| Organization | Atlas AI |
| Location | Nigeria, Democratic Republic of the Congo, Cameroon, Burkina Faso |
| Amount | $350,000 |
| Approval date | February 2026 |
| Funded by | Unrestricted Fund |
This grant will fund Atlas AI to develop subnational population estimates for Nigeria, Democratic Republic of the Congo (DRC), Cameroon, and Burkina Faso. Because population estimates are a key input in many of our cost-effectiveness analyses, we think that improving the quality of this data could allow us to direct funding more cost-effectively.
The challenge
Many of our cost-effectiveness estimates, including estimates for two of our Top Charities (Helen Keller Intl’s vitamin A supplementation program and Malaria Consortium’s seasonal malaria chemoprevention program), depend directly on population data to estimate the number of people reached by a program.1 We also use population data to double-check data provided by implementing organizations and government partners.2 However, population data for many low- and middle-income countries is unreliable, outdated, and inconsistent from source to source.3 Censuses are carried out infrequently: the most recent census in DRC, for example, was conducted in 1984, the most recent in Nigeria in 2006, and the most recent in Cameroon in 2005.4
What this grant will do
Atlas AI will estimate national and subnational populations for Nigeria, Democratic Republic of the Congo (DRC), Cameroon, and Burkina Faso by combining several data sources:
- First, Atlas AI will build a library of existing population estimates, identify the extent of disagreement between sources, and make recommendations for which sources GiveWell should rely on.5
- Second, Atlas AI will construct subnational population estimates for children under five and people over five by combining those initial national estimates with geospatial data (such as satellite data that can identify features associated with population like buildings, roads, vegetation, and light intensity) to estimate population distributions within a country.6
- Third, Atlas AI will engage in a bottom-up approach, drawing on data from coverage surveys commissioned by GiveWell or collected by GiveWell grantee partners, when available, and extrapolating that data to other parts of the countries.7
Atlas AI will also test Google’s AlphaEarth coding of satellite imagery features to predict populations and assess whether AlphaEarth should be incorporated into Atlas AI’s approach.8
Our reasoning
High-quality target population estimates are critical for accurately estimating the number of people reached by our grantees. GiveWell researchers recently identified getting better population estimates as a key strategy for improving our cost-effectiveness models.9 Cost-effectiveness estimates are directly linked to population estimates for programs such as vitamin A supplementation and seasonal malaria chemoprevention that depend directly on population data to estimate program reach, such that a change in population results in a corresponding change in estimated cost-effectiveness (e.g., a 1% increase in population results in a ~1% increase in estimated cost-effectiveness).
More than half of the grantmaking funds we directed in 2025 supported programs in Nigeria, DRC, Cameroon, and Burkina Faso, and we expect to continue directing significant funding to programs in those countries in the future.10 However, we have substantial uncertainty about the population estimates we use for those countries, three of which have not had censuses for at least 20 years.11
We believe that this grant could improve the accuracy of our models in several ways.
- The estimates Atlas AI provides will focus on the populations most relevant to our cost-effectiveness analyses, such as children under five, who would be targeted by, for example, vitamin A supplementation and seasonal malaria chemoprevention campaigns.12
- The library of existing estimates Atlas AI compiles will include their methodologies and recommendations on which GiveWell should use, allowing us to improve our judgment about how to weight these sources in our models.13
- Additionally, these estimates could potentially improve the decision-making of other organizations beyond GiveWell.
We are excited about using coverage survey and grantee data, and about Atlas AI’s plan to collect more data, because we think the main constraint on better population estimates is better data.
Creating cost-effectiveness estimates for value of information grants is more challenging than for direct delivery programs. We think there is a large possibility that this grant will not meaningfully change our grantmaking. However, given the amount of funding we direct in these countries, even a relatively low likelihood of moving a moderate amount of funding makes this grant cost-effective; for example, our rough back-of-the-envelope calculation indicated the grant would be substantially above our cost-effectiveness threshold based on an 18% chance that it would shift $9.5 million in grantmaking toward more cost-effective opportunities.
