Introduction
Recent advances in artificial intelligence (AI), particularly conversational agents and large language models, have generated growing interest in their application to mental health care. AI-based tools have demonstrated potential to provide scalable support and facilitate self-disclosure of sensitive information. Emerging research suggests that AI may support screening, symptom monitoring, and low-intensity therapeutic interventions, with some evidence of effectiveness in reducing symptoms of depression and anxiety (Yoon et al., Reference Yoon, An, Choi, Chang, Jang and Jeon2025). More broadly, AI has been proposed as a mechanism to extend care beyond traditional clinical encounters by offering continuous, on-demand support in the context of limited healthcare resources (Balcombe, Reference Balcombe2023).
For individuals with schizophrenia, AI may offer a novel channel to express experiences that are difficult to share with clinicians or caregivers. However, its application presents unique challenges. Concerns remain about the suitability of AI in this population, particularly due to the potential for ‘AI psychosis’ (Hudon & Stip, Reference Hudon and Stip2025), surveillance fears, or confusion about the nature of AI itself. Prior research has identified paranoid ideation as a barrier to engagement with digital mental health tools in individuals with psychosis, with some reporting that technology use may exacerbate paranoid symptoms (Eisner, Berry, & Bucci, Reference Eisner, Berry and Bucci2023).
Little is known about how people with schizophrenia, their families, and their providers perceive AI-enabled tools in the context of schizophrenia care. To ensure ethical, acceptable, and effective implementation, early-stage research must engage these stakeholders directly. Understanding their perspectives can illuminate potential risks, identify facilitators for adoption, and shape the design of treatment interventions. Through structured focus groups and surveys, this study explored the attitudes, expectations, and concerns of patients, caregivers, and providers regarding AI in mental health care among a small sample recruited from a single academic health center in New York City. These exploratory findings will guide the development of AI-based tools for individuals with mental illness and inform future, larger studies.
Methods
This study employed an exploratory, cross-sectional, qualitative design consisting of focus group discussions and surveys in a large, diverse academic health center in New York City. The primary objective was to assess stakeholder perceptions of a potential AI conversational companion tool designed to support individuals with mental illness. The study was approved by the Institutional Review Board at the Icahn School of Medicine at Mount Sinai (#25–00117).
Recruitment occurred through flyers, provider referrals, and recontact of participants from other internal studies. Participants were recruited into one of three stakeholder groups. Patient groups consisted of individuals diagnosed with a schizophrenia spectrum disorder and stable psychiatric status at the time of enrollment. Diagnoses were verified through medical record review or documentation from a treating physician, and stability was defined by current outpatient treatment adherence and no recent psychiatric hospitalization. Provider groups consisted of mental health care professionals. Caregiver groups consisted of individuals with experience supporting a person with a mental health diagnosis. Individuals who were under the age of 18, non-English speaking, incarcerated, or legally designated as wards of the state were excluded. No previous AI/digital tool experience was required for enrollment.
The research team obtained informed consent following standardized scripts and workflows. Prior to the focus group, participants completed a basic demographic form. Focus groups were facilitated by up to three study team members and were held either in person at Mount Sinai or via Zoom. A semi-structured focus group discussion guide was developed, consisting of predefined questions and prompts covering key topic areas such as experiences with mental health care, familiarity with AI, trust, privacy, and perceived usefulness. Each session started with a standardized review of AI, including examples of current healthcare applications. Participants discussed their familiarity with existing AI tools, like publicly available chatbots. The discussion shifted to a hypothetical HIPAA-compliant conversational AI companion designed to provide mental health support. Facilitators presented several potential implementation scenarios. Participants were asked to discuss perceived benefits, risks, and acceptability of these scenarios. Participants were asked not to disclose identifiable information outside of the group to mitigate potential privacy risks. Sessions lasted up to 90 minutes and were audio-recorded, then manually transcribed. Participants received $25 as compensation for completing the focus group. One day after the session, participants received an exit survey assessing their experience.
Audio recordings from focus groups were transcribed manually. Transcriptions were stripped of identifiers, then reviewed for reflexive thematic analysis (Ahmed et al., Reference Ahmed, Mohammed, Nashwan, Ibrahim, Abdalla, M. Ameen and Khdhir2025). Two members of the research team independently read each transcript multiple times and generated initial inductive codes that captured salient ideas and patterns within the data. The coders then compared and discussed their coding, consolidating related codes into preliminary themes. When discrepancies in interpretation arose, the coders discussed their perspectives to determine whether the differing interpretations enriched the analysis and should be incorporated into the developing themes or whether a consensus interpretation was more appropriate. The preliminary themes were subsequently reviewed and refined with two additional study team members to ensure coherence, consistency, and faithful representation of participants’ perspectives. Through iterative team discussions, overarching themes were finalized. Because an inductive analytic approach was used, themes emerged from the data rather than being defined prior.
