From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care.

From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care

Highlights

What are the main findings?•Among psychiatric outpatient attendees, 66.2% (CI 61.9–70.1%) reported prior use of generic AI chatbots for emotional concerns. Fear of negative judgment and burdening others were recurrent themes of barriers to sharing emotional concerns.

•“A broad psychiatric outpatient needs assessment and adult lived-experience co-design process informed CARES-MDD, a Hong Kong Cantonese, Cognitive Behavioral Therapy (CBT)-informed generative-AI conversational-agent prototype for adjunctive support in psychiatric care.”

What are the implications of the main findings?•Generative AI chatbots with specified features are needed. Content analysis identified positive and negative points in users’ experience during the development process.

•This study provides a stakeholder-informed intervention specification and design parameters for a subsequent controlled evaluation of CARES-MDD.

Abstract

Background/Objectives: The burden of major depressive disorder (MDD) calls for assessing the needs and design of innovative add-on interventions. Generative artificial-intelligence (AI) tools are increasingly used in everyday life. Evidence remains limited on how needs assessment and co-participatory design can inform AI chatbot trials for MDD. Methods: This study reports a mixed-methods design. Phase 1 comprised a needs assessment of 511 psychiatric outpatient attendees in Hong Kong. Phase 2 involved a participatory co-design and test-run experiences with 10 lived-experience experts with MDD and eight multidisciplinary clinical professionals. This process informed the development of CARES-MDD, a Hong Kong Cantonese, CBT-informed generative AI conversational-agent prototype intended to provide adjunctive support alongside usual psychiatric care. Results: In Phase 1, 66.2% reported using generic AI chatbots for emotional concerns; privacy, anticipated negative evaluation, and concern about burdening others were recurring themes in the qualitative data. Localized empathy, non-judgmental, and 24/7 accessibility were the main themes of a desired mental health generative AI chatbot. In Phase 2, CARES-MDD was co-designed and assessed against a 7-domain feedback rubric measuring perceived usefulness and acceptability. Overall, CARES-MDD received favorable feedback rubric ratings. Content analysis identified positive and negative experiences, which established parameters for future trial methodology. Conclusions: This study identifies needs and yields the co-design of the CARES-MDD purposed as adjunct support for MDD in psychiatric care in Hong Kong, providing the parameters to inform future controlled trials.

1. Introduction

Major depressive disorder (MDD) is a leading cause of disability worldwide [1], accounting for substantial disease burden, diminished functional capacity, and elevated suicide risk [2,3]. While empirically validated psychological interventions, most notably cognitive behavioral therapy (CBT), are recommended as first-line treatments, fewer than half of affected individuals receive adequate care [4,5].

Digital mental health interventions (DHIs) may increase access to information and self-management support; however, sustained engagement, cultural relevance, and integration with routine psychiatric care remain important challenges [6,7]. Critically, the overwhelming majority of mental health DHIs are developed without participatory input from clinical populations, producing tools that feel generic, robotic, and emotionally misaligned for individuals contending with severe depressive symptoms [8,9].

The emergence of generative artificial intelligence (AI) and large language models (LLMs) offers an opportunity to bridge this gap. Unlike rigid, rule-based algorithms, LLMs can generate responsive language that may be perceived as empathic or validating; however, such outputs are not equivalent to human understanding or clinician-delivered therapy and require empirical safety and efficacy evaluation [10,11,12]. When integrated with evidence-based psychotherapeutic frameworks—specifically CBT, which targets transdiagnostic cognitive and behavioral maintaining mechanisms (e.g., cognitive distortions, avoidance, and rumination [13,14])—generative conversational agents can deliver flexible, tailored micro-interventions in real time.

However, translating generative AI into a psychiatric adjunct treatment tool requires tuning, localized adaptation, and safety governance. This formative mixed-methods study aimed to identify user-defined requirements, stakeholder-identified failure modes, and governance conditions for the development of Conversational Artificial-intelligence for REducing Symptoms and REsponse for Major Depressive Disorder (CARES-MDD), which is a Hong Kong Cantonese AI-supported conversational agent for persons living with depression. This study was designed to inform a prespecified, staged validation program for CARES-MDD, including content curation, technical evaluation, safety assurance, and prospective feasibility testing before an efficacy trial. Specifically, this study sought to (1) characterize digital behavior, generic AI chatbot use, barriers to emotional disclosure, and desired chatbot features among people attending psychiatric outpatient services; (2) translate these findings, together with feedback from lived-experience experts and clinical professionals, into prioritized generative AI chatbot design recommendations; and (3) describe stakeholder perceptions of perceived usefulness and acceptability using feedback rubrics and qualitative interviews to collect their positive and negative points when testing CARES-MDD.

2. Materials and Methods

2.1. Study Design

This study employed a mixed-methods, participatory co-design framework, drawing upon established guidelines for DHIs [15,16,17]. Recognizing that the target demographic has important experience and opinions of generative AI, the investigation was structured into two interrelated phases:

Phase 1 (Needs Assessment): A cross-sectional quantitative and qualitative survey investigating baseline digital habits, help-seeking attitudes, self-stigma, and conversational AI acceptability/needs among adolescents and young people attending child and adolescent psychiatric outpatient services. This cohort was not restricted to individuals with MDD for real-world generalizability.

Phase 2 (Participatory Co-Design and Test-run Evaluation): A stakeholder co-design/feedback phase involved adult lived-experience experts (LEEs) advisers and multidisciplinary clinical professionals. The adult-only reference group was used to refine the initial prototype, identify design and governance considerations, and discuss sensitive issues relating to emotional disclosure and AI-supported mental health care. Phase 1 findings were used to identify candidate design priorities, including privacy and discretion, use pattern, low-burden access and barriers reported. Phase 2 included a workshop of LEEs to integrate the prioritized implementation ideas to co-design the generative AI chatbot (CARES-MDD) and then invited them for test-run CARES-MDD to evaluate its acceptability, perceived usefulness and perceived safety in mixed-methods evaluation for its potential for progression to a randomized controlled trial (RCT) design.

