Responsible Use of Artificial Intelligence in Refugee Status Determination
A rights-based policy framework and prototype for managing mixed-movement flows in Latin America — Leonardo Delmondes
Rationale
Asylum systems across Latin America are under sustained pressure from large-scale mixed movements, financial constraints, limited institutional capacity, and growing backlogs — affecting the timeliness, consistency, and perceived legitimacy of asylum processes. Well-designed and well-governed AI tools can help: improving access to information, reducing administrative burdens, and freeing eligibility officers to focus on the steps that require human decision-making.
But Refugee Status Determination (RSD) is a high-stakes, rights-based process where due process, data protection, confidentiality, transparency, and access to effective legal remedy are essential. Some elements, especially credibility assessment and other subjective components of a claim, cannot be automated without undermining fundamental guarantees.
The project starts from four premises:
AI can support more efficient processing of asylum claims.
AI should be used as a decision-support tool, only for clearly defined, bounded functions.
Based on the author's field experience, a substantial share of existing backlogs could be resolved through administrative proceedings or differentiated RSD procedures, without a full individual assessment.
In many developing countries, asylum infrastructure depends heavily on international cooperation (UNHCR funding, ECHO-funded initiatives, donor projects).
Key objectives
Map the asylum workflow in the Latin American context and pinpoint the stages where AI could bring operational improvements with manageable tradeoffs.
Assess the legal, ethical, and operational risks of AI-supported functions in asylum procedures.
Develop a policy framework defining minimum requirements, risk thresholds, and operational safeguards for responsible AI use in RSD.
Prototype concrete, rights-compatible use cases where AI supports asylum processing without replacing human judgment.
Scope: where AI helps, and where it must not go
Permitted, with safeguards
Country-of-origin information (COI) research — retrieval, organization, and summarization from validated sources such as Refworld, with full traceability
Translation assistance, with mandatory human review of substantive content
Fully automated final decisions on refugee status or inadmissibility, without a clear and accessible appeal mechanism
Automated assessment of credibility or other subjective elements, including emotion analysis, deception detection, or profiling
Anything that weakens access to explanation, contestation, or appeal
Systems lacking auditability or clear human accountability
Field experience
The project draws on the author's work as a former UNHCR Officer across five countries in the region, from the Southern Cone to Mexico. It is anchored in two of them — Peru and Mexico, which together accounted for ~14% of the asylum backlog in the Americas in 2025. They offer distinct dynamics — large Venezuelan caseloads, differentiated RSD procedures, and high operational pressure from regional mobility — and the contrast enriches the analysis.
Project focus
🇵🇪 Peru🇲🇽 Mexico
Also worked in
🇦🇷 Argentina🇧🇷 Brazil🇪🇨 Ecuador
Approach
Desk review of asylum procedures, backlog drivers and reduction strategies, legal and ethical concerns, and the infrastructure needed for on-premises AI.
Use-case analysis, each assessed against risk criteria: rights impact, error costs, bias, and operational viability in Latin America.
Joint build of the policy framework and prototype, iterating between technical feasibility and policy constraints.
Validation through semi-structured consultations with protection actors, legal practitioners, and data protection specialists, plus one structured validation session.
Workplan · 10 months
1 · Preparation for prototyping
Map the asylum workflow and what works against backlogs; review open-source language models and existing COI and case-processing tools; define the minimal on-premises infrastructure; plan the build.
2 · Risk assessment & safeguards
Minimum requirements on data protection, transparency, logging, and accountability; risk criteria and the "affordable compromises" of using AI where human resources are limited; safeguards for effective remedy and human oversight.
3 · Prototype & policy framework
Prototype the permitted use cases — COI support, translation assistance, administrative triage, complexity-based routing — and draft a concise framework with minimum requirements and red lines.
4 · Validation
A workshop with lawyers from or experienced in Latin America, humanitarian organizations, and other stakeholders; feedback integrated into the final outputs.
Main outputs
Policy framework
Minimum requirements for AI use in asylum procedures, and a risk assessment distinguishing permitted, conditional, and prohibited (red-line) uses.
Operational prototype
A prototype of an asylum system with embedded AI that could become an open-source solution for countries and organizations. It can be used whole or as stand-alone components, to fit each country's needs and regulations.
See a synergy, or want to compare notes?
Always glad to connect with people working on asylum systems, digital governance, or responsible AI in the public sector.