Machine Learning and Asylum Adjudications: From Analysis of Variations to Outcome Predictions

Individuals who demonstrate well-founded fears of persecution or face real risk of being subjected to torture, are eligible for asylum under Danish law. Decision outcomes, however, are often influenced by the subjective perceptions of the asylum applicant's credibility. Literature reports on correlations between asylum outcomes and various extra-legal factors. Artificial Intelligence has often been used to uncover such correlations and highlight the predictability of the asylum outcomes. In this work, we employ a dataset of asylum decisions in Denmark to study the variations in recognition rates, on the basis of several application features, such as the applicant's nationality, identified gender, religion etc. We use Machine Learning classifiers to assess the predictability of the cases' outcomes on the basis of such features. We find that depending on the classifier, and the considered features, different predictability outcomes arise. We highlight, therefore, the need to take such discrepancies into account, before drawing conclusions with regards to the causes of the outcomes' predictability.