It's nice to see an international application, but this is just a start.
A lot of talk about AI in the United States focuses on shiny new tech advances, threats of superintelligent robots, or the impact on businesses and workers in an already-polarized society. There's not much out there on what impact the technology can have on the rest of the world, particularly the developing world (though the issue of low-skilled jobs in the developing world being replaced by machines in the developed world pops up quite a bit). So I'm always on the lookout for more international stories.
One of those stories came out on Friday in Science. Researchers developed a computer program that can help resettle refugees in particular host countries. This is a great use of machine learning that could help improve people's lives and build global community. Still, it's worth proceeding with caution, not least because this technology is pitched as a quick fix to a thorny problem.
What the Researchers Did, and What the Model Can Do
The researchers use employment as the marker of a successful resettlement. And their method of algorithmic resettlement can in theory get more refugees jobs than the current approach to resettlement in the United States and Switzerland. That is, it's better at determining where in the countries refugees will be most likely to get jobs.
To create the model, the researchers used existing data on resettlement outcomes, and coded it according to refugee characteristics and employment success in a way that a computer program can understand. Then, they used that data to teach their algorithm what combinations of characteristics predict employment success. There's a lot more to it, but the bottom line is that algorithm can in theory be applied to new refugees to determine where they should be resettled in a way that increases their chances of getting a job.
The researchers point out that there are a number of things left to figure out to improve the accuracy of the model. And they call for the approach to be tested in the field with a controlled trial and random participants.
But based on a quick read, it seems like a few questions could use answering first.
Questions About Approach
The questions highlight issues that pop up in other discussions about machine learning and technological solutions to problems more broadly.
The researchers write, for instance, that "in contrast to more expensive interventions (such as language or job training programs) that are sometimes implemented long after refugees’ arrival, our approach is cost-efficient and implemented before refugees’ arrival, giving them the strongest foundation possible from which to integrate into host societies." OK—but why do we have to talk about contrasts? Isn't it possible that the optimal solution may well combine a number of interventions, and that this low-cost technology isn't the magic bullet? Let's slow down just a bit.
Additionally, the researchers note: "Because of the algorithm’s data-driven learning capacity, policy-makers do not need to invest in identifying the precise sources of those synergies—local economic conditions, social environments, resettlement office efficacy, etc.—to harness their benefits." But the "precise sources of those synergies" could point to underlying problems that are being ignored thanks to a focus on the algorithmic solution. Technological optimization without understanding can be dangerous.
And is employment the correct gauge of resettlement success? I don't have a background in this subject, but I'd be interested to hear if there is research demonstrating that employment is an acceptable stand-in for assimilation and well-being more broadly speaking. The researchers note that the algorithm can be targeted at any "quantifiable metric" of success, and this is an area that should be explored further. A closer look should also be given to the characteristics on which the algorithm was trained to see if they are the best and fairest indicators to use.
These are just initial reactions. But if there's anyone out there in the refugee space who wants to take a look at this approach, let me know what you think.
Questions About Applications
- Could this approach work for relocating chronically unemployed or underemployed people? What other factors would need to be addressed for such an approach to work—for instance, psychological interventions, systemic and generational changes, or education?
- Could the approach be flipped around and turned toward the country level, analyzing which countries would benefit from refugees the most and in what ways, or be taxed the most? This could help make refugee resettlement more efficient and also help build a sense of international community in tackling this global problem.
- How could you get countries that accept refugees to agree to join a common pool, and receive their refugee assignments based on an algorithm like this?