JMIR Mental Health

From Alliance to Nexus: Rethinking Digital Therapeutic Relationships

2026-07-02T11:45:02-04:00

In traditional human psychotherapy, the therapeutic alliance (TA) is regarded as a fundamental factor that describes the client-therapist relationship, mainly due to strong evidence demonstrating its impact on treatment outcomes regardless of theoretical orientation. More recently, advances in artificial intelligence (AI) and other technologies have led to the emergence of the concept of digital TA, used to characterize the relationship between clients and AI-based therapeutic systems. This approach replicates human dynamics but overlooks key differences between human therapists and digital agents. Prematurely translating the concept of TA into the digital context fails to address issues such as the sycophantic tendencies of current systems and the inherent limitations of algorithmic interaction. We propose the digital therapeutic nexus, a framework that recognizes these differences and provides a set of structured criteria for categorizing digital interactions into 3 progressive levels. This Viewpoint argues that only at the highest level can parallels be drawn to the human TA and stratifies the main risks associated with each nexus level. Transitioning from the concept of alliance to that of a nexus offers a more precise conceptual basis for describing and evaluating digital therapeutic relationships, with implications for research, design, and the ethical development of AI-based mental health interventions.

Governing Ethical Tensions in Youth Digital Mental Health Research

2026-06-30T16:15:10-04:00

As mental health research increasingly aims to generate societal impact, researchers operate at the intersection of innovation and ethical responsibility. Drawing on experiences from the cocreated NEON Young Norway Study on youth recovery narratives, this viewpoint identifies four ethical tensions that arise from the existing governance frameworks in youth digital mental health research: (1) balancing safeguarding against harm with youth participation, (2) protecting privacy without undermining authentic storytelling, (3) governing unpredictable outcomes of cocreated research, and (4) meeting ethical and legal standards while ensuring youth-friendly communication. These tensions highlight limitations in mental health research that adopts participatory and digital approaches, as this often struggles to accommodate iterative designs, narrative data, and cross-sector collaboration. We argue that responsible youth mental health research requires ethics to be understood as a dynamic, participatory practice that supports safe and equitable inclusion, rather than having a focus on risk prevention. Ethical governance, therefore, needs to evolve toward proportionate, context-sensitive approaches that can enable innovation while protecting young people’s rights, agency, and voices.

Use of a Conversational Agent for Training Mental Health Professionals in Suicide Safety Planning: Pilot Feasibility and Acceptability Study

2026-06-30T13:00:22-04:00

Background: Safety planning is recognized as one of the most effective interventions for reducing suicidal behaviors. The quality of safety plans strongly depends on professional training, and traditional methods, such as role-playing, are time-consuming and offer limited opportunities for repetition across diverse patient profiles. Generative artificial intelligence (GenAI) may provide innovative solutions by offering accessible, flexible, and realistic training environments. Objective: This pilot study aimed to evaluate the acceptability and feasibility of a GenAI-based simulator designed to train mental health professionals in safety planning. Methods: Twenty nurses and nursing assistants from psychiatric units in a French university hospital participated in a pre-post, single-session evaluation. After self-rating their ability, competence, and willingness to manage patients experiencing suicidal ideation, participants interacted individually with the text-based simulator for 20 minutes to perform a safety plan with a chatbot, then completed postsimulation acceptability items, and open-ended feedback. Composite scores were computed: acceptability (eg, helpfulness; 0‐40), realism (eg, looking like real interaction with patient; 0‐20), and challenge (eg, emotional challenge; 0‐30). Pre-post changes were tested (Wilcoxon signed-rank test), and age-group comparisons were performed. Results: Acceptability was high (mean 31.9/40, SD 5.3; median 32, IQR 7), realism moderate-to-high (mean 15.1/20, SD 4.1; median 15, IQR 5.25), and challenge manageable (mean 17.0/30, SD 8; median 18, IQR 12.5). Participants rated usefulness (mean 7.65/10, SD 1.57; median 8, IQR 1.57), perceived learning (mean 7.6/10, SD 1.79; median 8, IQR 2), recommendation to use the chatbot for training (mean 8.3/10, SD 1.59; median 9, IQR 2.25), and feedback quality (mean 8.35/10, SD 1.27; median 8.5, IQR 1.25) favorably. Willingness to actively manage patients experiencing suicidal ideation significantly increased postsimulation (.03). Younger participants reported higher acceptability (.04) and realism (.03). Participants reported minimal concerns regarding the simulator’s use. Conclusions: This pilot study demonstrates that a GenAI-based simulator for safety planning is feasible and highly acceptable among experienced mental health professionals. The findings are promising and warrant larger, controlled trials to assess impacts on training effectiveness and patient outcomes.

