The use of artificial intelligence in mental health services in Turkey: What do mental health professionals think?

Introduction

Mental health problems are a common and widespread problem (Vigo et al., 2016; World Health Organization [WHO], 2004). However, it is noted that the demand for mental health services is not being met. While there are nine psychiatrists per 100,000 people in developed countries, it is only 0.1 per 1,000,000 people in lower-income countries (Abd-alrazaq et al., 2019). WHO reports that 55% of people in developed countries and 85% of those in developing countries lack access to mental health services (Anthes, 2016). According to WHO (2020) the Covid-19 pandemic increased the demand for mental health services, which overwhelmingly hindered the access to mental health services.

Digital Interventions in Mental Health have been utilized for nearly 25 years (Aboujaoude et al., 2020; Hollis et al., 2015; Lattie et al., 2022; Rojas et al., 2019). The literature has proven the prevalent use of digital applications in mental health services during the Covid-19 pandemic (Connolly et al., 2021; Liu et al., 2020; Philippe et al., 2022; Stein et al., 2022; Wind et al., 2020; Zangani et al., 2022). AI-driven digital applications have recently gained much attention due to their potential to minimize the need for resources such as time, cost, and professional expertise (Blease et al., 2020; Lattie, 2022). Although the mentioned studies yield promising results for the use of artificial intelligence (AI) in mental health services, they are relatively new. There is still a debate on the pros and cons of such applications in mental health services (Broadbent, 2017; Gültekin, 2022). Taking into consideration that there are obstacles in accessing mental health services, high morbidity, and mortality in psychiatric patients, there is an urgent need for the integration of AI into mental health services (Lee et al., 2021). Nevertheless, health care professionals play a critical role in the acceptance and integration of this new technology into the field of mental health (Buck et al., 2022; Monteith et al., 2022; Sebri et al., 2020). In this regard, very few studies have investigated the views of mental health professionals (Blease et al., 2020; Sweeney et al., 2021). Even those few studies were conducted on experts who have not engaged in AI, and their opinions were primarily collected through questionnaires. In the literature, no study has examined experts’ perceptions of AI experiences through in-depth interviews.

In this study, we contributed to filling this gap in the literature by focusing on the opinions of mental health professionals using AI in therapeutic interventions in Turkey. Given that the use of AI in mental health services in Turkey is still very new (Bilge et al., 2020; Erebak, 2018; Usta et al., 2020), the insights of professionals engaging with this technology are essential to predict the future of AI in the field of mental health. Accordingly, answers to the following questions were sought:

RQ1: How do mental health professionals evaluate the advantages and disadvantages of using AI in health care services?

RQ2: What role do mental health professionals assign to AI in mental health services?

At the time of this study, AI technologies encompassed a range of advanced techniques, such as artificial neural networks, deep learning, reinforcement learning, machine reasoning (planning, scheduling, knowledge representation, search, and optimization), robotics (cyber-physical system integration of all techniques related to control, sensing, sensors, and actuators; The European Commission’s High-Level Expert Group on AI, 2018). It is noteworthy that the ChatGPT language model, which is based on deep learning (Cheng et al., 2023), represents a ground-breaking development (Xue et al., 2023), and has yet to be released.

The Use of AI in Mental Health Services

AI-driven digital applications have recently gained much attention for their potential to minimize the need for resources such as time, cost, and professional expertise (Blease et al., 2020; Lattie, 2022). Research has shown that AI technologies such as wearable technology, virtual reality glasses, chatbots, and smartphone applications are used in mental health services (Fiske et al., 2019). In this sense, the use of AI holds promising potential for early diagnosis (Kalmady et al., 2019), treatment (Fitzpatrick et al., 2017), training of mental health professionals (Rein et al., 2018), and drug development (Powell, 2017). For example, a web-based chatbot Woebot, grounded in the cognitive-behavioral model, tested students with depression and anxiety symptoms. Woebot engages users via smartphones and tablets, emphasizing that its service is not a therapy nor provided by a psychologist. Woebot’s algorithm is tuned for specific therapeutic skills (e.g., asking relevant questions, providing empathetic responses). Through conversations about moods, thoughts, and feelings (e.g., “How do you feel?”, “What are you doing?”), Woebot offers proper books, videos, and activities, focusing on detecting clients’ cognitive distortions by using some applications such as word games. Students in the Woebot group reported improvements in anxiety and depression symptoms (Fitzpatrick et al., 2017). Other similar applications have been developed to treat mental health problems as well. For example, Mobilyze, a mobile phone and internet-based application to treat depression and anxiety disorder (Burns et al., 2011); the PTSD Coach to treat post-traumatic stress disorder (E. Kuhn et al., 2014); Tess, a chatbot to cope with depression, loneliness and anxiety problems (Fulmer et al., 2018; Joerin et al., 2019); Wysa, an AI supported mobile application to treat depression (Inkster et al., 2018); SARA, an android application for the adolescents with substance addiction (Rabbi et al., 2017), and Deprexis, an application based on cognitive behavioral therapy model to treat depression are among those mediums with promising outcomes (Twomey et al., 2017). Several studies have proven the accurate and functional use of AI-supported applications in early diagnosing mental health problems, and developmental conditions, including autism (Doi, 2020), personality disorders (Carvalho & Pianowski, 2019), and Parkinson’s disease by “digital phenotyping” method (Capecci et al., 2016). The digital phenotyping method has mainly been preferred to treat depression and anxiety problems (D’Alfonso, 2020), and has proven to be an effective method in the early diagnosis of those disorders (Cao et al., 2017; Jacobson et al., 2020; Kamath et al., 2022; Mastoras et al., 2019; Melcher et al., 2021).

