This study used an exploratory qualitative design based on an online focus group, 4 semistructured interviews, and a brief questionnaire to gather contextual information. The focus group primarily generated interactive discussions around shared experiences and general expectations, while the individual interviews provided more nuanced and personal reflections, particularly regarding doubts about empathy, trust, and personal willingness to use CAs. The questionnaire was administered prior to the focus group and interviews to obtain participants’ independent views on CAs, uninfluenced by group dynamics, and to serve as a basis for more in-depth exploration during the discussions. The questionnaire data were used descriptively and for contextual triangulation of the qualitative findings, rather than for inferential analysis. Given the small sample size, the quantitative results (eg, mean scores) should be interpreted with caution and are intended to provide indicative insights only.

The qualitative and quantitative components of the study included 12 participants recruited between March 2025 and April 2025 from a Dutch online self-help program for alcohol use and the client advisory board of a large mental health care institution (Arkin, the Netherlands). The self-help program consists of 6 evidence-based modules based on both motivational interviewing and cognitive behavioral therapy. Other intervention features include a support forum, a personal diary, and encouragement for users to log their daily substance use and craving in accordance with cognitive behavioral therapy principles.

Inclusion criteria were participation in Jellinek’s digital self-help alcohol intervention, familiarity with the program, proficiency in Dutch, and willingness to participate in an online group discussion or interview. Due to the anonymous nature of the self-help program, participants’ contact details were not initially available. However, individuals who expressed interest in participating in the study by responding to a forum message or clicking a banner within the self-help program were asked to provide their phone number or email address, enabling the researchers to provide more information and formally invite them to the focus group. A total of 24 participants provided contact details, all of whom were contacted multiple times to schedule participation. Of these, 2 (16.7%) participants did not speak Dutch, 6 (50%) were unable or unwilling to join an online group discussion or interview, and 7 (58.3%) did not respond to phone calls or emails. Ultimately, 8 (66.7%) participants from the self-help community joined the focus group. In addition to the focus group, 4 semistructured interviews were conducted. Three interviews were held with members of Arkin’s client advisory board, and 1 was conducted with another self-help user who did not want to join a group discussion but was willing to participate in an interview. Because of the limited number of participants (n=12, 100%), formal data saturation could not be established; however, recurring patterns and overlapping themes emerged across the focus group and interviews. This suggests a degree of thematic consistency, even if full saturation was not attained [16].

All participants completed a questionnaire prior to the focus group and interviews and received a €45 (€1=US $1.14 as of June 23, 2026) gift voucher as compensation.

Data were collected between March 2025 and April 2025 through 1 online focus group session (n=8, 66.7%) and 4 online semistructured interviews (n=4, 33.3%). The focus group followed a semistructured discussion format exploring attitudes toward the integration of a CA. The focus group lasted approximately 3 hours, although it was planned for 2.5 hours. The same thematic areas were addressed in the individual interviews, allowing deeper reflection and more personal nuance. The interviews lasted approximately 1.5 hours. The focus group was moderated by an independent researcher, a female psychotherapist (MSc) with prior experience in moderating focus groups. She had no relationship with any participants. She was independent because she did not have assumptions or personal interests related to the study, which was important to minimize the risk of leading questions or undue influence. The primary researcher (first author) was present during the focus group in a nonparticipatory role. She introduced herself as the researcher of this project, explained why an independent person moderated the focus group, and explained the rationale for conducting this research. The researcher was also present to address potential procedural questions from participants and to become familiar with the data. This allowed the researcher to gain a contextual understanding of the discussions, including tone and group dynamics, which may not be fully captured in written transcripts. The researcher also made field notes during the focus group and interviews.

Prior to the focus group and interviews, participants were asked to complete a short, researcher-developed questionnaire about their prior experiences with CAs and their perspectives on how these could be used in the self-help program. All participants were asked for what purposes they would use a CA after being given several examples of possible functions. They were invited to write down as many desired functions as they wished. The questionnaire also included several demographic items (Table 1). The questionnaire was pretested by researchers involved in the study prior to data collection to ensure that it functioned properly and that the wording was comprehensible for individuals with varying levels of language proficiency. Several questionnaire items were also discussed during the focus group and interviews to gain deeper insight into participants’ responses and to stimulate group-level discussion (Checklist 1).

