This paper examines the functioning of large language models (LLMs) in the field of private international law. Its primary objective is to assess their effectiveness in determining the applicable law, particularly in the analysis of choice-of-law clauses, and to identify structural limitations arising from the architecture of these models. The study applies a dogmatic-legal and functional methodology, drawing on examples from legal practice, including due diligence processes and the use of AI systems in law firms. The central thesis is that the challenges associated with LLMs in private international law are structural in nature and stem from their probabilistic design, including tokenization, hallucinations, and semantic averaging. These features create a risk of misidentifying conflict-of-law rules, misinterpreting choice-of-law clauses, and introducing systemic bias toward dominant legal systems. The conclusions indicate that despite significant technological progress, particularly in reasoning models and retrieval-augmented generation (RAG), LLMs cannot replace legal professionals in applying conflict-of-law rules. Their use requires continuous human oversight, and the development of specialized, locally trained models may provide a more suitable path for private international law practice.