Coproduction Without Youth? Closing the Participation Gap in Digital Mental Health Research

2026-06-24T17:30:03-04:00

Young people are among the most intensive users of digital and generative artificial intelligence (GenAI)–enabled mental health tools, yet they remain underrepresented in the research and design processes that shape these technologies. Although participatory approaches such as co-design and patient and public involvement are widely endorsed as best practices, youth involvement in digital youth mental health (DYMH) research is often inconsistent, superficial, or limited to late-stage consultation. This participation gap risks producing interventions that are misaligned with young people’s lived experiences, priorities, and vulnerabilities, particularly in the context of rapidly evolving and scalable GenAI systems. This Viewpoint aims to reexamine the underlying drivers of the participation gap in DYMH research; clarify how participation is conceptualized and implemented across disciplines; and propose concrete, actionable recommendations to support more meaningful and consistent youth involvement across the research life cycle. We draw on interdisciplinary literature from digital mental health, human-computer interaction, child-computer interaction, and health research policy. Our Viewpoint integrates conceptual frameworks (eg, Lundy’s model of participation), existing reviews of co-design practices, and emerging evidence on GenAI in mental health. We adopt a life cycle–oriented perspective to examine how youth participation is distributed across stages of research and development, including problem formulation, design, implementation, and evaluation. We identify 3 interrelated drivers of the participation gap. First, conceptual and linguistic fragmentation obscures what participation entails in practice, with terms such as co-design, participatory design, user-centered design, and patient and public involvement used inconsistently across disciplines. Second, youth involvement is uneven across the research life cycle, with participation often concentrated in early ideation or usability testing but largely absent from upstream decision-making and downstream evaluation. Third, institutional barriers—including ethics review processes, consent requirements, funding constraints, and adult-centric research norms—systematically limit meaningful youth partnership. These challenges are amplified in the context of GenAI, where opaque “black box” systems, simulated therapeutic interactions, and rapid deployment cycles introduce distinct risks if youth perspectives are not integrated. We propose a set of minimum expectations to address these gaps, including explicit specification of participatory models, life cycle mapping of youth involvement, reporting of youth influence on decisions, dedicated funding for participation, proportional ethics frameworks, and mechanisms for youth-informed governance of GenAI systems. Closing the participation gap in DYMH research is both an ethical imperative and a practical necessity. Moving beyond aspirational commitments requires embedding youth participation as a standard, resourced, and accountable component of research, design, and governance. In the context of rapidly evolving digital and GenAI technologies, failure to do so risks producing interventions that are scalable but not safe, credible, or responsive to the needs of young people.

Functional Outcome Prediction in Young Adults With Mental Health Symptoms Using Machine Learning and Large Language Models: Longitudinal Observational Study

