We conducted a qualitative study to examine how to design a chatbot to treat depression in youth. The study included a questionnaire, a semistructured interview, and a concurrent think-aloud session with a chatbot prototype. We chose interviews because they allowed us to gather rich, detailed data on participants’ problems, coping strategies, attitudes and expectations, and chatbot design preferences.

Participants were eligible if they were between 14 and 17 years of age, owned a smartphone, and met diagnostic criteria for a current or remitted depressive episode. Participants with suicidal ideation or psychotic symptoms were excluded. Participants were recruited through a resident child and adolescent psychiatrist and the University’s newsletter between June and August 2021.

The target sample size was calculated based on the goal of data saturation [31], considering age group (14‐15 and 16‐17 y), depression status (remitted and acutely depressed), and gender (women and men), assuming homogeneity within each subgroup. Given that data saturation can be achieved at 6 cases per homogeneous group [31], the study aimed to recruit 48 participants. The sample was recruited through convenience sampling but the sampling was constrained by the predetermined recruitment period and available recruitment channels. The final sample consisted of 14 participants (12 women and 2 nonbinary).

After participants and caregivers provided informed consent, we assessed their eligibility to participate in the study. Next, enrolled participants answered a questionnaire on sociodemographic information and mental health on a laptop. We conducted semistructured interviews to explore participants’ experiences with depression, their adaptive coping strategies, attitudes and expectations toward chatbots for depression, and their design preferences. Finally, the participants interacted with a prototype chatbot using the concurrent think-aloud method.

The interview and think-aloud session were recorded using an audio recorder (Zoom Q2n [Zoom Corporation]) and were transcribed verbatim. Furthermore, 2 coders analyzed the semistructured interviews and concurrent think-aloud sessions using qualitative content analyses by Mayring and Fenzl [44] and Mayring [45] with the QCAmap software (Fenzl and Mayring) [46]. Similar to Zehetmair et al [47], we chose an inductive coding approach to achieve the most unbiased and thorough description of the data, which we deemed important for the exploratory nature of our study [44,45]. Both coders (Lilli Feline Metelmann and LB) coded the material independently, discussed their coding results, and jointly developed a category system. When disagreements between the coders could not be resolved, a third coder (FOK) was consulted. We chose this process because intercoder comparison is not feasible with inductive category development [44,45]. In an inductive approach, categories are created bottom-up from the material rather than selected before data analysis, as in deductive approaches [44,45]. We report the frequencies of the categories and codes to increase transparency and demonstrate the saliency of the codes in the data. The category system is provided in Multimedia Appendix 2. The questionnaires were analyzed using R (version 4.1.3; R Core Team). For PHQ-8, SCARED, and the potential capabilities questionnaire, we calculated descriptive statistics (mean and SD). For the potential capabilities, we also generated frequency distributions to visualize the response patterns.

The Institutional Review Board of the Medical Faculty of the University of Tübingen approved the study (project ID 595/2021B01). The participants and caregivers provided written informed consent (Multimedia Appendix 3) before screening for eligibility and enrollment. All participant data were deidentified before analysis by removing personal identifiers such as names, addresses, and contact information, and replacing them with unique study codes to ensure participant confidentiality. After the study, participants received a reimbursement of 30€ (equivalent to US $33) for taking part in the study. The study lasted between 1 and 2 hours.

In total, 14 youth between 14 and 17 years old (mean 16.1, SD 1.14 years) participated in the study. Furthermore, 12 participants identified as women and 2 as nonbinary. All participants reported a current or remitted depressive disorder. In addition, 10 participants had a PHQ-8 score of 10 or higher, indicating acute depressive symptoms [48]. Of the total, 12 participants had a SCARED score of 25 or higher, indicating current symptoms of anxiety [49]. Furthermore, 9 participants showed symptoms of both depression and anxiety. Out of 14, 9 (64%) participants were currently receiving psychotherapy, whereas 8 (57%) had received psychotherapy in the past. A comprehensive overview of the participant characteristics is illustrated in Table 2.