The Advantages and Disadvantages of Using AI in Mental Health Services

Although the mentioned studies yield promising results for the integration of AI in mental health services, it is still in its early stages., sparking debates on the pros and cons of such applications in mental health services (Broadbent, 2017; Gültekin, 2022). Scholars highlight some advantages of AI-supported applications, including better diagnosis, treatment, and care (Cresswell et al., 2018; Janssen et al., 2018; Luxton, 2016, pp. 27–51; McShane et al., 2012); less workload (Cecula et al., 2021; Fiske et al., 2019); easy access to mental health services and less stigmatization (Lucas et al., 2014; Luxton, 2014; Miner et al., 2019; Pham et al., 2022); less cost (Ahmed et al., 2021; Chew & Achananuparp, 2022); interesting and engaging therapies (Hudlicka, 2016); less risks in certain therapeutic interventions (Bilge et al., 2020; Luxton, 2014;); constant support for those living alone (Cavallo et al., 2018; Hung et al., 2019) and better personalization of the treatment for patients (Bzdok & Meyer-Lindenber, 2018). Conversely, some draw attention to ethical, legal, and practical concerns regarding using AI in mental health (Cosgrove et al., 2020; Turkle, 2010). For instance, the inability to establish empathy and therapeutic bonds (Scholten et al., 2017) as well as prejudice and discrimination problems (Fan, 2020, p. 76) are among the given disadvantages of such applications. Another issue is about the responsibility (Waldrop, 1987). Who would be responsible for an AI application’s misdiagnoses or biased treatment? Another problem is security. AI programs collect extensive information about a client’s cultural, social, economic, political, and health-related activities, which might lead to problems of confidentiality, privacy, and autonomy (Bennett & Doub, 2016; Broadbent, 2017). It is noteworthy that there are still no ethical guidelines for the safe use of AI in mental health (Fiske et al., 2019).

The ontological boundary problem further complicates the interaction between humans and AI technologies like social robots and digital assistants (Gültekin, 2022). Robots with humanoid features may lead users to perceive them as a new ontological category (Kahn & Shen, 2017; Pradhan et al., 2019; Tong et al., 2022), leading to misconceptions about cognitive and social attachment or isolation (Festerling & Siraj, 2021; Haggadone et al., 2021; Serholt et al., 2017; Sharkey, 2016). While AI can save time in diagnosis, offer personalized treatment and more accurate medical decisions in other health services such as pathology, radiology, surgery, dermatology, and ophthalmology, there are still concerns about privacy, confidentiality, stigmatization, and the possibility of wrong diagnosis and unemployment in health sector (Al-Medfa et al., 2023; Güvercin, 2020).

Mental Health Experts’ Perspective on the Use of AI

The limited evidence on mental health professionals’ views on the use of AI revealed a cautious stance among 791 psychiatrists, with 48.7% of the participants stating that AI would have little or no impact on psychiatrists in the next 25 years (Doraiswamy et al., 2020). In this global survey, only 3.8% of the participants believed AI would render their profession redundant. Another 47% believed that AI would have a moderate impact on their professions. In the same study, most participants (83%) reported that AI could carry out laborious tasks such as managing medical records and organizing and synthesizing documents.

By comparison, 83% stated that AI cannot replace an average mental health specialist in providing empathetic care. In a qualitative study by Blease et al. (2020), some participants emphasized that machines and humans would establish a “work-sharing” environment that complements each other and enriches the professional process, whereas others argued that it would increase administrative and bureaucratic workloads. The study also revealed opinions that AI might reduce the psychiatric skills of experts, cause over-dependence on technology, and underscore the significance for experts handling possible system errors.

Although present findings are promising, the adoption of digital interventions in mental health is not yet widespread (Lord et al., 2023). One of the factors is undoubtedly clinicians’ attitudes (Nogueira-Leite & Cruz-Correia, 2023). It is reported that obstacles to the widespread adoption of such technologies by therapists involve a lack of training, concerns about the quality of therapy, and limited experience (Feijt et al., 2023). For example, in recent studies on the opinions of mental health professionals, those who did not have AI experience tended to approach therapeutic interventions with AI with caution (J. Prescott & Hanley, 2023). Notably, experience in AI-supported psychotherapy may change cautious attitudes toward such practices (Aktan et al., 2022).

