This study was a 3-arm randomized controlled trial, with double-blinding between the AI-augmented iCBT (AI-iCBT) and iCBT groups, while the waitlist control group was unblinded. The intervention arms consisted of AI-iCBT, an unguided, fully self-administered iCBT program incorporating NLP feedback, and iCBT, an unguided, fully self-administered program without NLP feedback. These 2 arms are hereafter collectively referred to as “unguided iCBT.” The waitlist group served as the control condition.

Invitation emails were sent to all monitors registered with Nikkei Research Inc, with the aim of recruiting at least 900 participants. Interested individuals were directed to complete an online screening survey, which included the Patient Health Questionnaire-9 (PHQ-9) [33,34], to determine eligibility.

Inclusion and exclusion criteria are provided in Textbox 1.

Eligible participants were randomly assigned to 1 of the 3 groups. Immediately before the intervention, PHQ-9 scores were reassessed; individuals with scores <5 were excluded from the efficacy analysis set (EAS) but remained in the overall study population. The detailed definition of the EAS is provided in the Analytic Strategy section.

The AI-augmented iCBT program, developed by NEC Solution Innovators, Ltd (Tokyo), integrates an NLP module trained on 28,718 prior iCBT records from Japanese users. The entire program was delivered in Japanese, and Figure 1 presents an English-translated version of the original Japanese interface for publication purposes. The program includes a self-guided cognitive restructuring (CR) exercise, where users complete a 7-column thought record to address cognitive distortions. The NLP system processes user inputs, including situation, automatic thoughts, and feelings, referencing a corpus of past responses (Figure 1). It provides 2 types of automated feedback with text and phonation: (1) empathetic messages delivered through an animated character whose expressions, such as smiling or showing sadness, are synchronized with the message content, and (2) advisory messages offering guidance to refine inputs or direct users to appropriate exercises, including suggestions to revise content if the user’s input was unclear or misplaced (eg, a feeling given instead of a thought; Figure 2).

In contrast, the non-AI iCBT program retained the same structure but provided only neutral, noncontextual responses, such as generic phrases like “Uh-huh” with neutral facial expressions. Both AI-iCBT and iCBT programs were otherwise identical in content, using the same validated 6-session self-guided iCBT package. This iCBT program has previously demonstrated significant antidepressant effects in a randomized controlled trial among working adults (n=60 per group), compared with a waitlist control, with a medium to large effect size (Cohen d=0.65; P<.005) [37]. This package consisted of 6 weekly sessions, each including a 15-minute video-based psychoeducation module covering standard CBT principles such as behavioral activation and problem-solving, along with a weekly CR exercise in which users applied learned techniques. In this trial, the only difference was the addition of the NLP feedback system. The program was available on both smartphones and PCs. The AI-enhanced features, which were designed in advance to improve user engagement and response accuracy, exhibited high usability, with low dissatisfaction rates reported for both the empathetic (3/32, 9.4%) and advisory (1/24, 4.2%) feedback functions (Figure 3).

The registered participants were randomly and concurrently assigned to either the AI-iCBT, iCBT, or waitlist groups using a computer-generated random sequence provided by an independent third party. Stratified randomization was applied based on age (≤40 vs >40 years), sex, and baseline PHQ-9 score (≤9 vs >9), as baseline symptom severity has been shown to influence treatment outcomes in self-guided iCBT [23]. Participants in the waitlist group were aware of their allocation and were therefore unblinded, whereas those in the AI-iCBT and iCBT groups were told only that they would participate in iCBT using “the latest technology,” without disclosure of their specific group assignment. Accordingly, blinding was implemented between the 2 intervention groups.

Automated email reminders were sent to participants twice weekly during the 7-week intervention period. Each week, participants in the intervention groups were required to (1) view an online psychoeducational CBT module and (2) perform their allocated (AI-iCBT or iCBT) CR exercise at least once (6 times or more in total). Waitlist participants did not undergo any exercises during this period. All participants were required to complete assessments at baseline, postintervention (week 7), and follow-up (month 3 after postintervention). All intervention and assessment procedures, including attendance and outcome measures, were conducted online.

