In recent years, the uptake of digital mental health interventions has substantially increased (1). Digital mental health interventions, including internet and smartphone-delivered services, hold promise for overcoming significant barriers that are traditionally associated with face-to-face mental health care, such as stigma (2, 3), accessibility (4) and cost (5). Despite demonstrated efficacy (6) and potential to increase the reach of evidence-based care, research shows that digital mental health interventions are associated with relatively poor adoption and adherence (7, 8). Most notably, this is a problem for minimally guided and unguided smartphone interventions that do not involve any degree of therapist support (9–13). One reason that there is poor engagement and adherence with digital mental health interventions may be an insufficient therapeutic alliance (14).

The therapeutic alliance is one of the most robust mechanisms of change in psychotherapy interventions and can be broadly defined as a collaboration between the patient and therapist on the tasks and goals of treatment (15–18), along with an emotional bond. This is important, as it allows for the exploration of vulnerable issues by decreasing client's defensiveness and increasing their self-acceptance (19). The therapeutic bond is fostered through trust, acceptance, empathy and genuineness (20, 21), which are processes that are difficult to mimic via automated digital interventions. While this issue may be addressed through the augmentation of digital self-guided interventions with external human support [e.g., Internet-based cognitive behavioral therapy (CBT) paired with coaching calls (22)], the amount of human therapist involvement influences the cost and scalability of digital health interventions.

An alternative approach is to mimic human dialogue (23) with users via an artificial intelligence (AI) conversational agent. Conversational agents use machine learning and AI methods to simulate human-like behaviors and support (24, 25). Their potential for application in the digital mental health space is gaining traction and they have recently been implemented to help track medication and physical activity adherence (26, 27), provide cognitive behavioral therapy (CBT) (28), and deliver healthy lifestyle recommendations (29) across clinical and non-clinical populations. An increasing body of literature has demonstrated that such interventions are effective and feasible applications for the delivery of mental wellbeing to individuals with self-reported anxiety and depressive symptoms (30, 31).

Further, early evidence suggests that a therapeutic relationship may be established between humans and a conversational agent. For example, Bickmore et al. (32) demonstrated that adults established a therapeutic alliance with a health-related automated system while attempting to increase exercise. In another study, researchers found that within 5 days of initial mental health application (app) use, users reported a therapeutic alliance with a conversational agent, measured through the Working Alliance Inventory-Short Revised (WAI-SR), which was comparable to those in recent studies from the literature on traditional, outpatient, individual and group CBT (33). Recently, Dosovitsky and Bunge (34) examined a chatbot intervention for social isolation and loneliness and found that most chatbot users were satisfied and would recommend the intervention to a friend. Their results showed a pattern of personifying the chatbot and assigning human traits to the chatbot (i.e., being helpful, caring, open to listen, and non-judgmental). Additionally, they found that users were able to build a bond with a chatbot from asynchronous and exclusively text based conversations. While early evidence suggests that conversational agents are capable of establishing a bond with users (33), additional research is needed to understand the perceived aspects of therapeutic bond with CBT-based conversational agents. The literature on therapeutic alliance discusses the importance of caring, understanding, and acceptance in forging the therapeutic bond (19). Studies also highlight the value of empathy (35), genuineness, and positive regard as important aspects of the bond that influence the client's perception of the relationship (36).

In this study, the researchers examined user's perspectives of the therapeutic bond with a conversational agent. It introduces a mixed-methods investigation of the therapeutic alliance between a free-text, CBT-based conversational agent (Wysa) and users. Wysa is an AI-based emotionally intelligent mobile conversational agent app aimed at promoting wellbeing, positive self-expression, and mental resilience using a text-based conversational interface. Wysa is designed to be used by individuals above 18 years of age. According to the Google Play Store and Apple App Store, ~60% of users are between 18 and 34 years of age, with 55% of users identifying as women. A previous study that examined the real-world effectiveness and engagement levels of Wysa on users, researchers found that high-engagement users had significantly greater improvement in self-reported depressive symptoms as compared to low-engagement users in the app (30).

The present study aims: (1) To examine whether conversational agent users perceive a therapeutic alliance, (2) To examine changes in therapeutic alliance over time. We hypothesize that conversational agent users will perceive a therapeutic alliance that will be comparable to the alliance scores observed in human-delivered face-to-face psychotherapy with clinical populations. We also hypothesize that, as observed in other in-person studies, we will notice an improvement in therapeutic alliance scores over time. As an exploratory objective, we also explored user's perception of the therapeutic alliance with the conversational agent.