How we could be wrong
Our key uncertainties are:
- Whether the estimates we receive will be more accurate than the estimates we currently rely on. To address this concern, we are actively working with Atlas AI to triangulate their estimates against existing data sources and to learn from other organizations working on similar issues.
- Whether we should have asked Atlas AI to produce estimates for additional countries. If we find the deliverables from this grant to be high-quality, we may consider a follow-up grant to cover other countries.
Sources
- 1High-quality target population estimates are critical for accurately estimating the number of people reached by our grantees. The target population is the number of eligible people that the grantee is trying to reach. For many of our grantees, the relevant target population is the number of people in an area or the number of children under-5. For example, for Helen Keller Intl’s vitamin A supplementation program, the target population is a direct input used to determine the cost per child reached. First, it is used to estimate the number of children receiving supplements, as indicated in GiveWell, Helen Keller International cost per supplement, 2022, "Supplements delivered" tab. Then, that number is used in the corresponding country tab to estimate the cost per child reached.
- 2For example, we compare the number of people that administrative data show were reached against the percentage of people reached based on multiplying coverage by the population. "For each country, we multiply estimates of SMC coverage by the target population to estimate the total number of cycles delivered in each SMC campaign." GiveWell, Seasonal Malaria Chemoprevention, 2024, "Number of cycles delivered" section.
- 3For example, recent work by IDinsight and GiveWell researchers found differences of more than 30% between population estimates provided by organizations working in DRC, Nigeria, and Cameroon and other data sources. IDinsight, "GiveWell Estimating Populations Comparison Tables," 2026 (unpublished).
- 4"Cameroon last conducted a census in 2005 - 20 years ago." World Economics, 2025. The 2005 Cameroon Census is available at Cameroon Population and Housing Census 2005. "The last population census in DRC was held in 1984." African Development Bank Group, April 2, 2026. The 1984 DRC Census is available at Republic of Zaire General Census of the Population 1984. "In March 2006, Nigeria, for the first time, conducted a Population and Housing Census." National Population Commission, "The Previous Census." See also the dates, 2006, listed for the populations in the map.The 2006 Nigeria Census is available at Nigeria Population and Housing Census 2006.
- 5Atlas AI, "Sub-national population estimation to inform cost-effectiveness analysis," 2025 (unpublished)
- 6Atlas AI, "Sub-national population estimation to inform cost-effectiveness analysis," 2025 (unpublished)
- 7Atlas AI, "Sub-national population estimation to inform cost-effectiveness analysis," 2025 (unpublished)
- 8GiveWell and Atlas AI incorporated testing the use of Google’s AlphaEarth coding of satellite imagery features into the grant based on a recommendation from an external reviewer.
- 9This was identified during a recent effort to red-team the monitoring and evaluation processes for our grants. GiveWell, What We Learned From Red Teaming the M&E, Coverage, and Costs of Our Most-Funded Program Areas, 2026 (unpublished).
- 10More than $230 million of our total $418 million in grants approved during metrics year 2025 (February 1, 2025–January 31, 2026) were directed to programs in Nigeria, DRC, Burkina Faso, and Cameroon.
- 11We are including Burkina Faso, even though it has had a census in 2019, because we are uncertain about the accuracy of that census, adding another country to the grant is relatively inexpensive, and what we learn may improve upcoming decisions about grant opportunities there. We understand that DRC and Cameroon are planning to launch new censuses; we expect that any public data would not be available for several years and thus would not be able to inform our upcoming grant investigations until then.
- 12Atlas AI, "Sub-national population estimation to inform cost-effectiveness analysis," 2025 (unpublished)
- 13Atlas AI, "Sub-national population estimation to inform cost-effectiveness analysis," 2025 (unpublished)