In the past decade, several countries have begun experimenting with artificial intelligence (AI) as part of asylum procedures. Derya Ozkul’s studies show how digital technologies have taken foothold in European asylum and immigration systems and are used for a range of different purposes, from identity verification to digital evidence and lie detection. As in other public law domains, the use of AI in asylum is also moving closer and closer to actual case handling and decision support. Most recently, the United Kingdom has trialled an “Asylum Case Summarisation” tool, which summarises interview transcripts, and an Asylum Policy Search tool, which summarises country of origin information.

These developments can be seen as part of a wider push towards “technosolutionism” in public administration, built on the belief that automation and AI can promote both fairness and efficiency in case handling and human decision-making. There is also at least some argument that AI technologies could in principle prevent unequal treatment and reduce errors in asylum decision-making, as a domain repeatedly shown to be prone to bias and arbitrariness. At the same time, the use of AI in refugee status determination (RSD) also raises significant human rights and transparency issues. Even as EU law and courts begin more closely regulating such practices (for example, in 2023 the German Federal Administrative Court struck down the regular extraction of mobile phone data from asylum seekers), states continue to push for adopting new forms of AI technology in RSD which sit at the borderline of legality.

In this contribution, however, we take a step back and address a more foundational question: given what we know about how current generation AI models operate, how could automation in RSD give rise to new, foundational uncertainties? We take this as primarily an epistemological question and approach it by tackling the role of AI technologies in heightening human bias and epistemic problems in RSD. We then turn to ask whether the game is worth the candle, in other words, what – if anything – can be salvaged from AI going forward?

Artificial Intelligence and Human Bias

As a legal procedure, refugee status determination is characterised by limited availability of “external” or “objective” evidence. This means that the credibility assessment of claimants and their personal testimony have an outsized impact on the determination of a case’s outcome. The lack of clear and binding standards for assessing credibility and claimant testimony moreover opens the door for judicial discretion and bias. A growing number of studies have documented how RSD outcomes are intimately linked to judges’ identity, including political affiliation, experience, and gender, as well as a range of other factors linked to procedural choices, institutional design, and case handling. Other research shows how psychological assumptions about human memory, behaviour, and different types of bias relating to e.g., claimants’ religion and sexual orientation each impact credibility assessment, and by extension, RSD outcomes. So pervasive is this sense of bias and arbitrariness that several scholars have likened national asylum systems to a “luck of the draw” or a “refugee roulette.”

Against this background, introducing AI-support into decision-making may indeed have the potential to mitigate arbitrariness in asylum procedures, as it can account for systematic errors and perceive complex patterns in caseloads imperceivable to individual case workers. However, it also risks replicating pre-existing human biases and introducing new ones. When computational tools are trained on historical data for predicting human decisions, there are several points where bias may influence the model’s outcomes. First, the input data fed into the model may have embedded biases that are systemic, opaque, or obscured under proxy variables. Second, algorithms built on such input data may themselves be biased or skewed, and risk amplifying pre-existing biases. Studies show that even models with rich, accurate data can replicate past prejudiced patterns and provide inaccurate predictions if implemented with insufficient calibration. The fundamental challenge for implementing AI in RSD is that we have no “ground truth” to compare performance against; the arbitrariness across individual decision makers is simply too pervasive.

Moreover, AI not only replicates known and flawed patterns in human decision-making but also risks creating additional inscrutable biases through proxy discrimination. Proxy discrimination occurs when an AI model correlates a variable that is facially neutral, like zip code, with a legally protected attribute like gender or race. Proxy discrimination follows inextricably from the structure of AI models, which are programmed to find correlations between input data and target variables of interest. This facially neutral practice that disproportionately harms members of a protected group is not only shielded from human intervention but also created under the guise of the “objective” algorithm, illustrating the epistemic risk associated with introducing AI technologies in a field such as asylum decision-making.

Refugee Status Determination and Epistemic Uncertainty

The specific nature of asylum decision-making also raises questions about whether current-generation AI can legitimately simulate human decision-making. RSD suffers from an inverse operation of what legal epistemologists call the “proof paradox”; a conflict between legal and statistical reasoning whereby statistical evidence alone is insufficient to prove a fact, despite the likelihood of the evidential standard being satisfied. In many jurisdictions, asylum decision-making operates under an ingrained ‘culture of disbelief’ – a presumption of prima facie invalidity that results in suspicion of asylum narratives. This tends to have the effect of reversing the burden of proof, essentially requiring asylum seekers to dispel decision makers’ doubts. The paradox arises as statistics frequently show that most asylum seekers are found to be entitled to international protection, and thus this initial disbelief seems unfounded.