All primary and secondary outcomes were analyzed based on the intention-to-treat (ITT) population, which included all randomized participants.

Reporting of this trial followed the CONSORT (Consolidated Standards of Reporting Trials) 2010 statement [61] and the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) checklist [62] for internet-based interventions. The completed CONSORT-EHEALTH checklist is submitted as Multimedia Appendix 1.

This study was reviewed and approved by the Hiramatsu Memorial Hospital Ethics Committee (approval number 20150807). All participants provided informed consent electronically prior to enrollment after reading an online information sheet describing the study purpose, procedures, potential risks, and voluntary nature of participation. Participants were informed that they could withdraw at any time without penalty.

The trial was prospectively registered in the University Hospital Medical Information Network Clinical Trials Registry.

All data were anonymized prior to analysis to ensure confidentiality. No personally identifiable information was accessible to the research team. Participants who completed the final assessment received a ¥500 (US $4.5) gift voucher as compensation. No identifiable images or other personal data are presented in this manuscript.

A total of 1187 participants were eligible and randomly allocated to the AI-iCBT (n=396), iCBT (n=397), or waitlist (n=394) groups (ITT population; see Figure 4 for the CONSORT flow diagram). Baseline demographic and clinical characteristics are summarized in Table 1. The mean age was 43.50 (SD 9.85) years, and 699 (58.8%) of 1187 participants were male. Across the 3 groups, demographic and clinical characteristics were well balanced, with no significant differences in depressive symptom severity (PHQ-9: P=.56; QIDS-J: P=.74). No significant baseline differences were found between the AI-iCBT and iCBT groups, confirming the comparability of the 2 active interventions.

Figure 4 shows the flow of participants through the trial, including the numbers assessed for eligibility, randomized, allocated to each study arm (AI-iCBT, iCBT, control), completing follow-up assessments at week 7 and month 3, and included in the ITT analysis.

The primary outcome, the mean score on the PHQ-9, did not show statistically significant between-group differences compared with the control group at either week 7 or month 3 (AI-iCBT vs control: least squares mean difference –0.47, 95% CI –1.13 to 0.18; P=.16; Cohen d=–0.10; iCBT vs control: least squares mean difference –0.62, 95% CI –1.28 to 0.04; P=.07; Cohen d=–0.13; Table 2). No significant differences were observed between the AI-iCBT and iCBT groups. Nevertheless, both intervention groups showed significant within-group reductions from baseline at week 7 and month 3 (all P<.001), indicating that depressive symptoms improved over time in both groups.

For the secondary binary outcome of PHQ-9 ≥10, the overall proportion decreased over time across all groups (Figure 5). At month 3, the proportion was numerically lower in the AI-iCBT group compared with the control group, but between-group differences were not statistically significant in the ITT analysis (Multimedia Appendix 2).

Similar patterns were observed for other secondary measures. QIDS-J and GAD-7 scores improved significantly within both intervention groups but without significant between-group differences. SDS scores showed modest reductions but did not significantly differ from control. Full secondary outcome results are provided in Multimedia Appendix 3.

In evaluating factors associated with low adherence, participants who attended <3 sessions were more likely to be male (65.1% vs 48.4%; P<.001), older (mean age 44.2, SD 9.44 vs mean age 41.7, SD 9.55 years; P=.001), and employed (287/341, 84.2% vs 212/283, 74.9%; P=.02), compared with those who attended ≥3 sessions. No significant differences were observed for marital status, educational background, medical history, or mental health history (Multimedia Appendix 8).

As a further sensitivity analysis, when applying a stricter definition of depression severity—PHQ-9 score of ≥10 plus at least 1 core symptom and SDS score of ≥10—62.7% (178/284) of participants with a PHQ-9 score of ≥10 at baseline met this definition (AI-iCBT: 44/70, 62.9%; iCBT: 41/71, 57.7%; control: 93/143, 65%), with no significant group differences (Multimedia Appendix 9).

