Depression is a major global health care challenge that affects millions of people. The growing prevalence of depression strains mental health care systems, particularly due to a shortage of qualified professionals. The imbalance between rising demand and limited resources has prompted researchers to explore alternative solutions, such as using digital technologies to support diagnosis and treatment [1-3].

Artificial intelligence (AI) is widely discussed as a potentially powerful tool for depression treatment, given its capability to improve care access and quality. Early digital tools mainly offered basic remote therapy, while modern AI can offer more sophisticated and scalable mental health care interventions. AI-empowered chatbots, virtual therapists, and predictive algorithms are already being used to deliver cognitive behavioral therapy and other forms of support [4,5]. Another key strength of AI lies in its capacity to provide consistent diagnoses for identical symptoms, in contrast to human physicians who may offer differing interpretations of the same clinical signs due to variations in their knowledge, skills, and experiences [6]. This aspect of AI systems has been argued to suggest potential in avoiding diagnostic inaccuracies and standardizing patient care across various clinical settings. Nevertheless, AI is not a replacement for therapists. In clinical settings, this technology typically takes a supportive role in decision-making, improving workflow, and managing large volumes of patient data [7]. Human oversight remains essential throughout the process.

Despite rapid technical advances and increasing adoption of AI in mental health care, patients’ expectations of AI’s role in depression treatment remain largely unknown. Most of the published work focuses on system development or clinical integration [8,9], and only a handful of studies have explored the clinical experience of care recipients. This gap could be especially problematic in mental health care contexts, where the therapeutic relationship is deeply personal and the patient is not just a passive recipient but also a central agent in the treatment process. Therefore, neglecting the patient’s voice in the design of AI-based mental health care tools may lead to a misalignment between technological function and actual user needs, and such a mismatch can result in dissatisfaction, mistrust, and ultimately the nonadoption of AI systems.

Moreover, AI in mental health care must navigate a sensitive terrain where a tone-deaf response or clunky interface may be perceived as harmful or distressing by users. For someone in a depressive episode, a poor AI interaction might feel impersonal or, even worse, harmful. This study addresses the research gaps by adopting a patient-centered approach grounded in qualitative inquiry, drawing directly from the narratives of individuals who may have symptoms of depression. Specifically, this work explores patients’ actual expectations for AI functions and qualities as well as their concerns about potential risks associated with AI. Organizing these insights through grounded theory can align with patients’ needs and enable the development of AI tools that are both technically robust and emotionally attuned.

Existing studies have examined AI applications across mental health care, ranging from monitoring patient states and detecting early warning signs to supporting diagnosis and treatment.

AI can assist mental health care professionals in understanding a patient’s condition by collecting and analyzing real-time data, thereby enabling more accurate diagnoses and personalized treatment plans [10,11]. Through smartphones and wearable devices, AI continuously monitors a patient’s mental state by tracking diverse indicators, such as sleep patterns, physical activity, social interactions, and subtle behavioral cues [3,12,13]. These comprehensive data provide a complete picture of an individual’s mental state [11].

AI tools also facilitate the early detection of relapse or worsening mental conditions, thereby allowing for timely intervention and improved treatment outcomes [14]. AI supports early detection by analyzing diverse sources, including patient-reported data [15], clinical notes [16], mobile use patterns [17], social media content [18], and patient messaging data [19], hence allowing algorithms to predict depression risks.

AI can also support decision-making in mental health care [20-22]. When human professionals are unavailable, AI monitors patient data over time and adjusts clinical recommendations accordingly [23]. This tool can act as an assistant to clinicians by handling administrative tasks and enhancing efficiency [24] or provide an additional safeguard by supervising clinical processes [25] and detecting nuanced psychotherapeutic information [26]. In some cases, AI can surpass human capabilities by integrating information from multiple sources to avoid diagnostic inaccuracies [27].

Some platforms also use AI to simulate therapeutic conversations and provide personalized support. For instance, Woebot and Wysa deliver cognitive behavioral therapy for conditions such as depression and anxiety. Similarly, chatbots [4] and virtual therapists [5,28] assist users in navigating various mental health care challenges.

While a substantial body of research has documented the capabilities of AI in mental health care, these studies primarily reflect the perspectives of mental health care professionals or system designers and largely ignore the voices of patients. Existing research has yet to adequately explore how patients perceive AI psychotherapists, specifically the features they value, the risks they fear, and the functions they expect. This gap warrants further exploration as patients’ acceptance, satisfaction, and comfort play a crucial role in the perceived usability and acceptability of AI systems in mental health care. This study addresses this gap by grounding the investigation in patients’ own narratives.