2026-06-22T14:30:06-04:00

Background: Functional impairments associated with mental health conditions are on the rise. Predicting functional outcomes may improve the targeting of preventive interventions. While prognostic models have primarily focused on psychosis, early recognition services require a transdiagnostic approach. Objective: This study aimed to predict global functioning within a 2-year follow-up using baseline clinical and structural magnetic resonance imaging (MRI) data in a population-based sample of young, help-seeking individuals presenting with affective and anxiety symptoms as well as attention-deficit hyperactivity disorder. Methods: We classified 357 help-seeking individuals aged 18‐35 years recruited from 9 sites as “impaired” (Global Assessment of Functioning [GAF] ≤60; n=228) or “nonimpaired” (GAF>60; n=129) at year 1 and/or year 2 follow-up. GAF classification group status at follow-up was predicted using linear support vector machine (SVM), decision tree, and large language model (LLM) Llama-3 using clinical assessments and/or structural MRI. Leave-one-site-out (SVM) or external sample (LLM) was used for validation. Results: SVM achieved balanced accuracy of 69.2% using clinical features only. Items related to baseline occupational functioning, interpersonal relationships, cognitive functioning, psychotic and affective symptoms, as well as the presence of anxiety disorder, were most predictive. The decision tree further reduced the feature set to 5 predictive items, achieving balanced accuracy of 76.6%. Although amygdala and hippocampal subregions achieved balanced accuracy of 57.1%, structural MRI did not improve the overall prediction. Llama-3 performed comparably well to SVM (balanced accuracy of 72.6%). Conclusions: Machine learning demonstrated good performance in predicting global functioning. Interestingly, the out-of-the-box LLM performed comparably well without being trained or fine-tuned, highlighting the potential of leveraging free-text data for mental health prognosis.

Using AI to Detect Psychosis Relapse: Scoping Review

2026-06-16T14:30:16-04:00

Background: Psychotic disorder represents a leading cause of disability worldwide, and relapse in psychosis is common. Artificial intelligence (AI) is increasingly recognized as a method that could aid clinical monitoring for individuals experiencing psychosis. Objective: This review aims to map the existing literature on AI-based approaches—including machine learning, deep learning, and natural language processing—used to detect relapse in individuals with psychotic disorders. Methods: A systematic search strategy was conducted on PubMed, PsycINFO, and Embase up to January 7, 2026. Observational studies, randomized controlled trials, and quasi-experimental studies that used AI methods to detect relapse in psychosis were eligible for inclusion. Screening and data extraction procedures were conducted by at least 2 reviewers working independently. Findings were extracted, charted, and described using narrative synthesis based on data extraction and consensus meetings with the research team. The scoping review was prospectively registered with the Open Science Framework. Results: Relevant studies identified (N=10) included the use of digital tools such as smartphone- and smartwatch-based monitoring, ecological momentary assessment tools, social media activity, and internet searches. Digital phenotyping via smartphones and wearables emerged as the most common method for data collection. The efficacy of AI models varied with sensitivity (or recall) ranging from 0.25 to 0.77 and specificity (or precision) ranging from 0.06 to 0.88. The reported area under the receiver operating characteristic curve for models ranged from 0.63 to 0.78. AI models were heterogeneous across studies, and most study findings were not replicated. Conclusions: This scoping review highlights both the promise and the current limitations of AI in psychosis relapse detection. Passive digital phenotyping research in the detection of psychosis relapse has progressed, and personalized approaches with individual-level modeling show promise; however, further studies need to include larger numbers of participants and should incorporate methods such as large language models. Future studies will require large collaborations aimed at delivering AI methods for use in real-world clinical practice.

Prevalence and Predictors of Self-Reported Adverse Experiences in Digital Meditation Training: 2 Randomized Controlled Trials

