The rapid growth of AI use since 2022 has prompted regulatory responses ranging from cautious oversight to outright prohibition. Yet both effective regulation and meaningful accountability share a common prerequisite: AI use must be detectable. Detection, however, has proven challenging as it is a moving target that shifts with evolving practices and new generative models. Recent advances in authorship attribution and fraud detection have opened promising methodological avenues, but these tools were not designed for the legal domain and remain untested on legal text. This gap matters, because the legal domain is particularly exposed to this trend (OCDE, 2025) and, because judicial writing carries distinctive properties that pose specific challenges and opportunities for detection models. This article addresses this gap directly: we analyse the feasibility of detecting AI use in the legal domain by generating synthetic data and analysing different metrics and methods.