The rapid advancements in artificial intelligence (AI) have led to significant transformations across various sectors, including healthcare. AI's capacity to process extensive datasets and identify complex patterns has proven beneficial in fields like radiology, pulmonology, and dermatology. However, applying AI to mental healthcare, especially for individuals with serious mental illness (SMI), presents unique challenges. SMI, encompassing conditions such as schizoaffective disorders, bipolar disorders, and major depressive disorders, involves subjective and nuanced conditions. The historical context of mental healthcare, marked by practices like asylum confinement and lobotomies, has fostered mistrust in scientific and medical interventions, compounding the societal stigma surrounding mental illness. While AI offers the potential to expand access to mental health services, facilitate early interventions, and personalize treatment, its integration raises significant ethical and regulatory concerns. Currently, there is a scarcity of robust evidence regarding AI integration in mental healthcare, particularly for those with SMI. Given the escalating global burden of mental illness and resource limitations, exploring AI's potential and ensuring its responsible integration is crucial. This study aims to understand the perceptions of mental health stakeholders regarding the opportunities and challenges of integrating AI to support mental health care for individuals with SMI.

The existing literature highlights the potential benefits and ethical challenges of using AI in mental healthcare. Studies such as Graham et al. (2019) provide overviews of AI applications for mental health and mental illnesses, while Bajwa et al. (2021) discuss the transformative potential of AI in healthcare broadly. The successful integration of AI in somatic care, as demonstrated by Jha and Topol (2016) in radiology and pathology, contrasts with the unique complexities of applying AI to mental health conditions. Wang et al. (2023) explore users' experiences with online access to electronic health records in both mental and somatic healthcare, shedding light on the digital divide and user acceptance challenges. Olawade et al. (2024) and Naik et al. (2022) delve into the ethical and legal considerations of AI in healthcare, emphasizing the need for responsible development and implementation. Lee et al. (2021) provide a comprehensive analysis of AI for mental healthcare, including its clinical applications, facilitators, and barriers. These studies underscore the need for further research to understand the effectiveness, safety, and ethical implications of AI in supporting individuals with SMI.

This qualitative study employed individual interviews to gather perspectives from diverse mental health stakeholders in Norway. Participants were purposefully sampled based on their experience with digital health in mental healthcare. The sample included representatives from government agencies, hospitals, municipalities, universities/research institutions, health industry/clusters, and user organizations. Written consent was obtained from all participants, who were assured of their right to withdraw at any time. Data for this study originated from the corresponding author's PhD research and had not been previously published. The Norwegian Agency for Shared Services in Education and Research (SIKT) conducted an ethical review of the PhD study. Interviews were conducted via Microsoft Teams and transcribed verbatim. Data analysis involved thematic analysis using NVivo 14 and Microsoft Excel, following a six-step process: familiarization; generating initial codes; generating themes; reviewing themes; defining and naming themes; and writing up. Sentiment analysis was also performed using the Autocode Wizard in NVivo 14 to gauge the overall tone of the informants' responses (positive, moderately positive, moderately negative, and negative). It's important to note that the Autocode Wizard analyzes sentiment at the word level, not at the level of themes or overall response.

A total of 22 informants participated, with a slight male majority (55%) and most falling within the 40-59 age range (82%). The majority (59%) had healthcare backgrounds (psychiatry, psychology, nursing, social education). Sentiment analysis revealed that 75% of responses expressed moderately negative or negative sentiment towards using AI in supporting persons with SMI. Thematic analysis revealed four main themes: 1) AI's interaction with SMI: While some saw AI as a tool to help individuals break negative patterns and find their voice, concerns arose that AI might worsen social isolation. 2) Human-centered AI: The importance of flexible human care, safe and reliable AI, and personalized adaptation to individual needs was highlighted. 3) Improving service delivery: Enhancements in clinical decision support and resource management were seen as potential benefits, including AI-generated alerts for symptom thresholds and remote monitoring. 4) Building AI competence: The need for increased AI literacy and competence among all stakeholders, including users, professionals, and leaders, was emphasized, along with the need for more research to overcome skepticism and improve trust. Specific concerns included potential amplification of loneliness in already isolated individuals, the risk of AI guiding individuals towards harmful actions due to lack of human oversight, and the digital divide and accessibility challenges for vulnerable populations (e.g., older adults). The study also found that aligning AI with personal adaptations and ensuring safe and reliable AI usage were of high importance, especially given the potential for bias in AI models due to underrepresentation of vulnerable groups in datasets. Maintaining human connection and a trusting relationship between professionals and individuals with SMI was highlighted as crucial for successful AI integration.

The findings underscore the potential benefits of AI in mental healthcare, particularly in optimizing service delivery and resource management, but also highlight the crucial need for human oversight and personalized adaptation. The significant number of negative or moderately negative sentiments expressed by informants indicates a cautious approach toward AI integration is warranted. The discrepancy between the potentially positive tone in qualitative responses and the overall negative sentiment analysis might stem from the diverse backgrounds and experiences of the informants. This study emphasizes the need for a holistic approach that considers the interplay between AI and human connection, addressing concerns about exacerbating social isolation. The results highlight the ethical and practical challenges of AI implementation in mental healthcare, mirroring findings in previous studies. The limitations of AI in replicating human empathy and professional judgment highlight the need for cautious integration and robust ethical guidelines. Addressing the digital divide and promoting accessibility for vulnerable populations are critical for successful implementation. Future AI solutions must ensure culturally appropriate and personalized adaptations to address the diverse needs of individuals with SMI.

AI holds immense potential for supporting individuals with SMI, but its responsible integration requires addressing numerous ethical, regulatory, and practical challenges. The research underscores the importance of maintaining human connection, building AI competence among all stakeholders, and conducting further research to address existing limitations and build trust. Future work should focus on developing human-centered AI solutions that prioritize personal adaptation, accessibility, and safety, while mitigating risks associated with social isolation and algorithmic bias. Establishing clear ethical guidelines and regulatory frameworks, especially around transparency and data privacy, will be essential to ensure the safe and effective use of AI in mental healthcare.

The study's findings are based on qualitative data from a specific sample of Norwegian mental health stakeholders. The generalizability of the findings to other contexts may be limited. The sentiment analysis was conducted at a word level, which might not fully capture the nuances of the informants' overall perspectives. The study did not explicitly explore the experiences of individuals with SMI themselves, which could provide additional insights. Future research should involve a larger and more diverse sample, including perspectives from individuals with SMI, to strengthen the generalizability and validity of the findings.