2026-06-12T17:15:13-04:00

Background: Digital meditation-based interventions (MBIs) reach vast global audiences with millions of active users, yet concerns persist about the frequency and nature of adverse experiences (ie, AExs) occurring during meditation training. Some researchers have argued that AExs are substantially underdetected and reflect iatrogenic harm caused by meditation (ie, adverse effects [AEfs]). Others contend that these experiences largely reflect common stressors that would be experienced without meditation. These competing perspectives underscore the need for further research, particularly in the context of digital MBIs, the most widely used form of meditation training. Objective: This study examined the prevalence, predictors, and subjective evaluations of AExs during a digital MBI and tested whether reported experiences may be caused by meditation practice via comparisons between meditation-exposed and nonexposed participants. Methods: Data were drawn from 2 trials of the Healthy Minds Program. Exploratory study 1 (n=315) consisted of a sample of distressed US undergraduate students to estimate the prevalence of AExs and identify baseline predictors. Preregistered confirmatory study 2 (n=594) sampled distressed US adults from all 50 states to replicate findings from study 1 and to examine participants’ subjective evaluations of AExs. Study 2 additionally compared AEx rates between participants who did and did not complete guided meditations to assess whether AExs could be caused by meditation exposure. Study 3 (n=87) used qualitative methods to analyze study 1 participants’ responses to an open-ended question regarding their strategies for coping with AExs. Results: In studies 1 and 2, 27.9% (88/315) and 10.1% (40/396) of participants, respectively, reported at least one AEx during the study period, with 6.7% (21/315) and 3% (12/396) reporting functional impairment, largely aligning with previous research. Critically, in study 2, rates of AExs did not significantly differ between participants who did and did not complete guided meditations, suggesting that these experiences were not caused by meditation practice. Higher baseline depression, anxiety, loneliness, experiential avoidance, and perceived barriers to meditation predicted more frequent AExs. In studies 1 and 2, 89.8% (79/88) and 90% (36/40) of participants who reported AExs, respectively, indicated that they were glad to have learned to meditate. Qualitative analyses showed that participants used diverse coping strategies, often using skills learned through the Healthy Minds Program. Conclusions: AExs were relatively common but occurred at comparable rates among participants who did and did not meditate, challenging claims that such experiences were caused by meditation practice in distressed individuals. Although a small subset of participants reported some degree of functional impairment, most evaluated their AExs as tolerable and described their overall MBI experience as positive. Together, these findings highlight the importance of distinguishing AExs that likely reflect epiphenomena of preexisting distress or symptoms from iatrogenic harm attributable to MBIs. Trial Registration: Study 1: ClinicalTrials.gov NCT04741529; https://clinicaltrials.gov/study/NCT04741529; Study 2: ClinicalTrials.gov NCT06282523; https://clinicaltrials.gov/study/NCT06282523

Correction: Barriers and Facilitators in the Implementation of the Systematic Medical Appraisal, Referral, and Treatment (SMART) Mental Health Digital Intervention in Rural India: Mixed Methods Process Evaluation Study

2026-06-12T16:00:45-04:00

Examining the Social and Mental Health Benefits of Virtual and In-Person Physical Activity Intervention Among Postsecondary Students: Quasi-Experimental Study

2026-06-11T15:00:22-04:00

Background: Physical activity (PA) is a promising prevention approach for supporting mental health and enhancing social inclusion among postsecondary students. However, it is unclear whether similar outcomes are realized when PA programming is delivered in-person versus virtually. Objective: Using data from a multiphase research project, the purpose of the study was to examine the influence of on-campus PA programming (virtual and in-person delivery) on mental ill health symptoms (ie, anxiety and depression), social inclusion indices (ie, social connectedness, emotional ties, and social relationship quality), and well-being. Three objectives were addressed: (1) to assess pre-post change in symptoms, social inclusion indices, and well-being for virtual and in-person delivery; (2) to evaluate whether outcome change over time differed by delivery mode; and (3) to examine whether change in symptoms and social inclusion indices predicted change in well-being for both delivery modes. Methods: Physically inactive postsecondary students experiencing mental ill health participated in a 6-week structured and supervised PA program. Pre-post intervention data were collected across 3 phases, and the analytical samples included: 1. In-person delivery (n=87; 82%, 69/84 young adults; 86%, 74/86 women; 38%, 33/86 White; 20%, 17/86 Chinese; 86%, 75/87 with mental illness; 2. Virtual delivery (n=62; 69%, 42/61 young adults; 95%, 59/62 women; 34%, 21/62 White; 21%, 13/62 South Asian; 55%, 34/62 with mental illness), and 3. Data from students who received in-person or virtual delivery: (n=92; 67%, 61/91 young adults; 90%, 83/92 women; 32%, 29/92 White; 20%, 18/92 South Asian; 59%, 54/92 with mental illness). Data were analyzed using 2-tailed paired samples tests to address objective 1, a 2 (delivery mode) × 2 (time: pre-post) repeated-measures ANOVA to address objective 2, and hierarchical regression analyses to address objective 3. Results: Both virtual and in-person PA delivery were effective for symptom reduction and social inclusion improvements across all outcomes (<.001), with moderate-to-large effects. There was no significant time × delivery mode (=0.72, ²=0.04, =.60) interaction effect. Change in social inclusion indices explained unique variance in well-being, beyond covariates (gender, mental illness, and ethno-racial identity), and symptom reduction for virtual ( = 0.75, 008001) and in-person ( = 0.72, =0.16, <.001) PA delivery. Conclusions: Online distance learning is increasing across postsecondary settings worldwide, underscoring the need for accessible, technology-enabled mental health prevention interventions. The results provide support for the effectiveness of virtual and in-person PA programming for reducing symptoms of anxiety and depression, while also enhancing social inclusion indices and overall well-being. Social inclusion indices were also a key contributor to improved well-being, emphasizing the relevance of social factors in both virtual and in-person PA-based mental health prevention strategies for postsecondary students.

