Writing or speaking: a quantitative survey of communication modes with conversational agents to promote mental health among young adults in Germany

1.1 E-health on the rise

Aschbrenner et al. explored how young people utilize technologies such as smartphones and social media to address mental health concerns. ¹³ Participants were asked about their interest in connecting with other teenagers receiving treatment at a mental health centre through a private social media group. Results showed that 32 % expressed interest, 33 % showed no interest, and 35 % remained unsure. The concept of online peer-to-peer support holds promise in reducing stigma, fostering social connections, and ultimately improving the well-being of individuals with mental illness. However, qualitative interviews conducted as part of the study revealed that many participants expressed concerns about potential risks associated with joining a social media group of peers from a mental health center. ¹³

Lydia Seifert, the executive director of TelefonSeelsorge^® Deutschland e. V. (TSD), a free counselling service for people living in Germany, also fears “that we have already become a society of the lonely”. ¹⁴ Young people have expressed a tendency to rely on self-reliance when confronted with challenges and show reluctance to fully engage in the process of seeking help or accessing support services. ^{8 , 15} Consequently, digital health interventions (DHI), telepsychotherapy, and e-mental health services have emerged as a diverse and rapidly expanding field, encompassing a wide range of technological tools and platforms such as websites, games, computer-assisted programs, mobile apps, virtual reality, and text messages. ^{8 , 16 , 17 , 18 , 19} The field of mobile health (mHealth) is growing, with a notable increase in psychological tools within Human-Computer Interaction (HCI). ^{20 , 21 , 22} The range is broad and serves various purposes, spanning from identifying symptoms and helping individuals manage their mental health to promoting help-seeking behaviours and providing preventive or therapeutic interventions. ^{21 , 23} Potential advantages are enhanced effectiveness, accessibility, anonymity, prompt feedback, cost-effectiveness, applicability in real-life contexts, adherence to therapy, and reach of a broader population. ^{8 , 24}

1.2 Digital conversational agents for mental health

The proliferation of mobile applications addressing mental health issues is on the rise. Existing evidence regarding CAs in psychiatry indicates both potential benefits and associated risks. ^{7 , 13 , 25} CAs are programs designed to communicate with users to provide or collect information. They can be activated by natural language input such text, voice or image and respond with an output. ¹⁵ However, the preferences for these various modalities seem to be unknown. ²⁵

In the mid-1960s, ELIZA was the first prominent chatbot that enabled certain kinds of natural language conversation between human and computers. ²⁶ Josef Weizenbaum chose the psychiatric interview as one example of categorized dyadic natural language communication, where one participant could assume limited knowledge of the real world. The utilization of AI and machine learning (ML) algorithm in CAs allows for more complex dialogue management and greater conversational flexibility and individuality for users. ^{24 , 27}

Aguilera highlighted the significant growth in the use of digital technology in health applications, driven in part by the limitations of traditional one-on-one therapy in meeting the mental health needs of the population. There are calls to develop innovative ways to enhance the quality and reach of effective clinical interventions using tools like mobile phones and smartphones. ²⁸

On the other hand, security risks, data usage and ethical concerns are partly unclear, and the potential of this technology is supposedly not yet exhausted. Because this technology and its potential are quite new, no long-term studies exist. A study by Samantray found that digital cognitive-behaviour therapy (dCBT) intervention had a large positive impact on the overall symptoms of clients enrolled in a mental wellness care program. ²⁹ This is consistent with previous meta-analyses on the potential of digital mental health interventions to improve mental health outcomes by Lewis et al. ³⁰

Although the most popular chatbot, Chat Generative Pre-Trained Transformer (ChatGPT), is not designed to answer questions related to healthcare or mental health, users are willing to use it for self-diagnosis. ³¹ However, it is important to acknowledge that these opportunities come with safety concerns. There is a risk of incorrect self-assessment or therapy that could result in significant harm or even fatal outcomes. ^{27 , 32} Additionally, algorithm ethics raises issues of bias, responsibility, transparency, and explainability of algorithms, as well as validation and evaluation. Biased training data or unfiltered data from the internet can lead to biased results. ³³ Over-reliance with blind faith on ChatGPT may adversely affect patients’ health. Inaccurate content could potentially lead to incorrect self-diagnosis and even greater harm due to hallucinations of digital interlocutors.

