World: Standards and Assurance Framework for Ethical AI - Risks in humanitarian AI (May 2026)

# New SAFE AI Briefing Sets First Sector-Wide Guardrails for Humanitarian Artificial Intelligence Use As humanitarian organizations scramble to stretch ever-thinner resources amid record global displacement, escalating climate disaster response demands, and years-long protracted conflict crises, adoption of artificial intelligence tools across the sector has accelerated rapidly. For years, however, that rapid uptake has unfolded with little standardized, context-specific guidance to ensure these tools do not inadvertently harm the very populations they are designed to support. That gap is beginning to close with the May 19 release of a new public briefing from the CDAC Network, published via ReliefWeb, which lays out a structured set of risk categories and corresponding safeguards for humanitarian groups using or evaluating AI tools, tied to the organization’s existing SAFE AI Framework. The SAFE AI Framework is the AI-specific component of CDAC’s multi-year body of work building accountable information ecosystems in humanitarian contexts, a long-running initiative focused on ensuring information systems serve, rather than exploit or harm, crisis-affected populations. The new briefing sits at the intersection of two parallel CDAC workstreams: one addressing broader information integrity risks in humanitarian response, and another examining the real-world limits of community participation in AI design. Where the information integrity workstream addresses the wider information environment in which AI systems operate, and the participation workstream assesses what meaningful community input in AI development can and cannot achieve in practice, the SAFE AI briefing focuses specifically on risks inherent to the AI systems themselves, across their full lifecycle from pre-deployment design to post-implementation operation. At the core of the framework is the operationalization of a “right to know” for people whose lives are shaped by AI-influenced humanitarian decisions. The framework holds that crisis-affected populations have a clear right to know when automated systems are used to make decisions that impact them, a right to understand how those systems shape outcomes related to aid access, service delivery, or protection, and a right to contest decisions they believe are unfair or harmful. Every risk category outlined in the briefing is framed as an unmanaged failure mode that prevents this right from being realized if left unaddressed. The identified risks span the full scope of humanitarian AI use, with heightened stakes given the extreme vulnerability of the populations these tools serve. Data privacy, security, and infrastructure risks top the list: in humanitarian contexts, compromised personal data does not just result in a standard data breach— it can expose displaced people, marginalized groups, or conflict-affected civilians to persecution, exploitation, or targeted harm by bad actors. Bias and representation risks are also prioritized, as AI systems trained on incomplete or skewed data may systematically underallocate aid, exclude vulnerable groups from protection services, or misidentify individuals in need of assistance, entrenching existing structural inequities. The briefing also addresses participation as a governance tool, noting that while community input into AI design is a core ethical principle, it has practical limits in acute crisis contexts where rapid response needs preclude lengthy, resource-intensive consultation processes. Additional risk categories cover cascading harms across the wider information ecosystem: unvetted AI outputs may be amplified alongside misinformation or disinformation in crisis zones, worsening harm for already affected populations. The briefing also flags accountability gaps arising from the asymmetry between AI developers (often based in high-income countries with no direct ties to crisis-affected communities) and humanitarian deployers, who may lack the technical expertise to audit or modify off-the-shelf AI tools. It also addresses evolving risks that emerge after deployment, as on-the-ground context shifts, conflict dynamics change, or bad actors attempt to manipulate AI systems for harmful ends, and enshrines the principle of responsible refusal: the explicit requirement that humanitarian organizations retain the ability to opt out of AI use entirely in contexts where risks cannot be adequately mitigated, rather than feeling pressured to adopt tech for the sake of perceived efficiency. The briefing is intended as a practical, actionable resource for humanitarian organizations either scoping first-time AI use or reviewing existing deployments, and marks one of the first sector-specific ethical AI standards for the humanitarian space. As donor governments and private sector partners increasingly push for tech-driven efficiency in crisis response, the framework provides a shared baseline for assessing AI tools, aligning deployer and developer accountability, and ensuring AI adoption does not undermine core humanitarian principles of impartial