Artificial intelligence (AI) technologies are increasingly being developed to predict, diagnose, and treat disease. While still in the formative stages of development, AI implementation is increasing in pace across health systems and will soon be widespread. Within mental healthcare, a range of datasets has been used, from electronic health records and imaging data to social media and novel activity and mood monitoring systems, to predict outcomes (1). AI has also been used in a generative manner to summarize information, provide feedback for patient questions, or automate documentation. For example, predictive AI can evaluate the risk of postpartum depression in settings where screening is limited, and generative AI can deliver cognitive-behavioral therapy via chatbot (2, 3). Because rich information is stored in clinical encounter notes so often in mental health contexts, much of this work also leverages natural language processing, a set of computational methods to extract patterns from large sources of textual data (1). In the context of limited mental health resources and mental health professionals, AI technologies provide the opportunity to expand access for patients and offload the tasks clinicians find most burdensome so they may instead focus on the patient relationship that is so unique and essential for psychiatry (4).

Patient perspectives on the use of AI in mental healthcare are critical but missing. As patient engagement and shared decision-making continue to rise in importance, it is likely that patients will interact with AI-generated risk prediction models during the course of their care (5). Moreover, AI-informed chatbots are a potential modality for delivering therapy (3). In addition to gathering general perspectives, it will be important to examine differences in perspectives based on experiences with health and the healthcare system, such as experiences with mental healthcare, implicit bias resulting from one's identity, or social determinants of health (SDOH) that influence access to care and health outcomes.

Moreover, important ethical questions arise when using AI in mental healthcare relating to issues of bias, privacy, autonomy, and distributive justice (6). For example, the broad range of health-related data that may be used to train AI algorithms may alarm patients and undermine trust. For example, mental health researchers are already attempting to correlate keystrokes and voice data captured from smartphones to mood disorders (7), which patients may be unaware of. Moreover, these datasets may have systematic biases that perpetuate inequities in who is able to access mental healthcare or experience positive mental health outcomes (8). AI could also lead to harm when instructions on appropriate reliance are not communicated to patients; for example, researchers have raised concerns about patients developing a perceived therapeutic alliance with chatbots that could cause them to over-rely on chatbot guidance or undermine the therapeutic alliance they have with their mental health professionals (3). A synthesis of the ethics literature on responsible AI (6), consumer-generated data (9), AI in psychiatry (10), and maternal health (11, 12) reveals six important constructs to be considered in this context: autonomy, beneficence/non-maleficence, justice, trust, privacy, and transparency.

However, to date, AI research has focused almost exclusively on presenting model output and fostering trust in AI among clinicians (13). The information needs and ethical considerations among patients likely differ from those of clinicians, but may not be the same for all patients; they may differ by a patient's mental health history, demographic variables, and SDOH (14). Therefore it is important to provide insight to the psychiatry community regarding patients' nuanced perspectives on AI, so that it may be effectively integrated into clinical care in a way that does not degrade the patient relationship. Several prior studies have reported on patients' perceptions of AI in healthcare broadly, which have included a range of concerns and highlighted the importance of understanding and integrating patient views into AI development and deployment (15–17). Within mental healthcare specifically, patient perspectives on AI are comparatively understudied, but early research suggests patients are optimistic. For example, a recent scoping review of patient attitudes towards chatbots in mental healthcare reported that patients have positive perceptions but require high-quality, guideline-concordant, trustworthy and personalized interactions to feel comfortable using them in their care (18). However, differences in attitudes towards AI in mental healthcare by clinical or sociodemographic characteristics remain understudied.

The primary objective of this study was to determine whether a history of mental illness or current mental health status is associated with differences in attitudes towards AI being used in mental healthcare. Secondarily, we also aimed to investigate differences by demographic characteristics (age, gender, race, and ethnicity) and SDOH (financial resources, education, health literacy, and subjective numeracy).

In this analysis of 500 U.S. adults, significant differences in perceptions of AI in mental healthcare were found by mental health history, current mental health status, demographic characteristics, and SDOH. In general, participants with a history of mental illness, members of older generations (e.g., Baby Boomers), those identifying as women, Asian, Black, Latinx, those reporting “not enough” financial resources, those with less than a Bachelor's degree in education, and those with lower subjective numeracy expressed greater discomfort with various potential uses of AI in mental healthcare and expressed greater concerns about potential harms from AI in mental healthcare. Similarly, more participants with a history of mental illness, who rated their current mental health as poor or fair, members of older generations (e.g., Baby Boomers), those who identified as women, and those reporting “not enough” financial resources felt at least one of the bioethical constructs were “very important.” Paradoxically, participants with inadequate health literacy were more comfortable with AI, less concerned about its potential harms, and rated some of the bioethical constructs as “less important.”

