Place Matters: The Possibilities and Pitfalls of Machine Learning Algorithms for Refugee Resettlement

The global refugee population is dramatically increasing, posing a challenge for over-saddled refugee agencies who ultimately determine where these individuals should be placed. A machine learning algorithm called GeoMatch positions itself as a solution to this crisis by using historical data to identify communities where refugee families are most likely to succeed. In this paper, I aim to examine how the algorithm functions and has been implemented by resettlement agency personnel. Moreover, I pit GeoMatch's purported benefits against its risks, ethical concerns, and unaddressed limitations. The methodology includes analyzing quantitative research findings, supplementary documentation on the algorithm, and a previously published interview in Stanford's Momentum magazine with a senior resettlement agency staff member. I find that although GeoMatch has branded itself as a cutting-edge technology that simplifies the resettlement process, substantial concerns over data quality, modeling techniques, limited testing, and a hyper-fixation on maximizing shorter-term employment must raise alarm bells. The developers of GeoMatch need to address these shortcomings in order for GeoMatch's potential gains to translate to measurable, real-world improvements to refugee integration outcomes.