The use of clinical best practices, warranty of the accuracy, reliability, and validity of the content generated by these algorithms, and a code of ethics for the responsible use of ChatGPT is strongly recommended. ^{34 , 35} Additionally, emphasizing Fairness, Accountability, and Transparency (FAT*) in the corresponding context is essential. ³⁶ Approaches for prevention and early intervention should be explored for mental health issues to reduce restrictions in social and professional life.

A comprehensive understanding of the potential advantages and limitations of AI-based CAs, such as ChatGPT, in the context of mental health is crucial for optimizing their use and maximizing their potential to support well-being. The mode of communication can play an important role in promoting well-being and avoiding misunderstandings.

1.3 Changing communication behaviour

Over the past decade, there has been a remarkable surge in interpersonal communication facilitated by internet-connected mobile devices. ³⁷ Younger generations, including children, adolescents, and emerging adults, have grown up in an era saturated with digital communication tools. These tools have profoundly impacted the dynamics of interpersonal relationships within families and among friends. ³⁸ Almost everybody owns a smartphone and can be reached quite easily if desired. This trend has been accompanied by a significant shift in communication preferences, with mobile instant messaging applications like WhatsApp gaining prominence and, in some cases, even displace traditional communication channels such as phone calls. More recently, a new mode of communication has gained popularity: voice messaging. ³⁹ This emerging channel allows individuals to exchange recorded voice messages, offering a convenient and alternative method of conveying information and engaging in asynchronous interpersonal communication. According to a 2019 study, ³⁹ 69 % of Germans have engaged in sending voice messages through messaging services such as WhatsApp. The primary advantages of voice messaging include reduced effort compared to typing text messages, minimized chances of misunderstandings compared to written text, and the convenience of time independence compared to phone calls. Users can express themselves more naturally and convey a broader range of emotions. Notably, Haas et al. have indicated that nearly half of voice message interactions occur while individuals are engaged in activities such as walking, driving, or bicycling, highlighting the convenience and flexibility of this medium. ³⁷ Users have reported a heightened sense of closeness and joyfulness when using voice messages to connect with others. However, concerns related to privacy and the inconvenience of reviewing or editing recorded messages have also been raised, underscoring the need to address these issues for a more seamless and secure user experience.

In their in-the-wild study, ³⁷ the researchers found that the relationship between emotional stability and the frequency of voice message usage was not as straightforward, while extroversion appeared to align more closely with its use. To further analyse this phenomenon, the theoretical framework proposed by Nardi et al. can be applied. ⁴⁰ This framework distinguishes between interactional and outeractional message usage. Interactional message usage involves the rapid exchange of information, quick questions, clarifications and staying in touch with friends and family. On the other hand, outeraction refers to communicative processes that go beyond information exchange, where individuals reach out to others in explicitly social ways to foster connection. Human speech can evoke a sense of connection, even in situations where romantic partners or family members are temporarily separated, thus contributing to the concept of outeraction.

In 2022, Nummer gegen Kummer e. V., a free telephone counselling service in Germany for children, young people, and parents, observed a significant surge in chat requests compared to the previous year. This increase demonstrates a notable rise of 18 % in the demand for their chat services. ⁴¹ According to the statistics provided by TSD, there was a notable rise in contacts through mail and chat services between 2019 and 2020. ⁴² Mail contacts experienced a 28 % increase, while chatting witnessed a remarkable surge of over 70 % compared to the year 2019. Notably, these communication channels were predominantly utilized by younger individuals within the age range of 15–40. Furthermore, similar values were also recorded in 2021 compared with the previous year. Engaging in purely text-based communication, whether with a human coach or a supposed chatbot, can present challenges and occasionally result in misunderstandings. ^{43 , 44} As communication preferences shift, it raises the question of which mode would best support well-being.