The study was reviewed and approved by the Human Subjects Ethics Sub-Committee at City University of Hong Kong (project 11500322). Prior to participation, individuals viewed an informed-consent page outlining the study’s purpose, expected time commitment, voluntary nature of participation, types of data collected (survey responses and demographic measures), and the research team’s contact information. Only participants who provided consent were allowed to proceed to the questionnaire. No directly identifying information (eg, names, email addresses, or other personal identifiers) was collected as part of the survey. Survey responses were not linked to personally identifying information and were analyzed in deidentified form to protect participant confidentiality. Data were stored on the researcher’s local device, with access restricted to the research team. Each participant received US $5 via Amazon Mechanical Turk (MTurk) upon completing the study.

Participants were recruited through MTurk, yielding a closed convenience sample of adults who voluntarily opted into the study. Data were collected between March 2023 and May 2023. Participants read a scenario in which an AI system was used to assist in depression treatment and then completed a questionnaire. Our questionnaire consisted of 3 open-ended questions, along with a set of demographic, control, and contextual items. The questions were iteratively refined to improve clarity and face validity through internal review by the research team. The questionnaire was administered across multiple pages, with related questions grouped on the same page (eg, demographic items presented together). No item randomization or adaptive branching was implemented. Participants could review their responses before final submission, and item nonresponse was permitted for nonmandatory fields. The survey link was distributed to 493 individuals. A total of 452 online surveys were completed, yielding a completion rate of 91.7%. Because recruitment occurred via MTurk (a closed platform), view and participation rates based on unique site visitors were not applicable. Cookies, IP addresses, and log-file checks were not used, but duplicate participation was prevented by restricting to one submission per MTurk ID, and only completed surveys were included in the analysis. No completion-time cutoff or statistical corrections were applied because the study focused on qualitative open-ended responses. For further details regarding the study design, please refer to the CHERRIES (Checklist for Reporting Results of Internet E-Surveys) checklist provided as Checklist 1. The participants were asked the following 3 open-ended survey questions:

Although our Q2 used the term “physician,” all open-ended questions were explicitly framed within AI-enabled psychotherapy as defined in a vignette presented before the survey. We therefore interpreted Q2 responses as expectations for patient-valued care attributes and interaction qualities (eg, interactional behaviors, support functions, safety, and boundary setting). Our aim was to identify the desired physician attributes that are transferable to AI features. In addition to the 3 open-ended questions, we collected several control and contextual variables. We measured participants’ computer skills using a single question (“How would you rate your computer skills?”) [29], familiarity with AI using a 3-item measure (participants indicating their level of agreement with the following statements: “I am familiar with AI app”; “I am familiar with seeing a doctor using AI app”; and “I am familiar with the processes of using AI app to see a doctor”) [30], social stigma toward depression treatment using 5 questions (“I believe that receiving treatment for depression carries social stigma”; “I believe that it is a sign of personal weakness or inadequacy to receive treatment for depression”; “I think that people will see a person in a less favorable way if they come to know that he/she has received treatment for depression”; “I believe that it is advisable for a person to hide from people that he/she has been treated for depression”; and “I think that people tend to like less those who are receiving professional help for depression”) [31], trust in an AI doctor using 1 question (“To what extent do you trust an AI virtual doctor for depression treatment?”) [32], and privacy awareness using one single question (“In general, I care about my privacy and view the violation of my privacy seriously”) [33]. All of the above measures were assessed using 7-point Likert scales and adapted from the extant literature. We also measured participants’ current depression severity using the Patient Health Questionnaire-9 [34]. All continuous variables were calculated by averaging the ratings of the associated measurement items. Moreover, participants reported their demographic information, including sex, age range, and education level.

We conducted a qualitative analysis of the textual data, integrating each participant’s 3 open-ended responses to develop a holistic profile of an AI psychotherapist. More specifically, responses of Q1 were used to identify the clinical scenarios of use, responses of Q2 were used to elicit transferable care attributes, and responses of Q3 were used to capture perceived risks. The analysis followed the principles of grounded theory [35,36] and used an inductive, bottom-up approach to theme development. The first and third authors conducted the coding, while the second author randomly audited the coded data to ensure accuracy. Upon receiving the participants’ responses, coder 1 (the first author) conducted an initial open coding process, assigning preliminary labels to the raw textual data. The purpose of this step was to transform unstructured responses into manageable analytical units and to ensure that the labels comprehensively captured and accurately reflected the essence of the participants’ statements. Throughout the coding process, coder 1 set aside prior assumptions and focused on meanings emerging directly from the data [37]. This approach supported the discovery of meaningful themes suited to the exploratory nature of the study. Data saturation was reached when no additional labels emerged, and key terms and phrases related to scenarios, desired features, and perceived risks had been comprehensively identified. Subsequent responses largely repeated existing patterns, thereby reinforcing the established labels. Coder 1 then grouped related codes into broader thematic categories. After the themes were finalized, each theme was clearly defined to delineate its conceptual boundaries and clarify what content fell within its scope.

