Building Trustworthy AI to Support Migration Decisions

When helping refugees and asylum seekers settle into new communities, every placement decision carries high stakes. Lives, futures, and communities are shaped by the significant choices caseworkers make on behalf of incoming families and individuals.

Conscious of the importance of these choices, Stanford’s Immigration Policy Lab (IPL) developed an AI-powered tool – GeoMatch – to support governments and NGOs in making informed, data-driven placement decisions in service of newcomers. An interdisciplinary research lab based across Stanford and ETH Zurich, IPL is currently working with both the Dutch and Swiss governments to conduct pilots of the technology to understand how it can improve existing placement processes.

Because migration decisions are fundamentally about human beings, any algorithm meant to strengthen such decisions must be designed carefully and responsibly to ensure it contributes to improving the lives of the people it is meant to assist. That is why the research team has committed to rigorous safeguards with the aim of deploying technology that elevates the human context, rather than overshadowing it.

The lab’s approach is rooted in responsibility, transparency, and a drive to protect the dignity and humanity of the people resettlement agencies serve. “At its core, GeoMatch is designed to support human decisions, so that more newcomers have the chance to thrive where they are resettled,” explains Jens Hainmueller, IPL Faculty Co-Director. 

Centering Human Experience and Judgment

In most resettlement systems, government placement officers or nonprofit workers decide where a refugee family will live based on the available resources across their network of potential locations. 

To be appropriately equipped to enhance this process, IPL must understand what data is currently available to the agencies charged with determining resettlement placements and how that data informs day-to-day decision-making. So, before deploying even a single line of code, the research team collaborates closely with governments, nonprofits, and service providers to ask foundational questions like: How do caseworkers interpret the information they have? What data is missing that could improve placement outcomes? How could an AI tool provide recommendations in a way that allows caseworkers to trust and work alongside it?

These questions guide GeoMatch’s design at multiple levels—from high-level decisions, such as which integration outcomes the tool should aim to improve, to practical details, like how many recommendations it should display and in what format.

By starting with the needs of users whose knowledge has been informed by years of experience, IPL looks to avoid delivering a one-size-fits-all solution to every agency for every circumstance. And because GeoMatch was built to support human judgment, rather than supplant it, humans remain firmly at the center of every resettlement choice.

Indeed, the research team acknowledges that placement decisions require nuanced understanding that no AI model can fully capture. GeoMatch therefore always operates with a “human-in-the-loop”, without exception. The tool provides recommendations that placement officers may accept, modify, or disregard. Frontline workers therefore retain full authority over final placement decisions and can override any recommendation.

When the tool is deployed by an agency, IPL trains the placement officers using it on how to interpret GeoMatch’s recommendations. This guidance covers what the system can and cannot account for, as well as the kinds of information that may not be reflected in its outputs, so caseworkers can weigh the algorithm’s suggestions accordingly. 

“We work with our partners to design the training materials needed for each deployment so that we can adapt them specifically to what a particular government or agency might need,” says Hainmueller. “We always want to carefully explain both what the tool does, as well as what it does not do. Understanding its limitations is just as important as understanding the range of additional context it provides.”

Fairness and Feedback

Any system built on data is, of course, vulnerable to potential biases that could cause unintended harm. In order to mitigate this, IPL has embraced rigorous external accountability in partnership with the governments in the regions where the tool is deployed. For example, before beginning the pilot with the Central Agency for the Reception of Asylum Seekers in the Netherlands, the team completed multi-stage algorithmic audits, including an AI Impact Assessment.

These efforts are reinforced by research on group fairness that also informs the design of the algorithm. In order to be dynamic and adaptable to different scenarios, GeoMatch supports multiple fairness definitions, which allows placement teams to select a design that prevents disproportionate advantages for specific groups or locations.

“We think about every family individually, but we also step back and ask: is this system working fairly across the board? We regularly review outcomes across key groups to make sure the algorithm is not making some groups worse off. If we see signs of harm, we can adjust the algorithm”, says Michael Hotard, GeoMatch Director at IPL.

Once GeoMatch is deployed by an agency, IPL gathers continuous feedback from that agency through a combination of focus groups, surveys, and regular meetings with caseworkers who rely on the tool. The team also analyzes when caseworkers follow (or choose to override) recommendations, using those patterns as an important source of empirical insight.

“The best features of GeoMatch come from the people making migration decisions on a daily basis,” notes Hotard. “For example, in the United States, one agency we worked with needed to avoid sending too many families to the same location at once. If that happened, local staff and services could become stretched too thin. The agency staff started tracking recent placements in a spreadsheet outside of the tool. When we found out, we added that information directly to the GeoMatch interface, so users could easily see the number of recent placements at each location when they were reviewing recommendations.”

Of equal importance is the work of listening to refugees and their host communities – and being alert to broader macro trends. Through qualitative interviews, the research team studies how placements affect both sides of the resettlement equation, ensuring that the tool’s development is attuned to lived experience. And as populations, economic conditions, and local needs evolve, GeoMatch uses a modeling strategy that allows its machine-learning systems to intelligently adjust to new information about shifting demographic and socioeconomic patterns.

Lastly, the research team conducts rigorous randomized controlled trials (RCTs) to vigorously evaluate the AI. These often provide the strongest evidence of its impact on integration outcomes. RCTs allow the team to balance benefits against risks, compare performance across locations, and guide ongoing improvements to both placement strategies and system design.

Safeguarding Refugees and Migrants

Of course, not every setting is right for GeoMatch deployment. Before any new engagement, IPL works to understand whether the tool is a good fit: whether the right data infrastructure exists, whether placement teams are prepared to use it effectively, and whether an agency’s policy goals align with what the algorithm is designed to do. If those conditions aren't met, the project doesn't move forward.

During this vetting period, IPL also works with partners to ensure that required data privacy and protections are followed. GeoMatch operates using secure servers and data minimization practices to ensure user data is protected. These practices are particularly important with vulnerable populations like refugees.

Ultimately, GeoMatch is a demonstration that AI can be deployed in a way that places an emphasis on people – be they users, migrants, or host community members. By grounding the tool in fairness, transparency, and continuous learning, IPL aims to help governments and NGOs make choices that are not only more informed, but also more compassionate.

As global migration patterns shift and needs evolve, IPL is committed to continually testing and improving GeoMatch so that the tool contributes to a future in which every placement helps newcomers find stability, opportunity, and the best possible chance to thrive.