Sarah Holder is a staff writer at CityLab covering local policy, affordable housing, labor, and technology.
More than 65 million people are living in a state of displacement, the highest level in human history. Only a small fraction are successfully resettled into permanent homes. Is there a digital fix for this very human crisis?
In January, the U.S. State Department lifted refugee resettlement restrictions for 11 of the “high-risk” countries once targeted under President Donald Trump’s infamous travel ban. Now, displaced families fleeing unrest and oppression in areas on that list, likely to include places like Syria, Egypt, and Iraq, are again able to enter the U.S.—but only on a strict case-by-case basis. And as the number of refugees the country accepts shrinks—in 2017, the U.S. took in its lowest amount in more than a decade—the global refugee crisis grows.
According to the United Nations Human Rights Council, a global resettlement agency, the number of global refugees reached its highest level in measurable history by the end of 2016: Today, 65.6 million people live in some state of displacement. Most refugees in the last decade came from Syria, Iraq, and the Democratic Republic of the Congo. They’ve been recently joined by an estimated 800,000 Rohingya Muslims who have been expelled from their native Myanmar. Every minute around the world, 20 people are forcibly displaced.
The disparate impact of war and upheaval on civilians is often random, and always cruel. But researchers are attempting to use technology to ensure that the journey of resettlement, at least, doesn’t have to be.
The existing process of moving refugees from camps (many of which are located in Bangladesh or Jordan or Lebanon) into more permanent homes in places like the U.S. can be a time-consuming and often imperfect process. Once a week, U.S. resettlement agencies gather to determine the number of refugees that have been processed by the State Department who need to be resettled stateside. Those individuals are then allocated into cities, and placed under the jurisdictions of particular nonprofits.
All of this work is done by resettlement agency workers, who know refugees’ needs as well as what resources local organizations have at their disposal (such as hospitals, health clinics, housing, and local aid workers with knowledge of certain languages).
Over time, resettlement patterns have emerged, and pockets of refugee communities have grown around the country: Somali neighborhoods in Minneapolis, Minnesota; Nepali avenues in Burlington, Vermont; Rohingya enclaves in Kitchener, Ontario and Buffalo, New York. There, refugees can hopefully find the same social supports that have sustained generations of American immigrant communities—ethnic networks and local infrastructures in place to help them adjust.
It’s a very analog system. “[Agency workers] have a sense of each refugee profile, they have a great relationship with communities, and they know how to place people … manually, spreadsheet by spreadsheet,” said Alex Teytelboym, Oxford economics associate professor and co-founder of Refugees’ Say, an organization developing new refugee matching technologies. “You just think of the best place to put them and you trust your instinct.”
Now, however, researchers have begun to ask if if those instincts could be aided by data-driven algorithms. Stanford’s Immigration Policy Lab, for example, has developed a new tool for matching refugees to host cities, which sorts for where a given refugee will find the best employment potential. Teytelboym, meanwhile, is working on a different sort of algorithm with Refugees’ Say co-founder Will Jones. Theirs is designed to take into account a subtler interplay of factors—the real preferences of both refugees and the host countries that will eventually absorb them.
The hardest part of automating this kind of extremely high-stakes human decision-making: determining what a computer should prioritize, and in what order.
Jobs, jobs, jobs
Immigration policy organizations like Stanford’s have spent years developing improved education systems, job training programs, and language learning labs to give refugees a better chance at landing work. But such programs are costly and hard to scale. “So our attention shifted to this geographic allocation,” said Jens Hainmueller, a Stanford researcher and one of the algorithm developers. “That’s an area where if you can come up with a better way, you get these increased employment rates at basically zero cost.”
Stanford’s algorithm uses existing employment data to first analyze where refugees have been sent in the past, and whether or not they’ve been successful in their job searches. Then it uses this knowledge to match new refugees with similar characteristics. In test runs, a team of researchers used data on 30,000 refugees, ages 18 to 64, who were placed in the U.S. and Switzerland from 2011 to 2016. The majority were originally from Somalia, Myanmar, the Democratic Republic of the Congo, Iraq, Syria, and Afghanistan. They compared rates of actual employment to the chances of finding work via algorithm. With a robot on board, employment rates would have been 41 percent higher, they estimate. At the algorithm’s best, they say, employment opportunities for those it places could increase by up to 70 percent. “Some of these improvements are coming from the fact that there’s quite a lot of synergy between places and people,” said Hainmueller.
In the current allocation system, for example, a French-speaking refugee bound for Switzerland would have the same chance of landing in a German-speaking region of the country, rather than a French-speaking one. Their prospects of finding a job in the latter, however, are 40 percent higher.
The takeaway is obvious: Send refugees where they can speak and understand the language. Another easily programmed priority: Send people who can do certain jobs where those jobs are most needed.
“The nice thing is that there’s a multitude of these synergies,” said Hainmueller. “And the algorithm—by seeing how well [refugees have] done in the past—is learning all of those by basically picking up these patterns of where refugees of certain characteristics find work or not.”
But employment odds are not static, nor are labor supplies. Say the job market for meatpackers is saturated, or Afghans are starting to experience discrimination that cuts them out of the workforces where they’ve historically found support. The algorithm will identify these trends, learn from them, and update its predictions accordingly.
Early employment was chosen as a primary matching factor because it’s a tangible advance that has proven and positive trickle effects. “Employment is the pathway by which a refugee can get agency,” said Mike Mitchell, associate vice president of U.S. programs with HIAS, a Maryland-based refugee resettlement agency. “By getting a job, they can progress, and build an income and build assets and integrate into American society.”