Results
A total of 28 participants across 12 focus groups completed the study: 17 patients across 7 groups, 7 caregivers across 3 groups, and 4 providers across 2 groups. Given the modest sample size, the findings are considered exploratory. Demographic characteristics across the cohort are outlined in Table 1.
Participant characteristics by stakeholder group
Table 1. Long description
A table with four columns: Characteristic, Patients n = 17, Caregivers n = 7, and Providers n = 4.
* Age, mean S D: Patients 45.5 13.9, Caregivers 61.0 16.5, Providers 33.5 3.8.
* Gender, n percent:
- Male: Patients 10 58.8 percent, Caregivers 0 0 percent, Providers 0 0 percent.
- Female: Patients 6 35.3 percent, Caregivers 7 100 percent, Providers 4 100 percent.
- Other or Prefer not to say: Patients 1 5.9 percent, Caregivers 0 0 percent, Providers 0 0 percent.
* Ethnicity, n percent:
- Hispanic or Latino: Patients 5 29.4 percent, Caregivers 1 14.3 percent, Providers 0 0 percent.
- Not Hispanic or Latino: Patients 10 58.8 percent, Caregivers 6 85.7 percent, Providers 4 100 percent.
- Unknown or Prefer not to say: Patients 2 11.8 percent, Caregivers 0 0 percent, Providers 0 0 percent.
* Race, n percent:
- American Indian or Alaska Native: Patients 1 5.9 percent.
- Asian: Patients 2 11.8 percent, Providers 1 35.0 percent.
- Black or African American: Patients 6 35.3 percent, Caregivers 3 42.9 percent.
- White: Patients 4 23.5 percent, Caregivers 3 42.9 percent, Providers 3 75 percent.
- Other: Patients 2 11.8 percent, Caregivers 1 11.8 percent.
- Prefer not to say: Patients 2 11.8 percent.
* Employment status, n percent:
- Not employed: Patients 12 70.6 percent, Caregivers 1 14.3 percent.
- Part-time: Patients 2 11.8 percent, Providers 1 25.0 percent.
- Full-time: Caregivers 3 42.9 percent, Providers 3 75.0 percent.
- Retired: Patients 1 5.9 percent, Caregivers 3 42.9 percent.
- Other: Patients 2 11.8 percent.
* Education level, n percent:
- Some high school: Patients 1 5.9 percent.
- High school diploma or G E D: Patients 2 11.8 percent, Caregivers 1 14.3 percent.
- Some college: Patients 6 35.3 percent, Caregivers 3 42.9 percent.
- College degree: Patients 5 29.4 percent, Caregivers 1 14.3 percent.
- Some graduate school: Patients 2 11.8 percent.
- Graduate or professional degree: Patients 1 5.9 percent, Caregivers 2 28.6 percent, Providers 4 100 percent.
Thematic analysis
Five core themes were identified. Representative participant quotations are outlined in Table 2.
Themes with representative quotations
Table 2. Long description
The table consists of four columns: Theme, Description, Quote, and Stakeholder.
* Row 1: Theme is Available when no one else is. Description: A I can fill the gap between episodic care and continuous illness. Quote: ‘If I couldn’t get my doctor or therapist on the phone at the time and I just wanted to just hypothetically throw something out there just to get feedback from… something that’s… smart in that area, like A I’. Stakeholder: Patient.
* Row 2: Theme is What would it do? Description: A I’s role in mental health care remains ambiguous. Quote: ‘Would [the tool] be more of a regulating emotions kind of thing, or therapeutic? [Would we be] trying to provide therapy using A I? I don’t know’. Stakeholder: Provider.
* Row 3: Theme is How can i trust it? Description: Trust is the central barrier to adoption, shaped by privacy concerns and schizophrenia-specific vulnerabilities. Quote: ‘I would be scared that [A I] would leak [the] question that I’m asking to their mainframe and somehow have information on me in a computer. I don’t know how… the intelligence would interact with the information that I give it’. Stakeholder: Patient.