2.2. Participants and Recruitment Settings

Phase 1 Cohort. A clinical cohort of 511 outpatient psychiatric attendees was recruited from Child and Adolescent Psychiatric Outpatient Clinics at Queen Mary Hospital Service attending between 22 July and 21 August 2025 during routine clinical visits. Recruitment was conducted by trained research interns who were stationed in the clinic during selected weekday clinic sessions, according to intern availability. During staffed sessions, interns sequentially approached patients and, where applicable, their parent or guardian to introduce the study, provide study information, and invite voluntary participation. Participation or non-participation did not affect clinical assessment or treatment. The sample was a time–location convenience sample. Although recruitment was conducted sequentially among patients encountered during each staffed session, the recruitment opportunities depended on intern availability and the clinic sessions covered. The service did not operate diagnosis-specific clinics; therefore, patients attending the general child and adolescent psychiatric service during staffed sessions could be approached regardless of their presenting diagnosis. This cohort was not restricted to individuals with MDD, and depressive symptom severity was assessed using the Patient Health Questionnaire-9 (PHQ-9). Eligible participants possessed a personal smartphone with daily internet access and demonstrated the ability to read written Chinese and communicate in Cantonese. No formal a priori sample-size calculation was performed for Phase 1. As this was a formative, descriptive needs assessment, the sample size was determined pragmatically by the number of eligible patients who consented during the prespecified four-week recruitment period and staffed clinic sessions. All enrolled participants were included in the analysis.

Phase 2 Reference Group. The Phase 2 participatory co-design involved the interview of eighteen experts, comprising ten LEEs, alongside a multidisciplinary expert group of eight clinical professionals (two clinical psychologists, three counselors, one psychiatrist, and two psychiatric nurses). LEEs had a confirmed clinical diagnosis of major depressive disorder (MDD) based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM 5) [18], ascertained by case psychiatrists. This diagnostic criterion applied only to Phase 2 LEEs and did not apply to the broader Phase 1 outpatient needs-assessment cohort. The LEEs were all aged 18 years or above; restricting the reference group to legal adults was intended to ensure that participants had the developmental maturity to discuss lived experiences of depression, emotional disclosure, and digital mental health support in an interview setting. Exclusion criteria included acute suicidal risk, active psychosis, or concurrent substance use disorders. While the LEEs guided the humanistic and user-experience elements, the clinical experts provided feedback on the content and appropriateness of the interaction. Furthermore, the expert panel provided opinions on the safety features in practical use.

2.3. Phase 1: Needs-Assessment Measures and Data Collection

An electronic questionnaire was developed to capture quantitative metrics and qualitative insights across four primary domains:

Online behavior and AI chatbot use: Their daily digital screen time, generic conversational AI adoption rates, daily online interaction duration, and digital engagement frequency were asked. Depressive symptom severity was assessed using the validated Chinese version of the 9-item Patient Health Questionnaire (PHQ-9) as a clinical variable [19]. The PHQ-9 was used to quantify self-reported depressive symptom burden and to classify symptom-level groups for analysis; it was not used to establish a DSM-based diagnosis of MDD.

Help-Seeking and Self-Stigma: Standardized psychometric instruments measured help-seeking behaviors (Help-Seeking Scale, max score 28) and internalized mental health stigma (Self-Stigma Scale, max score 24, where higher scores denote lower stigma) from the Knowledge and Attitudes to Mental Health Scales (KAMHS) [20], which was recently validated in Hong Kong [21].

Barriers to emotional disclosure and help-seeking within existing psychiatric care networks: Participants also completed an open-ended item asking “What are the reasons that prevent you from sharing emotional difficulties with others?”

Desired AI Chatbot Specifications: To guide the system architecture and feature prioritization, a final open-ended item asked “What is the most desired feature of an AI chatbot within the existing psychiatric system?”

2.4. Phase 2: Co-Design Workshop, Beta-Runs, and Mixed-Methods Evaluation

To translate the needs-assessment findings into actionable generative AI chatbot development, a participatory co-design workshop was conducted with lived-experience experts (LEEs). The engagement protocol was grounded in established youth participatory action research (YPAR) principles [22] and guidelines for co-designing digital health interventions with young people [23], which used non-medicalized language, culturally relevant humor, and open body language (e.g., maintaining eye-level seating) to actively cultivate trust and rapport. Active listening techniques, such as reflective summarizing and open-ended prompting, were utilized to verify shared understandings, clarify conceptual misinterpretations, and empower LEEs to expand upon their ideas on the prioritization of the generative AI chatbot theme. This collaborative process directly informed the prioritized implementation strategies for the Conversational Artificial-intelligence for REducing Symptoms and REsponse for Major Depressive Disorder (CARES-MDD) intervention (detailed in Appendix A
Table A1), bridging the gap between clinical utility and user-centered design.

Following the initial co-design workshops, the expert group was invited for multiple test-run cycles conducted over a development window. The expert reference group attended several supervised, in-person test-runs throughout this period to report any good and bad points about the CARES-MDD. At the conclusion of this iterative cycle, participants completed a feedback rubric encompassing seven study-specific domains, which were generated directly from the prioritized co-design strategies (Appendix A
Table A1). The rubric was not a previously validated instrument, was not separately pilot tested, and was not intended to establish psychometric validity, formal consensus, or predefined feasibility criteria. Participants rated these domains on a 10-point visual analog scale (1 = completely inadequate, 10 = exceptional), as shown in Table 1. The seven-domain feedback rubric was designed as a joint stakeholder instrument rather than as separate patient-reported and clinician-reported measures, reflecting the co-design principle that acceptability and trial readiness should be triangulated across lived-experience and clinical-governance perspectives simultaneously, rather than evaluated as independent, potentially divergent outcomes. Given the small number of raters in each subgroup (10 LEEs; 8 clinical professionals), subgroup-stratified estimates would carry limited descriptive stability and were not considered informative at this sample size.