1.4 Research questions

To address the relevance and research gap, this study presents a quantitative investigation focusing on young adults aged 18 to 27 in Germany and their help-seeking behaviours for common mental health challenges such as depression, anxiety, and general emotional distress. The research examines the current usage and potential of AI-based CAs in promoting mental health, with a key aspect being the evaluation of preferences for spoken or written interaction. By collecting empirical data, the study aims to provide insights into the preferences and needs of this age group, contributing to a deeper understanding of how AI-based CAs can support mental health. The central research questions addressed are:

2.1 Participants

In Germany, young adulthood is defined as the age range between 18 and 27 years. This aligns approximately with the MIT young adult development project. ⁴⁵ According to the German Demography Portal, ⁴⁶ in 2021, there were an estimated 4,683,208 men and 4,298,535 women aged 18 to 27 in Germany, comprising roughly 10.7 % of the total population. This demographic has been exposed to mobile applications from a young age and is receptive to new technologies. Additionally, this life stage is often marked by significant changes and experiences, making it a crucial target group for the study.

To achieve the largest and most diverse sample size, identical questionnaires were distributed through three different URL access points, encouraging voluntary participation via advertising, an online panel, or social media distribution among acquaintances and study groups. To ensure high data quality, only responses meeting the following criteria were considered: participants answered all required questions, were within the specified age range, and either resided in Germany or held German citizenship. Participants from the online panel were also required to pass an attention test.

According to the vendor, paid digital and print advertising on Facebook and Instagram reached a total of 18,048 people. However, only data from two participants recruited through this method met the quality criteria and were included in the analysis. The collection distributed under personal responsibility had an access rate of 651 visitors, with 193 participating in the survey and 163 successfully completing it. Data from 135 participants (84 female) met the quality criteria and were included in the analysis. The final collection from the paid online panel had a view count of 100, with 88 participants and 81 successful completions. Data from 81 participants (36 female) met the criteria for further analysis.

Since the questions in the three surveys were identical and compensation did not influence the research question, the usable results from all three surveys were combined into a single dataset of 216 participants (120 female).

2.2 Materials

This section outlines the materials and methods employed to conduct the study, addressing the research questions and ensuring the quality of all measurements.

2.3 Design and procedure

To enhance the utility and transparency of data collection in this research paper, the Checklist for Reporting the Results of Internet E-Surveys (CHERRIES) is employed. ⁶⁰ To address both research questions, a quantitative, anonymous, web-based online survey was conducted among young adults aged 18–27 living in Germany, adhering to the principles of the Declaration of Helsinki. ⁶¹ Although an ethics application was not submitted prior to conducting the study, the authors are committed to ensuring the responsible and ethical conduct of research. Notes on the survey aim, contact details of the responsible scientists, an estimated duration of 5 min, and guidance for those seeking help and advice were provided in the introduction. Participation was voluntary, anonymous, and could be aborted at any time without giving reasons. Every participant had to confirm a declaration of consent before starting, and it was ensured that answers could not be traced back to any individual. Finally, the contact details of the researcher responsible were provided. Technical measures taken by the tool provider prevent the same person from submitting multiple contributions within the framework of the online tool used.

From July 30, 2023 to August 5, 2023, a pre-study was conducted with 21 volunteers from within the known network to assess usability, technical functionality, comprehension of the questions, and timing. The main survey, available in German and English, was carried out from August 6, 2023, to September 17, 2023, and was made publicly accessible via a specific link. Three different collections of links, each with a unique URL, were created to facilitate the distribution of the online survey through various media and allow for result comparisons. The different links were disseminated through:

1. The social media platforms Facebook and Instagram, WhatsApp, university email groups, the SurveyCircle research platform

2. The online panel provider ClickWorker

3. Paid digital advertising on social media (Facebook and Instagram) and distributed as leaflet in Munich

It’s important to note that only individuals recruited through ClickWorker were compensated in accordance with the statutory minimum wage in Germany for their participation. There was no competitive element or other financial incentive for survey engagement.