Methods

Study Design

In this study, we employed a qualitative design to explore the views of mental health professionals regarding the use of AI in mental health services. Accordingly, we aimed to elicit participants’ insightful and comprehensive opinions and thoughts about the topic, through their own expressions.

Participants

We used the purposive maximum variation sampling method (Creswell, 2012) to select the participants and conducted the study in Turkey. The criteria for inclusion in the study comprised working as a mental health professional in Turkey and incorporating AI practices into their professional life. The reason for this criterion was to benefit from the insights of those with knowledge and experience in AI practices. Hence, we aimed to obtain in-depth information about participants’ AI experiences. We conducted semi-structured interviews with 13 mental health professionals who met the inclusion criteria. To identify potential participants, we firstly scanned the internet for open sources and listed the mental health professionals in Turkey who have academic papers on AI or apply it in mental health care. Then, we sent an informative invitation email explaining the purpose, scope, and method of the study. We interviewed the volunteers, and we also received suggestions for alternative experts in the field. We contacted by email. As seen, we also used the snowball sampling method.

The participants encompassed a diverse group, including nine psychologists at varying levels of expertise, three psychiatrists, and one psychological counsellor and guide. All participants have had hands-on experience with AI in their professional lives, engaging in technology-assisted psychotherapies (e.g., virtual reality, augmented reality, cognitive-behavioral therapy-based digital applications), human-robot interaction, machine learning, Gerom technology, predictive medicine, and self-help interventions (see Table 1). Data diversity was achieved by involving individuals interested in different facets of AI.

Table 1. Participant Characteristics.

Data Collection Process

We designed a semi-structured interview form following the review of the relevant literature on the use of AI in mental health. As suggested by Glesne (2011), we took feedback from three experts and conducted a pilot interview to finalize the interview form. We included open-ended questions to identify the perceived advantages and disadvantages of AI in mental health services and ascertain the optimal stages for applying AI in practice. The interviews centered around five questions:

1. Why do you prefer using AI-assisted applications in your professional life?
2. What do you think about using AI in mental health services? What are the pros and cons of using such applications?
3. What do you think about the therapeutic relationship between a client and AI-assisted applications?
4. What do you think about the potential applications of AI-assisted technologies in mental health services?
5. In the event of a conflict between an AI-assisted application and a human mental health professional’s diagnosis or treatment plan, how would you evaluate this situation?

Before beginning the study, we received ethical approval from Afyon Kocatepe University Social and Human Sciences Research and Publication Ethics Committee (2020/137).

We conducted semi-structured interviews via ZOOM with willing participants, and we recorded the interviews with the consent of the participants. The interview phase spanned from July 2020 to June 2021, with sessions lasting between 43 and 75 minutes. Of the 13 interviews, six were conducted by the first researcher, and seven by the second researcher. Participation was voluntary, and no incentives were offered.

Data Analysis

We prepared the exact transcriptions of the recorded interviews, which fostered our familiarity with the study data (Braun & Clarke, 2006). The transcriptions were reviewed and confirmed by the participants. We referred to each participant as P1, P2 (see Table 1).

We followed Creswell’s (2012) suggestions: (1) We meticulously read and noted key points, (2) we delved into the content to decipher underlying meanings, (3) we created relevant codes, (4) we categorized the codes for systematic analysis, (5) we reviewed our codes again, and (6) we generated themes and sub-themes by combining similar codes and categorizing hierarchically. Especially in the third and fourth stages, considering the research questions, we created codes to describe and label the data in the analysis process (Saldaña, 2015), and we coded using symbolic abbreviations to represent the data. For example, while trying to figure out what participants think about the role of AI in mental health services, one participant said, “…I think technology-assisted applications support therapy and facilitate our work… So, they can be used as a support for therapy instead of being used alone”. Here, participants especially emphasized the auxiliary role of AI-assisted applications in clinical intervention and psychotherapy, so we created a code labelled “Assistance to interventions.” Then, we created a category called “expert assistant” and grouped any codes implying assistant roles that fell under this category. As Glesne (2011) states, we gradually extracted the codes from the data.

After the coding phase, our research team collaboratively reviewed and refined the codes for clarity and consistency. Some statements were cross-referenced into more than one code. Therefore, we organized the codes that associated similar meanings into categories involving common themes in order to achieve semantic integrity. For example, we created the category of “advantages for clients”, encompassing service satisfaction, affordability, easy accessibility, non-stigmatization, transparency, and risk control. Then, we reviewed all the data to minimize the overlap and redundancy in the codes. We conducted a content analysis using the interview transcripts and the MAXQDA program. The employed analysis program contributed to a more rigorous analysis thanks to its ability to manage extensive and complex data, provide quick and easy access to the codes created by the researchers, and allow for closer scrutiny of the data (Creswell, 2013). Thus, our work became more systematic, refined, and meticulous (Bazeley & Jackson, 2013). We also benefited from the program to visualize our findings.

