AI for Humanitarian Action and Social Good

Humanitarian organisations serve people affected by conflict, disaster and displacement — among the most vulnerable people on Earth. There are over 120 million forcibly displaced people worldwide according to UNHCR's recent global figures. Resources are always insufficient; information is often incomplete; decisions are urgent. AI can help — to anticipate crises, allocate scarce resources, understand needs and communicate in many languages. But in humanitarian contexts, errors and data misuse can cost lives and endanger people. This lecture explores both the promise and the responsibilities.

Applications#

Anticipation and early warning#

Forecasting displacement and needs from conflict data, climate indicators and economic signals (e.g. research tools such as Project Jetson, piloted by UNHCR, for predicting movements in Somalia).
Anticipatory action: triggering funding and assistance before a forecast flood or drought peaks, based on predictive models.
Disease outbreak forecasting and surveillance.

Mapping and situational awareness#

Satellite and aerial imagery analysis: detecting shelters in camps, estimating population, mapping flood extent and building damage after earthquakes and cyclones. Community mapping initiatives (e.g. Humanitarian OpenStreetMap Team) increasingly combine volunteers with AI-assisted mapping.
Crisis informatics: analysing social media and hotline messages during emergencies to identify needs and misinformation.

Services and communication#

Multilingual information services: chatbots and helplines answering questions about registration, services and rights in affected people's languages.
Translation and speech technologies for low-resource languages.
Document processing: OCR and extraction to reduce data-entry burden.

Operations#

Supply-chain and logistics optimisation, demand forecasting for relief items.
Targeting and prioritisation support for assistance programmes.
Fraud and duplication detection in registration and distribution systems.

Principles#

Humanitarian action is guided by humanity, neutrality, impartiality and independence, and by the imperative to do no harm. For data and AI, sector guidance (e.g. the ICRC Handbook on Data Protection in Humanitarian Action, UN data-protection and privacy principles, IASC operational guidance on data responsibility) emphasises:

Data minimisation and purpose limitation — collect only what is necessary.
Protection of people: assess risks that data or models could be used to harm, target or discriminate against affected people.
Informed consent where feasible, recognising power imbalances (people may feel unable to refuse when assistance is at stake).
Accountability to affected populations: people should understand how data about them is used and have ways to give feedback and complain.
Security proportional to the sensitivity of data.
Context and local knowledge: involve affected communities and local staff in design.

Specific risks#

⚠️Common Pitfall

Biometric and identity data: databases of fingerprints or iris scans of refugees raise serious concerns if shared with authorities or breached; the consequences can include persecution. Some high-profile controversies have concerned the sharing of refugee data with governments of origin.
Group harms and "demographically identifiable information": even aggregated data can endanger groups (revealing locations of an ethnic minority).
Automated targeting errors: an excluded family may go without food; appeals must be accessible.
Surveillance creep: systems built for assistance repurposed for control.
Bias from unrepresentative data: models trained elsewhere may fail for different populations, languages or housing types.
Techno-solutionism: deploying AI because it is fashionable rather than because it addresses a real need better than simpler options.
Power and dependence: reliance on private technology providers; data sovereignty questions.

A responsible project checklist#

Start from the need, defined with affected communities and field staff — not from the technology.
Consider simpler alternatives (better forms, a spreadsheet, a phone line).
Conduct a data protection and human rights impact assessment.
Minimise and protect data; prefer aggregation, on-device processing, pseudonymisation and differential privacy.
Validate locally: evaluate on data from the actual population and context, with subgroup analysis.
Keep humans accountable for decisions affecting individuals; design accessible appeals.
Pilot small, monitor outcomes and harms, and be willing to stop.
Share learning — including failures — with the sector.

python

# Example: releasing only safe aggregates from a needs survey
import pandas as pd

def safe_aggregate(df, by, value, min_cell=10):
    """Aggregate and suppress small cells that could identify individuals or small groups."""
    agg = df.groupby(by)[value].agg(["count", "mean"]).reset_index()
    agg.loc[agg["count"] < min_cell, ["count", "mean"]] = None       # suppress small groups
    return agg

survey = pd.DataFrame({"camp": ["A"] * 40 + ["B"] * 6, "needs_water": [1, 0] * 20 + [1] * 6})
print(safe_aggregate(survey, "camp", "needs_water"))

How students can contribute#

Volunteer for humanitarian mapping and data projects (e.g. mapping tasks, data challenges) — learning while contributing.
Build language technology for under-served languages you speak.
Join organisations' data and innovation teams, internships and fellowships.
Do research with partners on real problems, with ethics review and community involvement.
Most importantly, bring humility: listen to field staff and affected people; their knowledge is essential.

#humanitarian ai#social good#refugees#do no harm#responsible data

Written by

Janin A Apurba

B.Sc. in CSE, AUST · Advanced ICT Officer, CNRS-UNHCR. Teaching AI, ML and Deep Learning to the next generation of engineers and researchers.