In addition to the homepage, the survey contained five pages to guide respondents through a total of 20 questions (overview in Appendix A. Survey Structure). The questions were not randomized, and each page contained four to six questions. Only the mental health questions were optional; all others were mandatory. Respondents could review and change their answers at any time during the process until final confirmation on the closing page.

The collected data were examined for correlations to gain insights into various aspects of mental health, acceptance of CAs, and preferred forms of interaction. Patterns between current mental health disorders and the mode of communication with the respective CA were identified. Internal consistencies, descriptive statistics, and t-tests were calculated using Excel (Microsoft^® Excel^® for Microsoft 365 MSO (version 2308 Build 16.0.16731.20182) 64 bit). To compare the means of two or more groups, t-test or ANOVA were used with a level of significance set at (alpha) p < 0.05.

3.1 Communication behaviour and addressing concerns and discomfort to human

The first question was about current messaging behaviour to get an overview of how often text messages; voice messages and calls are made. Respondents could choose between daily, several times a week, several times a month, less often and never. 89 % of respondents use text messaging daily, 7 % several times a week and 2 % each several times a month and less often. 31 % use voice messages daily and 30 % several times a week, 10 % never use voice messages. Phone calls are made by 26 % of respondents daily and 36 % several times a week, 22 % several times a month, 13 % less often and only 1 % never. There is no difference in the frequency of text messaging between men and women (t(214) = −1.72; p = 0.09), whereas women send voice messages more often (t(214) = −2.71; p = 0.007) and men use phone calls more often as a synchronous communication medium (t(214) = 2.46; p = 0.01) (details in Appendix C. Messaging Behaviour).

To gather more specific information about whether people prefer to write or talk when communicating their feelings and concerns to a person, participants were asked to rate their communication preference. The results show that 48 % of respondents prefer to talk to a person about their feelings and concerns, 29 % favour both talking and writing, 19 % prefer to write, and 4 % would rather neither talk nor write to share their feelings and concerns with a person. Preferences are identical for both genders, with no differences found within the sociodemographic data.

Communication also involves receiving messages and feedback. To this end, respondents were asked to indicate their preference for receiving verbal or written feedback from another person after expressing their feelings and concerns. The results showed that 56 % of respondents favour verbal feedback after expressing their feelings and concerns, 28 % prefer both verbal and written feedback, 15 % prefer written feedback only, and 1 % of respondents would rather not receive either verbal or written feedback. No difference in preference was found between men and women. The importance of immediate feedback on a scale of 1 (unimportant) to 7 (very important) shows a mean of 5.54 (n = 216; SD = 1.32). Women have a mean of 5.73 (SD = 1.26) and men 5.29 (SD = 1.36). From this it can be concluded that immediate feedback is important as soon as concerns and discomfort have been expressed, as well as that this should preferably be done in verbal form.

To find out if young adults prefer to keep their feelings and worries to themselves, they were asked this question using a 5-point Likert scale. 22.2 % prefer to keep their feelings and worries to themselves, 39.4 % prefer to keep them rather, 16.2 % indicate neither yes nor no, 20.8 % say rather no, and 1.4 % say no. Gender specific details are visualised in Figure 1. The distribution strongly indicates that men are more likely to keep their worries to themselves (t(214) = 3.19; p = 0.002), whereas women have a higher tendency to share them.

Figure 1:

Young men (n = 95) and women (n = 120) deal differently with keeping worries to themselves.

3.2 Usage and communication with AI-based conversational agents

Additionally, this study explores whether digital conversation partners, such as ChatGPT or Siri, are or would be used to share worries, fears, or depression. Participants were asked if they had ever used a CA, whether they used it for mental health support, and how likely they might be to do so. Ratings on a 5-point Likert scale provided insights into this, also in combination with mental health status. The next section of questions examines whether communication behavior differs between human-to-human and human-to-CA interactions, focusing on communication mode, mental health status, and potential impact.