All participants (N=14, 100%) experienced problems categorized as depressive symptoms, including lack of motivation and energy, depressed mood, and self-doubt. Furthermore, 2 participants explained their main problems:

Some participants (9/14, 64%) reported comorbidities, such as panic attacks, anxiety, and excessive alcohol consumption, while 4 participants (28%) reported physical problems, such as tension, headaches, or stomach pain. All 14 participants (100%) also reported interpersonal problems, particularly withdrawal from social relationships, stress caused by parents, and problems with friends. Furthermore, 11 participants (78%) expressed concerns about school or their future. They shared negative experiences at school and experienced pressure to plan their future, which were reflected by statements like:

In total, 6 participants (42%) reported experiencing stigma associated with depression. They mentioned that adults stigmatized or trivialized their problems or feared negative reactions when sharing their problems. One participant was concerned that disclosing that they are seeing a psychotherapist would lead to others perceiving them as “sick in the head.” Most participants reported encountering difficulties with mental health care. Furthermore, 12 individuals (85%) reported experiencing treatment barriers, including attitudinal and structural barriers. Attitudinal barriers included fear of being judged, self-assessment that problems are not severe enough to deserve support, and negative experiences during psychotherapy (n=9, 64%). Participants reported ineffective techniques, difficulty with disclosing personal information, pressure to perform, trust violations, and condescending treatment from therapists. Structural barriers reported are long waiting lists or being deemed not sufficiently severe by healthcare providers or parents to qualify for professional support.

All 14 participants (100%) identified social support as an adaptive coping strategy. The majority received social support from friends, family members, partners, teachers, or people they interacted with online. In total, 13 participants (93%) engaged in activities to distract themselves or to have a positive experience such as spending time outdoors or consuming media. A total of 9 participants (64%) reported receiving professional treatment and 9 (64%) used cognitive strategies, including positive self-talk, self-reflection, and reducing rumination. Furthermore, 6 participants (42%) implemented a daily structure including forming habits, getting up early, scheduling positive activities, and setting goals. In addition, 7 (50%) prioritized their needs as a coping mechanism, such as intentionally allocating time to self-care or avoiding stressful social situations. Of the total, 3 individuals (21%) engaged in mindfulness practice, such as breathing techniques and meditation, and 2 individuals (14%) sought online information regarding their depressive symptoms, interrelated problems and further information.

Most participants had positive attitudes toward and expectations of chatbots to treat depression. In total, 12 participants (85%) stated that they would be less anxious about using a chatbot than seeing a human therapist. They pointed out that using a chatbot would be a suitable option for discussing sensitive topics that they would not share with others, primarily because they would not fear negative reactions. In addition, texting was considered less intimidating than speaking with a therapist. Furthermore, 11 participants (79%) pointed out the unlimited capacity and flexibility of chatbots:

In addition, participants indicated that a chatbot is more accessible and requires less effort than seeing a human therapist. A total of 8 participants (57%) expressed confidence in the effectiveness of chatbots for the treatment of depression. They indicated that such a chatbot would be capable of addressing a wide range of issues, and welcomed the idea of having a helpful everyday chatbot. All participants demonstrated keen interest in using a chatbot to treat depression either because they expected it to be effective or because they were curious about using it. One participant stated that a daily usable chatbot could alleviate feelings of loneliness. Several others noted that a chatbot’s personal and human-like nature would increase the motivation to use it, particularly compared with other less interactive DMHIs.

However, participants also reported several concerns regarding the use of chatbots to treat depression. Some participants (n=9, 64%) were skeptical about the chatbot’s intelligence and natural language capabilities. They expressed concern that the chatbot would not be able to address individual, diverse, or unusual problems effectively, and feared being disappointed if the chatbot was unable to do so. Participants were particularly worried about inappropriate answers to emotional and intimate topics or inappropriate advice for their problems. Some participants (n=9, 64%) were concerned that a conversation with the chatbot would not feel like a conversation with a human therapist, due to a potential lack of emotional intelligence or overly robotic or analytical responses. However, chatbots that appeared too human were believed to be uncanny. In total, 2 participants (14%) were skeptical about sufficient data security and privacy. Finally, 2 participants (14%) worried that their symptoms of depression, such as little motivation or forgetfulness, would result in low engagement with the chatbot, thereby preventing them from effectively using it. They also pointed out that the lack of social pressure when using a chatbot, compared to seeing a human therapist, could contribute to low engagement, which might not be solvable by the chatbot.

This study explored the problems youth face with depression, their adaptive coping strategies, and their attitudes and design preferences for chatbots to treat depression. Our findings indicate that youth experience diverse problems beyond core depressive symptoms, employ a range of coping strategies, hold predominantly positive attitudes toward chatbots for depression, and express diverse, sometimes contradictory, design preferences. In the following section, we discuss the results according to the four research questions guiding this study: (1) the problems youth with depression face, (2) their adaptive coping strategies, (3) their attitudes and expectations toward chatbots to treat depression, and (4) their design preferences for chatbots to treat depression. As presented in Table 3, our findings across all research questions offer specific guidance for designing chatbots tailored specifically to youth experiencing depression.

