended to reduce depressive symptoms, but after a more conservative analysis, this effect lost statistical significance, and studies remained heterogeneous.

No stable or statistically significant differences in user adherence were observed across chatbot types. This suggests that chatbot type alone may not explain variation in engagement patterns across studies. In addition, user adherence and clinical effectiveness did not show a stable one-to-one relationship.

By contrast, the exploratory analyses of interaction features revealed more informative patterns. Emotional responsiveness, feedback strategy, and interaction frequency showed more consistent associations with user adherence, whereas their relationships with clinical effectiveness were more mixed and heterogeneous. Dialogue depth, self-disclosure encouragement, and user agency showed weaker or more context-dependent associations. Overall, these findings suggest that interaction design may offer more explanatory value than chatbot type alone in understanding variation across studies.

The findings of this review are broadly consistent with previous research on DMHIs. Previous studies suggest that technology-assisted interventions may reduce depressive symptoms to some extent, but their effect sizes are often small to moderate and vary across studies [53-55]. A similar pattern was observed in the present review. Although the pooled effect on depressive symptoms remained in a favorable direction, it was no longer statistically significant after applying the more conservative HKSJ method, and substantial heterogeneity remained. The wide PI further suggests that the true effect in a comparable future setting could vary considerably, potentially including no meaningful benefit. Taken together, these findings suggest that the overall clinical effectiveness of AI-driven chatbot interventions remains uncertain and may be influenced by differences in intervention design, implementation context, and study populations [56,57]. This cautious interpretation is also consistent with the GRADE assessment, which rated the certainty of evidence for clinical effectiveness as very low.

Another important finding of this review is that user adherence and clinical effectiveness did not show a stable or directly corresponding relationship. In this review, some studies reported relatively high levels of sustained engagement at the descriptive level [46,48,50], but this pattern did not consistently correspond to greater overall symptom improvement. This finding is consistent with previous research showing that user engagement is an important condition for the success of digital interventions, but it does not reliably predict clinical benefit [58,59]. In the present review, the overall adherence analysis also showed no significant difference between intervention and control groups. Heterogeneity remained high, and the PI was wide. This suggests that adherence outcomes may also vary substantially across studies and implementation settings. In this sense, adherence may be understood as an important condition for intervention success but not a sufficient one [55,59]. At the same time, the findings of this review suggest that chatbot type alone does not provide a stable explanation for variation in effectiveness or adherence across studies.

Building on this distinction, the present review further examined how different interaction features may relate to these divergent patterns. The exploratory analyses suggest that different interaction design elements may relate differently to user engagement and symptom change. In particular, the associations between interaction features and user adherence appeared to be clearer, whereas their relationships with clinical effectiveness were more mixed and heterogeneous. Taken together, these findings suggest that variation across studies may be better understood by focusing on interaction design features rather than chatbot type alone. In other words, differences across studies may be more closely related to how interactions are designed and implemented than to the underlying technical category itself [55,56,59]. At the same time, given the low certainty of evidence for the main outcomes, these interpretations should remain cautious. A more fine-grained understanding of interaction design may help explain variability across studies and inform future system development.

The exploratory analyses suggested that interaction features may provide a more informative lens than chatbot type alone for understanding variation across studies. To clarify these findings, the potential design implications are discussed separately for clinical effectiveness and user adherence. Given the heterogeneity of the evidence and the low certainty of the main outcomes, these implications should be understood as cautious and exploratory rather than prescriptive design recommendations.

The exploratory analyses suggested that the relationships between interaction features and clinical effectiveness were mixed and heterogeneous. Across the 6 interaction features, no single feature showed a uniformly stable association with symptom improvement. Instead, different features appeared to relate to treatment outcomes in different ways, and their potential value seemed to depend on intervention context, therapeutic structure, and user characteristics. Taken together, these findings suggest that interaction design may contribute to clinical effectiveness, but the current evidence does not support simple or universal design conclusions.

