This systematic review investigated the experiences and perceptions of patients and clinicians regarding artificial intelligence applications in mental healthcare. A systematic review methodology was adopted to ensure the comprehensive synthesis of existing evidence from qualitative, quantitative, and mixed-methods studies. The process adhered to the PRISMA 2020 guidelines and involved systematic identification, screening, appraisal, and synthesis of relevant literature published between 2018 and 2025. Databases including PubMed, Cochrane Library, and Google Scholar were searched using structured Boolean combinations of keywords related to artificial intelligence, mental healthcare, and user experiences. The quality of the included studies was critically appraised using the Mixed Methods Appraisal Tool (MMAT). As the study relied exclusively on previously published research, no ethical approval was required. The review was registered with PROSPERO under the registration number CRD420251164840.

The initial database search yielded a total of 5,400 studies. After the removal of 2,224 duplicate records, 3,176 unique studies remained for further screening. Of these, 1,152 were excluded as they focused on other than AI-based digital tools. Additional 1,823 studies were removed due to involvement of treatments of critical conditions like AIDS, hepatitis, cancer, etc. This left 201 studies eligible for full-text assessment. Upon closer examination, 124 were excluded as they were secondary research articles such as reviews or meta-analyses, 34 were removed for being study protocols without reported findings, and 23 were excluded as they were editorial letters without original research data. Following this screening process, a total of 20 studies met the inclusion criteria and were included in the final synthesis (Fig. 1).

Ethical considerations were carefully acknowledged in the conduct of this systematic review. Since the study involved the collection, appraisal, and synthesis of data from previously published research rather than direct interaction with human participants or animals, no ethical approval was required. All sources of evidence were obtained from peer-reviewed journals and credible databases, ensuring that the included studies had themselves undergone appropriate ethical scrutiny during their original conduct. Furthermore, efforts were made to maintain integrity and transparency by accurately reporting methods, critically appraising the quality of evidence, and avoiding bias in interpretation. Proper citation and referencing practices were adhered to throughout, thereby respecting intellectual property rights.

The included studies varied in design and participant characteristics, encompassing mixed-methods, qualitative, and quantitative approaches. Mixed-methods studies included Inkster et al. (2018), which involved 108 high-use users and 21 low-use users, Blease et al. (2020) with 791 psychiatrists, Götzl et al. (2022) with 8 young people and 5 expert interviews, and Blease et al. (2024) with 138 psychiatrists. Qualitative studies were represented by Prakash and Das (2020), which analyzed 1,826 user reviews, Reis and Maier (2022) with 15 mental health professionals, Malik et al. (2022) using 7,929 descriptive English-language user reviews, Zhang et al. (2023) with 20 individuals including practitioners, social workers, mental health nurses, and educator scientists, Ding and Barbic (2024) with 23 participants with mental health issues, Cross et al. (2024) with 107 community members and 86 mental health professionals, Gültekin and Şahin (2024) with 13 mental health professionals, Alanezi (2024) with 24 outpatients, Stroud et al. (2025) with 42 physicians, Petersson et al. (2025) with 25 young adults with mental health issues, Hiller et al. (2025) with 27 young participants with mental health issues, and Beg and Verma (2025) with 17 participants with mental health issues. Quantitative cross-sectional studies included Kleine et al. (2023) among 206 participants comprising psychology students and psychotherapists, Benda et al. (2024) among 500 adult participants with mental health issues, Aamer et al. (2025) among 125 mental health professionals, and Sharif et al. (2025) among 251 mental health professionals.

When utilising AI-based mental health application, it was found that the most visible issue in the mind of a user was the focus on trust, privacy, and confidentiality, which was perceived as a decisive element in continued app use. Malik et al. (2022) observed that the lack of registration demands or the lack of personal data collection in particular were mentioned multiple times by the users as essential to engaging with the app, particularly due to the sensitive nature of mental health disclosure. The promise of privacy and information protection generated a feeling of safety that increased the dependence of the user on the application and made it stand out amid the other online platforms that are usually labelled as having invasive data policies. This emphasis on privacy was also a sign of a more underlying concern of integration whereby the user unconsciously assessed whether these technologies could actually integrate into their personal lives without the threat of being misused. Although the majority of reviews praised this protective design, the low percentage of negative comments showed that not everyone believed the app to be equally helpful and it may not be able to address more complex or serious mental health requirements.

