AI system
- System
-
Novissi rural targeting pipeline: satellite-imagery poverty maps for canton ranking plus supervised machine learning on anonymized mobile-phone call detail records for individual eligibility
- Model
-
LightGBM, used for gradient-boosted regression.
- Release
-
Exact release not established.
- Application developer
-
Innovations for Poverty Action, CEGA and UC Berkeley (Aiken and Blumenstock, UC Berkeley; Bellue, University of Mannheim; Karlan and Udry, Northwestern University) for the Government of Togo
- Role
-
The models estimated poverty to rank areas and determine eligibility for payments.
- Configuration
-
Researchers implemented gradient-boosted regression with Microsoft’s LightGBM software.
- Inputs and outputs
-
Not assessed
- Autonomy
-
Not assessed
- Workflow
-
Satellite-imagery poverty map ranking cantons. Consumption model trained on mobile-phone metadata with a recent phone survey as ground truth. USSD self-registration and mobile-money payment rails. Government and GiveDirectly funding decision.
- Human oversight
-
Not assessed
- Relation to AI
-
Decisive for who was paid: canton ranking and individual eligibility were both set by machine-learning predictions, while humans fixed the budget, thresholds and registration rules. The Nature evaluation is co-authored by the team that built the targeting method. The World Bank/Government of Togo case study is independent of the developers.
- AI contribution
-
Decisive for who was paid: canton ranking and individual eligibility were both set by machine-learning predictions, while humans fixed the budget, thresholds and registration rules. The Nature evaluation is co-authored by the team that built the targeting method. The World Bank/Government of Togo case study is independent of the developers.
Benefits
- Outcome identifiers
-
Not assessed
- Benefit
-
reduced exclusion errors by 4–21% relative to geographic alternatives
- Category
-
Increased Scalability
- Status
-
Not assessed
- Beneficiary groups
-
People. Organizations. Society.
- Beneficiaries
-
Not assessed
- Reach
-
Not assessed
- Distribution and access
-
Not assessed
- Duration
-
Not assessed
- Expected?
-
Yes
- Costs and harms
-
Not assessed
Measurement
- Studies
-
Not assessed
- Study design
-
Not assessed
- Measurement
-
Not assessed
- Setting and sample
-
Not assessed
- Measures
-
reduced exclusion errors by 4–21% relative to geographic alternatives
- Comparison
-
the geographic targeting options considered by the Government of Togo
- Results
-
Reference: errors of exclusion of 59%–78% for the feasible alternatives. Result: errors of exclusion of 53% for the phone-based approach, a reduction of 4–21%.
- Additional value from AI
-
Not assessed
- Limitations
-
Not assessed
- Denominators
-
Not assessed
- Materials for replication
-
Not assessed
Money
- Kind of value
-
Not assessed
- Amount
-
Not assessed
- Currency and prices
-
Not assessed
- Period and unit
-
Not assessed
- Realization
-
Not assessed
- Recipients and payers
-
Not assessed
- Gross or net
-
Not assessed
- Included costs
-
Not assessed
- Comparison
-
Not assessed
- Method
-
Not assessed
- Return and payback
-
Not assessed
People and places
- Participants
-
Developer: Innovations for Poverty Action, CEGA and UC Berkeley (Aiken and Blumenstock, UC Berkeley; Bellue, University of Mannheim; Karlan and Udry, Northwestern University) for the Government of Togo. Operator: Government of Togo, Ministry of Digital Economy and Transformation, with
GiveDirectly. - Jurisdictions
-
Benefit: Rural cantons,
Togo. Developer: 
United States. Operator: Togo. - Ownership and control
-
Not assessed
- Dependencies
-
Not assessed
- Cross-border connections
-
Not assessed
- Geopolitical implications
-
Not assessed
Follow-up
- Conditions for replication
-
Target population owns and uses mobile phones (non-subscribers are excluded by design). Operators provide anonymized call detail records. Recent ground-truth survey available for training. Mobile-money infrastructure in place.
- Replication attempts
-
1 follow-up entries. Benefit recurred: 1.
- Later findings
-
Not assessed
- Availability
-
Not assessed
- Institutional decision
-
Not assessed
- Grounds
-
Not assessed
- Action and result
-
Not assessed
- Success criterion
-
Not assessed
Evidence
- Source numbers
-
AIBD-0093-source-0001–0003
- Citations
-
Numbered references below.
- Language
-
Not assessed
- Source relationships
-
Developer, Operator
- Original evidence
-
Not assessed
- Access and versions
-
Not assessed
- Supported findings
-
Not assessed
- Evidence links
-
Not assessed
- Evidence strength
-
Not assessed
- Missing evidence
-
Not assessed
Grades
- Assessed benefits
-
People. Organizations. Society.
- Depth
-
People: 2. Organizations: 2. Society: 3.
- Reasons
-
People: Five mobile-money installments of 8,170 CFA (women) or 7,000 CFA (men) to each of 138,531 rural beneficiaries during COVID-19 restrictions (World Bank study, Table 1). Organizations: The Government of Togo delivered shock-responsive rural assistance without a social registry. The World Bank documents the model as a reusable case study. Society: Material to many: 138,531 rural beneficiaries reached, and the Nature evaluation reports exclusion errors 4–21% lower than geographic targeting.
- Observed or intended
-
Not assessed
- Rubric
-
Not assessed
- Assessor and date
-
Not assessed
- Alternative assessments
-
Not assessed
- Fill
-
The deployer's account
Entry details
- Kind
-
Not assessed
- Related entries
-
World Bank-backed poverty-targeting algorithm for cash transfers oversimplified household vulnerability, excluding poor Jordanians; Novissi used the same machine-learning targeting of the poor.
- Field
-
Social protection and humanitarian cash transfers. Government social assistance.
- Purpose
-
Prioritize the poorest cantons and the poorest mobile subscribers within them for COVID-19 emergency cash transfers in the absence of a social registry.
- Dates
-
Date basis unassessed: 2020 (precision unassessed). Updated: 22 October 2026.
- Stage
-
Not assessed
- Version
-
SHA-256: 77448671472d…
References
- Other identifiers
-
AI Incident Database: 546.
- Category mappings
-
Not assessed
- Brief
-
Not assessed
- Revisions
-
8 revisions. Latest: 22 October 2026.
Sources
-
Peer-reviewed study Developer
Aiken, E., Bellue, S., Karlan, D., Udry, C., Blumenstock, J. E. (2022). Machine learning and phone data can improve targeting of humanitarian aid. Nature 603, 864–870 (the quotation is the Extended Data Fig. 1 caption, panel a, with its reference markers omitted).
Satellite imagery of Togo is used to construct micro-estimates of poverty (middle), which are overlayed with population data to produce canton-level estimates of wealth.
-
Report Deployer
Lawson, C., Koudeka, M., Cárdenas Martínez, A. L., Alberro Encinas, L. I., George Karippacheril, T. (2023). Novissi Togo: Harnessing Artificial Intelligence to Deliver Shock-Responsive Social Protection. World Bank and Government of Togo, Social Protection and Jobs Discussion Paper No. 2306, September 2023 (month only; first day used).
Subsequently, the program was extended on November 2, 2020, to rural areas with a high prevalence of poverty (henceforth referred to as model 2).
-
Press release Deployer
GiveDirectly (2021). Innovations in Togo and the future of cash distribution. GiveDirectly, The Future of Cash, published 20 April 2021 (implementing partner and funder, not the AI developer).
Leveraging machine learning in combination with satellite imagery and cell phone metadata, the program was able to identify those living in extreme poverty