Description of the CARES-MDD Chatbot

The CARES-MDD application is a text-based, multi-threaded chatbot application developed using a large language model and designed to communicate in natural Hong Kong-style Cantonese. Its conversational content was developed by psychiatrists and clinical psychologists and reviewed by members of the clinical team with reference to evidence-informed, primarily CBT principles. To address clinical safety, CARES-MDD has a keyword-based crisis-signposting module screening for expressions associated with suicidal ideation, death wishes, self-harm, hopelessness, and possible acute risk, including terms such as “suicidal”, “death”, and “giving up”. Detection of these triggers initiates an automated safety interface,’ temporarily suspends the generation of routine chatbot content, and displays a static emergency interface with 24/7 direct-dial crisis hotlines. The research team also conducted periodic clinical review of chatbot content and system responses to identify potential safety concerns on a weekly basis. CARES-MDD was designed to support compliance with applicable data-protection requirements. Appropriate measures were implemented to protect participant information and restrict access to study data to authorized members of the research team. It was not intended to diagnose MDD, provide emergency care, replace professional assessment, or autonomously manage suicidal risk. The CARES-MDD system identity, intended use, and technical configuration, together with its safety-screening and crisis-response pathway, are summarized in Appendix A
Table A2, Table A3 and Table A4.

2.5. Analysis

We used descriptive statistics and inferential tests to examine closed-ended questions regarding participants’ digital behaviors, help-seeking patterns, and the evaluation rubrics. Quantitative data was organized and checked in Microsoft Excel, and statistical analyses were conducted in R statistical software (version 4.3.0; R Foundation for Statistical Computing, Vienna, Austria). In Phase 1, inferential analyses were restricted to comparisons of two clinically relevant psychosocial outcomes—help-seeking score and self-stigma score—between participants with elevated versus lower depressive symptoms. Depressive symptom groups were defined using the Patient Health Questionnaire-9 (PHQ-9): participants with a PHQ-9 score of ≥10 were classified as having elevated depressive symptoms, while those with a PHQ-9 score of <10 were classified as having lower depressive symptoms. This symptom-based grouping was used for analysis only and does not represent a DSM-based diagnosis of major depressive disorder. Between-group comparisons of continuous outcomes. Welch’s unequal variances t-tests were used to compare mean help-seeking and self-stigma scores between the two PHQ-9 symptom groups according to distributional and variance considerations because equal variances were not assumed. Analyses used available cases for each outcome; no missing-data imputation was performed. Continuous variables were examined for distributional characteristics before summarization. Variables with clearly right-skewed or zero-bounded distributions, including digital-engagement durations, were reported as medians with interquartile ranges (IQRs). Approximately symmetric psychometric and usability measures were summarized as means with standard deviations. Categorical variables, such as generic AI adoption rates and interaction frequencies, were reported as frequencies and percentages. Inferential analyses remained restricted to the prespecified comparisons of help-seeking and self-stigma scores between participants with elevated and lower depressive symptoms. These comparisons were conducted in R’s base function stats::t.test() with var.equal = FALSE, corresponding to Welch’s unequal-variances t-test. Welch’s test was selected a priori because the two symptom groups were of unequal size and equality of variances was not assumed; no Student’s equal-variance t-tests were used. Analyses used available cases for each outcome, and no missing-data imputation was performed. No formal adjustment for multiple comparisons was applied because the inferential analysis was restricted to two prespecified clinically relevant psychosocial outcomes rather than a broad set of exploratory tests.

For qualitative data derived from Phase 1, open-ended responses were generally brief phrases or one-to-two-sentence comments. Short-text responses were coded to categorize non-mutually exclusive recurring themes (e.g., privacy concerns, fear of judgment). To provide cultural context, frequently occurring Chinese terms within these responses were mapped and visualized using word clouds. The qualitative feedback of Phase 2 was conducted as an inductive, data-driven iterative process [24]. Due to limitations within the dataset (short phrases or sentence fragments), full thematic analysis was not appropriate [25]. Two researchers (M.S.Y.Y. and K.Y.H.L.) independently coded approximately 30% of responses using inductively generated, non-mutually exclusive descriptive codes. Multiple codes could be assigned when a response contained more than one meaning. They compared coding decisions, discussed discrepancies, and reached negotiated agreement to refine the coding framework, which was then applied to the remaining responses. No formal inter-rater reliability statistic was calculated because coding was inductive and non-mutually exclusive, and the purpose was descriptive categorization rather than validation of a fixed coding instrument. Codes were grouped into recurring categories and summarized by frequency. For Phase 2, brief spoken and written feedback on positive and negative aspects of the test-runs was analyzed using inductive descriptive content analysis. The same researchers generated codes directly from the data, discussed coding differences, and reached negotiated agreement on final categories. Related categories were grouped into broader themes, to identify prototype-refinement priorities.

2.6. Ethical Approvals and Consent Procedures

The assessment received institutional review board approval from the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW25-299). Informed consent was obtained from adult participants and their parents (for minors) prior to engagement. Participant confidentiality was secured by decoupling contact information from survey data and assigning de-identified alphanumeric study IDs.

3. Results

3.1. Phase 1 Results: Clinical Needs Assessment for Barriers and AI Chatbot Use

The cohort comprised 511 young people (mean age 14.97 ± 1.86 years; 54.0% female, 46.0% male). Analysis of online behavior revealed a significant divergence based on psychological distress levels. No patient-level information was logged for individuals who were not approached, were not eligible, or declined participation. Accordingly, no participant-versus-non-participant comparison was possible. Participants recorded a significantly high digital footprint, spending an average of 296.3 min per day online. Nearly two out of three young people (66.2%; 95% CI: 61.9–70.1%) are already familiar with and have navigated generic AI chatbots, implying the clinical population is already primed for conversational agents. While adoption is high, frequency data highlights a critical gap: a total of 56.4% of active users have only had sporadic, surface-level interactions (fewer than 10 conversations total). This suggests that, while young people are curious about AI chatbots, current generic AI chatbot tools are not providing the sustained, meaningful engagement required for psychological intervention for the clinical demographic (Table 2).

