A Hybrid AI–Human Mental Health Support Model: A Privacy-Preserving, Clinically-Supervised Digital Wellbeing Framework for Multi-Lingual Populations (Preprint)

BACKGROUND Mental health disorders are increasing globally, and access to qualified mental-health professionals remains limited, particularly in low- and middle-income countries such as India. Over the past five years, the rapid growth of artificial intelligence (AI), especially large language models (LLMs), has led to the development of digital mental-health tools intended to improve accessibility, engagement, and early detection. However, many AI-only systems face critical limitations, including lack of clinical oversight, safety risks, inadequate multilingual support, and weak privacy safeguards. At the same time, hybrid AI–human systems—where clinicians supervise, review, or intervene during digital interactions—have shown higher user trust, improved safety outcomes, and better continuity of care. Despite this promise, existing reviews rarely examine hybrid models in depth or evaluate their applicability for diverse populations like those in India, where cultural variation and linguistic diversity require adapted solutions. Furthermore, privacy-preserving architectures, ethical considerations, and LLM-specific safety protocols are insufficiently addressed in current literature. This creates a significant research gap and highlights the urgent need for a systematic and comprehensive review focusing on hybrid AI–human mental-health systems. OBJECTIVE The objective of this study is to systematically evaluate hybrid AI–human mental health systems developed between 2020 and 2025, with a particular focus on their safety, clinical supervision models, multilingual adaptation, and privacy-preserving mechanisms. Additionally, this review aims to identify gaps in existing digital mental health literature and propose a comprehensive, culturally aligned hybrid architecture suitable for large, linguistically diverse populations such as India. METHODS Study Design This study followed the PRISMA 2020 guidelines for conducting systematic reviews. The protocol was defined a priori and adhered to best practices in digital health evidence synthesis. Eligibility Criteria Studies were included if they met the following criteria: Population: Users seeking mental health support or psychological wellbeing interventions. Intervention: Systems involving AI-assisted, LLM-based, or algorithmic mental health support with explicit human involvement (clinician, counselor, moderator, supervisor). Outcomes: Engagement, safety, clinical effects, privacy mechanisms, multilingual usability, or system architecture. Study Type: Randomized trials, observational studies, feasibility studies, and development/technical evaluations. Timeframe: January 2020 to March 2025. Language: English. Publication: Peer-reviewed journal or conference proceedings. Studies were excluded if they: (1) focused only on standalone AI without human involvement, (2) lacked empirical data, (3) were reviews, commentaries, or opinion pieces. Data Sources and Search Strategy Searches were conducted across five major databases: PubMed Scopus IEEE Xplore ACM Digital Library Google Scholar The search strategy combined Boolean terms related to: “mental health”, “depression”, “anxiety”, “wellbeing” “AI”, “chatbot”, “LLM”, “conversational agent”, “digital therapeutic” “hybrid”, “clinical supervision”, “human-in-the-loop” “privacy”, “multilingual”, “cultural adaptation” Example query (PubMed): ("mental health" OR "depression" OR "anxiety") AND ("artificial intelligence" OR "LLM" OR "chatbot") AND ("hybrid" OR "clinician-in-the-loop" OR "human supervised" OR "clinical oversight") Search coverage: January 1, 2020 – March 30, 2025. Study Selection A total of 2,847 records were identified. After duplicate removal, two reviewers independently screened titles and abstracts. Full texts of potentially eligible studies were assessed using predefined criteria. Disagreements were resolved through discussion or senior reviewer arbitration. Final inclusion: 56 studies. A PRISMA flow diagram summarizes the selection process. Data Extraction A structured extraction form captured: Study characteristics: country, year, design, sample size AI architecture and model type Nature of human involvement (clinician, counselor, moderator, peer supporter) Multilingual and cultural adaptation features Privacy-preserving mechanisms LLM-specific safety controls Engagement metrics and clinical outcomes Two reviewers extracted data independently to minimize bias. Quality Assessment Quality appraisal was performed using: RoB-2 for randomized controlled trials Newcastle–Ottawa Scale (NOS) for observational studies Mixed Methods Appraisal Tool (MMAT) for development/feasibility