* Row 4: Theme is A tool, not a therapist. Description: A I is acceptable only as a supplement to, not a replacement, for human care. Quote: ’[A I] shouldn’t replace an actual psychiatrist… I don’t think we should make the patient believe they’re talking to a therapist when they’re not’. Stakeholder: Caregiver.
* Row 5: Theme is Keep it low-risk. Description: Acceptable uses of A I are constrained to low-risk, supportive functions. Quote: ‘I usually use [A I] for very low-level things… if it was something serious… I wouldn’t use [A I] as the answer’. Stakeholder: Provider.
Theme 1: Available when no one else is
Participants across all three stakeholder groups described mental health care as episodic and insufficient for managing a continuous illness. Within this context, there was broad agreement that an AI conversational agent was conceptualized as a potential tool to fill the gap between episodic care and continuous illness by providing ongoing support outside of clinical encounters and supporting ongoing self-management.
Patients described managing symptoms primarily outside of clinical encounters, relying on self-directed strategies such as social interaction, music, journaling, podcasts, and online content to cope with day-to-day experiences. Within this context, some patients expressed interest in AI as an additional source of support during moments when clinicians were unavailable. They described scenarios in which they might seek feedback or reassurance in real time, particularly when experiencing symptoms between visits. AI was viewed as a potentially helpful, on-demand tool in these situations, even if its benefit was limited.
Caregivers similarly described the need for continuous support in the context of unpredictable and fluctuating symptoms, and some viewed AI as a potential adjunct during periods when they were not available.
Providers echoed this perspective, noting that AI could offer limited between-visit support, while emphasizing the importance of maintaining clear boundaries.
Theme 2: What would it do?
Participants across the stakeholder groups demonstrated heterogeneous familiarity with AI technologies. While some reported regular use of tools such as ChatGPT or AI-assisted search engines, others had minimal exposure or expressed uncertainty about how AI functions. Participants reported currently using AI conversational tools mostly for task-oriented contexts (e.g. trip planning, information retrieval, studying, and general problem-solving). Participants consistently described these uses as practical and low risk rather than personal or emotionally driven. Across groups, there was uncertainty about whether AI conversational tools should function as a companion, a practical tool, or an extension of the care system.
Patients expressed cautious openness to using AI for support, particularly in situations where reaching out to another person felt difficult. Some described the potential for AI to serve as a nonjudgmental listener or to facilitate discussion of sensitive topics.
Caregivers viewed AI as a potential Supplementary Material support system that could provide reassurance or assistance during periods when they were unavailable, particularly given the unpredictability of symptoms.
Providers were the most cautious and clearly differentiated between acceptable and unacceptable uses. AI was generally viewed as appropriate for administrative or structured tasks, but not for therapeutic engagement. Several providers emphasized that current use of AI in clinical contexts was limited to functions such as documentation or organization.
Theme 3: How can I trust it?
Participants across all three stakeholder groups expressed concerns about data privacy, security, and lack of regulation. Although privacy concerns were nearly universal, the reasons underlying those concerns differed by stakeholder group. Many described hesitation to share personal information with an AI conversational tool or general AI systems due to uncertainty about how data would be stored, used, or accessed.
Patients expressed discomfort with the idea of AI systems that are continuously present or ‘always listening’. Some participants described AI as intrusive or unsettling, while others expressed general unease with its role in everyday technology.
Caregivers expressed concern that AI could interact negatively with psychosis symptoms (e.g. delusions), particularly if individuals misinterpret the system or its intentions.
Providers explicitly linked their concerns to clinical features of schizophrenia (e.g. paranoia). They emphasized that trust is already a central challenge in therapeutic relationships, and that introducing AI systems could further complicate this dynamic.
Theme 4: A tool, not a therapist
Participants across all stakeholder groups had a strong consensus that AI conversational tools are acceptable only as a supplement to, not a replacement, for human care.
Patients valued relationships with clinicians, peers, and family members, even when expressing openness to using AI for certain types of support. Human interaction was viewed as essential to understanding and managing their condition.
Caregivers raised concerns that reliance on AI could contribute to increased isolation or disengagement from real-world relationships. Given existing challenges related to social withdrawal, this was viewed as a potential risk.
Providers were particularly explicit in rejecting the idea of AI replacing clinical care. They emphasized that therapeutic relationships, emotional understanding, and clinical judgment could not be replicated by automated systems.
Theme 5: Keep it low-risk
Participants across stakeholder groups converged on a relatively narrow set of acceptable roles for AI in mental health care. AI tools were considered most appropriate for supportive functions, such as reminders, scheduling, psychoeducation, and journaling.