3.3 Mental health status

Measurement of GAD-2 revealed that 30.2 % of respondents (n = 65/215, mean = 2.03, SD = 1.47) scored three or higher, indicating a high proportion of individuals with an anxiety disorder. The internal consistency of the two questions was adequate, with a Cronbach’s alpha of 0.713. Distinguishing between female (n = 40/65) and male (n = 24/65) participants showed that 62 % of those with an anxiety disorder were women. Thus, female respondents showed a higher tendency for anxiety (t(213) = 2.212, p = 0.028) than males.

According to K6, 13.9 % of respondents (n = 30/216, mean = 7.88, SD = 4.92) experienced psychological distress. A Cronbach’s alpha value of 0.863 shows a good internal consistency. Among these, 73 % of participants with psychological stress are women (n = 22/30), compared to men (n = 7/30). Female respondents were more likely to experience psychological stress (t(213) = 2.713, p = 0.007).

For the Low SOC measurement, scores ranged from one (high resilience) to seven (low resilience). The overall mean score for all three questions from 216 participants was 10.53 (SD = 3.85). Breaking down the scores for comprehensibility (n = 216; mean = 3.48; SD = 1.46), meaningfulness (n = 216; mean = 3.31; SD = 1.63), and manageability (n = 216; mean = 3.74; SD = 1.70) showed moderate results, with manageability having a slightly lower SOC. The internal consistency for these three questions was reasonable, with a Cronbach’s alpha of 0.724. Comparing female (n = 120; mean = 10.86; SD = 3.75) and male (n = 95; mean = 10.11; SD = 3.94) data shows no significant difference between genders (t(213) = 1.422, p = 0.156).

3.4 Sociodemographic data

During the 42-day period, the survey links were accessed by 912 people. Of these, 287 participated, 248 completed the survey, and 216 data were usable. Participants included 120 females (55.6 %), 95 males (44.0 %) and one non-binary person (0.5 %). The mean age was 23.6 years (SD 2.58), with 64 participants (30 %) aged between 18 and 22 years, and 152 (70 %) aged between 23 and 27 years.

In terms of education, 181 participants (83.8 %) had a high level, 32 (14.8 %) had a medium level, and three (1.4 %) had a low level. Regarding occupational status, 19 (9 %) were currently in education, 102 (47 %) were undergraduate or graduate students, 73 (34 %) were employed, 13 (6 %) were self-employed, four (1 %) were unemployed, and five (2 %) reported other employment types such as working student jobs.

All participants lived in Germany, 207 (96 %) had German citizenship, and eleven (5 %) used the English version of the survey instead of the German version.

4.1 Principal findings

Communication is constantly evolving. Currently, text messages are the primary mode of long-distance communication among young adults using mobile phones. However, immediate responses cannot always be expected, even with voice messages, which offer richer communication due to the potential for more personal conversations based on voice and expression. This richer communication can also help reduce misunderstandings. The participating young males tend to prefer traditional telephone calls more frequently than females. In contrast, women are slightly more inclined to use voice messages, as also shown in a survey by the German Bundesverband Digitale Wirtschaft e. V. ³⁹

Gender differences are evident in how individuals cope with grief and worry. Men tend more to internalize and discuss their concerns with themselves, while women tend to share them with others. When expressing discomfort to another person, both typically prefer verbal communication. This study did not identify significant differences between genders, nor were there variations in the threshold values for GAD-2 and K6 as measured in the study. However, to address RQ2, a more detailed analysis of the results reveals that as anxiety or psychological distress increases, so does the preference for written communication or a combination of written and verbal communication, whether with human or artificial counterparts. Respondents exhibiting these patterns may also display greater insecurity or lower resilience, as indicated by Low SOC scores. These findings suggest a link between SOC and mental health behaviors in young adults, as also observed by Henrique da-Silva-Domingues et al. ⁶²

In contrast to communication with humans, respondents in the random sample showed a preference for written communication or even no communication at all when interacting with CAs. The reasons behind this choice were not explored in this survey. Unlike written communication with people, users can typically expect immediate feedback when engaging with CAs, although the potential for misunderstandings remains. It’s important to note that the current state of AI technology is not yet fully mature, and voices generated by AI can often sound artificial. As this new technology continues to advance and becomes more established, the future path of communication will become clearer.