This study is among the first to describe differences in attitudes towards AI by mental health history; to our knowledge, this has not been previously reported. However, our findings do align with prior research on patient perspectives on AI in healthcare more broadly, in which individuals reporting good to excellent health are less concerned and more comfortable with AI in healthcare compared to those with poorer health status (24). Our findings on differences in attitudes by other characteristics generally align with the small but growing body of literature in the area. Specifically, in prior studies, women, older and racially and ethnically minoritized adults, and those with lower education, socioeconomic status, and/or digital literacy, express greater concern about the use of AI in mental healthcare, particularly surrounding privacy, accuracy, lack of transparency, loss of human empathy and connection, and financial costs (24, 25).

These findings highlight significant concerns among multiple vulnerable populations about the use of AI in mental healthcare. At the same time, they are unsurprising in the context of historical failures to consider and integrate the needs of vulnerable populations when emerging technologies are developed and integrated into healthcare systems (26). For example, there are known and persistent disparities in the adoption and use of patient portals (27), telehealth (28), and wearable technologies (29, 30) by many of the groups of patients who expressed concerns and discomfort with AI in this study. In fact, the finding that lower health literacy participants did not perceive these threats suggests that these concerns could be realized, as the patients least able to find, understand, and use health-related information are the least aware or concerned about AI-related harms. Future research should examine the intersectionality of multiple vulnerabilities and its impact on perceptions of AI in mental healthcare, given that important differences along a matrix of personal characteristics may exist that were not uncovered in this analysis.

In this study, participants’ perceptions of who is responsible or to blame for harm resulting from AI, who is responsible for ensuring the safety of AI, and the impacts on trust with mental health professionals were inconsistent across vulnerable groups. In general, more participants who identified as Asian or Black, reported “not enough” financial resources, or had inadequate health literacy assigned responsibility or blame to hospitals or clinics, regulatory governmental agencies, mental health professionals, and/or other groups for harm resulting from AI. More participants who rated their current mental health as poor or fair, and those with lower subjective numeracy, would also reduce trust in their mental health professional if their assessment disagreed with AI. However, more participants with lower education were unsure about who was responsible, fewer participants with a history of mental illness would assign responsibility or blame to hospitals or clinics, regulatory governmental agencies, and AI companies, and fewer participants with lower subjective numeracy would blame AI companies for harms caused by AI.

Given the speed with which AI is being embraced across many disciplines within healthcare, including mental healthcare, and the perspective among some groups of patients that mental health professionals are responsible, hold blame, or are less trustworthy as a result of harm or conflicting information from AI, AI literacy among clinicians is imperative. AI literacy calls for clinicians to understand data governance, basic statistics, data visualization, and the impact of AI on clinical processes (31). Mental health professionals will need to be able to continue practicing in an effective, efficient, safe, and equitable way in the context of AI-enriched healthcare systems. Perhaps more importantly, it will be essential that AI will not be used, inadvertently or otherwise, to extend existing biases that create disparities between patients in timely and accurate mental health diagnoses and care. Their perspectives will also be critical as those developing AI systems for mental healthcare aim to create more accurate systems, for example using deep learning methods, while combating the “black box” problem of limited explainability (1, 3). Similarly, reviews have demonstrated that AI in mental healthcare is in an early, proof-of-concept stage; in future work, the engagement of mental health professionals, along with patients and caregivers, will be critical to successful clinical implementation in ways that do not overburden clinicians (1).

AI-related competencies for clinicians have been proposed; Russell et al. (32) identified six competencies including basic knowledge of AI, social and ethical implications, workflow analysis for AI-enabled systems, evidence-based evaluation of AI-based tools, conducting AI-enhanced clinical encounters, and continuing education and quality improvement initiatives related to AI. While both continuing professional development (32) and medical school education (33) have been proposed venues for implementing these competencies, strategies for training the mental health workforce for impending AI interactions and influences remain to be determined (4).

Additionally, the findings of this study suggest mental health professionals may need to be especially prepared for vulnerable patient populations to express greater reticence about AI being used in their care. These conversations will be critical as failure to engage these populations in technologies that have a demonstrated benefit with respect to mental health outcomes could actually widen disparities (26). For example, we found that older generations expressed greater concern and discomfort with specific uses of AI in mental healthcare. This is salient because of the known underdiagnosis of mental illnesses (particularly racial and ethnic minorities), distinct symptom presentation, and lower utilization of mental health services among older adults (34). While AI may be poised to address these issues through improved disease detection, caregiver support, and novel solutions to reduce loneliness, failure to engage older adults could further exacerbate these disparities (34).

Limitations of this study include the limited generalizability of the sample, especially with respect to race and ethnicity. Oversampling of racially and ethnically minoritized adults may be important in future work to better understand differences in attitudes toward AI. Moreover, we only examined overall differences and did not apply corrections to account for multiple comparisons. Further, we did not report on group post-hoc comparisons due to small sample sizes in some cells.

Multiple vulnerable subgroups of U.S. adults perceive potential harms related to AI being used in mental healthcare, place importance on upholding bioethical constructs related to AI use in mental healthcare, and would blame or reduce trust in multiple parties, including mental healthcare professionals, if harm or conflicting assessments resulted from AI. Future research examining strategies to uphold bioethical constructs in AI implementation and support clinician AI literacy is critical for optimal patient and clinician interactions with AI in the future of mental healthcare.