Subsequently, coder 1 assigned binary codes to each response for every theme (1=“theme present” and 0=“theme absent”). A codebook encompassing all identified scenarios, features, and risks was developed to facilitate subsequent cross-checking between coders. When coding “features,” coder 1 only categorized expectations as AI psychotherapist features when the response described a capability or interactional behavior that could plausibly be delivered by an AI system (or an AI component in hybrid care), rather than assuming that desirable human clinician qualities are automatically transferable to AI. Coder 2 was then provided with the raw responses and the finalized list of themes along with their definitions. Independently, coder 2 assigned binary codes to each response based on these definitions. Intercoder reliability was assessed by calculating the average matching rate across all responses (a match was defined as both coders assigning a value of 1 to the same theme). The Cohen κ coefficient yielded a value of 0.80, which falls within the range for substantial agreement (0.61‐0.80) and supports the overall consistency and validity of the coding process [38].

In addition to extracting general desired features and perceived risks, we examined how these features and risks varied across participant profiles based on (1) social stigma toward depression treatment, (2) current depression severity, (3) trust in an AI doctor, and (4) privacy awareness. Participants were first categorized into groups for each profile variable. For stigma, trust, and privacy awareness, participants were grouped into high, medium, or low bands using 1‐SD cutoffs around the mean; depression severity followed established Patient Health Questionnaire-9 criteria [34] to categorize participants into 5 levels (none, mild, moderate, moderately severe, and severe). Within each subgroup, we recalculated the frequency of each desired feature and perceived risk. These frequency distributions allowed us to identify the prioritization patterns of expected features and perceived risks across different participant profiles.

The sample was relatively balanced by sex, skewed younger, and predominantly well-educated. The participants also reported moderately high levels of computer proficiency. Detailed demographic characteristics are presented in Table 1.

Five envisioned scenarios of AI psychotherapist applications emerged from the collected responses, including diagnosis, treatment, consultation, self-management, and companionship (Table 2). The most frequently mentioned scenario was diagnosis, followed by treatment and consultation, indicating a clear emphasis on the clinical and decision-support functions of AI psychotherapists. Self-management reflects expectations for AI tools to assist users in independently managing their mental health over time. While less prominent, companionship suggests that some participants envision AI playing an emotional role. Taken together, these scenarios suggest that the participants imagined AI functioning across a broad spectrum of care, ranging from diagnostic assessment to emotional companionship.

Expected features were identified through interpretive coding. Some of these features were explicitly stated (eg, “I want [feature]” or “It should [do something]”), while others were inferred from negative statements (eg, “I would avoid an AI doctor that [does something]”). In such instances, undesirable attributes were conceptually reframed as their positive counterparts to inform design features aimed at mitigating user resistance. These features reflect participant expectations rather than empirically validated requirements. A broad range of expected features was identified using a criterion of at least 5% of the 452 participants (ie, mentioned ≥23 times). This threshold was used to highlight the most salient features that may warrant prioritization in AI psychotherapist implementation. These features are summarized in Table 3. Sample quotes for all the expected features are provided in Multimedia Appendix 1.

Participants’ expectations can be mapped onto a continuum of desired attributes, ranging from those traditionally associated with human physicians, through a middle ground of qualities that both human physicians and AI psychotherapists can potentially possess, to attributes that primarily apply to AI psychotherapists (Figure 1). Importantly, attributes located in the shared middle region do not imply equivalence between AI and human physicians; rather, they indicate domains in which AI psychotherapists may approximate valued aspects of clinical interaction through deliberate design, structured workflows, and adaptive responses. Despite this continuum, we retain and discuss the full set of desired attributes because participants framed them as a holistic profile of “good care,” which we henceforth treat collectively as expected features of an AI psychotherapist. In this sense, even attributes that sit at the human physician end of the continuum remain informative for AI psychotherapist design and implementation: they can be applied in hybrid AI-human care models, where the human component handles relational aspects, while AI contributes complementary strengths.

Perceived risks were also identified through an interpretive coding process. In addition to explicitly naming risks, the participants described specific situations they feared. In such cases, the concrete examples were abstracted into higher-level categories based on their underlying nature (whether rooted in the diagnostic process, the therapeutic interaction, and so on). As with the analysis of expected features, the present focus is limited to risks mentioned by at least 5% of the participants to highlight the most salient concerns. These risks are summarized in Table 4, and sample quotes corresponding to each perceived risk are provided in Multimedia Appendix 1.

To translate participants’ expected features and perceived risks into a conceptual prioritization structure, we applied the MoSCoW framework [42] and used mention frequency as an indicator of salience. Features mentioned by at least 20% of participants were classified as “Must-have,” those mentioned by 10% to 20% as “Should-have,” and those mentioned by 5% to 10% as “Could-have.” In addition, the “Won’t-have” category captures risks and undesirable aspects that the system should explicitly avoid to prevent exacerbating users’ primary concerns. The resulting general MoSCoW prioritization is summarized in Figure 2.

