Ethical Guidelines
Ethical Guidelines for Humanitarian AI Systems Principles for Ethical Design, Development, and Deployment in Health Supply Chain Optimization
Version 3.0 | November 2025
Executive Summary
Artificial Intelligence is reshaping how humanitarian health supply chains anticipate, procure, and deliver medicines in Uganda and across East Africa. This transformation introduces both technological promise and deep ethical responsibility.
These Ethical Guidelines establish the framework for all design, development, and deployment of AI tools for supply chain optimization in humanitarian and fragile settings. Developed by IFRAD under the Elrha Humanitarian Innovation Fund, they govern the AI Framework for Health Supply Chain Optimization currently being implemented across ten facilities in Karamoja and Southwestern Uganda.
First of Its Kind
This is the first comprehensive ethical framework for humanitarian AI supply chain systems developed in Africa for African contexts.
Alignment
These guidelines align with:
Uganda's National Development Plan IV
Sphere Humanitarian Charter
Uganda's Data Protection and Privacy Act (2019)
WHO Ethics and Governance of AI for Health (2021)
Purpose and Scope
These guidelines accompany the AI Framework for Health Supply Chain Optimization, forming its ethical foundation. They define minimum ethical conditions under which AI may be used to inform decisions affecting medicine availability, nutrition commodities, and lifesaving supplies.
Applies to: Governments, NGOs, software developers, donors, and research institutions creating or funding AI systems for humanitarian supply chains.
Governs: The three-tier forecasting system:
Tier 1: Rule-based forecasting (offline, facility level)
Tier 2: Hierarchical statistical forecasting (district servers)
Tier 3: Machine learning forecasting (central level)
Humanitarian Ethical Mandate
Humanitarian technology must begin from humanity, not efficiency.
Every AI system deployed under humanitarian mandate must uphold the four humanitarian protection principles from the Sphere Handbook (2018):
Avoid exposing people to further harm
Provide access to impartial assistance
Protect people from physical and psychological abuse
Assist people to claim rights and recover dignity
For AI systems, these translate into: transparency in automated decisions, equity in data representation, and the right to human oversight.
Guiding Principles
1. Do No Harm
All AI activities must undergo pre-deployment ethical risk assessments evaluating potential harm to individuals, communities, or institutions.
The system must never autonomously execute a recommendation without validated human approval at decision points appropriate to the tier and context.
Tier-Specific Safeguards:
Tier 1 (Facility)
Recommendations are advisory only. Staff must manually approve all orders through weekly inventory review
Tier 2 (District)
Statistical forecasts supplement but cannot override facility decisions
Tier 3 (Central)
ML predictions require human validation before redistribution decisions
2. Equity and Inclusion
Models must be trained and tested on data reflecting facility types, refugee settlements, and low-connectivity regions.
Concrete Equity Mechanisms Implemented:
Storage Capacity Adjustment
Very inadequate storage: 50% cap. Inadequate: 70% cap. Addresses r = -0.695 correlation
Refugee Facility Prioritization
1.2x weighting in allocation algorithms for refugee settlements
Nutrition Program Integration
Integration with WHO/WFP/UNICEF/UNHCR Nutrition Appointment Platform
Inventory Frequency
Auto-triggers weekly count recommendations for facilities with >4 stockouts annually
Cold Start Protocols
New facilities use stratified baselines matched by type, region, and refugee-serving status. Never initialize rural facilities using urban averages
3. Transparency and Explainability
Every AI recommendation must be intelligible to human operators.
Tier-Specific Requirements:
Tier 1
Plain language: "Based on 3-month average (800 tablets) + Malaria Season adjustment (+200) = 1,000 calculated. Capped at 700 due to Inadequate storage rating."
Tier 2
Model confidence intervals and data quality flags when records incomplete
Tier 3
Feature importance rankings showing which variables most influenced predictions
4. Accountability and Oversight
Responsibility for outcomes rests with human actors and institutions, not algorithms.
Oversight mechanisms include:
Governance committees
Audit trails
Incident-reporting protocols
5. Local Edits Always Win
FOUNDATIONAL PRINCIPLE
When conflicts arise between system recommendations and user adjustments, the facility worker's local edit is preserved.