Swift employment is also touted as a shortcut to cultural immersion. “People meet other people at work, they learn the language,” said Hainmueller. Relying solely on in-group networks without forging ties to host communities has been shown to hinder future financial stability.
There’s another good reason to focus on employment metrics: they’re the statistics governments are actively collecting. “We know 90 days after arrival whether they’ve found work or not, because this is the declared goal of [the U.S.’s] refugee resettlement program—to transition refugees into quick self-sufficiency,” said Hainmueller.
But employment can be a touchy issue among the communities that accept refugees, thanks to fears that hordes of newcomers will snap up jobs that existing residents should be doing. Matching refugees to labor networks that need them can work to dissolve the misconception that refugees are job-stealers—they’re filling gaps in an existing labor market, not duplicating American skills.
But it also feeds into the notion that in order to be an American you have to be “skilled”—and that the worth of immigrants is tied to their productivity, not their humanity.
According to Mitchell, finding a job is less important than finding the right job: one with room for growth; that will help you build skills and save for retirement. “If refugee A is working in a CVS, stocking, and refugee B is working in a computer company on a path to getting a certification at Microsoft, that makes all the difference,” he said. A better algorithm would include factors like the kind of job, length of commute, and quality of the family’s school district. “That’s all going to contribute to the stability of a refugee’s family, which is going to help ensure integration.”
Teytelboym agrees: “Being employed within three months is not necessarily a measure of success—it’s just an easy political gain for the host countries, and it puts pressure on refugees to end up in low-paying jobs. We would prefer a holistic view.”
The bigger picture
The Refugees’ Say algorithm is designed to measure that thus-far immeasurable (or unmeasured) whole.
Before placement, refugees spend days in meetings with State Department officials and attending cultural orientations, getting briefed on their placement options—but rarely their opinions. Teytelboym proposes surveying refugees in the camps, allowing them to tell placement agencies exactly where they’d prefer to build a life. Then they’ll build an algorithm around it.
“Some things are kind of hard to observe. That’s exactly where precise preference comes in,” said Teytelboym. “It’s the ultimate arbiter of what good welfare is: to what extent people’s own individual, idiosyncratic, perhaps strange or unusual preferences are taken into account.”
And it’s not just about preference: Every family’s needs are multidimensional. One member of a family might have specific medical needs, or speak a rare language, and another might have unusual or important skills. Their ultimate destination should be a place that can support all or most of those realities. Those idiosyncrasies are mirrored at the city level. A host city like Burlington might have better hospitals than those in Kitchener, or more Afghani restaurants. In the algorithm Teytelboym is building, those location metrics are weighed, too.
Asking host cities for their “preferences” could have some less desirable effects: It might conceivably make racial profiling easier, opening the door for cities to enforce their own local-level versions of President Trump’s “Muslim ban.” But Teytelboym and Mitchell say that while the administration has slowed the flow of refugees at the federal level, local agencies are far less reticent.
“By and large we find the opposite,” said Teytelboym. “These communities want to take a lot more refugees than the State Department is willing to move.”
That’s what Mitchell says, too. “Our experience at HAIS is that the communities where we resettle refugees are really welcoming,” he said. “Even if there’s controversy because of the Trump administration, Americans on the whole tend to be open.”
Of course, there is no such thing as an “American whole.” A September study by Dartmouth researchers complicates this narrative: They found that even liberals who support refugee resettlement on the national level balk at accepting new families into their communities.
To avoid discrimination, however, Teytelboym’s algorithm codes “preferences” as linked to resource availability. Esoteric medical needs? Don’t choose us. Speak Afghani? We’re ready for you.
A digital tool for a human crisis
Refugees’ Say is partnering with refugee resettlement organizations to collect preference data themselves, while Stanford’s program is limited based on existing data. But if jurisdictions began to track other, more intangible indicators of refugee well-being, Stanford’s algorithm could easily be tweaked to include more human dimensions. “We’d very much encourage governments to [collect more data] in order to get a broader read of refugee integration success,” said Hainmueller. For now, however, measurable emotional considerations (like keeping relatives together) are built-in as algorithm overrides.
As adaptable as both algorithms are, they’re not meant to entirely automate government’s placement processes. Agencies can’t simply send thousands of refugees to the place an algorithm identifies as the best labor market, without taking into consideration the shares and capacities of local offices. And a human touch is still beneficial, especially when dealing with such a human crisis.
“This is just one way of considering how to make decisions on behalf of the people we serve,” said Mitchell. “This is probably the start of something that will transform the decision-making we do for clients that is needed—not only in refugee resettlement, but in the nonprofit sector.”
Hainmueller says Stanford’s algorithm could be implemented immediately, wherever there is enough historical data for a system to generate future projections. Swiss and U.S. officials have already expressed interest in rolling out a pilot, he says, and Canada, Germany, and the United Kingdom all have enough data to start. Teytelboym is working with one of the three largest U.S. resettlement agencies, which he did not name because of mounting federal antagonism to refugee resettlement, to roll out a pilot in May.
In 2016, only 189,300 of the millions of global refugees were officially resettled, mostly to cities in the U.S., U.K., Canada, Australia, Germany, Switzerland. By the third fiscal month of 2018, only 5,000 were admitted into the U.S.—far fewer than years before.
“It’s a difficult environment at the moment, because a lot of the refugee numbers are being reduced,” said Teytelboym. “But at the same time this is an amazing way to save time and headaches they create by placing refugees in the wrong places. In a way, there’s no better time to do this—it would be great if the numbers were larger, but we’re trying to place the best numbers we can.”
CORRECTION: An earlier version of this post misspelled Alex Teytelboym’s name.