In the study, we did not consider participant frequency as an indicator, and only one participant’s opinions were used to create codes. Instead, we prioritized how many participants mentioned a code to have a general overview of the data distribution (Miles & Huberman, 1994). The relevant figures show the number of participants who mentioned the codes by using the symbol N.

Credibility and Dependability

As Merriam (2009) suggested, several researchers play a role in ensuring the credibility of the findings. As researchers, we often conducted collaborative session to discuss the codes and categories and agreed on them all throughout the analysis process. We devoted sufficient time (13 months) to collecting data and made a detailed description of the study group characteristics and data collection and analysis processes. We enhanced sample diversity by interviewing a range of mental health specialists interested in AI.

In the final coding stage, we employed an external audit and asked a field expert to check the coding to enhance credibility. Lastly, we utilized the reliability formula of Miles and Huberman (1994), and it was .94. We used direct quotations in the findings section to reflect the participants’ views. Additionally, we reduced the margin of error by performing computer-assisted analysis, and digitally stored the transcripts, data, and code contents to ensure confirmability.

Results

Findings were presented under four themes, considering the responses to the research questions. Accordingly, the themes of the perceived advantages and disadvantages of using AI-assisted applications in mental health were generated following the responses to those questions: Why do you prefer AI-supported applications in your professional life? What do you think about using AI in mental health services? What are the pros and cons of using such applications? What do you think about the therapeutic relationship between a client and AI-assisted applications?. The theme of the perceived role of AI in mental health services was generated following the responses to the questions of What do you think about the possible ways to use AI-assisted applications in mental health services? When an AI-assisted application conflicts with a human mental health professional’s diagnosis or treatment plan, how would you evaluate this situation? Apart from them, the theme of the future of AI in mental health services was generated according to the overall assessment of the responses.

The Perceived Advantages of Using AI in Mental Health Services

According to the analysis results, the advantages can be grouped into three categories: the advantages to clients (N = 12), professional advantages (N = 12), and advantages to mental health care experts (N = 11). The “the professional advantages” category covers the effects related to the nature and practice of the mental health profession, the “the advantages to mental health care experts” category covers the effects related to the personal and professional experience of mental health experts, and lastly, the “the advantages to clients” category describes the effects that the service recipient may experience. The codes in the given categories are shown in Figure 1.

Figure 1. The Categories and Codes of Advantages.

Code descriptions and direct quotations related to each code are presented in Table 2.

Table 2. The Categories, Codes, Code Descriptions and Example Quotations Related to the Perceived Advantages of Using AI in Mental Health Services.

The Perceived Disadvantages of the Use of AI in Mental Health Services

According to the experts’ opinions on the possible disadvantages of using AI applications in mental health services, we generated three categories: professional disadvantages (N = 13), the disadvantages to clients (N = 10), and the disadvantages to experts (N = 8). Figure 2 below shows the codes and categories.

Figure 2. The Categories and Codes of Disadvantages.

Code descriptions and direct quotations related to each code are shown in Table 3 below.

Table 3. The Categories, Codes, Code Descriptions and Example Quotations Related to the Perceived Disadvantages of Using AI in Mental Health Services.

The Perceived Roles of AI in Mental Health Services

We found three fundamental roles of AI in mental health services: expert assistant (N = 13), mental health expert (N = 8), and care-support services provider (N = 4; Figure 3). The “expert role” of AI refers to using AI applications as subject or agent, while the role of a mental health specialist is either absent or auxiliary. “Expert assistant role” refers to using mental health experts as the primary subjects and AI applications as a tool in the role of expert helper. In the care-support services provider role, AI applications can offer social, emotional, and cognitive support to the elderly, children, or adults, which signals its additional roles in preventive mental health services. Figure 3 below shows the codes and categories.

Figure 3. The Categories and Codes Regarding the Perceived Roles of AI in Mental Health Services.

Participants mostly believed that AI applications would assist experts. For AI’s roles as an expert assistant, the participants stressed assistance to interventions (N = 11).

Regarding AI’s roles as a mental healthcare expert, they emphasized the potentials of non-human interventions (N = 4) and non-human diagnosis systems (N = 2) thanks to the long-term developments in AI technology. However, some were hesitant about the complete expert role of AI under all circumstances and stated that it could depend on the extent of the problem, the title of the mental health specialist, and the development level of the AI technology (N = 6). Although the participants mentioned the possibility of AI’s expert roles, they emphasized the differences between the virtual service and the one offered by a human mental health specialist: “But of course, they cannot give feedback like a human and convey the empathy feeling.” (P6). Such a point of view can be considered as a mechanism to protect professional self-esteem. The category, code, and sample statements regarding the roles of AI in mental health services are presented in Table 4.