Of the 511 participants, 65 (12.7%) provided additional comments in response to the question “What are the reasons that prevent you from sharing emotional difficulties with others?” Comments were generally brief, comprising a phrase or one to two sentences (total word count = 834). Iterative thematic analysis identified Privacy concerns (n = 22), characterizing emotional disclosure as “privacy” or a “personal matter” and expressing concern that their concerns can be gossiped about to others; Fear of judgement (n = 16) encompassed anticipated misunderstanding, criticism, stigma, or negative evaluation, as reflected in comments such as “other people’s views and prejudices” and “afraid of being misunderstood”; and Avoidance of burdening others (n = 8) captured concerns that disclosure might negatively affect or inconvenience others, illustrated by “do not want to affect other people’s mood” and “do not want to disturb others”. These are visualized in the word cloud in Figure 1.

A total of 42 (8.2%) provided additional open-ended comments in response to the question “What is the most desired feature of an AI chatbot within the existing psychiatric system?” Comments were generally brief, comprising short phrases or single sentences (total word count = 586). Iterative thematic analysis identified four non-mutually exclusive desired features. The category most frequently identified among the 42 respondents who provided comments was 24/7 immediate accessibility (n = 18), with participants valuing on-demand emotional support between formal clinical appointments (e.g., “always available” and “immediate reply when I cannot sleep”). Privacy and non-judgmental space (n = 12) reflected the preference to share vulnerable emotions without fear of human evaluation, characterized by comments like “no human judgment” and “completely confidential”. Localized empathy (n = 8) captured the desire for naturalistic, Hong Kong-style Cantonese interactions that feel culturally resonant rather than robotic (e.g., “sounds like a local friend”). Finally, practical psychological guidance (n = 4) referred to the need for actionable coping strategies, exemplified by “helps me reframe my negative thoughts”. These are visualized in the word cloud in Figure 2.

3.2. Phase 2 Results: Co-Participatory Design and Expert Evaluation

Following the Phase 1 needs assessment, 10 LEEs (mean age 22.9 ± 2.9 years; 50% female, 50% male) and eight clinical experts were invited to the initial workshop for co-design (Appendix A
Table A1). They were then invited to test-run CARES-MDD to evaluate its acceptability, perceived safety, and feasibility of progression to a definitive trial of CARES-MDD. Only the LEEs had the pre- and post-observations adapting the outcome assessors. Descriptive ratings from the study-specific seven feedback domains are summarized in Figure 3.

Within this small stakeholder reference group, participants provided generally favorable ratings across all CARES-MDD feedback domains. Empathy and HK-Cantonese understanding, Accuracy and useful CBT info, and Balance of privacy and clinical governance achieved the highest median scores. These findings indicate positive preliminary stakeholder perceptions of the prototype’s culturally resonant communication, perceived usefulness and acceptability, and value of clinician oversight.Ratings regarding acceptability of a future randomized evaluation suggested that participants considered the proposed trial approach broadly acceptable. However, no prespecified feasibility criteria or formal consensus threshold was used; therefore, these ratings should not be interpreted as demonstrating feasibility of a future trial. Safety features and adverse digital events alongside active listening and Socratic questions showed wide interquartile ranges, suggesting variation in stakeholder perceptions, while others found automated overrides abrupt and that Socratic questioning could be demanding during distress. Finally, Supportive pacing and intervention duration yielded the lowest median score, capturing challenges foreseeable if it is positioned as an open-ended, indefinite intervention.The descriptive observations of the CARES-MDD test-run were summaraised in Table A5. Additionally, Table 3 thematically organizes the qualitative feedback underlying these quantitative evaluations.

Among the positive experiences, LEEs valued CARES-MDD as a private “tree hole” for emotional disclosure and perceived it as more private than commercial applications because it was not perceived to use disclosed content for targeted advertising that would pass the information to vendors. The use of natural Hong Kong-style Cantonese was viewed as familiar, emotionally accessible, and appropriate to local expressions. CBT-informed content was generally regarded as clinically appropriate and distinguishable from generic emotional support after several active exchanges. To ensure privacy, participants advocated for using pseudonyms to prevent data misattribution. However, they simultaneously supported backend clinician oversight, noting that human supervision of their conversations provided necessary reassurance against potential AI unreliability. Rather than limitations, their sharing of negative experience establish parameters for future trial methodology. Specifically, the cognitive demands of a text-based modality indicate that the intervention is optimally targeted toward a specific demographic: namely, younger, digitally native populations with a higher baseline tolerance for typing. Furthermore, the eventual onset of digital fatigue suggests that CARES-MDD is most viable as a time-limited intervention instead of an open-ended, indefinite intervention. Technologically, the LLM occasionally exhibited inconsistent longitudinal memory, failing to seamlessly recall contextual details from very early sessions, a stark contrast to human psychotherapists. Automated safety overrides were sometimes perceived as abrupt, particularly when triggered by non-suicidal colloquial exclamations such as “死啦” (“oh no”). Some participants expressed residual concern about possible AI hallucinations, despite not reporting such experiences during the study.

4. Discussion

This study reported a mixed-methods study: first, to characterize the online digital behavior, usage of generic AI chatbots for emotional concerns, its barriers, and needs assessment of desired AI mental health chatbot specifications; second, to co-design and evaluate the iterative test-run of the CARES-MDD among experts living with MDD. By bridging clinical needs-assessment data (Phase 1), we identified (1) barriers of self-stigma and fear of human judgement for sharing emotional concerns in add-on support within the existing psychiatric care framework; (2) high technological readiness with generic AI chatbots for emotional concerns; and (3) the desired features for development of an AI mental health chatbot for this specific clinical demographic. Phase 2 reported an iterative expert co-design process of CARES-MDD involving LEEs and a clinical expert reference group. They prioritized and shared their positive experience regarding CARES-MDD’s features of local style communication, CBT-curated RAG instead of generic content, and the balance of privacy and safety monitoring. Rather than limitations, their sharing of their negative experience regarding LLM-related risks and intervention pacing highlights key areas requiring reference parameters for future trial methodology.