System logs conflicts for district review
Cannot retroactively override without explicit facility consent
Acknowledges frontline workers have ground truth
Human-in-the-Loop by Context:
Routine replenishment
Facility staff manual approval during weekly review
Emergency redistribution
Explicit district-level approval required (no automation)
Outbreak response
Clinical officer must activate outbreak mode
Delivery delays
District supply officer reviews redistribution recommendations
Override Logging:
User ID, timestamp, original recommendation, final decision, mandatory reason
High override rates flag forecasting errors, not user penalties
No-blame principle applies
6. Privacy and Data Protection
Data collection and processing must comply with Uganda's Data Protection and Privacy Act (2019).
Offline Data Protection:
AES-256 encryption using Android Keystore System
Keys generated in hardware-backed secure enclaves
Never appear unencrypted in memory or file storage
12-month active retention, archived after upload confirmation
7. Frontline Worker Protection
Stakeholder validation identified frontline worker burden as a critical ethical concern.
Health workers are already overburdened—digital systems must reduce, not increase, their workload.
Worker Protection Mechanisms:
Administrative Burden Reduction
Auto-save, single-tap adjustments, auto-generated paper forms matching HMIS 105
Battery Optimization
Auto-sync only when battery >20% AND device not in active use
Cognitive Load Reduction
Single-column layouts, 48dp touch targets, 5-inch minimum device support
Training Minimization
Embedded video guides and progressive interface complexity
Device Charging
Offline-first architecture, low computational requirements
Data Governance
Developers and implementing partners must:
Collect only necessary data (no patient-level health data)
Obtain explicit informed consent for primary data collection
Use anonymization for all storage and transmission
Maintain 7-year retention schedules
Conduct annual data protection audits
Never transfer data to third parties without written authorization
Algorithmic Integrity
Documentation Requirements
Each model must document:
Algorithm type and selection rationale
Feature engineering and preprocessing
Known limitations and failure modes
Validation methodology and metrics
Override patterns and refinement history
Quarterly Algorithmic Audits
Audits assess:
Forecast accuracy by facility type (refugee vs. non-refugee, HC II vs. RRH)
Stockout reduction patterns across regions
Override rates and reasons by facility
Emergency procurement frequency changes
Storage capacity constraint effectiveness
Audit results must be transparent to donors and government partners.
Risk Management
Ethical Risk Register
Implementers must maintain a register identifying:
Data leakage risks
Biased forecasts disadvantaging specific facilities/regions
Misuse of analytics for political manipulation
Over-reliance on automation
System failure during critical decisions
Increased frontline worker burden
Incident Response
Reporting Timeline
Within 24 hours of discovery
Offline Adjustment
Clock starts when facility syncs or reports via phone
Report Contents
Root cause analysis, affected facilities, immediate mitigation, corrective measures
Serious Incidents
Require public disclosure to affected communities
Governance Structure
Forecasting Governance Committee
Convenes quarterly to review compliance, approve updates, and manage grievances.
Membership:
Ministry of Health (Pharmacy Division)
District Health Officers (rotating)
Facility users (nominated by peers)
Technical partner (IFRAD or successor)
Academic validator (Kyambogo University)
Responsibilities:
Quarterly review of model performance and stockout trends
Analysis of override patterns
Approval of algorithm changes
Recommendations for framework refinement
Advocacy for addressing infrastructure inequities
Alignment with Frameworks
Uganda National Development Plan IV
Full alignment
Uganda Data Protection and Privacy Act (2019)
Compliant
National eHealth Policy Framework (2013)
Aligned
African Union Continental AI Strategy (2024)
Aligned
WHO Ethics and Governance of AI for Health (2021)
Aligned
OECD AI Principles (2019)
Aligned
Sphere Humanitarian Charter (2018)
Foundational
Conclusion
Ethical innovation is not constraint but foundation for sustainable technological progress.
These guidelines translate ethical principles into operational parameters:
Storage capacity → algorithmic constraint
Human oversight → tier-specific approval workflows
Equity → concrete weighting factors
Transparency → plain-language explanations
Worker protection → battery optimization and auto-save
By grounding AI in humanitarian ethics informed by African operational realities, these guidelines assert both technical capability and moral leadership in defining how intelligent systems serve human need.
Document Information
Version
3.0 (Post-Validation)
Date
November 2025
Organization
International Foundation for Recovery and Development (IFRAD)
Technical Validator
Ojok Ivan, Kyambogo University Department of Data Science, Networks & AI
License
CC BY 4.0
Last updated