Table 4. The Categories, Codes, Code Descriptions, and Example Statements on the Role of AI in Mental Health Services.

Regarding the expert role of AI in mental health services, we observed discrepancies between participants’ anticipations about the future of AI and their desires. The possibility and desirability of AI’s role in mental health services are shown in Table 5.

Table 5. The Possibility and Desirability of Using AI in Mental Health Services.

All participants expressed expectations and wishes regarding using AI in an assistant role in mental health services. However, the willingness of some participants, who mentioned the potentials of AI in an expert role, differed in this regard, and they seemed doubtful and showed negative attitudes towards using it without expert support. The majority of the participants were unwilling or undecided about the use of AI applications without expert support. Most participants were optimistic about the care-support roles of those applications. In light of the findings, we found that the controversy in using AI in mental health services surrounded the expert-role definitions of such applications rather than their assistant or care-support roles.

The Future of AI in Mental Health Services

As discussed in the findings above, AI plays the roles of expert, assistant, or care support in mental health services. Along with these roles, all participants agreed that using AI would be practical and effective in the future of mental health services. “A new and inevitable era is beginning… one day, I think that a psychiatrist would prescribe the AI application instead of drugs.” (P4). We generated two themes regarding the future of mental health services using AI.

The Barriers to Use AI: The Ontology of Mental Health Services and Experts’ Resistance

Although AI-assisted applications would be effective in the future of mental health services, some participants (N = 7) assumed that the field of mental health, especially psychotherapy, would be the last field to embrace AI developments.

“AI may dispel the demand for many professions. For example, you would not need to hire a craftsman, AI would make the optimization automatically instead (…), but I think mental health seems to be one of the professions that would be affected last.” (P11)

“AI may begin to replace people in certain areas, which I agree with and may lead to unemployment as well. But psychology will be one of the last fields that AI would impact, that is for sure” (P5). Such viewpoints acknowledge that AI will eventually be effective in mental health services, but they emphasize that this impact will occur relatively later compared to other service areas. Here, the focus is on the ontology of mental health services and an evaluation is made regarding the content of these services. Another obstacle highlighted by all participants is the resistance from mental health experts. When evaluating the participants’ opinions, it is believed that the resistance from mental health experts will slow down the progress. One participant stated. “I know they are somewhat hesitant, the mental health community in general, when it comes to technological advances, virtual reality. I actually did some research, and there is a certain hesitation” (P10). Another participant indicated,

“Naturally, they will oppose this and try to prevent it at some point. They will emphasize their belief that it will be an inadequate or unsuccessful attempt. However, there are some things that cannot be prevented, so no matter how much we reject, I am sure we cannot change the fact that it will eventually happen.” (P1)

These statements reflect the participants’ opinions regarding expert resistance. When trying to understand the origins of the resistance towards AI, it was observed that the perceived disadvantages of AI mentioned in the previous findings were emphasized. These include obstacles arising from the experts themselves (e.g., job insecurity, lack of knowledge, and conservatism), obstacles related to the perception of AI (its potential for harm, empathy and responsibility issues, and trust problems), and environmental obstacles (cost-related issues, and team building problems).

The Areas Where the AI Has Potential to Impact Mental Health Services

According to participants’ views, two categories have emerged regarding the areas where AI can potentially impact mental health services. These categories are the transformation in the role of experts/professionals (N = 10) and the emergence of new working areas and methods (N = 7; Figure 4).

Figure 4. The Categories and Codes of AI Impact of Mental Health Services.

Code descriptions sample expressions reflecting each category are presented in Table 6.

Table 6. The Categories, Codes, Code Descriptions, and Example Quotations for AI Impact on Mental Health Services.

Discussion

In this study, we examined the opinions of mental health experts with experience in AI regarding the use of AI in mental health services. By conducting in-depth interviews with participants who have expertise in AI and related fields, we obtained detailed answers to our research questions. In this section, we will discuss the key findings of our study in relation to the field of mental health research and practice.

The experts we interviewed generally displayed a positive but cautious attitude towards the implementation of AI in mental health services. The participants highlighted several advantages of AI, including increased client satisfaction (Fiske et al., 2019), improved accessibility to mental health services (Kim, 2019), reduced stigma concerns (Luxton, 2014), and the ability to facilitate advanced analysis, objective evaluation, and dissemination of mental health services (Luxton, 2016; Senders et al., 2019). They also emphasized the potential of AI to reduce the workload of experts and minimize errors made by professionals (Doraiswamy et al., 2020). However, the participants also expressed concerns regarding the use of AI in mental health services. Some of the concerns highlighted the inability of AI to meet therapeutic conditions, reliability issues, problems related to responsibility, privacy and security concerns, potential harm to clients, and resistance from experts. These issues indicate three key challenges associated with AI: philosophical/ontological, ethical, and practical.

Ontological Problem: Can Machines Understand Humans?

