What Is in There for Artificial Intelligence to Support Mental Health Care for Persons with Serious Mental Illness? Opportunities and Challenges

The paper situates AI as a transformative technology in healthcare, noting major advances in somatic fields but unique challenges for integrating AI into mental health care for serious mental illness (SMI: schizoaffective, bipolar, and major depressive disorders). Challenges include the subjective and nuanced nature of mental health, historical mistrust due to non–trauma-informed systems, and societal stigma. Despite potential benefits—broader access, early intervention, and personalized treatment—ethical and regulatory concerns persist and evidence specific to SMI remains scarce. The study’s purpose is to understand how diverse mental health stakeholders perceive opportunities and challenges of integrating AI to support persons with SMI, to inform responsible development and implementation.

Design: Qualitative study using individual semi-structured interviews with multiple stakeholder groups in Norway (government, hospitals, municipalities, universities/research institutions, health industry/clusters, user organization). Sampling and recruitment: Purposive sampling based on experience with digital health in mental health. Ethics: Written informed consent obtained; participation voluntary with right to withdraw. The overall PhD project was assessed by the Norwegian Agency for Shared Services in Education and Research (SIKT; ref. 269350). Data collection: Interviews conducted via Microsoft Teams in autumn 2024; audio recorded and transcribed verbatim. Interview guide covered: (1) ways AI can support persons with SMI; (2) AI as a feature in digital mental health solutions (e.g., apps); (3) how AI may better meet needs or improve quality of life; (4) concerns about using AI to support persons with SMI. Analysis: Thematic analysis (NVivo 14 and Microsoft Excel) following six steps—familiarization, generating initial codes, generating themes, reviewing themes, defining/naming themes, writing up. Sentiment analysis performed using NVivo 14 Autocode Wizard to classify sentiment (positive, moderately positive, moderately negative, negative) based on machine learning applied to words/phrases per predefined criteria. Note: the Autocode Wizard analyzes sentiment of individual words and does not classify content by sentiment or rate on a Likert scale.

Sample characteristics: 22 informants; majority male (12/22, 55%); most aged 40–59 (18/22, 82%); over half with healthcare backgrounds (13/22, 59%). Sectors represented: government (2/22, 9%), hospitals (6/22, 27%), municipalities (6/22, 27%), universities/research institutions (6/22, 27%), health industry/clusters (3/22, 14%), user organizations (1, 5%). Sentiment analysis: 0 positive, 5 moderately positive (25%), 15 moderately negative (50%), 5 negative (25%); three-quarters showed moderately negative or negative sentiment toward AI supporting persons with SMI. Themes (4) and subthemes (6): 1) When AI meets SMI: (a) Break the downward spiral—AI pattern recognition and chatbots could help interrupt negative routines, support goal-setting and provide timely prompts; (b) To connect, or to be lost in connection?—risk of harmful guidance without professional oversight; potential to exacerbate loneliness and isolation; concerns for older adults’ ability to discern AI vs human interactions and accept technology. 2) Human-centered AI for humanity: (c) Personal adaptation and safety—need accessibility and adaptation to age, function, impairments, and local linguistic/cultural contexts; address data bias and underrepresentation of vulnerable populations; (d) Maintain human touch—AI cannot replace therapeutic relationships; trust, being seen and understood remain central to recovery. 3) AI to improve service delivery for SMI: (e) Enhance efficiency of clinical decision support—AI-generated alerts/thresholds for symptoms, monitoring while awaiting care, integration of physiological parameters (e.g., heart rate, breathing, movement) to inform assessment; (f) Better resource management—triage and allocation to prioritize the most seriously ill, remote follow-ups/self-help support to reduce unnecessary visits. 4) Building AI competence to support SMI: Need to increase AI literacy among users, clinicians, and leaders; actively use and learn AI tools; current evidence in mental health (especially for severe/acute cases) remains limited, fostering skepticism and highlighting the need for further research.

Stakeholder perspectives indicate that AI holds promise for augmenting mental health services for SMI—improving decision support and resource allocation—while underscoring that AI cannot replicate human empathy or professional judgment. The findings directly address the research question by mapping concrete opportunities (e.g., real-time monitoring, thresholds, triage, personalization) and significant challenges (e.g., potential for harmful advice, exacerbation of isolation, digital divide, data bias, cultural/linguistic fit). Ethical and regulatory considerations are central; the authors highlight the importance of explainable and transparent AI, aligned with frameworks like the EU AI Act, to ensure safety, accountability, and user autonomy in high-risk contexts such as SMI. Addressing digital inclusion is critical: accessible, universally designed solutions and engagement of peer support workers can help tailor AI to diverse user needs and foster trust. AI may enhance, not replace, human empathy when used thoughtfully; early clinical use cases (e.g., drafting mental health documentation) suggest feasible entry points that complement human care. Overall, realizing AI’s potential for SMI requires robust evidence, human oversight, cultural/personal adaptation, and clear governance.

AI offers substantial opportunities to support persons with serious mental illness, particularly in optimizing service delivery, clinical decision support, and personalized assistance. However, integration must prioritize human-centered care, maintain therapeutic relationships, and ensure safety, transparency, and fairness. The study contributes stakeholder-derived insights into opportunities, risks, and prerequisites for responsible AI use in SMI care. Future work should: (1) build evidence for efficacy and safety in SMI populations; (2) develop explainable, culturally and linguistically adapted AI; (3) address digital divide and accessibility via universal design; (4) engage users and peer support workers in co-design; and (5) establish ethical and regulatory frameworks and implementation guidance to support trustworthy deployment.

The sentiment analysis relied on NVivo’s Autocode Wizard, which assesses sentiment at the word/phrase level without contextual interpretation or Likert-scale ratings, potentially misrepresenting overall tone. A discrepancy between the predominantly negative sentiment output and more positive qualitative interpretations may be attributable to heterogeneity among informants (differences in education, profession, and patient contact). The perspectives reflect Norwegian stakeholders interviewed in autumn 2024, which may limit transferability to other settings.