The present study offers a mixed methods approach to understand the therapeutic alliance amongst users of a conversational agent for mental health. The app studied (Wysa) leverages evidence-based CBT techniques to build mental resilience and promote mental wellbeing. In line with our hypotheses, we found evidence that user's therapeutic alliance scores sustainedover time and were comparable to ratings from previous studies on alliance in human-delivered face-to-face psychotherapy with clinical populations (Tables 3, 4) (47, 48, 52). These findings support recent studies suggesting that the therapeutic relationship can be established between humans and conversational agents in the context of mental health (33). While some mobile mental health applications allow free text responses by the user, few can handle complex and diverse free-text input. In this setting, alliance perhaps thrives because of similar relational principles as have been observed in human dyadic interactions (54). In a free-text setting, a user can communicate their sense of a bond, or their goals and tasks with the conversation agent. This can take place as a constant, dynamic interaction, thereby mirroring the capacity within human interactions for reciprocity, engagement and acknowledgment (55). For example, in a smartphone-based CBT trial for pain self-management, patients criticized the lack of free text answers and an overly static flow of interaction (56). These findings suggest that individuals may prefer interacting with mental health free-text CAs, which is perhaps due to the flexibility and autonomy offered by diverse free input text, similar to conversing with humans (57). Research indicates that fostering individual autonomy is central to the development of the therapeutic relationship. Although speculative, our results indicating comparable WAI-SR scores to in-person therapy, may be due, in part, to user's feelings of autonomy from the free dialogue. Research indicates that fostering client's autonomy is a central experience to good outcomes in psychotherapy (58).

Qualitative research has also found that clients' experiences of honesty, safety and authentic caring play a critical role in alliance development, especially through therapist's supportive behaviors, such as encouraging statements, friendliness, respect, and validation (59–63). More specifically, clients report that a relational connection with the therapist facilitated vulnerable explorations and self-disclosure (19). Similar to these meta-analytic qualitative findings on traditional in-person therapy, recent research suggests that Wysa also offers therapeutic elements of a comfortable, safe, and supportive environment (37). Similarly, the present study findings from Wysa user's free-text input indicate experiences of honesty, safety and comfort with Wysa.

We found that users expressed gratitude and appreciation for the conversational agent. These findings are in line with past research, which suggest that expressions of safety and comfort in a medical setting are indicative of having developed a therapeutic alliance (64). Additionally, past research in a mental health context found that the therapeutic alliance and therapeutic bond are associated with a continued experience of positive emotion, such as feelings of gratitude when thinking about therapy or the therapist (65). The qualitative analysis revealed themes indicating that users commented about their perceived positive benefits. This is further supported by the meta-analysis that demonstrated the linkage between alliance and outcomes in adult psychotherapy (15).

The findings also indicated aspects that negatively affected the bond, such as dissatisfaction, perception of limitations of the conversational agent and perceived negative impact of the alliance with the conversational agent. The dissatisfaction and perceived ineffectiveness is consistent with extant literature which suggests that users express frustration when conversational agents misunderstand or give irrelevant responses (66). This indicates the need to further improve the AI algorithms and striking the balance in achieving human-like qualities.

A unique finding of this study indicates that users accept the limitations of the conversational agent when conversing with it, evidencing the experienced relational bond. Much like the ruptures in a therapeutic alliance that offer critical indicators for exploration and the potential for strengthening the relationship (67), these ruptures with conversational agents point us to the possibilities for improvement with iterations.

Notably, in the present study, many participants also made comments that personified the bot. These findings are aligned with results indicating that users of a chatbot intervention for social isolation and loneliness personified the chatbot and assigned human traits to it, such as being helpful, caring, open to listen, and non-judgmental (34). It is possible that user's personification of Wysa also indicates greater engagement. For example, one study found that as users' social interactions with a conversational agent increased, a greater level of personification occurred, which was associated with increased product satisfaction (68). Also, it is possible that Wysa's flexible conversational interface increased relational capacity building with the bot, leading to greater user engagement and resultant conversational agent personification.

The findings of this study should be interpreted in the context of its limitations. In order to maintain complete privacy and confidentiality with users, Wysa is an anonymous app and our data were anonymized. Thus, we did not collect demographic data, such as age or gender, from participants. Future work should investigate whether therapeutic alliance with a CBT conversational agent varies between adolescent and adult populations. The participants were not required to complete the follow-up assessment, which resulted in a large number of participants dropping out. Thus, the study was limited by the sample size. As digital mental health is a nascent field, this study was limited by the studies available for comparison, and had to rely on studies that include more specific health conditions in its analysis. Future studies should account for the use of quantitative approaches and should have control groups for comparison. Also, the study was limited by the short timeframe, and future studies could employ a longitudinal approach to assess changes in outcomes in conjunction with changes in the therapeutic alliance. Additionally, it is important to consider that although qualitative analysis is a strength of this study, and contributes significantly to the field, the possibility of researcher bias remains even though an attempt was made to mitigate this through a balance of internal and external researchers within the team.

The present study provides insight into the establishment of a therapeutic alliance with a conversational agent and its growth over time, as well as what a therapeutic alliance looks like with a conversational agent. This provides critical direction for user-bot partnerships in future digital mental health initiatives. One such opportunity of effective mental health interventions would be to set tasks for each session, as well as establish weekly and overall treatment goals. This may further enhance the goal and task subscores of therapeutic alliance, thereby contributing to the development of a more impactful therapeutic experience.

The datasets presented in this article are not readily available because the assessment data can be reviewed, but the conversational data used within the article for content analysis is protected under the Privacy Policy of the app consented to. Requests to access the datasets should be directed to chaitali@wysa.io.

Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. The patients/participants provided their written informed consent to participate in this study.

TM and CS designed the study on the app. CS and SM worked on the quantitative and qualitative data analysis aspects of the study, contributed to the review, and final presentation of the manuscript. CB and TM wrote the first and final versions of the manuscript. CS completed and prepared the final review of the manuscript. All authors contributed to the article and approved the submitted version.

TM and CS are employed by Wysa Inc. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.