Although some of the experts we interviewed adopted a more cautious stance, they generally expressed concerns that AI-based technologies might struggle to establish a therapeutic alliance with the client. This concern is rooted in the essential difference/ontological contrast between human and machine intelligence. For instance, one interviewee commented on AI, stating, “I do not think there will be a structure that embodies concepts like unconditional acceptance, transparency, and respect.” (Professional Disadvantages/Inability to meet therapeutic needs). In contrast, a different participant said, “I think that such a relationship will never be established because we are inherently social beings who prioritize human connections. While Japan has made significant advancements with humanoid robots that resemble humans and exhibit facial expressions, I don’t think they can fully replicate emotional connection.” (Professional Disadvantages/Inability to meet the therapeutic needs).

Although the definition of therapeutic alliance varies, it generally refers to the relational bond between the therapist and the client, which includes empathy, trust, congruence, and unconditional respect. It plays a central role in the success of therapy (Ardito & Rabellino, 2011). The concern that most participants associate with AI as a “helper tool” can be attributed to this issue. This concern is likely rooted in the anthropocentric paradigm of psychology. Traditional mental health services are built upon the assumption that the provider and the recipient are both “human”, assuming that meaningful ontological relationships can be established between people (Joy, 2005). This perspective, embedded in modern psychology, implies that animals (J. L. Kuhn, 2001; Pots, 2010) and machines (Turkle, 2010) cannot be the subjects of a therapeutic relationship. Some of the interviewees highlighted this issue. For instance, one participant stated, “However, the drawback of AI is that it can never be a human (…) it seems to me that people will never approach it as they would with another human.” (Professional Disadvantages/Inability to meet the therapeutic needs). Another participant similarly said, “No matter how advanced and powerful their software is, I believe that the human-to-human relationship has a unique dimension.” (Professional Disadvantages/Inability to meet the therapeutic needs). In a study by Doraiswamy et al. (2020), 83% of the participants stated that AI could not replace an average mental health professional in providing empathetic care. Other studies have also identified a lack of emphatic incompatibility between web-based applications and users (Scholten et al., 2017). In our study, the participants frequently highlighted the inability of AI to meet the therapeutic needs, which involves empathy, emotional reflection, and context-specific awareness.

At the core of the debate lies the question of whether machines can “empathize.” Several studies indicate that people can empathize with machines (Darling et al., 2015; Rosenthal-von der Pütten et al., 2013) and perceive their behaviors as empathic (Leite et al., 2014). Some experts approach AI or robot empathy from a “functional” perspective, arguing that it is not necessary for AI to empathize but rather be “empathetic” to the client. Consequently, the design of AI and robotic technology aims to create products that humans can perceive as empathetic (Malinowska, 2021). In our study, a participant expressed support for this viewpoint by saying, “…empathy, unconditional acceptance- these are all within a client’s perception. If the client feels accepted, they will eventually feel the acceptance extended from the other party.” (Expert Role/Non-human interventions). Some experts engage in philosophical discussions and argue that genuine empathy is possible in human-machine interaction. Accordingly, AI machines are social-linguistic constructs (Coeckelberg, 2011). The interaction between humans and machines is “inter-subjective,” lacking genuine empathy (Damiano & Dumouchel, 2018, 2020). In our study, one participant stressed the importance of whether the client “healed or not”, supporting this viewpoint by saying, “I have observed that in the applications that, whether it is a face-to-face relationship or an occasional virtual reality experience, this relationship remains central.” (Professional Disadvantages/Inability to meet the therapeutic needs).

On the other hand, some experts have criticised these perspectives and argued that AI technology, despite its ability to interact meaningfully with humans, fundamentally involves a “deception” (Sharkey, 2016). Turkle (2010, 2018) objects to the AI and robotics industry’s production of devices that act as “therapists” and refers to this technology as “cheat technology.” According to Turkle, machines can replicate thought and emotion but cannot think or experience genuine emotions. Emphasizing that honesty is one of the fundamental principles of psychotherapy, Sedlakova and Trachsel (2023) also question the acceptability of chatbots interacting “as if they were empathetic.”

The integration of AI technology into mental health services sparks a profound philosophical discourse on whether machines can empathize, revolving around the potentiality of establishing a meaningful connection between the inherently different entities: machines and human beings. Although it may be impossible to entirely bridge this ontological gap, there is potential for reducing it gradually, which could have an impact on the role of human clinicians in mental health services.

The AI: An Expert or an Assistant/an Agent or a Tool?

In our study, the participants generally perceived AI as a “tool” that assists human clinicians. For instance, one respondent stated: “People innately seek connection with other people (...) I see them as tools. I consider myself as a standard therapist who also utilizes technology as a tool, in supportive roles.” (Expert assistant/Assistance to interventions).