Among the subgroup of Phase 1 respondents who provided voluntary open-ended comments, recurring categories included privacy concerns, fear of negative judgement, and concern about burdening others. These findings provide contextual insight into potential barriers to emotional disclosure even among this vulnerable cohort already attending psychiatric outpatient clinics [26]. These barriers may reduce willingness to seek help even when clinical need is present, particularly when the clinical population fears that their emotional disclosures will be misunderstood, shared beyond their control, or imposed on others [27,28]. A generative AI chatbot has the potential as an add-on micro-intervention to foreground immediate accessibility without burdening people, employing a conversational register that approximates support interaction rather than formal assessments.

It was further supported by the findings that most outpatient attendees had high technological readiness, with widespread surface acquaintance with generative AI chatbots. Users who have interacted briefly with commercial chatbots may hold implicit expectations derived from those platforms (rapid factual responses, task completion, minimal affect) that are poorly matched to the longitudinal, reflective, emotionally containing functions that a therapeutic chatbot must serve [24]. Among the 42 respondents who commented on desired features, recurring categories included immediate availability, a private and non-judgmental conversational space, natural local style communication, and practical psychological guidance. These comments informed the desired priorities of the subsequent process design priorities for the subsequent co-design process but should not be interpreted as population-level preferences.

The Phase 2 evaluation indicated generally favorable perceptions across the seven CARES-MDD feedback domains. Empathy and local style understanding, the accuracy and usefulness of CBT-informed content, and balance of privacy and safety received the highest ratings. Participants valued CARES-MDD as a private “tree hole” for emotional disclosure. The preference for unique identifiers or pseudonyms further illustrates the importance of maintaining confidence that individual disclosures will not be confused with those of another user. Participants appreciated that CARES-MDD responded in colloquial Cantonese rather than formal written Chinese, and several noted that this linguistic calibration created a sense of relational proximity uncommon in health-technology products [29,30].

Participants did not frame privacy as the absence of clinical oversight. Rather, they supported proportionate human oversight when data governance arrangements were transparent and sensitive personal details were protected. This finding is important for the design of clinically oriented conversational AI: trust may depend on clearly communicating what is monitored, who can access information, what will be documented, and the circumstances under which confidentiality may be overridden because of safety concerns.

CBT-informed content was perceived as more useful than generic reassurance when interactions progressed beyond short exchanges. Participants and experts considered the chatbot more helpful when it could validate distress, clarify the immediate concern, and only subsequently introduce structured reflection or cognitive restructuring. This supports a staged conversational design: initial empathic listening and emotional validation should precede problem solving or Socratic questioning. The interactions were perceived as more useful after several active exchanges, particularly when a session involved at least several active exchanges. This observation addresses a methodological challenge in DHI research: operationalizing intervention “dosage” outside of traditional psychiatric frameworks. Unlike conventional psychotherapy, which relies on structured 50 min appointments, unguided digital tools face tremendous variability in user engagement duration. Because meta-analyses indicate that an effective clinical dosage for digital interventions ranges from 30 to 60 min per week [31], future RCTs can formally define an “active digital session” as several active exchanges.

Several LEEs also reported that abrupt safety override responses, while appropriate from a risk-management standpoint, interrupted the conversational flow at moments when continued empathic engagement may have been more therapeutically productive. Clinical professionals specifically noted that, while algorithmic safety gates are non-negotiable from a duty-of-care standpoint, their current implementation in generative AI systems typically operates as a binary interrupt—abruptly redirecting conversation to crisis resources—rather than as a graduated, therapeutically calibrated transition. In human clinical practice, the decision to escalate, contain, or stay present with distress is a moment-to-moment clinical judgement informed by relationship history, non-verbal cues, and dynamic risk appraisal.

A central question explored during the co-design interviews was how clinical users conceptualize the identity of CARES-MDD. Instead, users perceived the CARES-MDD prototype as a portable, on-demand conversational support and skills-coaching tool during unassisted hours (e.g., late nights or between outpatient visits). Multiple LEEs affirmed their willingness to recommend CARES-MDD to peers who ‘aren’t ready for traditional therapy yet.’ By allowing distressed individuals to practice cognitive restructuring and emotional articulation in an unobserved, safe environment, the conversational agent serves as an accessible entry point that may lower psychological barriers and build readiness for future formal human psychiatric care.

The integration of generative AI into psychiatric populations raises broader ethical and clinical considerations for the field, particularly regarding unhealthy dependence, hallucination risks, and longitudinal memory constraints [32,33,34]. Their qualitative feedback highlighted a known vulnerability inherent to current LLMs [35]. LLM models unable to dynamically contextualize the psychotherapeutic clinical formulation of the engaging client unlike a human therapist may be better served as a complementary tool for “augmenting” pre-existing therapeutic alliance with patients, rather than serving as the therapist alliance in a standalone role [36]. In the emerging era of AI, we should be more cautious in the role of generative AI chatbots and monitor possible safety issues and adverse digital events. Although the LEEs did not experience these, they mentioned the risk of AI hallucinations and psychosis that they overheard. The LLM risks identified in the process are summarized in the subsequent Table 4, aligning with the guideline for a generative AI chatbot architecture ready for future trials [37].

Participants’ comments regarding long-term engagement should be interpreted as design considerations for future evaluation rather than as evidence of an optimal intervention duration. Participants anticipated digital fatigue if CARES-MDD was available as an open-ended or indefinite intervention. In light of attrition commonly reported in unguided digital mental health interventions, a time-limited, on-demand format may offer a pragmatic starting point for prospective feasibility testing. An approximately 8-week intervention window could allow participants sufficient opportunity to engage with CBT-informed self-management content while limiting prolonged exposure and the potential for overreliance on the conversational agent [38,39]. To mitigate the digital fatigue issue, further RCT trials could integrate the ecological momentary assessment technique, which has shown evidence of boosting adherence through self-monitoring and reminder effects [40,41].