The qualitative data were analyzed using inductive thematic analysis [17,18]. This method was selected for its suitability for generating in-depth, exploratory insights into participants’ experiences, perceptions, and attitudes regarding complex and potentially ambivalent topics such as AI in mental health care. In addition, the inductive approach enables researchers to maintain an open perspective toward the data. Alternative qualitative approaches, including grounded theory, content analysis, and framework analysis, were considered in consultation with a qualitative data supervision group before thematic analysis was selected as the most appropriate method, given the exploratory aim of the study.

The focus group and interviews were audio-recorded and transcribed verbatim. Transcripts were anonymized prior to analysis and were not returned to participants. The focus group and interviews were conducted in Dutch, and the quotes were translated into English and checked by a native speaker for accuracy.

In the first phase, the primary researcher (first author) familiarized herself with the data by repeatedly reading the transcripts of the focus group and interviews. An initial phase of open, inductive coding was then conducted, in which segments of the data relevant to the research question were identified. Given the nature of focus group and interview data, nonrelevant conversational content was not coded. Codes were generated at a semantic level, staying close to participants’ own words and meanings without imposing predefined categories.

After completion of the initial coding, codes were systematically compared and grouped based on similarities and patterns, resulting in the development of candidate themes. These themes were iteratively reviewed and refined in relation to both the coded extracts and the full dataset. The coding process and theme development were discussed with a second independent researcher (M Begeman). Discrepancies in interpretation were resolved through discussion until consensus was reached. A documentation process was maintained by documenting coding decisions, code definitions, and changes to the coding structure, supported by analytic memos and versioning within MAXQDA 2022 (VERBI Software GmbH). Participants did not provide feedback on the findings.

The questionnaire data were analyzed descriptively using means and SDs to provide contextual information about participants’ familiarity with and attitudes toward CAs.

All participants provided informed consent and received €45 as compensation for participating. The Medical Research Ethics Committees United determined that the study was not subject to the Dutch Medical Research Involving Human Subjects Act (Dutch: wet medisch-wetenschappelijk onderzoek met mensen) and granted approval (2024.1147). To minimize potential bias, the focus group was moderated by an independent facilitator who was not involved in the research project. The primary researcher has a background in clinical psychology and addiction research, with an interest in digital mental health. She was not involved in the development of the self-help program under study but is involved in planned future adaptations and improvements to the program. To minimize potential bias, the researcher adopted a nondirective role during the focus group, avoiding leading questions and allowing participants to freely express their perspectives.

Participant characteristics are presented in Table 1. The sample consisted of 8 (66.6%) men and 4 (33.3%) women, with a mean age of 61.1 (SD 8.3) years. Previous research among users of the same self-help program has shown that the average age in this population is relatively high, at around 51 years [19]. Most participants had completed higher education, and half reported prior experience with digital interventions (excluding the Jellinek digital alcohol self-help intervention).

The thematic analysis of the focus group data identified 2 overarching themes related to the use of CAs: motivating factors that encouraged participants to consider CAs as supportive tools and hindering factors that limited their willingness to engage with such systems. Within each theme, several subthemes captured both functional and experiential dimensions of how participants envisioned interacting with CAs in the context of a self-help program to reduce or quit their alcohol use.

Many of the themes raised by self-help participants during the focus group were echoed in the interviews with members of the advisory board, indicating a high degree of consistency across the 2 participant groups.

Participants’ knowledge of and experiences with AI and CAs are presented in Table 2. The desired functions of a CA are presented in Figure 1.

The most frequently desired functions were helping users find information, providing advice, and offering motivation when struggling or after a relapse.

On a 5-point scale, participants reported moderate knowledge of AI (mean 3.0, SD 1.2) and perceived the usefulness of CAs for reducing alcohol use as fairly high (mean 3.8, SD 1.1). Comfort and trust levels were largely neutral (mean 3.0, SD 1.2 and mean 2.8, SD 0.9, respectively), while willingness to use a CA was mixed (mean 3.3, SD 1.3).