Likewise, the views of psychiatrists in 22 countries, as delivered by Blease et al. (2020), offered some insightful information on the optimism and scepticism about AI role in mental care. One overwhelming reason why many participants raised it was that AI could not displace the relational essence of psychiatry, especially empathy, trust, and therapeutic alliance, which was considered as a key to effective care. Although psychiatrists did not deny the potential efficiency advantages of AI through less administrative work and more access to intricate clinical information, they also cautioned that they might undermine clinical finesse and that they risked causing more harm than benefit when used without appropriate supervision. In the meantime, Prakash and Das (2020) also disclosed that, although mental health chatbots were accessible and anonymous, the adoption of them was strongly influenced by the issue of privacy, safety, and trust. Several reviewers said they were reluctant to provide sensitive personal information because they were afraid of a breach or abuse and were experiencing general concerns about digital surveillance in healthcare. Those fears were also reinforced by the uncertainties about the reliability of chatbot responses in case of a crisis or emergencies when machine-based interactions were viewed as the poor alternatives to professional assistance. Trust consequently was a significant obstacle which was not only affected by the perception of technical reliability but also the absence of transparency in the data storage and usage.

Conversely, the study by Kleine et al. (2023) established ease of use was not associated with the feedback tool but it had negative associations with the intention to utilise the treatment recommendation tool, thus suggesting that perceived credibility in high-stakes decision situations could be impaired by oversimplification. This contradicts the traditional thinking of the technology acceptance models and leads to the necessity to balance between usability and professional depth and seriousness. Furthermore, cognitive readiness was also found to be a facilitator where familiarity and comfort with technology predicted adoption and AI anxiety always inhibited willingness to use either tool. On the same note, Ding and Barbic (2024) also raised the issue of the capacity of AI to serve in complex or high-risk situations with respondents raising doubts about its suitability in simulating empathy, clinical judgement, and subtle human interactions. These concerns were reinforced by the fears of the threat to privacy, uncertain outcomes of the fast development of AI, and the absence of the regulatory frameworks.

In the meantime, Benda et al. (2024) emphasised the high societal demands concerning what values are expected to be the foundation of AI in mental health. Transparency and explainability were regarded by more than 80% of respondents as a necessity, which clearly indicates the desire to know how algorithms work and arrive at conclusions. Fears of accuracy, bias, confidentiality and errors prevailed in the responses, further supporting the notion that trust in AI depends on its promise of reliability and safety. Interestingly, 81.6% of the respondents thought that health professionals should be held accountable in the event of any misdiagnosis committed with the help of AI, indicating a low readiness to transfer the responsibility to the machine. The differences between human professionals (77.8% comfort) and AI chatbots (47.6% comfort) are statistically significant and support the necessity to retain relational and human-centred care.

Cross et al. (2024) also recorded a high level of scepticism despite the perceived benefits; almost half of the respondents who reported harms or concerns using AI. These patterns were reflected in the attitudinal differences with the MHPs demonstrating more positive attitudes toward AI than CMs did, probably because they viewed AI as an efficiency-enhancing complementary practise instead of a substitute. It is also worth noting that 68% of MHPs highlighted saving time and more productivity, as compared to CMs who were more reserved and approved fewer use cases that involve direct client care. CMs, in their turn, were more worried about loss of human support, the risk of misdiagnosis, and data privacy issues, which reflects more profound fears that AI is replacing human interactions with mechanised reactions. Sentiment analysis also showed that only a quarter of the answers express optimism regarding AI increasing accessibility, personalization, or integration, and there is a small but significant segment of people who think AI is groundbreaking.

In Beg and Verma (2025) and Alanezi (2024), the mental health patients were also expressing doubts about the reliability and accuracy of AI App responses, as they feared that poor measurements or cheques and balances could lead vulnerable people astray. Furthermore, cultural and language peculiarities were also revealed as obstacles as some of the participants felt that the support was weak when subtleties in the language or cultural backgrounds were not well-represented. In the same vein, Gültekin and Şahin (2024) and Beg and Verma (2025) also focused their attention on the viewpoint of the professionals who mostly perceived AI not as a substitute of a clinician but as a supportive tool under their supervision. This viewpoint represents long-standing scepticism regarding the failure of AI to reproduce empathy, therapeutic alliance, and delicate aspects of clinical judgement to which effective mental health care is inherently linked. Although the participants indicated that AI is beneficial in diagnosis and treatment plan, they emphasised that human supervision was required in the application of AI to guarantee ethical and safe practise, which is consistent with other research reporting concerns that AI is not a replacement but a complement in psychiatry.

