Can AI Determine Credible Fear? A Critique of the State’s Use of Text Analytics in Asylum Adjudications

Author: Jeremy A. Rud (University of California Davis, U.S.A.)
Speaker: Jeremy A. Rud
Topic: Critical Linguistic Anthropology
COMELA 2021 General Session


In line with the long history of scholarship on language in contexts of migration and asylum (Canagarajah, 2017), in this study I bridge European and North American discussions of language analysis in asylum policy (Patrick, Schmid, & Zwaan, 2019) and artificial intelligence (AI) in migration management (Beduschi, 2020) by examining the use of “text analytics to look for boilerplate language” to detect “fraud” in asylum applications. This practice, a planned function of the United States Asylum Vetting Center currently in development by US Citizenship and Immigration Services, is just one example of the increasing use of AI at borders around the world. In order to preemptively unpack this black-box process, I apply both narrative critical discourse analysis and lexicon-based sentiment analysis, a method of text analytics that numerous scholars across Europe applied to media data in response to the 2015 “refugee crisis” (Nerghes & Lee, 2019, inter alia), to a corpus of 20 former refugee narratives in order to compare human and algorithmic readings of these high-stakes linguistic performances. Specifically, I examine how credible fear in an entextualized asylum seeker narrative could be (mis)determined by AI and conclude that the nature of the training data, the composition of the sentiment dictionary used, the accuracy of sentiment scores, and the ideologies of the practitioners all raise serious concerns for the use of sentiment analysis for automatic decision-making in asylum proceedings. Overall, I argue for greater international coordination of linguistic anthropological scholarship that takes an active, rather than reactionary, stance against the unchecked entrenchment of AI in asylum policy and I advance the line of inquiry of the politics of listenership that extends the borders of linguistic anthropological analyses of asylum to concerns of aurality, listening, and artificial intelligence.

Keywords: asylum, credibility, artificial intelligence