This qualitative study examined how users of a Dutch preventive alcohol self-help program and members of a client advisory board perceive the potential role of CAs in digital alcohol prevention. Across both qualitative and survey data, findings indicate a general perception of CAs as potentially useful, albeit accompanied by notable reservations regarding trust, authenticity, and willingness to adopt them.

Participants broadly valued the accessibility and functionality of CAs, particularly their around-the-clock support and practical guidance during moments of craving or relapse. They emphasized the appeal of an always-available, nonjudgmental companion that could offer both information and motivation without the social barriers often associated with human help. These qualitative perceptions were supported by survey results showing relatively positive evaluations of usefulness for reducing alcohol use (mean 3.8, SD 1.1), suggesting that participants recognize the potential functional value of CAs in preventive contexts.

Despite the perceived usefulness, both datasets consistently indicated ambivalence regarding personal uptake. Although survey responses indicated only moderate willingness to use a CA themselves (mean 3.3, SD 1.3), participants in the focus group and interviews expressed ambivalence about empathy and authenticity. This tension suggests that perceived usefulness does not directly translate into intended use.

Concerns about empathy and authenticity were central in both methods. In interviews and focus groups, participants questioned whether CAs could provide meaningful emotional understanding or whether responses might feel too “scripted” or “pleasing,” particularly in vulnerable situations. These concerns were also reflected in the survey data, with neutral trust ratings (mean 2.8, SD 0.9) and moderate comfort levels when discussing alcohol-related issues with a CA (mean 3.0, SD 1.2), indicating cautious attitudes toward relational aspects of CAs despite recognition of their functional benefits. Privacy concerns, in contrast, were mentioned less frequently and were generally not perceived as a major barrier.

Several factors that may facilitate engagement include trust, consistency, and flexibility in design. Participants viewed trust as something that must be earned gradually through reliable, transparent, and contextually appropriate interactions. The combination of practical functionality and emotional credibility appeared central to the willingness to use a CA in preventive care. Taken together, these findings provide a more nuanced picture of user perspectives on CAs in alcohol self-help: interest exists, but readiness to adopt CAs is conditional.

The study adds to existing knowledge by situating these insights specifically in a preventive context, where users are not seeking treatment but exploring support to maintain control over their drinking. This distinction helps explain the observed ambivalence, as CAs are perceived as promising complements rather than substitutes for human contact or established self-help methods.

The finding that accessibility, neutrality, and broad knowledge are seen as major advantages aligns with prior digital mental health research [20,21]. Our results extend this work by showing that these qualities may be particularly relevant in preventive contexts, where individuals often hesitate to seek formal treatment because of stigma or low perceived need. In such cases, the neutrality of a CA may not be a limitation but rather a functional strength, lowering the threshold for disclosure and support.

Personalization emerged as a strong desire in our sample, ranging from tailored interaction styles to flexible interface design. This resonates with broader evidence that adaptive systems can foster engagement and adherence [22]. At the same time, personalization brings to the forefront a central tension in digital health: the more interventions are tailored to individuals, the more data collection is required. Prior research has shown that increased data demands can create privacy concerns and reduce trust [15]. Interestingly, most of our participants downplayed privacy as an obstacle in hypothetical discussions. This aligns with previous research on the self-help intervention, which showed limited concern for ethical aspects of applying machine learning to adapt the digital intervention [23].

This discrepancy points to a possible gap between imagined use and real-world adoption. Our survey data support this interpretation. Although most respondents described AI as useful for reducing alcohol use (5=“very useful” and 3=“quite useful”), only a minority expressed strong intentions to use a CA themselves (3=“definitely yes”), with trust and comfort ratings largely neutral. These findings suggest that enthusiasm for CA functionalities in principle does not automatically translate into personal uptake.

Our findings also add nuance to debates about the emotional capacity of CAs. Previous studies have questioned whether they can provide empathy and called for human supervision [11]. Some participants in our study echoed these doubts, but others highlighted neutrality as preferable to human interaction, which can evoke feelings of disappointment or judgment. This suggests that the role of CAs may not be to replicate human warmth but to provide a complementary form of support in moments where neutrality is an advantage.