This study has several unique features. First, it directly compared a fully self-administered CR exercise delivered via iCBT, with and without AI-based NLP functionality, under a randomized (partially masked) design. Notably, the addition of AI led to a statistically significant improvement in engagement—an effect, to our knowledge, not previously documented. As AI-based interventions have become increasingly sophisticated and deeply integrated into intervention programs, disentangling the specific contribution of AI has become difficult. In particular, establishing control conditions that differ only in the presence or absence of AI functionality requires substantial resources, and prior studies have therefore often relied on heterogeneous control conditions. By applying a more robust design—feasible in part because the technology was still in a transitional phase—this trial provides new insights into how AI may enhance fully self-administered iCBT and offers a timely perspective for advancing scalable mental health care.

Self-administered interventions are known to be modestly to moderately effective for depression [15,23], but adherence remains a major limitation [15,18,19,21,55,63]. Systematic reviews indicate that approximately one-third to one-half of participants drop out before completing the program [19,64]. In this context, the engagement improvement observed in this study represents a potential step toward overcoming this barrier.

With regard to clinical effectiveness, no additional benefits of NLP feedback were observed in the ITT population. Recent evidence has reported that greater antidepressant effects are associated with lower dropout rates [15,20,65]. In contrast, although no between-group differences in antidepressant effects were found here, the addition of AI feedback was associated with a statistically significant increase in adherence. This suggests a novel engagement-enhancing mechanism, distinct from the traditionally assumed link between larger clinical effects and lower dropout rates.

Exploratory analyses further indicated that the “empathic function” of AI feedback was significantly associated with improved adherence, whereas the advisory function showed no significant effect. Participants who received empathic responses during their first exercise subsequently demonstrated higher adherence. While self-disclosure was not directly measured, the sense of being supported may have facilitated persistence. These findings align with prior evidence that empathic conversational agents and chatbots support therapeutic alliance and sustained engagement [66-69]. Research in behavioral change has likewise shown that AI-mediated feedback can promote sustained self-management [70], supporting the plausibility of these findings. Such effects of human-AI interaction may have been less visible in prior studies using heterogeneous control conditions but became evident here through the structured randomized design.

Regarding antidepressant effects, no significant between-group differences were observed in the ITT population. In the EAS, results—while requiring cautious interpretation—indicated that at week 7 only the iCBT group showed significant improvement in both the mean PHQ-9 score and the proportion of participants with a PHQ-9 score of ≥10 compared with the control group, whereas the AI-iCBT group did not.

This suggests that the AI function may have attenuated or failed to enhance short-term antidepressant effects. However, this short-term benefit in the iCBT group disappeared at long-term follow-up. For the continuous outcome in particular, the short-term difference was –1.1 points on the PHQ-9, below the minimal clinically important difference (approximately 3 points) [35,71], suggesting limited clinical significance.

By contrast, the short-term dichotomous outcome in the iCBT group represented about a 29% reduction in the proportion of PHQ-9 scores of ≥10 cases relative to the control group. This implies that part of the potential benefit may not have been realized when AI was introduced. At long-term follow-up (month 3), however, only the AI-iCBT group showed a significant reduction of about 15% compared with the control group. These findings highlight the absence of the expected short-term effect in AI-iCBT and the unique long-term effect observed only in AI-iCBT.

The fact that AI-iCBT ultimately demonstrated an effect at long-term follow-up is noteworthy. Although exploratory, this suggests a potential contribution of AI-iCBT in reducing the proportion of participants exceeding a clinically significant severity threshold. One possible explanation is that approaches emphasizing empathy as a core therapeutic skill—such as interpersonal psychotherapy or family therapies—often show slower onset but more enduring gains compared with CBT, lasting well beyond the end of treatment [72-77]. It is possible that the empathy-related function of AI contributed in a similar way, although the underlying mechanisms remain unclear.

