The world’s refugee population has exploded over the last decade, straining the resources of international resettlement agencies as staff struggle to keep up with manually matching millions of refugees and asylum seekers to host countries.
Artificial intelligence and machine learning—which have transformed several industries over the past two years—may offer solutions for overwhelmed resettlement programs run by both nongovernmental organizations and governments, according to a study by Harvard Business School Assistant Professor Elisabeth C. Paulson. Her recent article proposes using two new algorithms for matching refugees and asylum seekers to host countries based on their likelihood of finding successful employment.
“Can we build algorithms that will help find better matches that will allow people to integrate more easily?”
A working solution is needed because the current process can be painfully slow and difficult—and the decade-long surge of refugees and asylum seekers has left staff members scrambling to make timely placements, Paulson says. The war in Ukraine, coupled with ongoing conflicts and instability in the Middle East, Africa, and South America, have flooded these systems with asylum seekers, while fragile economies, political tensions over immigration, and low unemployment rates have raised the stakes higher.
“It’s a very time-consuming and detail-oriented process that requires looking at a lot of different datasets to find places for everyone to go,” Paulson says. “What we’re asking is, can we build algorithms that will help find better matches that will allow people to integrate more easily?”
The paper presents data from Switzerland and the United States that showed promise in using machine learning to assist with refugee placement. One of the algorithms dramatically outperformed existing assignment methods—improving on the status quo by up to 50 percent with respect to predicted employment—while the other successfully placed refugees in a balanced way across available host country sites, while also improving employment outcomes.
In addition to easing the transition for refugees, the results could help businesses, too, Paulson says, since finding a better way to anticipate where refugees will thrive and how to support them in success potentially offers fresh sources of talent.
The daunting challenge of placing refugees
Paulson developed and studied the performance of the algorithms with Kirk Bansak, an assistant professor at the University of California at Berkeley. The work builds from their partnership with the Immigration Policy Lab at Stanford University, which uses data science to improve resettlement outcomes through its GeoMatch software application.
The team built algorithms that predicted places where refugees were most likely to find work, which they considered one of the best indicators of a family’s success in assimilating. Their second goal: To make sure that those assignments allowed for a relatively even distribution of refugees across resettlement sites throughout each year, so that no single location was either overwhelmed or underutilized during the year.
That’s a steep challenge, as the population of refugees and asylum seekers worldwide has grown to about 50 million as of May, according to United Nations High Commissioner on Refugees. That’s roughly 2.5 times larger than it was a decade ago. About 75 percent of refugees come from Afghanistan, Venezuela, Syria, Ukraine, and Sudan.
Agency employees face the staggering challenge of evaluating various factors, such as work history, education, and cultural background, to help families start new lives safely.
AI significantly improves matches
Algorithms seek to streamline that work. Paulson and Bansak analyzed patterns in two countries that accept humanitarian immigrants: Switzerland and the United States.
- In the US, the research examines United Nations High Commissioner for Refugees settlers, whose number is determined by an annual cap set in advance. These refugees are distributed to 10 resettlement agencies, each of which runs a network of local affiliates that have their own predetermined capacity restrictions. The authors drilled in on 2015 and 2016 data for about 2,000 refugees between the ages of 18 and 64 at about 30 locations to see if they are employed after 90 days.
- In Switzerland, the research examines about 4,500 adult asylum seekers in 2015 and 2016 who are assigned by the Swiss State Secretariat for Migration to one of the 26 Swiss cantons, each of which can only accept a certain number of arrivals based on their population sizes. Because of differences in the European labor market, the researchers investigated whether the asylum seekers in Switzerland had found work within three years.
The researchers applied two algorithms to help distribute refugees:
The first aims to help refugees find jobs fast. In the United States, this employment-focused algorithm achieved 96 percent efficiency, compared to about 70 percent without AI. Looking at data from Switzerland, where the algorithm is part of a pilot program that began in 2020, this algorithm performed at 98 percent efficiency, versus about 65 percent without using AI.
But it came with a downside: It left many local affiliates either underutilized at different points throughout the year, wasting precious resources, or overutilized, creating an overwhelmed and congested system.
The other algorithm, focused on balancing where refugees settle, seeks to maintain an even distribution over time. This algorithm was able to ensure that the placements to each location are smooth and steady throughout the year. That way, volunteers and resources at the localities that help families resettle were utilized at a steady rate, while only reducing eventual employment by 2 percent.
Tapping into AI in additional countries
Finding ways to strike a balance between an unknown future flow of refugees and local capacity constraints presents one of the biggest challenges to the international resettlement community, Paulson says.
“If I knew in the future that some people are going to arrive who had a high chance of finding a job” in a particular area, Paulson says, “then maybe I would know to save some slots. But not knowing who is going to arrive makes it very difficult.”
“It’s not like it’s a one-size-fits-all solution that we can apply to every country.”
With the pilot tests underway, Paulson and Bansak are hopeful that the US and Switzerland will continue using their new algorithms for case assignments. At the same time, the team and their Immigration Policy Lab colleagues are looking to apply what they have learned toward developing solutions in other countries, such as Netherlands and Canada.
“There are always differences in country contexts. It’s not like it’s a one-size-fits-all solution that we can apply to every country,” Paulson says. “It’s a very lengthy process, and we continue to look to bring on new partners.”
While the refugee and asylum-seeker algorithms have a particular application, Paulson says the methodology can be adapted to solve resource allocation challenges in other contexts in business, health care, and beyond. For instance, Paulson says she recently had a conversation with an executive MBA student who is designing a platform for matching patients with mental health care providers and wanted to learn from the Immigration Policy Lab’s experience.
“I think in any instance where you have a limited resource, and you need to distribute that resource to people with different skills and backgrounds, these types of tools can be immensely useful,” Paulson says.
You Might Also Like:
Feedback or ideas to share? Email the Working Knowledge team at hbswk@hbs.edu.
Image: Image by HBSWK with assets from AdobeStock/Yulia Raneva, AdobeStock/Siarhei, and generated with Midjourney, an artificial intelligence tool.