Finally, our results speak to the broader challenge of adherence in digital self-help programs. Dropout and inconsistent use remain persistent problems in unguided interventions [10]. The motivating factors identified by our participants—24×7 availability, nonjudgmental stance, and personalized feedback—mirror elements that prior research has associated with higher adherence in digital interventions, particularly studies showing that guidance and personalization increase completion rates [5,9,10,22].

The findings highlight both practical and ethical implications for the design and implementation of CAs in preventive alcohol self-help. From a design perspective, participants expressed diverse preferences regarding visual appearance and interaction style. Some favored a simple, text-based interface, while others appreciated the idea of a more humanlike avatar or personalized appearance. However, across these differences, participants consistently emphasized that the tone of communication mattered most. The CA should respond in an empathetic, respectful, and nonjudgmental way, without becoming overly pleasing or superficial. Participants valued honesty and even confrontation when appropriate, but only once trust had been established. This suggests that CAs in this context should be capable of adjusting their interactional style over time, matching their tone and feedback to the user’s stage of engagement and level of trust.

Beyond design, and building on participants’ emphasis on 24×7 availability and low-threshold access, CAs may serve as complements to existing self-help or prevention programs by offering immediate support during moments when human help is unavailable. Integrating such systems into alcohol prevention and broader eHealth strategies could potentially increase accessibility, although this remains to be tested in practice. At the same time, the very features that make CAs attractive, such as their constant availability and responsiveness, may also introduce new ethical challenges. Participants’ concerns about undue influence and the lack of genuine empathy suggest that CAs are not merely neutral tools but may function as moral technologies that shape decision-making and redistribute responsibility within care relationships. For example, when users rely on a CA for advice during moments of vulnerability, questions arise regarding who holds responsibility for the outcomes of such interactions—whether this lies with the user, the developers, or the system itself.

Several limitations should be acknowledged. The sample was small, relatively homogeneous, and skewed toward older, higher-educated participants, which limits generalizability. However, previous research among users of the same self-help program has shown that the average age in this population is approximately 51 years [19], suggesting that the sample is broadly consistent with the program’s user base. Another limitation concerns recruitment. Because participants were informed that the study involved discussing AI-based tools, it is possible that the sample included individuals who were more interested in or open to technological innovations in health care than the average self-help user. This may have resulted in a group with relatively well-developed or positive opinions about CAs, potentially limiting the range of perspectives represented.

Finally, although recurring patterns and overlapping themes were observed across the focus group and interviews, formal data saturation could not be confirmed because of the limited number of participants. Future studies with a larger and more diverse sample could build on these findings to further explore differences in user expectations, trust development, and CA design preferences.

Future research should replicate this study in larger and more diverse samples to verify whether these findings hold across different populations and drinking profiles and to achieve stronger data saturation. Observational and experimental studies involving actual interaction with CAs would allow researchers to examine how trust, comfort, and engagement evolve over time. Directly observing user-CA exchanges, for example, within existing preventive self-help platforms, could provide valuable behavioral data and reveal when and how users disengage.

It would also be valuable to explore how specific design features, such as text-only vs avatar-based interfaces and conversational tone, affect engagement. Finally, longitudinal research could clarify whether CAs can meaningfully support adherence and sustained behavior change over time, beyond initial curiosity or short-term use.

This study shows that users see potential for CAs in preventive alcohol self-help, mainly because of their constant availability, practical guidance, and neutral stance. However, CAs are not perceived as replacements for human contact. Moreover, enthusiasm in principle does not automatically translate into willingness to use such tools, as doubts about empathy, authenticity, and trust remain. CAs should therefore be viewed as complementary support within broader self-help systems. Future design efforts should emphasize personalized design; an appropriate conversational tone, supportive yet not overly agreeable; and transparent communication about data use to enhance engagement and acceptance. Their future value will depend on design, user involvement, and real-world evaluation to strengthen trust and engagement.