Götzl et al. (2022) found out that the attitude of young people towards AI-informed mHealth applications was rather open and pragmatic, but their acceptance was heavily determined by data privacy and sensitivity concerns. Although a significant number of the participants were willing to provide non-sensitive information, such as the number of steps, sleep habits, or physiological indicators to enhance the personalization of the apps, they were unwilling to give out the information that is highly personal, i.e., conversations or emotional reveal. Such discriminating generosity highlights a subtle attitude within the youth generation, who distinguish between useful information that improves the functionality of available apps and personal information that is likely to violate personal boundaries. Significantly, the focus group discussions revealed that youth did not unanimously oppose the idea of data sharing but the concerns were transparency and being in control of the user as the preconditions of trust. This duality represents a broader generational ambivalence of digital health between comfort/technology and anxiety over excessive surveillance and lack of protection of personal data.

The significance of app design and professional supervision in the formation of the usability and safety of AI-informed mHealth tools was also emphasised by Götzl et al. (2022). Both the participants in the youth category and the professional department have underlined personalization, simplicity, and adaptability as the major factors that led to sustained use, but not the competitive aspects like progress comparisons. The demand to make the functions more flexible and individual is connected to the need of the young people, who prefer tools that can be adapted to various and changing mental health needs and do not necessarily give attention to one particular issue. Scholars supported the following priorities by indicating the need to implement safeguards, like the use of help-me buttons and sensitive text wording, especially concerning suicidality or self-harm because they were aware of the dangers associated with unsupervised interactions with AI. A restrained yet positive attitude can be seen in the common belief promoted by them that AI-informed applications should not replace in-person mental health care, but be used as an addition to it.

As Zhang et al. (2023) and Sharif et al., (2025) found, mental health professionals expressed a rather low level of familiarity with AI technologies, which resulted in a certain level of scepticism and apprehension regarding their inclusion in care. The interviews showed that clinicians were interested in learning about AI but did not have well-organised training, mentorship, and available educational resources to develop confidence in AI use. The potential of AI to depersonalise care was not the only issue raised, as there was concern that a perceived mechanistic system would not be trusted by clients themselves. Educator scientists were more open to it, but practising clinicians tended to focus on the necessity of reassurance provided by evidence, training, and peer-led programmes that could normalise the use of AI. It was also reported by Zhang et al. (2023) and Sharif et al., (2025) that insufficient funding, existing clinical priorities, and lack of standardised accreditation models are major obstacles to integration of AI. Experts expressed serious concerns regarding biases in AI datasets, emphasising that their use of poorly validated or context-specific algorithms may keep reinforcing disparities in the provision of care. The need to incorporate federally approved and standardised tools is a cry of desperation in seeking governance tools that hold accountability, transparency, quality assurance. In the absence of such frameworks, clinicians do not trust or simply encourage AI tools in the practise setting.

There are a number of limitations associated with this systematic review. To begin with, only English studies that were published in the last 2018–2025 were used, which could have given a bias against other relevant evidence in the past or non-English literature, which could affect the thoroughness of findings. Second, various study designs were evaluated; however, the diversity of populations, interventions, and outcomes measures restricted the opportunity to conduct a meta-analysis and make the studies less comparable. Third, there is still the risk of publication bias since grey literature and unpublished research were not included which may have biassed the results towards better findings. Nevertheless, the limitations notwithstanding, the strengths of the review are a clear and strict screening process that narrowed the initial records to 20 high-quality ones, the application of Mixed Methods Appraisal Tool (MMAT) to guarantee the methodological credibility of the studies of all types, and the synthesis of quantitative, qualitative, and mixed-methods findings to present a comprehensive synthesis of both patient and clinician perspectives.

This review has significant implications of research, practise, and policy. Among researchers, there is an urgent necessity of large-scale randomised controlled trials that not only evaluate the symptom reduction but also such outcomes as long-term assessments as therapeutic alliance and relapse prevention. To clinicians, digital literacy and the inclusion of AI use in practise guidelines should be the focus of training programmes to overcome scepticism and improve confidence. To policymakers, the findings illustrate the need to come up with governance structures able to focus on data sovereignty, biassed algorithms and answerability when it comes to equitable access in diverse populations. Through these dimensions, AI applications could transform into experimental tools to mainstream, ethically sound, and clinically useful aids in mental health care.