The EAS, however, was more restrictive than the ITT population. Among ITT participants with a baseline PHQ-9 score of ≥5, only 149/312 (47.8%) in the AI-iCBT group and 134/312 (42.9%) in the iCBT group attended at least 3 sessions. Furthermore, although a PHQ-9 score of ≥10 is widely recognized as a proxy for “major depression equivalent” in research [33,39], concerns have been raised that it may not be sufficient for diagnostic purposes and may contribute to overdiagnosis [56,78,79]. As a sensitivity analysis, therefore, we used a stricter definition requiring a PHQ-9 score of ≥10 plus at least 1 core symptom (depressed mood or anhedonia) [56,57], together with an SDS score of ≥10 as an indicator of functional impairment [49,58-60]. Results confirmed that only about 62.7% (178/284) of participants who met the PHQ-9 score of ≥10 at baseline also met this stricter definition, highlighting the importance of cautious interpretation (Multimedia Appendix 9).

Taken together, regardless of whether it corresponds to major depression, the finding of a significant reduction in the proportion of participants with clinically meaningful depressive states (PHQ-9 ≥10) compared with the control group may have clinical significance, particularly given the fully self-help nature of the program. From a public health perspective, such a difference could also carry implications for the scalability of self-help programs that do not require therapist involvement.

AI communication, including generative AI, continues to advance rapidly. However, the development of appropriate control programs has often been constrained by logistical and cost-related factors, limiting opportunities to rigorously investigate the antidepressant and anxiolytic effects of AI. Beyond psychiatry, maximizing the effectiveness of self-administered interventions while enhancing engagement remains a critical challenge across health care and welfare domains. This study represents a step toward addressing this challenge.

First, most participants were recruited from a research panel with high affinity for digital interventions. Only 10.9% (129/1187) were actual users of mental health services, which is consistent with the national average in Japan, but caution is required in generalizing these findings to broader populations. Second, high dropout rates were observed, with only 37.6% (149/396) of the AI-iCBT group and 33.8% (134/397) of the iCBT group in the ITT population meeting EAS criteria. Additional analyses indicated that low adherence was associated with being male, older, and employed (Multimedia Appendix 8). Consistent with recent studies, time constraints [22], particularly among employed men [21] and older adults [20,21], were confirmed as key barriers to engagement. Third, although the AI-iCBT group consistently showed 5%-10% higher adherence rates than the iCBT group throughout the trial, both groups had already dropped by half to 50% (396/793) at the very first session and continued to decline over time, remaining at only 30%-40%.

Fourth, missing data were substantial, and MMRM and generalized estimating equations were applied to minimize bias. However, these approaches assume data are missing at random. In this study, since attrition occurred according to participant attributes, the possibility of missing not at random cannot be ruled out, and estimates may remain biased. Fifth, a significant group × time interaction was observed in adherence in the AI feedback group, suggesting a potential role of AI in improving engagement. However, this conclusion is based on a single trial and requires replication in different populations and designs, as well as further elucidation of the underlying mechanisms. Sixth, due to technical issues, session-by-session data on depressive symptoms (Overall Anxiety Severity and Impairment Scale) [80] and anxiety (Overall Depression Severity and Impairment Scale) [81] were lost, precluding more detailed analyses. Future studies should implement automatic data saving and backup systems to prevent such loss. Overall antidepressant effects were small. In addition to limitations of the program itself, this may reflect a ceiling effect due to the predominance of participants with mild depression. Several meta-analyses [23,82,83] report that treatment effects may be less pronounced in cases with mild baseline depression compared to moderate or severe cases. Another limitation is related to our stratified randomization. While stratification by age, sex, and baseline PHQ-9 severity increased internal validity by balancing key prognostic factors, it may also restrict the generalizability of our findings to populations with different distributions of these characteristics.

Despite these challenges, this study provides valuable insights into the potential of AI-enhanced self-help digital interventions, particularly in relation to participant behavior dynamics. Importantly, no major adverse events were reported, underscoring the safety of this innovative approach and its potential to significantly advance mental health care practices.