Participants generally considered randomization, including allocation to a psychoeducation-only active digital control, to be fair and acceptable. Future trial design should minimize participant burden, particularly for providing appropriate reimbursement, online modes for assessments; and include follow-up sufficient to examine engagement, acceptability, safety-related events, and longer-term outcomes. While this formative study provided evidence on the architectural feasibility, cultural acceptability, and safety governance boundaries of CARES-MDD, it was not designed to evaluate randomized clinical efficacy or comparative attrition. Therefore, rather than demonstrating clinical superiority, these findings provide the critical structural parameters, such as the intervention dosage, the system and clinical measures to counteract the potential LLM risks, e.g., pre-consented clinical escalation pathways, required to safely initiate a definitive RCT.

Strengths, Limitations, and Future Clinical Directions

A strength of this investigation is its mixed-methods methodology, uniting a relatively large clinical sample (N = 511) with more in-depth mixed-methods evaluation of LEEs throughout the development lifecycle, and CARES-MDD shows acceptable ecological and cultural validity. However, several limitations must be acknowledged. First, the pre-to-post observations from the uncontrolled beta-testing phase are strictly descriptive; they could not deduce therapeutic efficacy, and further clinical trial is needed. Second, it was a convenience sample drawn from a single professional and community network, which limits the generalizability. Without a screening log, we cannot determine response rates or compare non-participants. Due to potential selection bias from voluntary enrollment, these findings should be viewed as formative data rather than population-representative prevalence estimates. Nevertheless, given the uncompensated time commitments required for these phases of co-development, leveraging a network with pre-established clinical rapport was essential to maintain experts and accurately capture user-driven needs within existing resource constraints. Third, CARES-MDD intentionally avoids highly experimental technological features. This design choice was driven by the co-design participants, who prioritized clinical reliability and evidence-based safety over technological novelty, which might lead to “blackbox” concerns. While this limits the platform’s technological boundary pushing, it ensures the intervention maintains the rigorous clinical safety required for vulnerable psychiatric populations. Fourth, the Phase 1 needs assessment included adolescents aged 14–19 years, whereas the Phase 2 co-design group comprised adults aged 18 years or older with confirmed MDD. Adult LEE feedback may inform early prototype design but cannot establish acceptability, usability, safety, or clinical relevance for minors. This distinction is particularly important because some younger participants may have restricted smartphone access during school hours and because generative AI deployment among minors requires additional safeguards relating to assent, parental consent, privacy, crisis escalation, and school-context implementation. Future research should undertake adolescent-specific co-design and a prospectively approved feasibility study before extending CARES-MDD to people aged under 18 years.

Future work in the development of generative AI chatbots for mental health purposes must prioritize three interrelated agendas to address the universal challenges. Technically, the field should move beyond aligning global clinical guidelines and standardizing protocols for LLM applications across the disciplines [42,43]. Clinically, the next appropriate evaluative step involves prospective designs, such as full-scale RCTs, to systematically examine the clinical effectiveness. Digital adverse events should be continually assessed in future trials. Conceptually, the pronounced help-seeking barriers and high self-stigma identified in Phase 1 motivate a critical agenda for comparative research: investigating whether initial AI-mediated emotional disclosure can serve as a transitional bridge within a stepped-care pathway, lowering psychological barriers and facilitating subsequent engagement with human-delivered psychiatric care [44]. Addressing these universal challenges will be paramount for integrating generative AI as a safe, regulated, and continuously audited adjunct to global mental health services [11].

5. Conclusions

In conclusion, the broad needs assessment demonstrates that, while the use of generic AI chatbots for emotional support is already prevalent within a heterogeneous cohort of young people attending psychiatric outpatient services in Hong Kong, barriers—namely fear of judgment and burdening others—continue to impede traditional help-seeking behaviors within the traditional psychiatric care system. The needs assessment was not restricted to people with MDD; rather, it reflected the diagnostic complexity and comorbidity encountered in routine child and adolescent psychiatric practice. This underscores the necessity that generative AI chatbots with specified features are needed for individuals living with MDD, even among those already engaged in psychiatric care. By utilizing a needs-driven participatory framework with LEEs, a CBT-grounded generative AI chatbot for MDD has been co-developed, with mixed-methods findings supporting its feasibility to end-users and clinical experts. The findings provide preliminary, clinically grounded design parameters and feasibility information for subsequent evaluation. A future trial should compare CARES-MDD plus usual care with an appropriate active digital control plus usual care, incorporate predefined safety and adverse-event monitoring, and evaluate engagement, long-term safety, and clinical outcomes. This mixed-methods study does not establish the efficacy, safety, or therapeutic equivalence of CARES-MDD; these questions require subsequent prospective controlled trials.

References

1.Chan J.K.N. Solmi M. Lo H.K.Y. Chan M.W.Y. Choo L.L.T. Lai E.T.H. Wong C.S.M. Correll C.U. Chang W.C. All-Cause and Cause-Specific Mortality in People with Depression: A Large-Scale Systematic Review and Meta-Analysis of Relative Risk and Aggravating or Attenuating Factors, Including Antidepressant Treatment World Psychiatry 2025 24 404 421 10.1002/wps.21354 40948054 PMC12434377

2.Malhi G.S. Mann J.J. Depression Lancet 2018 392 2299 2312 10.1016/S0140-6736(18)31948-2 30396512

3.World Health OrganizationDepression and Other Common Mental Disorders: Global Health Estimates World Health OrganizationGeneva, Switzerland2017

4.Rush A.J. Trivedi M.H. Wisniewski S.R. Nierenberg A.A. Stewart J.W. Warden D. Niederehe G. Thase M.E. Lavori P.W. Lebowitz B.D. Acute and Longer-Term Outcomes in Depressed Outpatients Requiring One or Several Treatment Steps: A STARD Report Am. J. Psychiatry 2006 163 1905 1917 10.1176/ajp.2006.163.11.1905 17074942

