Unstable ground: Automated dialect recognition, ground truth, and AI regulation in asylum procedures

Abstract While problems of AI systems are often attributed to technical shortcomings, some systems fail because they are built on ontologically and epistemologically “unstable ground.” Ground truth, or the datasets against which machine-learning systems are trained, validated, and tested, renders a particular representation of the world learnable by translating complex qualitative phenomena into mathematical formalisms. In doing so, it can stabilize tenuous truth claims while concealing fundamental flaws in how those phenomena are modeled, leaving systems unable to address the problems they are designed to solve. We examine this problem through a case study of Germany’s Dialect Identification Assistance System (DIAS), an automated language assessment system used in determining asylum applicants’ countries of origin. Drawing on Ground Truth Tracing (Kang 2023), we identify three reductions through which DIAS links speech to country of origin: treating language as code, mapping dialect onto bounded territory, and inferring nationality from this mapping. From linguistic and legal perspectives, these assumptions lack sufficient empirical grounding. DIAS may therefore produce internally coherent outputs while lacking the capacity to fulfil its intended task. We further argue that the EU AI Act is ill-equipped to detect such ontological flaws, as its risk-based framework evaluates systems against internally constructed ground truth rather than interrogating its aptitude to solve real-world problems. In asylum procedures, where such outputs may affect decisions on protection and return, these failures carry profound consequences.