This study leverages the mention frequency of features and risks to build a MoSCoW prioritization framework specifically for AI psychotherapists. To the best of our knowledge, it is among the first to systematically capture and synthesize patient expectations and perceived risks related to AI psychotherapists and to propose such an implementation hierarchy. Importantly, the findings reflect participants’ expectations and concerns elicited in a hypothetical context, rather than evaluations based on direct interaction with AI psychotherapists. Accordingly, the proposed MoSCoW framework should be interpreted as a foundational, user-informed prioritization schema, intended to inform future system prototyping and empirical validation, rather than as a set of validated or prescriptive design requirements.

Given the applied relevance of participants’ perspectives, the findings of this study offer several design-relevant insights for the design of AI-based tools for depression care. A key strength of this study lies in the breadth of clinical scenarios represented. As shown in Table 2, the participants envisioned AI capabilities spanning the continuum of mental health care, including diagnosis, treatment, consultation, self-management, and companionship. This versatility points to how future systems could be adapted to diverse use cases, including home-based self-assessment tools, previsit screening aids, postvisit follow-up interfaces, or in-session assistants for documentation and care planning.

Furthermore, the MoSCoW prioritization framework can serve as a conceptual reference point for designers and researchers when formulating and testing AI psychotherapist prototypes. In addition to the general prioritization applicable to broad user groups, project managers targeting specific populations with well-defined attributes can draw on the extended MoSCoW framework to tailor system features. These actionable insights offer a valuable and practical foundation for system developers and clinical stakeholders seeking to align design decisions with user-specific needs.

Ultimately, aligning AI functionalities and design with patient expectations is hypothesized to foster trust and engagement, which should be evaluated in future empirical studies. Grounding features and safeguards in user values makes the system more intuitive, credible, and meaningful, which may inform future investigations of adoption and retention.

This study has several limitations. First, the study used cross-sectional online survey data. While it captured user narratives about opportunities, desired features, and perceived risks in AI-based depression treatment, it provided limited insights into the underlying reasons. The design also precluded a follow-up assessment of how satisfaction or user experience might change if an AI psychotherapist met expectations and mitigated concerns. Thus, this study provides a useful foundation but leaves user outcomes unexamined. Future research should extend this work by using qualitative methods to explore the motivations and contextual factors in depth and by using longitudinal or mixed methods approaches to evaluate the impact of these prioritized features and risk mitigations on satisfaction, trust, engagement, and behavioral outcomes. Moreover, although AI capabilities have improved since data collection, AI applications in depression treatment remain limited in real-world deployment and constrained in clinical scope. As such, the expectations and concerns captured in this study primarily reflect enduring values related to trust, safety, emotional support, and human oversight, which are less sensitive to short-term advances in model performance. These perspectives, therefore, remain relevant for informing responsible design and hybrid AI-human care models.

Furthermore, the sample was recruited via MTurk. Although MTurk is a prevalent platform for data collection in medical and psychological research, its participant demographics tend to be younger, more educated, and exhibit greater technological proficiency compared to the general population. This may limit the generalizability of the findings, particularly to older adults or individuals from diverse cultural and socioeconomic backgrounds. Future research should aim to recruit more diverse samples, including individuals from underrepresented or high-stigma groups, to enhance the external validity of the findings.

Third, the analysis used a binary coding scheme (theme present or absent) to systematically categorize open-ended responses. While this approach is easy to quantify, it may not capture the full range of themes within participants’ narratives. Moreover, we included only themes mentioned by at least 5% of participants in the primary analysis to establish a clear priority framework. This ensured that we focused on the most core user perspectives but may have overlooked some rare, yet potentially significant, insights. Future research should use in-depth qualitative methods, such as semistructured interviews or focus groups, to explore a broader range of themes.

This study provides a patient-centered perspective on the use of AI in mental health care, identifying diverse scenarios envisioned by participants for the application of AI psychotherapists and prioritizing expected features and concerns based on user input. The findings highlight which qualities participants believe an effective AI depression psychotherapist should embody and emphasize the need to extend beyond clinical accuracy to address interpersonal support, emotional engagement, accessibility, and privacy protection. Designing systems in alignment with these user-informed priorities is critical to creating AI psychotherapists that are functionally effective, trusted, engaging, and responsive to the needs of users. Future work is needed to evaluate whether systems designed around these priorities lead to improved satisfaction, adherence, or clinical outcomes. As AI technologies continue to advance in mental health care contexts, embedding patient voices and values into their development will be critical to supporting their ethical integrity and guiding future real-world impact.