Switzerland is about to pilot a new algorithm which could boost refugee employment by up to 30%. Built on big data from tracking previous refugees, the tool analyses asylum seekers’ characteristics to recommend the region where they’re most likely to find work. Launching in summer 2018, the program will be the first of its kind in the world.
As things stand, only 15% of asylum seekers in Switzerland manage to find a job after three years. Increasing this number would have “dramatic societal consequences,” said Dominik Hangartner, a professor at ETH Zurich who co-created the algorithm. Even just one refugee being in employment saves the Swiss state an average of $35,000 in one year.
Since the refugee crisis erupted in 2015, European countries have been searching for ways to resettle and integrate refugees. Though it’s early days, the algorithm has the potential to be used far more broadly – even at European Union level, Hangartner suggests.
• For more like this, see our digital government newsfeed.
So how does the algorithm work, and what kind of results can we actually expect?
Building the machine
The algorithm was built and tested on historical data by a team from ETH Zurich and Stanford University, using machine learning to match geographic employment outcomes with refugees’ personal characteristics, such as gender, age and nationality.
The researchers began by gathering data from more than 30,000 refugees resettled in the US between 2011 and 2016. They found that using the algorithm would have led to 41% more refugees finding jobs, compared to current assignment practices. A similar process was repeated for Switzerland and the results were even more impressive: the algorithm would have boosted the employment rate by 73% for those arriving in 2013.
“We are not changing the law, we are just trying to be more efficient within it”
Alongside the Swiss State Secretariat for Migration (SEM), the research team is now translating that study into a real-life pilot. Some 1,000 asylum seekers will be placed by the algorithm across Switzerland’s 26 cantons, or states. Their outcomes will be compared to a control group, distributed according to the existing system of random quotas. Importantly, the algorithm will be fed with new data on a frequent basis: the relationship between refugee characteristics and the locations can change over time, explained Hangartner.
The final decision on where refugees go will still be a human one made by SEM staff; the algorithm is a tool to assist them. Family members will still be reunited, and asylum seekers with special medical conditions still matched to cantons with the relevant university hospitals. And importantly, refugees will still be distributed to the country’s cantons proportionally, according to a legally-defined key based on population numbers. “We are not changing the law, we are just trying to be more efficient within it,” said Lukas Rieder, a spokesperson at SEM.
The black box
The use of artificial learning in public policy raises important questions of transparency and fairness. Should governments be able to use a “black box” in a room to make decisions about people’s futures? In this case, Hangartner argues that it’s “easy to explain what goes in and comes out” of the algorithm: data of past refugees goes in, and a recommendation of the asylum seeker’s best employment chances comes out. In the interests of full transparency, the algorithm’s code is available online.
Some also have reservations about the algorithm’s design, such as why language isn’t included. However, most asylum seekers don’t speak a Swiss language, and it could clash with Switzerland’s law to avoid ghettoisation: refugees from the same country are distributed to different cantons so they don’t concentrate in one place. This rule-of-origin law reduces its effectiveness in boosting employment by around 20%. SEM and the academic team are also exploring how to introduce educational achievement and trajectory to the algorithm, but this requires further data collection.
A model for Europe?
Expectations for the pilot’s employment impact are lower than the academic study’s results. As well as legal restrictions on country of origin, the pilot includes asylum seekers with good chances of a right to stay – some of whom may be rejected – whereas the original study focussed exclusively on provisionally admitted persons. Also, the study was based on data from 2013, but since 2015 Switzerland’s huge influx of refugees has made finding employment even harder.
“There’s not much to lose”
But any increase in asylum seeker employment would likely be worth it. “There’s not much to lose,” said Rieder. Apart from the time it takes to give staff basic training and for them to use the software – which takes just seconds – the only cost associated with the pilot is research time, which is covered by the universities. At scale, the software is unlikely to be free, but it wouldn’t take many refugees to find employment for that investment to pay for itself.
What’s more, the algorithm is thoroughly scalable. Policymakers can choose which characteristics and outcomes they want, and the algorithm can be rebuilt in any country. The only limit is data: Switzerland has a particularly rich dataset on refugees, which is not the case across the whole continent. In data-rich countries such as Sweden – which has accepted more refugees per capita than any other European nation – the algorithm could easily be implemented within existing institutional structures.
And the potential to use this tool across borders is even greater. “The bigger the differences across the resettlement locations,” say between European countries as opposed to Swiss cantons, explained Hangartner, “the bigger the potential gains of assigning refugees to those places.” If the European Union – or another supranational body – were ever to reach a more substantial resettlement agreement for asylum seekers, the algorithm could have dramatic implications for refugee integration.
(Picture credit: Flickr/Sebastian)