Patients expressed cautious openness to using AI for low-intensity support or reflective use, particularly in structured or self-guided contexts. While some participants described potential benefits in using AI for practical tasks or limited emotional support, they were generally hesitant to rely on AI for more complex or clinically meaningful aspects of care.
Caregivers similarly emphasized the importance of limiting AI to supportive roles. Rather than rejecting AI outright, they expressed concern about how it might be used in more complex or sensitive situations, particularly in the context of psychotic symptoms, and emphasized the need for clear boundaries and appropriate use.
Providers reinforced these boundaries more explicitly. They emphasized the importance of clearly defining the scope of AI systems and ensuring appropriate safeguards and integration into existing care structures. Many described current clinical use of AI as limited to administrative functions.
Exit survey findings
Twenty-six participants completed the study exit survey. Overall satisfaction with the focus groups was high, with most participants rating their experience as excellent. Most respondents felt that the discussion was relevant to their experiences and were comfortable expressing their opinions.
Participants reported that the groups increased their interest in the role of AI in mental health care. Several respondents indicated that the discussion provided new perspectives on how AI might support mental health interventions, including the potential for AI systems to tailor support based on an individual’s behavior, preferences, and past interactions. Some participants expressed curiosity about the future development of AI-based therapeutic tools, while others emphasized that AI technologies remain in early stages of development and require further refinement before widespread clinical use.
Several participants reported that the discussion influenced their perspective on AI in mental health. Some providers and participants who were initially skeptical of AI noted that the discussion broadened their views about potential applications of AI for individuals with schizophrenia if implemented carefully and with appropriate safeguards. However, several respondents indicated that their views remained unchanged.
Participants also raised several ethical and implementation considerations. These included concerns about accountability if AI systems provide harmful advice or fail to recognize crises, the importance of human oversight and safety monitoring, and the need for further research on how AI tools could effectively support individuals with schizophrenia and their caregivers. Caregivers highlighted the significant unmet needs faced by families caring for individuals with schizophrenia and expressed interest in innovative approaches that could improve access to support.
Finally, participants reported a variety of motivations for participating in the focus groups, including interest in learning about emerging AI technologies, contributing to mental health research, and sharing personal or professional experiences related to schizophrenia and mental health care.
Discussion
This report presents, for the first time in the medical literature, the perspectives of individuals with schizophrenia on the use of AI in clinical care. Across groups, findings revealed consistent tension: the proposed conversational AI tool was perceived as both a potential therapeutic bridge and a psychopathological risk. On one hand, AI may reduce barriers to disclosure, provide nonjudgmental support, and increase access to care between clinical encounters. On the other hand, its constant presence may exacerbate mistrust or reinforce delusional thinking. This paradox represents a central challenge for the development of AI-based interventions in mental health care, particularly with respect to psychotic disorders. Participants also emphasized the irreplaceable role of human relationships in providing emotional understanding, clinical judgment, and trust.
Several limitations should be considered. The sample size was small, and participants were recruited from a single site, limiting the generalizability of findings. Thus, findings presented here are exploratory and do not necessarily reflect broader consensus. Selection bias may have caused nonrandom inclusion of individuals with greater interest in mental health care or technology. Focus group dynamics and social desirability bias may have influenced participant responses, potentially limiting the expression of dissenting or highly sensitive views. Finally, the research team’s thematic analysis may reflect their own perspectives and assumptions.
Despite these limitations, this exploratory study presents important preliminary insights into the implications for the design and implementation of AI-based mental health tools. First, they must be explicitly designed to support, rather than replace, existing care structures. Second, building trust with users is critical to ensure adoption. Finally, interventions must account for the ways in which AI may be interpreted within the context of psychosis, including the potential for misattribution of intent. While AI tools may offer meaningful opportunities to extend care, their successful implementation in schizophrenia will require careful alignment with the clinical realities of the disorder. Future research should prioritize co-design with patients, caregivers, and clinicians, as well as empirical evaluation of how AI systems are perceived and used in real-world contexts.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726105728.
Acknowledgments
All study participants are acknowledged for their contribution to science. This work was led at the Icahn School of Medicine at Mount Sinai by the Jeff and Lisa Blau Adolescent Consultation Center for Resilience and Treatment.
Funding statement
This work was funded by the Icahn School of Medicine at Mount Sinai through the Jeff and Lisa Blau Adolescent Consultation Center for Resilience and Treatment.
Competing interests
The authors declare none.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