Among the 6 features, dialogue depth showed one of the clearest patterns in relation to treatment outcomes. Deeper dialogue was associated with more consistent symptom improvement, whereas lower dialogue depth was linked to more variable treatment effects. These findings suggest that dialogue depth may support therapeutic benefit when it is appropriately structured. Existing research provides mixed evidence regarding the value of open-ended dialogue in DMHIs [51]. Many AI-driven chatbots rely on structured and guided conversational flows to maintain clarity and therapeutic focus [12]. While more open dialogue may increase perceived empathy and human-likeness [52], it may also increase cognitive load or lead to topic drift if not carefully designed [60]. Evidence from cross-cultural studies suggests that reflective and emotionally expressive dialogue can be beneficial, but its impact depends on contextual relevance and timing [61]. From a design perspective, dialogue depth may be better understood as an adaptive feature rather than a fixed attribute [62-64]. Integrating deeper dialogue within structured therapeutic components, such as journaling or behavioral activation tasks, may help maintain alignment with therapeutic objectives [65]. Overall, these observations highlight an important tension between expressive interaction and cognitive manageability. Dialogue depth may contribute to therapeutic alliance and perceived empathy under certain conditions [66], but its effectiveness likely depends on user characteristics, emotional state, and intervention structure.

Other interaction features showed less consistent relationships with symptom change. Emotional responsiveness was more consistently associated with user adherence and, to a lesser extent, with clinical outcomes. Interventions that incorporated more consistent and contextually appropriate emotional feedback tended to show more stable effects, but increasing emotional expressiveness without moderation is unlikely to improve outcomes directly. Research suggests that user engagement is influenced not only by emotional tone but also by the relevance and structure of therapeutic content [67]. Experimental work further indicates that improvements in emotional response mechanisms may enhance user trust and cognitive restructuring processes [68]. Emotional responsiveness may therefore be better understood as a process of calibration rather than intensity [69,70].

Feedback strategies also appeared to have a less stable relationship with clinical effectiveness than with adherence. Interventions that incorporated structured and personalized feedback tended to show more stable engagement patterns, but their influence on symptom improvement likely depends on how feedback is implemented and integrated within the intervention. Within internet-delivered cognitive behavioral therapy, individualized feedback has been associated with lower dropout even when symptom change is comparable [71]. This suggests that feedback may support treatment delivery, but its direct clinical impact is likely to vary across contexts.

Interaction frequency similarly showed different patterns for symptom change and sustained engagement. Lower-frequency interventions were associated with more consistent symptom improvement, whereas higher-frequency contact did not necessarily correspond to better short-term outcomes. Some studies have found that increased conversational exchange is associated with symptom improvement [72], while others report that gains may stabilize or diminish over longer periods of exposure [73]. This suggests that the effects of interaction frequency may not be linear. From a design perspective, interaction frequency should therefore be considered alongside timing, tone, and user context, and adaptive scheduling may be preferable to fixed high-frequency contact.

By comparison, self-disclosure encouragement and user agency showed weaker and more context-dependent relationships with clinical effectiveness. Encouraging self-disclosure was not consistently associated with improvements in symptoms. This contrasts with prior evidence indicating that self-disclosure is a key mechanism for building therapeutic alliance and enhancing engagement [74-76]. In face-to-face care, disclosure helps reduce stigma and promotes help-seeking, and digital interventions have attempted to replicate these processes through structured prompts for emotional expression and narrative sharing [77,78]. Taken together, these findings suggest that self-disclosure may not function in the same way across digital and in-person settings. Similarly, user agency was not consistently associated with clinical outcomes. Previous research supports the importance of perceived control in digital mental health systems [41,79], but agency may shape how users experience and engage with the intervention rather than directly improve outcomes.

Taken together, the exploratory findings suggest that interaction features may contribute to clinical effectiveness, but their relationships with symptom improvement are mixed and strongly shaped by context. Dialogue depth appeared to show the clearest potential relevance to therapeutic benefit, whereas emotional responsiveness, feedback strategy, and interaction frequency showed less stable associations with clinical outcomes. Self-disclosure encouragement and user agency showed weaker and more context-dependent patterns. These observations suggest that potential design implications for clinical effectiveness should be interpreted cautiously. At present, the evidence is better suited to generating conceptual implications than to supporting fixed design recommendations.

