Google and Rockefeller Helped Fund the Stanford AI Recommending Which U.S. Communities Receive Refugees

Google.org and the Rockefeller Foundation helped fund GeoMatch, an AI-assisted placement tool developed by Stanford University's Immigration Policy Lab in collaboration with ETH Zurich. The system analyses past resettlement data to recommend locations where refugees, asylum seekers and other migrants may have stronger predicted integration outcomes.

GeoMatch does not make binding placement decisions. The lab says placement officers and resettlement workers can accept, modify or disregard its recommendations, retaining final authority over where individuals and families are placed.

How GeoMatch Uses Historical Data

GeoMatch uses machine learning to identify patterns between migrant characteristics, potential destinations and prior integration outcomes. Its published factsheet says the system can use information including country of origin, education and gender, alongside outcomes such as time taken to find work, job type and later moves between locations.

The tool can be configured around different objectives, including employment or income, and can incorporate local capacity and individual preferences. It is designed to operate within an organisation's existing placement process rather than replace the judgment of caseworkers.

Stanford describes GeoMatch as a 'human-in-the-loop' system. Officers using the tool receive training on what its recommendations can and cannot account for, while placement teams can override recommendations where additional context points to a different choice.

The lab says an early US partner needed to avoid sending too many families to the same location at once because local staff and services could be overstretched. It added recent-placement information to GeoMatch's interface after staff had begun tracking those figures separately.

US Research Estimated Employment Gains

The first major US analysis drew on records for more than 30,000 refugees aged 18 to 64 who were placed between 2011 and 2016. In a paper published in 'Science' in January 2018, researchers estimated that algorithmic assignment could have increased average employment by about 41% compared with the placements that occurred.

The historical analysis estimated an average employment rate of 34% under the actual placements and 48% under algorithmic assignments. That represents a 14% difference, or a relative increase of roughly 41%.

Those figures were a back-test using historical data. They were not evidence that a live US deployment subsequently produced the same gain.

The Immigration Policy Lab began working with the resettlement organisation now known as Global Refuge, formerly Lutheran Immigration and Refugee Service, on a prototype in 2022. The initial version was designed to help placement staff consider capacity and individual circumstances across the organisation's local-affiliate network.

By 2023, project material said GeoMatch was being prepared to provide employment-based recommendations for refugees without existing ties to the US. The available project documents do not establish how many refugees received recommendations through the US pilot or whether it remained active after that period.

Google.org and Rockefeller Funding

A 2023 GeoMatch factsheet lists Google.org and the Rockefeller Foundation among the project's funders. Other listed supporters were the Abdul Latif Jameel Poverty Action Lab's European Social Inclusion Initiative, Schmidt Futures, Stanford Impact Labs and the Stanford Institute for Human-Centered Artificial Intelligence.

Funding is not evidence that Google.org, the Rockefeller Foundation or other supporters selected communities or made individual resettlement decisions. The project's published material describes GeoMatch as a tool used by resettlement agencies and governments, with frontline workers retaining authority over individual placements.

Stanford Impact Labs said in March that the Immigration Policy Lab was working with government partners in Switzerland and the Netherlands. It described safeguards including data minimisation, placement-officer training, fairness assessments and reviews of when workers followed or overrode the tool's recommendations.

The lab says it conducted an AI Impact Assessment before the Dutch pilot with the Central Agency for the Reception of Asylum Seekers. Independent researchers at Utrecht University later criticised the Dutch deployment, raising concerns about transparency, ethnicity-related variables and potential unequal outcomes. Their assessment did not establish that unlawful discrimination had occurred.

Swiss Trial Offers Live Evidence

A preprint submitted on 28 September 2026 reports a double-blind randomised trial involving 2,000 refugee cases handled by Switzerland's State Secretariat for Migration between January 2020 and June 2023. Each case consisted of an individual or family.

Cases were randomly assigned either an employment-optimised recommendation for one of Switzerland's 26 cantons, or a recommendation designed to resemble existing placement procedures. Placement officers and refugees were not told which recommendation type had been assigned.

The study reported a 2.2% increase in the average share of months in employment among adult case members over three years. The researchers said this amounted to about a 10% relative increase from the control group's average.

At 36 months, the share of cases with at least one employed adult was 5.2% higher. That household-level measure differs from an individual employment rate.

The study is a preprint and has not yet been published in a peer-reviewed journal. It does not establish effects on wages, job quality, long-term stability, well-being, or family reunification.

The earlier US study was a historical estimate, while the Swiss research tested recommendations in a live placement setting. Both forms of evidence suggest that placement can affect employment outcomes, but neither establishes that an algorithm should determine where a refugee lives.