However, it remains unclear how to formulate the relationship between AI and mental health specialists (Miner et al., 2019). The question of whether AI is merely a tool or an agent/subject is raised within the discussion. Sedlakova and Trachsel (2023) also highlight the ambiguity surrounding the role of AI applications in the mental health field, noting that while they can serve as valuable tools, they should not be regarded as an agent/subject. In our study, the uncertainty regarding the status of AI is evident in the responses of some participants. An interesting response reflecting concerns about the potential displacement of human clinicians by AI is as follows: “Because it challenges human narcissism, there may be a sense of inadequacy. I have also experienced this in seminars. If it can do the job, what do we need?” (Transformation in the role of experts/professionals/Requiring to plan, manage, and create content for AI). According to Sedlakova and Trachsel (2023), the use of AI as an “agent” with a psychological existence may give rise to unrealistic expectations and potential harm for clients. Another participant’s response supports this concern, stating: “Another dimension to consider is the potential harm to people. To mitigate this risk, service providers should clearly define the capabilities and limitations of AI and ensure that users have realistic expectations.” (Client Disadvantages/The possibility of being harmed).

Maintaining realistic expectations seems challenging due to human’s tendency to anthropomorphize non-human beings (Epley et al., 2007; Guthrie, 1993, p. 62). To facilitate meaningful interaction between AI and humans, AI systems must be designed to resemble humans (Duffy, 2003; Fong, 2003). As the resemblance between humans and machines increases, humans tend to attribute human-like characteristics to machines (de Visser, 2016), blurring the ontological boundaries between humans and machines (Gültekin, 2022; Guzman, 2020; Haggadone et al., 2021). Some studies indicate that children establish a new ontological status, placing social robots between humans and machines (Kahn et al., 2012; Severson & Carlson, 2010). According to scholars, assigning psychological status to inanimate machines emerges as a central ethical concern in human-social-robot interactions (T. J. Prescott & Robillard, 2020).

An Anxious Optimism: How Will AI Impact Mental Health Services?

In our research, the participants mostly expressed that the experts would approach the use of AI in mental health services with scepticism. They cited the inability of AI to possess therapeutic skills, such as empathy, transparency, and emotional reflection as a reason for this. This viewpoint is supported by studies investigating experts’ opinions on the use of AI in mental health (Blease et al., 2020; Doraiswamy et al., 2020). While the participants emphasized “expert resistance,” (The Ontology of Mental Health Services and Experts’ Resistance), they acknowledged that the impact of AI on mental health is inevitable. They believed that it would lead to a shift in the identity of experts and the emergence of new areas of study (The Areas where the AI has potential to Impact Mental Health Services). Researches have highlighted to the increasing use of digital mental health interventions within the past two decades (Werntz et al., 2023). However, the exact implications of this increase on mental health services remain unclear. In other words, there is a sense of anxious optimism in mental health. One underlying concern is that AI represents the “most disruptive technology” ever developed (Păvăloaia & Necula, 2023). Disruptive technologies have potential to radically transform the paradigms, norms, legislation, and service delivery methods (Hopster, 2021). AI and robotic technology are disruptive innovations that can revolutionize the paradigm of mental health services (Bickman, 2020; Sarris, 2022). Reflecting on the last 50 years of mental health services, Bickman (2020) argues that AI holds promises in addressing problems such as diagnostic issues and treatment inadequacies. Additionally, it is posited that the most significant impact of AI will be on the role of clinician, playing a central role in transforming mental health services.

However, several ethical, legal, and practical concerns arise from the ontological incompatibility between AI technology and humans as well as its structure (Broadbent, 2017; Fiske et al., 2019; Liyanege et al., 2019; Luxton, 2014, 2016), which limits the potential of AI to transform the field of mental health. Braga and Logan (2019) argue that AI cannot gain self-awareness or replace human intelligence. They contend that no matter how sophisticated it becomes, AI cannot be relied upon independently, but it can provide the desired efficiency when employed with human intervention.

Studies focusing on three main areas—diagnosis, prognosis, and treatment—highlight numerous technical (such as black box problems), ethical (such as biased data generation), and logistical (such as time and cost constraints) limitations of AI (Cecula et al., 2021). Furthermore, Cecula et al. (2021) stress a gap regarding “patient experience” and identify potential issues that may arise from a patient’s contact with the human clinician. While Bickman’s (2020) vision suggests that AI can bring about a paradigmatic transformation in mental health services, Brown (2020) argues that algorithms may harm the clinician-client relationship. While Horn and Weisz (2020) agree with Bickman’s (2020) idea, they focus on the challenges of using AI in mental health and offer some caveats. The authors note that AI primarily focuses on biological measures such as medicine and pharmacology, while psychotherapy places more emphasis on a client’s cognitive, emotional, and behavioral characteristics. Therefore, the effectiveness of AI in other fields of medicine, which are more suitable for objective measurement, may not necessarily apply to mental health services. Graham et al. (2019) indicate that AI can classify mental health diseases more objectively than the DSM-5 as it still continues to evolve. However, the authors caution that, due to technical and ethical reasons, careful consideration should be given to its use in clinical practice for diagnosis, treatment, and prognosis. In our study, the participants held the viewpoint that mental health services would be among the sectors least affected by AI.