The exploratory analyses suggested that interaction features showed clearer and more consistent patterns for user adherence than for clinical effectiveness. Across the 6 interaction features, emotional responsiveness, feedback strategy, and interaction frequency appeared to be more consistently associated with sustained engagement, whereas dialogue depth, self-disclosure encouragement, and user agency showed weaker or more context-dependent relationships. Taken together, these findings suggest that interaction design may be particularly important for understanding continued participation in chatbot-based interventions.

Emotional responsiveness showed one of the clearest relationships with user adherence. Interventions that incorporated more consistent and contextually appropriate emotional feedback tended to show more stable effects. However, this does not suggest that increasing emotional expressiveness without moderation will necessarily improve outcomes. Rather, emotional responsiveness may support engagement when it is calibrated appropriately, whereas excessive or poorly timed amplification may increase emotional burden or cognitive load and thereby undermine sustained engagement [80-82]. Qualitative studies have shown that users value personalized emotional support delivered at an appropriate pace [83]. Concerns about fully automated systems often extend beyond privacy and safety to include whether the system responds in a socially and emotionally appropriate manner [84]. Personalization therefore remains important, and adjusting tone, timing, and response frequency in relation to recent mood patterns may enhance usability and satisfaction [85].

Feedback strategy also appeared to play an important role in supporting user adherence. Interventions that incorporated structured and personalized feedback tended to show more stable engagement patterns. Timely and context-aware prompts can increase short-term engagement [86], whereas generic reminders may be less effective for sustaining engagement in real-world settings [87]. Broader research on guided digital interventions indicates that formats incorporating responsive elements or human support tend to achieve better retention than unguided approaches [88-90]. Methodological reviews further identify adherence and attrition as central determinants of overall effectiveness in DMHIs [91,92]. Feedback may therefore serve as a reinforcement cue that helps stabilize engagement over time, although excessive or poorly timed prompts may contribute to notification fatigue [93].

Interaction frequency similarly showed a clearer relationship with sustained engagement than with symptom change. Higher-frequency contact appeared more closely linked to continued participation. Interaction frequency may operate as both a structural and behavioral cue. A predictable rhythm of contact can reduce decision burden and support habit formation by transforming prompts into routine action cues [94-96]. From a design perspective, interaction frequency should be considered alongside timing, tone, and user context. Adaptive scheduling based on user behavior or mood patterns may be preferable to fixed high-frequency contact, and allowing users to adjust contact frequency may further support autonomy and reduce fatigue.

By comparison, dialogue depth showed a less stable relationship with sustained engagement. Although deeper dialogue was associated with more consistent symptom improvement, its relationship with continued use was less clear, and lower dialogue depth was linked to inconsistent adherence patterns. Existing research provides mixed evidence regarding the value of open-ended dialogue in DMHIs [51]. Many AI-driven chatbots rely on structured and guided conversational flows to maintain clarity and therapeutic focus [12]. While more open dialogue may increase perceived empathy and human-likeness [52], it may also increase cognitive load or lead to topic drift if not carefully designed [60]. Evidence from cross-cultural studies suggests that reflective and emotionally expressive dialogue can be beneficial, but its impact depends on contextual relevance and timing [61]. Research on digital behavior change interventions also suggests that early interactions should minimize cognitive demands to support initial engagement [97], and providing users with options to regulate conversational depth may reduce interaction fatigue [98].

