Introduction¶

AI can speed aid and accountability, but field results depend on context, data quality, and governance. These cases highlight what worked, what failed, and how to improve.

AI for crisis response¶

• Damage assessment from satellite/air imagery accelerates resource allocation but can miss informal settlements; pair with local validation teams.
• Demand forecasting for supplies reduces stockouts yet struggles with fast-changing ground truth; keep human override and rapid re-training loops.
• Key safeguard: transparency about model confidence and clear escalation when predictions conflict with field reports.

Supply-chain tracing in reconstruction¶

• Ledger-based provenance for construction materials can deter diversion but requires reliable on-ramps and tamper-resistant IDs.
• Risk scoring vendors helps spot corruption but may unfairly penalise small local firms; include appeals and manual review.
• Key safeguard: publish criteria, avoid black-box scoring, and rotate auditors to prevent capture.

Data-based human rights monitoring¶

• Crowdsourced incident reporting scales coverage but invites misinformation; use verification tiers and geolocation checks.
• Automated pattern detection can surface hotspots but risks false positives; blend OSINT with trusted local sources.
• Key safeguard: protect witnesses with redaction, consent gates, and secure storage; delay publication if safety risks persist.

Cross-cutting lessons¶

• Start in shadow mode to calibrate against human assessments before acting on outputs.
• Invest in data quality pipelines and feedback loops; retire models that drift beyond agreed thresholds.
• Maintain public transparency notes summarising methods, limits, and mitigation steps.

Conclusion¶

Responsible AI in crises and recovery demands humility, validation, and continuous oversight. Treat models as decision aids - not decision makers - and build in contestability from the start.