However, to address RQ1, the results of the online survey indicate that CAs are currently used by most young adults (82 %) and 17 % used them to consult for help or advice when experiencing mental health challenges. However, it should be emphasized that the tendency to use this new technology as a communication partner appears to be significantly higher among people with pre-existing anxieties, mental health problems or low resilience.

Presently, the available evidence suggests that written communication is the preferred mode when interacting with a CA about grief. As technology continues to advance the preferred mode of communication will require further exploration and examination of using AI-based CAs for mental health purposes.

4.2 Comparison with previous work

Compared to previous studies, the findings from this survey confirm that asynchronous voice messages are increasingly common in Germany. As of 2019, 69 % of 1,051 Germans had sent voice messages through messaging services like WhatsApp. ³⁸ In this current study, over 90 % of young adults in Germany are already using voice messages. Like trends noted in the survey of 2019, ³⁹ women in our study also utilize voice messages more frequently than men (t(214) = −2.71; p = 0.0072). According to the TSD 2021 statistics, ⁴² the slight increase in preference for writing about worries can also be attributed to the rise in written traffic related to sorrow. The concept of outeractional message usage to evoke a sense of connection seems to hold true when communicating with people, but this does not appear to apply to digital counterparts. ⁴⁰ In their empirical study, Zhu et al. utilized the Theory of Consumption Values (TCV) to observe satisfaction and continuance intention towards mHealth chatbots during the COVID-19 pandemic in China. Contrary to their expectations, voice interaction between citizens and the chatbot Xiaolv did not play a significant role. ⁶³

Regarding keeping worries and problems to themselves, our survey aligns with earlier research. Approximately 22 % prefer to keep their concerns to themselves, and 39 % lean towards self-reliance in handling their worries. ^{8 , 13 , 15}

The prevalence of mental health problems among children and young people is a significant concern, with reported rates ranging from 10 % to 20 % worldwide, and these rates are still on the rise. ^{1 , 2 , 64} Large-scale epidemiological studies have shown that women may have a two to three-fold higher incidence of GAD compared to men. ⁶⁵ In this study, most respondents with increased GAD-2 and K6 scores were female. However, no significant difference was observed between female and male participants (t(213) = 1.422; p = 0.156) regarding measured Low SOC.

Furthermore, in their study, Abd-alrazaq et al. assessed 41 chatbots for mental health, most of which provided written text input and mixed media (written, spoken, and visual) as output. ⁶⁴ The reasons for this decision were not provided, but it can be assumed that it is due to simpler technical implementation. Additionally, the study mentioned that 92.5 % of the chatbots relied on decision trees to generate their responses for therapy, training, and screening. The authors note that these chatbots have less AI integration compared to chatbots in other fields, such as customer service, possibly because they are less error-prone and more secure.

In a study by Shahsavar et al., 78.4 % of the 475 respondents expressed a willingness to use ChatGPT for self-diagnosis in health-related topics. ³¹ Although this study did not specifically focus on mental health-related topics, it is worth noting that, in contrast, approximately 72 % of the 216 German participants indicated a disinclination to use a CA for mental health-related topics. In their online survey conducted in 2021 with a total of 100 mental health professionals from Northern Ireland, Ireland, Scotland, Sweden, and Finland, Sweeney et al. found that as years of experience increased, there was a corresponding increase in the belief that chatbots in the mHealth environment could help clients better manage their own mental health, improve quality, access, and timeliness of mental healthcare, and assist mental healthcare workers and professionals in their daily occupational roles. Additionally, a total of 80 % of the professional respondents indicated that they would be very likely or somewhat likely to prescribe a chatbot for mental health in the next five years. ⁶⁶ The results of the previous conducted study in 2019 with a total of 100 practicing physicians across the United States from Palanica et al. were less optimistic. ⁶⁷ A total of 44 % of the respondents believed that chatbots for mental health would become very likely or somewhat likely in the next five years. Furthermore, the majority (70 %) of the physicians expressed concerns about risks associated with chatbots for mental health. They worry that patients may abuse the use of chatbots and self-diagnose too often, understand the diagnose inaccurately, and that the digital assistant cannot provide detailed clarification on patient assessment. It becomes evident that experts’ perceptions and views vary based on culture, time, and knowledge of innovations and may differ from the view of their patients.

