AI Chatbots for Refugees: A Field Guide to Deployments, Outcomes, and Failures

Why chatbots became the dominant refugee-information interface

The single most common refugee question is not 'when will I get aid' but 'what are my rights and what is the next administrative step'. Answering at scale, in the asker's language, with content that varies by location and legal status, is exactly the problem a well-designed chatbot solves. Since the Signpost project launched in 2017 (a partnership of the IRC, Mercy Corps, and Zendesk), conversational interfaces — first scripted, then retrieval-augmented LLMs — have become the default first-line information channel for refugees and asylum seekers in dozens of country contexts.

The deployments with the most published evidence

Signpost runs context-specific instances (CuentaNos in Central America, InfoMigrants partner content in Europe, Khabrona.info in Arabic) with millions of annual interactions and a moderator team that reviews flagged conversations. UNHCR's Help platform provides country-specific guidance in a structured chat format. Mercy Corps' Refugee.info in Europe pioneered the model in 2015-2016. Newer LLM-backed pilots include Iraq and Colombia deployments using retrieval-augmented generation over a curated knowledge base.

What chatbots actually deliver

Three measurable outcomes recur. Information reach scales by orders of magnitude over a hotline-only model, especially in the first weeks of an emergency when human capacity is the binding constraint. 24/7 availability matters for asylum seekers whose only safe window to ask sensitive questions is at night. Pseudo-anonymity lowers the threshold for asking about gender-based violence, LGBTQ+ protection, and trafficking — categories that helplines consistently underserve. Where chatbots are evaluated against a human-hotline baseline, satisfaction scores are typically high for informational queries and lower for case-specific ones.

Where chatbots fail and why

The failure modes are predictable. Out-of-scope questions about a specific case ('will my asylum claim succeed') are answered confidently but uselessly; well-designed bots route these to a human. Outdated content is the silent failure: a chatbot trained on 2023 procedures answering 2026 questions will be wrong without flagging it. Language quality drops sharply outside the top ten supported languages. Protection sensitivity: bots that store conversation logs without explicit consent are incompatible with humanitarian data principles.

The retrieval-augmented generation upgrade and its risks

The transition from scripted decision trees to RAG-backed LLMs since 2023 expanded coverage at the cost of new failure modes. A RAG bot is only as accurate as its knowledge base; if the curation pipeline lapses, hallucinations enter. The defensible architecture in 2026 pairs a curated, versioned knowledge base with a constrained generation step that cites the source paragraph, plus a human moderation queue for flagged or out-of-distribution conversations. This is the Signpost pattern and is increasingly the sector default.

A short checklist for assessing a refugee chatbot

• Is the knowledge base versioned and dated?
• Does every substantive answer cite a source?
• What is the human moderation cadence on flagged conversations?
• What languages are supported, and what is the per-language quality score?
• What is the data retention and consent policy?
• Is there a documented escalation path to a human caseworker?

Further reading and primary sources

• Signpost: https://www.signpost.ngo/
• UNHCR Help: https://help.unhcr.org/
• Refugee.info: https://www.refugee.info/
• InfoMigrants: https://www.infomigrants.net/
• IASC data responsibility guidance: https://interagencystandingcommittee.org/
• ICRC data protection handbook: https://www.icrc.org/en/data-protection-humanitarian-action-handbook