Encouraging self-disclosure and increasing user agency also showed weaker and more context-dependent relationships with adherence. Although prior evidence indicates that self-disclosure can support therapeutic alliance and engagement [74-76], the present findings suggest that its effects in digital interventions may depend more strongly on timing, pacing, and context. Digital interventions have attempted to introduce structured prompts for emotional expression and narrative sharing [77,78], but willingness to disclose sensitive information may vary across regions and populations [99,100]. Privacy and ethical concerns remain important barriers, as fear of data misuse or personal information leakage can directly undermine trust and weaken adherence [101-103]. Similarly, user agency was not consistently associated with sustained engagement, although providing an appropriate degree of choice and control may still contribute to perceived engagement and satisfaction. Previous research supports the importance of perceived control in digital mental health systems [79]. At the same time, excessive freedom may increase interactional burden, whereas overly constrained interaction may reduce perceived control and engagement [104]. A balanced approach may therefore be more acceptable across different users and contexts [105].

Taken together, the exploratory findings suggest that interaction features may offer greater explanatory value for user adherence than for clinical effectiveness. In particular, emotional responsiveness, feedback strategy, and interaction frequency appeared to be more consistently related to sustained engagement, whereas dialogue depth, self-disclosure encouragement, and user agency seemed more dependent on timing, structure, and user readiness. However, given the heterogeneity of the evidence and the low certainty of the main outcomes, these patterns should be interpreted cautiously as conceptual implications rather than prescriptive design rules.

This systematic review has several limitations that should be considered when interpreting the findings. First, the number of included studies was relatively small (n=11), particularly for subgroup analyses by AI chatbot type. This may have limited statistical power and reduced the generalizability of the findings. Second, substantial heterogeneity was observed across studies. Variations in intervention duration, chatbot design, delivery format, and target populations may have contributed to the variability and uncertainty in effect estimates, despite the use of random-effects models. The wide PIs in the main analyses further suggest that effects may vary across comparable future settings.

Third, interaction features were extracted and scored based on descriptions reported in the included studies rather than direct inspection of chatbot behavior. Although a structured coding protocol was applied and ratings were conducted independently by 2 human-computer interaction experts, with adjudication by a third expert, some degree of subjectivity in feature interpretation was unavoidable. The feature-level findings should therefore be interpreted as exploratory. Finally, this meta-analysis relied exclusively on published studies, which may have introduced publication bias, as studies reporting nonsignificant or negative results are less likely to be published. In addition, statistical assessment indicated evidence of small-study effects for clinical effectiveness, and trim-and-fill adjustment attenuated the pooled estimate. However, given the substantial heterogeneity across studies, funnel plot asymmetry may partly reflect genuine between-study variability rather than publication bias alone. Taken together, these limitations indicate that the main findings should be interpreted cautiously.

Future research should adopt more standardized reporting of interaction features, clinical outcomes, and adherence measures. Studies that directly test the causal impact of specific interaction strategies, ideally within comparable therapeutic frameworks, are needed to clarify how different design elements relate to symptom change and sustained engagement. In addition, adaptive and personalized interaction models warrant further investigation to better accommodate diverse user needs, intervention contexts, and patterns of use. More transparent reporting of chatbot interaction design and more standardized documentation of intervention characteristics would also improve comparability across studies and support stronger evidence synthesis in the future.

The main contribution of this systematic review and meta-analysis is that it not only evaluated the clinical effectiveness of AI-driven chatbots for depression but also examined user adherence and interaction features within the same analytic framework. This allowed the review to move beyond the question of whether AI-driven chatbots may work and to explore possible reasons why findings vary across studies. The results showed a favorable trend for depressive symptom reduction, but the overall evidence remained uncertain. In addition, chatbot type alone did not provide a stable explanation for differences in user adherence. By contrast, interaction features, especially those related to sustained participation, appeared to offer a more informative perspective for understanding user engagement. Compared with previous reviews that mainly focused on overall effectiveness or differences between chatbot types, this study places greater emphasis on the role of interaction design in explaining variation in both outcomes and adherence. In this way, it offers a more fine-grained interpretive framework for the field. In practical terms, the findings suggest that the value of AI-driven chatbots for depression depends not only on the underlying technical architecture but also on how interactions are designed, structured, and supported over time. Future system development, clinical evaluation, and real-world implementation should therefore consider clinical outcomes and sustained engagement together, rather than relying only on short-term symptom change or chatbot type as the main basis for evaluation.