First, participants reported various problems with depression beyond the symptoms described in the DSM-5 [50], including mental health and somatic comorbidities, interpersonal issues, and concerns about school and the future. These findings align with previous research showing that youth with depression face an increased risk for anxiety disorders, substance abuse, and physical health issues [51], as well as social withdrawal, lack of friendship [52], diminished academic achievement, school dislike, and pessimism [52]. To address these problems, engaging and effective chatbots should (1) assess both core depressive symptoms and related problems repeatedly to inform therapeutic decisions and strengthen the therapeutic relationship by understanding the users’ individual challenges, and (2) curate a diverse content database to address the wide range of problems, aligning with Li et al [53]. Second, participants reported significant barriers to receiving professional treatment, including poor mental health literacy (eg, awareness of symptom severity), attitudinal (eg, perceived stigma), and structural barriers (eg, waiting times). These findings align with research showing that youth struggle to recognize symptoms [54], hesitate to disclose personal information to professionals due to fear of judgement or not being taken seriously [7], and question the effectiveness of professional treatment [7], a concern supported by evidence that 60% of youth with depression do not benefit from psychological treatments [55]. To address these barriers, chatbots should (1) encourage emotional disclosure by providing a nonjudgmental conversational environment, something DMHIs cannot fully achieve due to their static, nonconversational design, and (2) clearly communicate their scope and limitations to set realistic user expectations and direct to external support when needed.

Participants reported various adaptive coping strategies, including seeking social support, establishing positive activities, using cognitive strategies, and receiving professional treatment. Many also emphasized the importance of focusing on their needs and establishing a daily structure. Most of these strategies fall into the CBT strategies behavioral activation, cognitive restructuring, and problem solving [56], which are effective and first line treatments for depression in youth [57].

The prevalence of CBT-based coping strategies may be due to the previous psychotherapy experience of most participants. Similarly, among youth with nonsuicidal self-injury, those with experience in dialectical behavioral therapy have suggested incorporating it into a DMHI designed for them [58], showing how previous therapy experience influences preferences for DMHIs. The use of CBT-aligned adaptive coping strategies is important, as research shows that youth with depression use adaptive cognitive strategies less frequently than healthy controls [59] and that applying these strategies is associated with fewer depressive symptoms [60]. Chatbots for youth with depression should incorporate these strategies to create engaging and effective content that aligns with users’ existing coping strategies and effective therapeutic techniques. For example, the chatbot could facilitate social support by helping users identify supportive individuals and suggesting personalized ways to reach out, such as a video call with a friend who is good at cheering them up or a message to a family member who provides good advice. Chatbots can even draft customizable messages to support users who struggle to reach out.

Participants predominantly held positive attitudes toward chatbots to treat depression. Many participants reported feeling less anxious about using a chatbot than seeing a human psychotherapist, especially when discussing sensitive topics. This finding extends research on post-traumatic stress disorder, showing that digital agents lead to greater disclosure of sensitive or stigmatized information [61,62], and supports evidence that youth value the privacy and anonymity offered by DMHIs [63]. Given these findings, chatbots could help users become comfortable with sharing sensitive information and practicing therapeutic conversations, with potential benefits for subsequent sessions with psychotherapists. However, some participants were skeptical about the chatbots’ ability to address individual problems and provide appropriate advice on sensitive topics, which is supported by evidence that commercial chatbots for adults with depression often fail to match user inputs, understand messages, or respond appropriately [64]. Although privacy and data security are frequently cited as primary risks of DMHIs for young people [65], few participants raised these concerns. Instead, they viewed chatbots as privacy-enhancing compared with human therapists. Nevertheless, robust privacy and data security remain essential, and it is unclear whether participants were genuinely unconcerned or simply assumed chatbots would have strong protections in place.

Our study revealed important insights into design preferences for chatbots for youth with depression. First, participants expressed diverse preferences for dialogue topics, with three areas standing out: (1) chatting about daily life, (2) discussing urgent problems and receiving advice, and (3) working through therapy exercises that address specific problems. Interestingly, while youth with nonsuicidal self-injury [58], emotional problems [66], and from the general population [63] included being connected to others facing similar challenges as a key feature of their DMHI, our participants ranked such social connection as their lowest priority, highlighting how design preferences may vary across mental health conditions. In summary, a chatbot to treat depression in youth should reflect these preferences. However, available applications have not implemented chats about daily life, and relied on rule-based approaches for problem-solving [14,15], likely due to insufficient conversational capabilities. Structured therapy techniques, such as behavioral activation or cognitive restructuring, are standard components of DMHIs [38] and chatbots [67]. However, their current implementation remains predominantly static [68], without personalized advice or feedback during therapeutic exercises. Large language models (LLMs) such as OpenAI’s GPT or Anthropic’s Claude promise to overcome current limitations of DMHIs and chatbots. LLMs enable chatbots to better address our participants’ design preferences, such as natural conversations about daily life and advice on urgent problems. In addition, LLMs can enhance static implementations of therapeutic exercises by offering personalized guidance and feedback. In behavioral activation, LLMs can personalize the explanation of the relationship between behavior and feelings and evaluate the proposed activity plan for feasibility and therapeutic appropriateness. In cognitive restructuring, Sharma et al [69] demonstrated that LLMs can assist users in identifying thinking traps and generating reframed thoughts. Although these enhanced capabilities have the potential to improve understanding and skill development, key challenges remain. LLMs can generate false or harmful messages [70], posing risks to vulnerable users. Future research needs to explore how capitalize on their advanced conversational abilities while ensuring therapeutic quality and safety [71].