4.3 Limitations

It’s worth noting that current technology, including platforms like ChatGPT, is still relatively young and in an immature stage of development. A significant limitation of this study is the lack of differentiation between AI-based CA technologies. While chatbots like ChatGPT utilize LLMs to provide more comprehensive and personalized responses, voice assistants such as Siri, Alexa, and Cortana do not. These voice assistants primarily rely on predefined commands and lack the advanced conversational capabilities of LLM-based CAs. Despite this crucial technological difference, voice assistants were included as an example in this study to capture a broader range of user experiences with AI-based CAs. We recognize that users may interact with different types of CAs in their daily lives, and this survey was not about the technology behind it, but rather the fact that there is technical intelligence at the other end instead of a human. In addition, LLM-based CAs did not offer natural and s communication via voice at that time. However, we acknowledge that this approach may have introduced variability in responses, as participants’ experiences with LLM-based CAs and non-LLM-based voice assistants are inherently different. Like our follow up study in 2024, ⁶⁸ future research should focus specifically on LLM-based CAs to provide more targeted insights into their potential for mental health support. In addition, the advantages and disadvantages of asking a digital counterpart for help with grief should be identified.

An additional influence could have happened during the survey introduction. There, the research question regarding the ability of CAs to provide written or verbal support for grief was presented relatively clearly. However, this clarity may have deterred individuals who have a fear or aversion to chatbots or similar technologies. This also might be a reason for the small size of participants, successfully finished the survey. Cause of this, our findings must be interpreted with caution. Achieving the right balance between transparency and impartiality is a challenge.

Although an ethical application was not made in advance, we are convinced that this is safe and therefore responsible research for the respondents involved. For this we took several steps to ensure the responsible and ethical conduct of our research. These measures included:

1. Informed Consent: All participants were provided with detailed information about the study’s aims, the voluntary nature of participation, and the right to withdraw at any time without consequences. Participants had to confirm their consent before proceeding with the survey.

2. Anonymity: The survey was designed to be completely anonymous, with no personally identifiable information collected. This ensured that responses could not be traced back to individual participants.

3. Support Resources: We provided contact information for mental health support services in case participants felt distressed during or after the survey.

4. Data Security: All data were stored securely, and access was limited to the research team to protect participant confidentiality.

We are committed to upholding the highest ethical standards in our research and believe that these measures ensured the responsible conduct of this study. Nevertheless, an ethics application was successfully submitted for the subsequent study in 2024 before it was carried out. ⁶⁸

A potential limitation of our study is the reliance on self-reported data, particularly from participants recruited through paid advertising channels – 100 of which 81 met the quality criteria. To mitigate the risk of dishonest or invented responses, we implemented several quality control measures:

1. Attention Checks: Participants from the online panel were required to pass an attention test to ensure they were engaged and attentive during the survey.

2. Consistency Checks: We examined the internal consistency of responses using established scales (K6 and GAD-2) and found adequate to good reliability (Cronbach’s alpha of 0.863 for K6 and 0.713 for GAD-2).

3. Data Cleaning: We excluded responses that did not meet the quality criteria, such as incomplete or inconsistent answers.

While these measures help ensure the reliability of our data, we acknowledge that the risk of dishonest responses cannot be eliminated. Future research could benefit from additional validation methods, such as cross-referencing self-reported data with clinical assessments. Despite these limitations, the internal consistency of our results and the alignment with existing literature suggest that our findings are valuable and provide meaningful insights into the use of AI-based CAs for mental health support.

