description
The paper addresses the challenges of artificial intelligence (AI) in the justice system, particularly in criminal justice, and proceeds from the premise that the first task is the correct classification of the system. The AI Act (Regulation (EU) 2024/1689) is a product-safety instrument with an added fundamental-rights dimension, built on a risk-based approach that allocates obligations among providers, importers, distributors and deployers. Systems for assessing evidence, profiling and supporting judicial decision-making fall within the high-risk category, yet the boundaries between prohibited practices, exceptions and permitted systems remain unclear, while preparatory tasks are excluded; this gap is partly filled by the Council of Europe Framework Convention and by the HUDERIA methodology and model, which tie obligations to interference with fundamental rights. Data protection is a key component, since the anonymity of a model cannot be taken for granted, and the initial unlawfulness of collection may be transferred to the deployer, with additional risks arising from transfers to third countries, which calls for the anonymisation of material before input and the use of systems storing data within the EU. