AI for Mental Health Support in Refugee Populations: Evidence and Ethical Limits

The MHPSS gap is large and persistent

Estimates from WHO and IASC MHPSS Reference Group put the prevalence of common mental disorders in conflict-affected populations at roughly 22% in any given period. The supply of trained clinicians is orders of magnitude below need in most refugee settings. AI-supported mental health interventions have entered this gap with both genuine promise and considerable risk.

What the published evidence supports

Three intervention types have meaningful evidence in refugee or refugee-adjacent populations. Guided self-help with AI-assisted content (psychoeducation, behavioural activation prompts, structured exercises) shows modest but measurable benefit in adapted formats, building on the WHO Self-Help Plus and Problem Management Plus protocols. Triage and screening chatbots that route users to appropriate care levels have evidence of acceptable accuracy in adult populations. Clinician-support tools (transcription, note generation, decision support) reduce the administrative load on the few clinicians available, indirectly expanding capacity.

Where the evidence does not support deployment

Three uses remain ahead of the evidence. AI as primary therapist for severe disorders is not supported and is increasingly prohibited by professional regulators. Crisis intervention (suicide risk, acute psychosis, child safeguarding) requires human escalation paths that several deployed chatbots have failed to provide reliably. Children under adolescence are a population for which acceptability and safety evidence is sparse; defaulting to human-mediated MHPSS is the operative standard.

Cultural and linguistic adequacy

Most evidence on AI mental health interventions comes from English-speaking, high-income populations. Direct transfer to refugee contexts is not safe. Adaptation requires: cultural validation of constructs (depression, trauma, grief) that do not translate uniformly; language quality in the actual languages of the served population, not the languages the vendor markets in; and supervision by clinicians who share or understand the cultural context. The WHO mhGAP Intervention Guide and the Inter-Agency MHPSS minimum service package are the operative frameworks.

Privacy and protection

Mental health data is among the most sensitive categories handled by humanitarian organisations. AI-supported interventions involving cloud APIs that retain inputs for model improvement are incompatible with the ICRC data protection handbook and with most national health-data regulations. Practical mitigations: on-device or self-hosted models, contractual no-retention agreements, explicit informed consent in the client's language, and clear data-deletion rights.

A short evaluation checklist for any AI MHPSS deployment

• Is the intervention aligned with a recognised MHPSS protocol (WHO SH+, PM+, mhGAP)?
• Is there a documented escalation path for crisis cases, and is it tested?
• Has the content been culturally and linguistically adapted by clinicians from the population?
• Are clinicians supervising the deployment, and at what cadence?
• What is the data governance, and does it meet humanitarian standards?
• Has an external evaluation been planned, funded, and published?

Further reading and primary sources

• WHO mental health in emergencies: https://www.who.int/health-topics/mental-health-in-emergencies
• IASC MHPSS Reference Group: https://interagencystandingcommittee.org/iasc-reference-group-mental-health-and-psychosocial-support-emergency-settings
• WHO Self-Help Plus: https://www.who.int/publications/i/item/9789240035119
• WHO Problem Management Plus: https://www.who.int/publications/i/item/9789241598064
• WHO mhGAP Intervention Guide: https://www.who.int/publications/i/item/9789241549790
• MHPSS Minimum Service Package: https://www.mhpssmsp.org/