Through its ability to categorise and sort large bodies of data, AI technologies hold the potential for deepening the knowledge required for asylum determinations by forecasting the likelihood of certain events transpiring. However, AI tools cannot make good predictions on bad data, and they can only make accurate predictions when the future is sufficiently regular to be predictable. Scholars have long argued that RSD, as a normative field beset by uncertainty and irregularity, is not particularly amenable to this kind of judgment. In practice, RSD thus tends to operate based on abductive reasoning processes, i.e. inference to the best explanation from alternatives, as opposed to more regular rule-based reasoning and subsumption.

More fundamentally, while current generation AI models are apt to identify non-negligible risks across large datasets (such as financial portfolios), they fare much more poorly when it comes to qualifying individual predictive outcomes with a low probability threshold, as is required in RSD. A problem further compounded by the fact that large language models (LLMs) are generally developed to risk “hallucination” rather than admitting uncertainty. Whilst research is currently underway to help quantify such uncertainties, LLMs may never be apt to deal with fine-grained problems such as assessing the “real risk” of persecution in individual asylum cases (for a related example, see here).

Conclusion: Support, not Substitution

Conceiving RSD as characterised by epistemic uncertainty suggests that the process is “knowable in principle” and any ignorance surrounding it can be mitigated at least in part. However, significant issues arise when a fact-heavy and intersubjective legal procedure is subject to algorithmic calculation, and when the grounds for calculation are far from obvious. This post has argued that beyond regulatory demands for adopting AI in a high-risk domain such as asylum, AI-assisted RSD entails a range of more technical challenges that may ultimately make it unsuitable for assessing whether an asylum claimant faces a “real risk” of being subjected to persecution or irreparable harm if returned. Ultimately, any deployment of AI in RSD requires a clear legal basis, published model documentation, independent validation against legally relevant outcomes, and safeguards against automation bias. Responsibility must remain with migration authorities, and claimants must have access to meaningful reasons and effective remedies.

This does not mean, however, that new technologies cannot play a more constructive role if conceived as supplementary critical tools more peripheral to the legal decision-making process itself. Exactly because existing asylum decision-making appears to be so prone to bias and arbitrariness, AI holds the potential to improve procedural fairness of RSD by zooming in on and providing reflexive support in these very areas.

First, AI can be used to nudge asylum decision makers. In law, nudging tools can help alert decision makers to more carefully consider elements of a case commonly overlooked, or to flag individual decision-making patterns (e.g., multiple rejections in a row) that significantly deviate from the average pattern. Nudging tools can also work to alert decision makers to different types of bias linked to specific types of claims or applicant profiles. Vice versa, computational research on asylum decision-making is increasingly documenting such bias and arbitrariness. Explainable AI (XAI) can further make visible inherent uncertainties embedded in decision-making processes. This could take the form of pro-hoc explanations, which offer alternative explanations for each possible outcome of a case, and essentially require decision makers to second guess their chosen outcome. Drawing attention to alternative explanations and prompting decision makers to pause, reconsider and/or provide more extensive reasoning in their decisions could also be used as “training wheels” for newly appointed decision makers, as experience is yet another factor shown to impact outcomes.

Second, the use of AI tools has significant potential to level the playing field between asylum claimants and the receiving state, a relationship which remains subject to significant disparities in resources and power. In Australia, researchers have used generative AI to develop a platform for asylum-seekers to understand their legal rights and help lodge formal complaints. More generally, lawyer-facing AI tools that help identify similar cases, analyse successful arguments and help prepare legal briefs could play a significant role in terms of widening access to legal representation, which in turn is known to have a significant impact on outcomes. AI tools could similarly play an important role when it comes to issues around evidence in RSD. For example, multi-lingual AI tools could be used to search for and identify potential translation errors as a commonly suspected source for claims being rejected. AI tools can also point to gaps where external evidence, such as country-of-origin information, should be updated or rests on limited or questionable sources.