Second, our findings show that engaging and effective chatbots for youth with depression need personalization beyond the conversational capabilities of LLMs. Drawing on the framework by Cohen et al [72] for psychotherapy personalization, effective chatbots require personalization in content selection and interaction style. Content selection ensures that chatbots address the diverse problems and dialogue topics our participants reported, aligning with findings from Li et al [53] and Ludlow et al [66]. A personalized interaction style accommodates the preferences for different social roles and language use. While the benefits of personalizing the chatbot persona have been shown [73,74]. LLMs enable the personalization of interaction styles via prompt instructions, reducing the need to craft different responses for each style manually [75]. Despite the clear need for personalization, participants disagreed on who should control it [76], preventing a clear design recommendation. While some preferred user-led personalization, aligning with Kenny et al [63], others favored chatbot-led personalization or a hybrid approach. Given this ambiguity and the lack of empirical evidence [77], future research is needed to guide chatbot personalization.

Our study has 2 main limitations. First, our sample size was limited, and the participants were predominantly women. Although depression is more prevalent in women [78-80], the absence of men limits the generalizability of our findings across genders. The gender imbalance resulted from convenience sampling through a single youth psychiatrist. One man was informed about the study but declined to participate. While our study included 2 nonbinary participants, an underrepresented demographic in research, recruiting men would have required targeted efforts beyond our study’s constraints. In addition, our sample consisted entirely of participants who had actively sought mental health support, most of whom had previous psychotherapy experience. As a result, perspectives from those who avoid professional help or resist treatment may be underrepresented. Due to these sampling limitations, we likely did not achieve full data saturation. Although qualitative research can reach saturation with as few as twelve interviews [31], our sample composition suggests that some perspectives may have been missed. However, our findings on youth depression and related problems replicated findings from studies with much larger sample sizes [81], indicating that our results are comprehensive. Future studies should recruit a more diverse sample, specifically including men, individuals reluctant to seek professional treatment, and those without previous psychotherapy experience, to validate and extend our findings.

Second, our study captured attitudes, preferences, and hypothetical usage scenarios rather than actual usage behavior. While participants interacted with a prototype during the think-aloud sessions, this brief, controlled interaction may not reflect real-world interactions. The gap between stated preferences and actual behavior is well-documented [82], and our study design does not allow us to determine whether the design preferences of the participants would translate into long-term engagement or effective use. Future research should follow the next steps of an iterative human-centered design process [83,84] with functional prototypes to assess whether the identified design preferences lead to actual user engagement and therapeutic benefits. Despite these limitations, our study provides in-depth findings on the design of chatbots to treat depression in youth, highlighting the value of qualitative research in the iterative development of DMHIs [83,84].

Our study provides valuable insights into the problems and coping strategies of youth with depression, and their attitudes, expectations, and design preferences for a chatbot to treat depression in youth. We found complex user needs, predominantly positive attitudes toward chatbots, and various design preferences, including the need for diverse dialogue topics and personalization. Our findings led to concrete design recommendations that lay a crucial foundation for developing engaging and effective chatbots to treat depression in youth.

Despite these contributions, several research directions remain. First, future studies should validate and extend these findings with larger, more diverse samples to ensure broader representation of youth with depression. Second, examining actual usage patterns and long-term engagement with functional chatbot prototypes will help assess the effectiveness of our design recommendations. Third, investigating the responsible integration of LLMs is important, including the development of robust safeguards and evaluating dialogue quality and therapeutic outcomes. Finally, researchers must design and evaluate effective personalization features, particularly regarding whether users, the chatbot or both should control personalization, while balancing implementation efforts and impact. By addressing these research directions, we can further improve chatbots to treat depression in youth and ultimately contribute to more accessible, engaging, and effective mental health support for this vulnerable population.