Between Help and Harm: An Evaluation Study of Mental Health Crisis Handling by Large Language Models

2026-06-11T13:30:15-04:00

Background: The use of large language models (LLMs)–powered chatbots has reshaped how people seek information and advice, including for emotional and mental health support. While LLMs can offer scalable support, their ability to safely detect and respond to acute mental health crises—including suicidal ideation, self-harm, and violent thoughts—remains poorly understood. Progress is hampered by the absence of unified mental health crisis taxonomies, annotated benchmarks, and empirical evaluations grounded in clinical best practices. Objective: We addressed these gaps by introducing (1) a unified taxonomy of 6 clinically informed mental health crisis categories; (2) an evaluation dataset of over 2000 user inputs drawn from 12 publicly available conversational mental health datasets, classified into crisis categories; and (3) an expert-designed protocol for assessing response appropriateness. We also used LLMs to automatically identify crisis-indicative inputs and conducted an auditing study of 5 LLMs to evaluate the safety and appropriateness of their responses. Methods: We developed a taxonomy of mental health crisis categories informed by clinical experts and established literature. From over 239,000 mental health–related user inputs collected from 12 Hugging Face datasets, we curated 2252 examples (206 for validation, 2046 for testing) covering all taxonomy categories. We evaluated 3 LLMs on their ability to classify inputs into crisis categories, selecting the model with the strongest agreement with human annotators as the judge to label the test set. We then audited 5 LLMs on their ability to generate safe and appropriate responses to the 2046 test examples. Response quality was measured using a clinically informed 5-point Likert scale (1=harmful and 5=fully appropriate), relying on an LLM-as-a-judge validated against human expert feedback. Results: Several LLMs exhibited high consistency and generally reliable behavior when responding to explicit crisis disclosures, but significant risks remain. A nonnegligible proportion of responses was rated as inappropriate or harmful, particularly in the self-harm and suicidal ideation categories. Substantial performance differences were observed across models: gpt-5-nano and deepseek-v3.2-exp achieved very low harmful response rates, whereas gpt-4o-mini, Llama-4-Scout-17B-16E-Instruct, and grok-4-fast-non-reasoning generated markedly higher rates of unsafe outputs. All models exhibited systemic weaknesses, including poor handling of indirect or ambiguous risk signals, reliance on formulaic responses, and frequent misalignment with user context. Conclusions: These findings underscore the urgent need for enhanced safeguards, improved crisis detection, and context-aware interventions in LLM deployments and highlight the central role of alignment and safety engineering—beyond model scale or openness—in determining crisis response reliability. Our taxonomy, dataset, and evaluation framework lay the groundwork for ongoing research in artificial intelligence–driven mental health support, helping to minimize harm and protect vulnerable users.