5.Thornicroft G. Chatterji S. Evans-Lacko S. Gruber M. Sampson N. Aguilar-Gaxiola S. Al-Hamzawi A. Alonso J. Andrade L. Borges G. Undertreatment of People with Major Depressive Disorder in 21 Countries Br. J. Psychiatry 2017 210 119 124 10.1192/bjp.bp.116.188078 27908899 PMC5288082

6.Linardon J. Torous J. Firth J. Cuijpers P. Messer M. Fuller-Tyszkiewicz M. Current Evidence on the Efficacy of Mental Health Smartphone Apps for Symptoms of Depression and Anxiety. A Meta-analysis of 176 Randomized Controlled Trials World Psychiatry 2024 23 139 149 10.1002/wps.21183 38214614 PMC10785982

7.Linardon J. Fuller-Tyszkiewicz M. Firth J. Goldberg S.B. Anderson C. McClure Z. Torous J. Systematic Review and Meta-Analysis of Adverse Events in Clinical Trials of Mental Health Apps npj Digit. Med. 2024 7 363 10.1038/s41746-024-01388-y 39695173 PMC11655657

8.Chinsen A. Berg A. Nielsen S. Trewella K. Cronin T.J. Pace C.C. Pang K.C. Tollit M.A. Co-Design Methodologies to Develop Mental Health Interventions with Young People: A Systematic Review Lancet Child Adolesc. Health 2025 9 413 425 10.1016/S2352-4642(25)00063-X 40311649

9.Veldmeijer L. Terlouw G. Os J.V. Dijk O.V. Veer J.V.’t Boonstra N. The Involvement of Service Users and People With Lived Experience in Mental Health Care Innovation Through Design: Systematic Review JMIR Ment. Health 2023 10 e46590 10.2196/46590 37490326 PMC10410372

10.Volpato R. DeBruine L. Stumpf S. Trusting Emotional Support from Generative Artificial Intelligence: A Conceptual Review Comput. Hum. Behav. Artif. Hum. 2025 5 100195 10.1016/j.chbah.2025.100195

11.Torous J. Linardon J. Goldberg S.B. Sun S. Bell I. Nicholas J. Hassan L. Hua Y. Milton A. Firth J. The Evolving Field of Digital Mental Health: Current Evidence and Implementation Issues for Smartphone Apps, Generative Artificial Intelligence, and Virtual Reality World Psychiatry 2025 24 156 174 10.1002/wps.21299 40371757 PMC12079407

12.McIntyre R.S. Nemeroff C.B. Rasgon N. Implementing Artificial Intelligence (AI) to Facilitate Health Outcomes in Mood Disorders: Application versus Aspiration J. Affect. Disord. 2026 403 121444 10.1016/j.jad.2026.121444 41698618

13.Schefft C. Heinitz C. Guhn A. Brakemeier E.-L. Sterzer P. Köhler S. Efficacy and Acceptability of Third-Wave Psychotherapies in the Treatment of Depression: A Network Meta-Analysis of Controlled Trials Front. Psychiatry 2023 14 118997010.3389/fpsyt.2023.1189970 37867779 PMC10585267

14.Schaeuffele C. Meine L.E. Schulz A. Weber M.C. Moser A. Paersch C. Recher D. Boettcher J. Renneberg B. Flückiger C. A Systematic Review and Meta-Analysis of Transdiagnostic Cognitive Behavioural Therapies for Emotional Disorders Nat. Hum. Behav. 2024 8 493 509 10.1038/s41562-023-01787-3 38228727 PMC10963275

15.Gonsalves P.P. Ansari S. Berry C. Gonsalves F. Iyengar S. Kashyap P. Mittal D. Pal S. Razdan E. Michelson D. Co-Designing Digital Mental Health Interventions with Young People: 10 Recommendations from Lessons Learned in Low-and-Middle-Income Countries Health Educ. J. 2025 84 373 384 10.1177/00178969241297821

16.Peters S. Guccione L. Francis J. Best S. Tavender E. Curran J. Davies K. Rowe S. Palmer V.J. Klaic M. Evaluation of Research Co-Design in Health: A Systematic Overview of Reviews and Development of a Framework Implement. Sci. 2024 19 63 10.1186/s13012-024-01394-4 39261956 PMC11391618

17.Warraich H.J. Tazbaz T. Califf R.M. FDA Perspective on the Regulation of Artificial Intelligence in Health Care and Biomedicine JAMA 2025 333 241 247 10.1001/jama.2024.21451 39405330

18.So E. Kam I. Leung C. Chung D. Liu Z. Fong S. The Chinese-Bilingual SCID-I/P Project: Stage 1—Reliability for Mood Disorders and Schizophrenia Hong Kong J. Psychiatry 2003 13 7 18

19.Kroenke K. Spitzer R.L. Williams J.B.W. The PHQ-9 Validity of a Brief Depression Severity Measure J. Gen. Intern Med. 2001 16 606 613 11556941 10.1046/j.1525-1497.2001.016009606.x PMC1495268

20.Simkiss N.J. Gray N.S. Dunne C. Snowden R.J. Development and Psychometric Properties of the Knowledge and Attitudes to Mental Health Scales (KAMHS): A Psychometric Measure of Mental Health Literacy in Children and Adolescents BMC Pediatr. 2021 21 50810.1186/s12887-021-02964-x 34774022 PMC8590271

21.Tsoi J.K.K. Yuen S.S.Y. Kwok C. Xu N.S. Tsoi X.L. Hui C.C.Y. Yuen P.T.Y. Cheng A.W.C. Chu W.L. Hao J.J.X. Translation and Validation of the Traditional Chinese Version of the Knowledge and Attitudes to Mental Health Scales (KAMHS) in Hong Kong Adolescents East Asian Arch. Psychiatry 2026 36 196 204 42802396 10.12809/eaap2694

22.Cammarota J. Fine M. Revolutionizing Education: Youth Participatory Action Research in Motion RoutledgeLondon, UK2010 978-1-135-91324-3

23.Blease C. Tibbs M. Balaskas A. Liverpool S. Hagström J. Fitzgerald A. Coproduction Without Youth? Closing the Participation Gap in Digital Mental Health Research JMIR Ment. Health 2026 13 e91739 10.2196/91739 42342243 PMC13347077