4.4 Future research

AI against worries? Future research should shed more light on the reasons for the relatively high reluctance to use CAs for sorrow compared to health-related purposes. Furthermore, extending the study to different age groups may lead to further insights into this phenomenon. It would be interesting to compare whether AI-based CAs can provide added value over traditional psychotherapeutic methods or pose a real danger as things stand. The potential of CAs as supportive therapists needs to be further explored, considering aspects such as safety, acceptability, control, quality, and cost. Furthermore, ethical considerations in the application of AI in healthcare highlight the need to address ethical issues such as transparency, accountability, bias, privacy, and responsibility when using AI in healthcare. ^{69 , 70 , 71} These ethical concerns must be carefully considered to ensure the safe and responsible integration of AI into healthcare practices for both patients and specialists.

Furthermore, a comparison to teletherapy, which promises already increased access, quality, and cost reduction, should be evaluated from an ethical and effectivity perspective. Especially socioeconomic and educational barriers may limit technology use and therefore access to healthcare resources. ⁷² By collaborating with healthcare workers, scholars can apply their qualitative and quantitative skills to address a variety of ethical issues in healthcare.

Beside transparency, accountability, and trust, as well as values of fairness, justice, and equality are necessary to integrate AI not only in medical topics but also into daily practice. In their systematic review of published empirical studies of medical AI ethics, Tang et al. aimed to map the main approaches, findings and limitations of existing scholarship to inform future practice considerations. During this they figured out that generally clinicians report more ethical concerns related to the use of AI in health than patients. ⁷³ To address questions of legal responsibility, Müller et al. have developed 10 easy understandable commandments as practical guideline, which were agreed upon in an online survey of 121 computer experts (47 %) and medical doctors (33 %). Among other things, they emphasize the need for recognizability regarding which part of the communication is performed by a CA and that humans should never be deceived by these systems. ⁷⁴ Therefore, future research should also consider collecting opinions, information, and considerations from both patients and specialists. This can help ensure a comprehensive understanding of the challenges and inform the development of appropriate guidelines and frameworks.

Certainly, only long-term studies can provide insight into whether an AI can really assist with mental problems or other related issues, such as relapse prevention for alcohol, drug or other addictions. ⁷⁵ Furthermore, evaluation on the effectivity for promoting resilience and well-being with the use of this technology combined with SOC could be considered. ^{76 , 77}

However, digital communication partners must not and can never replace friends, family, therapists, coaches, or other trusted persons and specialists. Nonetheless, perhaps they can serve as valuable writing or speaking partners for those people who are struggling with worries and individually address the problems to prepare ways out of the negative spiral of thoughts or bridge the gap between therapy slots. Various possibilities to integrate AI-based CAs in the patient-expert interaction should be rated for acceptance by both involved groups to gain insights about possibilities to improve mental health. Although, in a scholarly publication from 2015, both professionals and drug-dependent patients confirmed the high relevance of information system – like teletherapy – research, related to relapse prevention. ⁷⁵ Potentially, the Action Research (AR) approach can help to move from ‘pilotitis’ to sustainable solutions, thereby reducing the current limited resources for therapy. ⁷⁸ Nevertheless, the responsibility for an AI decision, action, or communicative process must be explainable and transparent, and it must be taken by a competent physical or legal person. ⁷⁴ Possibilities of combining artificial intelligence with human intelligence should be considered in future research instead of providing standalone AI mental health applications.

Particularly worrying is the fact that young adult men tend to keep their worries to themselves (75 %) compared to young women (52 %). This is most relevant considering that the suicide rate among young men aged 15–29 years in Germany in 2021 is significantly higher (77 %, n = 609/794) than among women. ¹¹ The lack of access to mental health services, not only in low-income countries, may contribute to suicidal behaviour, resulting in increased mortality rates. ⁶⁴ Better education on mental health topics and social acceptance are needed, especially among young men who find it harder to share their concerns and seek help. This cannot be achieved through technology alone, but possible approaches combining human and artificial intelligence can help to widen bottlenecks and provide preventative or remedial support.