Risks in humanitarian AI

A SAFE AI briefing note for humanitarians scoping AI use or reviewing existing deployments

This briefing sets out the main categories of risk that arise when humanitarian organisations deploy AI, and shows how the SAFE AI Framework and its tools respond to each.

SAFE AI is the AI-specific layer of CDAC’s longer-running work on accountable information ecosystems in humanitarian contexts. Where SAFE AI governs the AI systems themselves, CDAC’s Information Integrity workstream addresses the wider information environment in which those systems operate, and CDAC’s work on participation and AI examines what meaningful community involvement in AI design can and cannot achieve. This briefing sits between them.

The right to know is the norm the SAFE AI Framework operationalises across the AI lifecycle. People whose lives are shaped by AI-influenced decisions have a right to know when automation is in play, a right to understand how it shapes decisions about them, and a right to contest those decisions. The risks set out in this briefing note are the failure modes that, left unmanaged, prevent that right from being real.

The categories of risk include:

• Data privacy, security, and infrastructure

• Bias and representation

• Participation as a governance instrument, with limits

• The wider information environment

• Accountability and the deployer/developer asymmetry

• How AI risk changes after deployment

• Responsible refusal