24.Braun V. Clarke V. Using Thematic Analysis in Psychology Qual. Res. Psychol. 2006 3 77 101 10.1191/1478088706qp063oa

25.Marks D.F. Yardley L. Research Methods for Clinical and Health Psychology SAGEThousand Oaks, CA, USA2003 978-1-4462-3232-3

26.Thornicroft G. Stigma and Discrimination Limit Access to Mental Health Care Epidemiol. Psychiatr. Sci. 2008 17 14 19 10.1017/S1121189X00002621 18444452

27.Lee J. Lee D. Lee J. Influence of Rapport and Social Presence with an AI Psychotherapy Chatbot on Users’ Self-Disclosure Int. J. Hum.–Comput. Interact. 2024 40 1620 1631 10.1080/10447318.2022.2146227

28.Lee Y.-C. Yamashita N. Huang Y. Fu W. “I Hear You, I Feel You”: Encouraging Deep Self-Disclosure through a Chatbot Proceedings of the Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems Association for Computing MachineryNew York, NY, USA2020 1 12

29.Grové C. Co-Developing a Mental Health and Wellbeing Chatbot With and for Young People Front. Psychiatry 2021 11 60604110.3389/fpsyt.2020.606041 33597898 PMC7882508

30.Sit H.F. Ling R. Lam A.I.F. Chen W. Latkin C.A. Hall B.J. The Cultural Adaptation of Step-by-Step: An Intervention to Address Depression Among Chinese Young Adults Front. Psychiatry 2020 11 65010.3389/fpsyt.2020.00650 32733296 PMC7359726

31.Ma Q. Shi Y. Zhao W. Zhang H. Tan D. Ji C. Liu L. Effectiveness of Internet-Based Self-Help Interventions for Depression in Adolescents and Young Adults: A Systematic Review and Meta-Analysis BMC Psychiatry 2024 24 60410.1186/s12888-024-06046-x 39243081 PMC11380195

32.Elyoseph Z. Levkovich I. Beyond Human Expertise: The Promise and Limitations of ChatGPT in Suicide Risk Assessment Front. Psychiatry 2023 14 121314110.3389/fpsyt.2023.1213141 37593450 PMC10427505

33.Keshavan M. Torous J. Yassin W. Do Generative AI Chatbots Increase Psychosis Risk? World Psychiatry 2026 25 150 151 10.1002/wps.70017 41536114 PMC12805049

34.Paulus M.P. Torous J. LLMs as Clinical Instruments—Toward Verifiable Reasoning JAMA Psychiatry 2026 83 896 897 10.1001/jamapsychiatry.2026.2211 42525421

35.Tie J. Yao B. Li T. Fang H. Ahmed S.I. Wang D. Zhou S. “Should I Give Up Now?” Investigating LLM Pitfalls in Software Engineering ACM Trans. Softw. Eng. Methodol. 2026 10.1145/3801972

36.Dayan R. Uliel B. Koplewitz G. Age against the Machine—Susceptibility of Large Language Models to Cognitive Impairment: Cross Sectional Analysis BMJ 2024 387 e081948 10.1136/bmj-2024-081948 39706600 PMC12128858

37.Freyer O. Wiest I.C. Kather J.N. Gilbert S. A Future Role for Health Applications of Large Language Models Depends on Regulators Enforcing Safety Standards Lancet Digit. Health 2024 6 e662 e672 10.1016/S2589-7500(24)00124-9 39179311

38.Lipschitz J.M. Pike C.K. Hogan T.P. Murphy S.A. Burdick K.E. The Engagement Problem: A Review of Engagement with Digital Mental Health Interventions and Recommendations for a Path Forward Curr. Treat. Options Psychiatry 2023 10 119 135 10.1007/s40501-023-00297-3 38390026 PMC10883589

39.Li Y. Luk T.T. Cheung Y.T.D. Zhao S. Zeng Y. Tong H.S.C. Lai V.W.Y. Wang M.P. Engagement With a Mobile Chat-Based Intervention for Smoking Cessation: A Secondary Analysis of a Randomized Clinical Trial JAMA Netw. Open 2024 7 e2417796 10.1001/jamanetworkopen.2024.17796 38922618 PMC11208971

40.Tucker I. Bertotti M. Hanafiah A. Hossain S. Watts P. Ahad M.A.R. Developing a Digital Ecological Momentary Assessment Tool for ‘Real Time’ Evaluation in Implementation Science: Testing through Evaluation of a Novel Digital Social Prescribing Intervention Front. Public Health 2026 14 171830210.3389/fpubh.2026.1718302 42254647 PMC13236872

41.Chirokoff V. Tessier A. Serre F. Dupuy M. Auriacombe M. Chanraud S. Berthoz S. Fatseas M. Misdrahi D. Relevance of Ecological Momentary Assessment for Medication Adherence in Clinical Settings: A Precision Psychiatry Approach Br. J. Clin. Psychol. 2025 64 692 701 10.1111/bjc.12532 39960195 PMC12334975

42.Knott M. Krebs M. Kerscher A. Large Language Models in Healthcare Quality Management: A European Perspective on Process Automation and Compliance Front. Digit. Health 2026 8 176164110.3389/fdgth.2026.1761641 41970523 PMC13062252

43.Ong J.C.L. Ning Y. Liu M. Ma Y. Zhao L. Singh K. Chang R.T. Vogel S. Lim J.C.W. Tan I.S.K. Innovating Global Regulatory Frameworks for Generative AI in Medical Devices Is an Urgent Priority npj Digit. Med. 2026 9 364 10.1038/s41746-026-02552-2 41857339 PMC13168469

44.Esmaeilzadeh P. Mirzaei T. Dharanikota S. Patients’ Perceptions Toward Human–Artificial Intelligence Interaction in Health Care: Experimental Study J. Med. Internet Res. 2021 23 e25856 10.2196/25856 34842535 PMC8663518