Undoubtedly, the rise in psychological health issues, the shortage of specialists, difficulties in accessing mental health services, and structural, methodological, and practical challenges faced by the mental health field in diagnosis, treatment, and prevention have sparked optimism regarding the potential contributions of AI. However, as Nobel Prize-winning economist Robert Solow highlighted, “You can witness the computer age everywhere but in the productivity statistics”, this optimism is outpacing the actual efficiency of AI technology (Monteith et al., 2022).

AI still falls short in meeting the clinical application conditions outlined by the Technology Readiness Level (TRL) scale developed by NASA. For example, machine learning in intensive care medicine lags behind in terms of technology readiness scores (van de Sande et al., 2021), and even in the field of radiology, where medical AI is most advanced, the performance of AI remains poor (Allen et al., 2021). In their research on AI in psychiatry, Monteith et al. (2022) suggest that while AI technologies will serve as an important source of information for physicians, they will continue to be the foundation for mental health services provided by human clinicians. This viewpoint is supported by the majority of participants in our study, who were reluctant to attribute the role of an expert to AI.

On the other hand, developments in the field of AI persist at a rapid pace. For instance, some experts consider the newly launched OpenAI chatbot ChatGPT-4 on 30^th November 2022 ground-breaking (Gordjin & ten Have, 2023). Although ChatGPT-4, which is trained through Reinforcement Learning from Human Feedback (Mann, 2023), has great potential for mental health services (Singh, 2023), studies have pointed out shortcomings in developing therapeutic alliances with clients, especially in terms of empathy (Carlbring et al., 2023).

Conclusion

It can be suggested that the interviewees mostly shared similar views on four central issues. Firstly, the participants had a positive perspective on the use of AI in mental health services. Although they did not ignore its disadvantages, they prioritized the advantages. Secondly, although technological developments would make it possible to attribute roles to AI in the future, they mostly underlined the use of AI under the supervision of a human clinician. They stressed that AI would be insufficient to establish mainly emphatic or therapeutic bond with a client. They described AI applications mainly as a tool to assist human clinicians. Additionally, they expressed the belief that while AI would bring about noteworthy changes in mental health services, they were sceptical about the possibility of a dramatic transformation. They argued that either it was not currently feasible or that achieving the necessary technical maturity for the safe utilization of AI in the field of mental would health require a significant amount of time. Fourthly, we observed that the participants made limited comments on the ethical and legal issues such as data ownership, black box problem, algorithmic bias, and discrimination.

According to the research results, we foresee a technology-supported paradigmatic shift in traditional mental health care approaches. The literature suggests that human-robot interaction could play an important role in the delivery of mental health services. Further studies are required to show their effects on mental health services, especially on mental health education and management.

Limitations and Implications for Future Research

In this study, the interviews were conducted on Zoom, which might be considered a limitation to the research due to the limited interaction on a digital platform. Besides, the participant experts demonstrated varying degrees of interest in AI technologies, leading them to have different viewpoints and concerns. Another limitation to bear in mind is that the participants were proficient in AI. It is important to acknowledge the potential presence of a positive bias toward AI among these participants. To develop a more comprehensive understanding, it would be valuable to incorporate viewpoints from experts from different cultures, diverse schools of thought, and those who do not currently utilize AI in their practices.

This study addressed the advantages and disadvantages of AI in three dimensions: “client,” “professional,” and “expert.” It should be noted that the mentioned advantages and disadvantages may differ according to the perceived role of AI. For example, empathy may be perceived as a significant disadvantage in the “expert role of AI, but an advantage in the “assistant” role. Future studies can categorize and compare the perceived advantages and disadvantages, considering the roles attributed to AI.

On the other hand, the roles attributed to AI and the perceived advantages and disadvantages may vary depending on experts’ attitudes and epistemological beliefs toward innovative technology. For example, people who often use digital technologies and are more open to technological innovations would attribute different roles to AI than others or would have different perceptions of advantages and disadvantages. Likewise, the discrepancies in epistemological beliefs could lead to varying opinions about the credibility of the AI algorithms. Further research may discuss the attitudes toward technology and epistemological assumptions.

Conflict of Interest

The authors have no conflicts of interest to declare.

Acknowledgement

Ethical Approval: Approval was obtained from the ethics committee of Afyon Kocatepe University (n.137/2020). The procedures used in this study adhere to the tenets of the Declaration of Helsinki.

Informed Consent: Informed consent was obtained from all individual participants included in the study.

Data availability: The protection of the privacy rights of the participants and confidential handling of the data were requirements demanded by the ethical bodies. The data consequently are not publicly available but anonymized version of the data can be available by authors under reasonable request.

Acknowledgment: We extend our sincere appreciation to all participants for their attendance in the research.