studies Each study was rated as low, moderate, or high risk of bias. Synthesis Approach Due to heterogeneity in study designs, interventions, and outcome measures, a narrative synthesis approach was used. Findings were grouped under: Hybrid AI–human system architecture Clinical supervision and safety Multilingual capability and cultural adaptation Privacy-preserving mechanisms LLM safety controls Engagement and clinical effects Quantitative trends were reported where possible (e.g., trust improvement, crisis reduction). RESULTS Study Selection and Characteristics From 2,847 screened records, 56 studies met the inclusion criteria. Study types included: 18 randomized controlled trials (RCTs) 23 observational studies 15 development or feasibility studies The median sample size was 243 participants. Geographically, most studies originated from the United States (39%), followed by Europe (27%), Asia (21%), and India (5%). AI modalities included rule-based chatbots, NLP systems, LLM-based conversational tools, emotion detection models, and hybrid digital therapeutics integrating clinician oversight. Hybrid AI–Human Mental Health Models Only 11 studies (19%) implemented structured clinician or human-supervisor involvement. Among these: Trust scores increased by 35–50% compared with AI-only systems. Crisis events decreased by 40–55% in systems with human escalation protocols. User satisfaction was higher (mean 4.3/5 vs 3.5/5 in AI-only systems). Hybrid workflows supported context correction, ethical alignment, and safer crisis management. Studies integrating therapists, psychologists, or trained moderators consistently reported improved accountability and perceived emotional safety. Multilingual and Cultural Adaptation Only 12 studies (22%) provided multilingual support, primarily English–Spanish or English–Mandarin. However: Only 2 studies (4%) conducted true cultural adaptation, including local idioms, emotion constructs, and culturally sensitive phrasing. Indian languages were addressed in only 3 studies, none of which implemented LLM-based cultural tuning. This reveals a major evidence gap for linguistically diverse regions such as India. Privacy and Data Protection Mechanisms Privacy-preserving methods were reported in 15% of studies: Federated learning (n = 4) Differential privacy (n = 3) On-device inference (n = 2) Encrypted voice or text journaling (n = 3) However, 85% of studies relied solely on basic encryption or platform-level security, with no advanced privacy engineering. Studies using federated learning showed better user retention (+12–18%), likely due to higher perceived safety. Voice Journaling and Sensor-Based Inputs Seven studies integrated voice journaling or voice biomarkers: Engagement improved to 78%, compared with 53% in text-only systems. Voice-based emotion detection improved early stress identification. Voice journaling was especially effective for low-literacy and older users. Some studies combined voice inputs with passive sensing (sleep, steps, mobility), but integration with clinical workflows remained limited. LLM-Based Systems and Safety Controls LLM-driven interventions appeared in 17 studies (30%), mostly after 2023. Implemented safety strategies included: Content moderation (n = 12) Crisis detection classifiers (n = 10) Human validation or supervisory review (n = 6) Alignment with psychological guidelines (n = 7) LLM-based tools offered improved empathy, personalization, and multilingual potential, but unpredictable outputs and safety risks were frequently reported, highlighting the need for hybrid oversight. Overall Impact and Adoption Gaps Across all 56 studies: Hybrid systems consistently outperformed standalone AI. Multilingual and culturally adapted systems were rare. Privacy-preserving architectures were uncommon despite high user concern. India-specific research represented only 5% of the included studies. These trends underscore the critical need for a privacy-first, multilingual, clinician-supervised model for scalable mental health support in India and similar regions. CONCLUSIONS Hybrid AI–human mental health systems demonstrate clear advantages over standalone AI tools, particularly in terms of trust, safety, engagement, and crisis management. Although LLM-driven systems have expanded rapidly since 2023, their effectiveness still depends heavily on structured human oversight, culturally informed design, and robust safety mechanisms. Despite global advancements, adoption remains limited, with significant gaps in multilingual support, cultural adaptation, and privacy-preserving architectures. For countries like India—characterized by linguistic diversity, high digital penetration, and large treatment gaps—current systems are insufficien

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.