Artificial intelligence is increasingly entering refugee status determination through tools such as dialect recognition, name transliteration, speech transcription, transcript summarisation, document analysis, country-of-origin research, and case matching. These tools are often presented as instruments of efficiency, consistency, and administrative support. Yet, in asylum law, they operate within a field where proof is already fragile and credibility is often decisive. Applicants may flee without documents, lose evidence during displacement, face trauma-related memory gaps, or remain unable to obtain corroboration from unsafe States. In such conditions, AI-shaped evidence may not merely assist decision-makers. It may silently influence how truth, identity, origin, and risk are understood. This paper examines the legal reliability standard that should govern AI-assisted evidence in asylum adjudication. It argues that technical accuracy alone cannot justify evidentiary reliance. Asylum decisions require legal trustworthiness, which must include explainability, traceability, data quality, contestability, human oversight, and protection against sole or decisive reliance on automated outputs. The paper analyses the doctrinal foundation of credibility assessment, the benefit of doubt principle, evidentiary vulnerability of asylum seekers, and the risks of administrative over-reliance on technical tools. It further evaluates the EU AI Act, especially its classification of asylum-related AI systems as high-risk, and considers its relationship with asylum law safeguards, non-refoulement, individual assessment, and the right to an effective remedy. The paper concludes that AI may assist asylum decision-making, but it must never replace human legal judgment. In refugee protection, technology must remain subordinate to fairness, reasons, and the duty to protect people from persecution and serious harm.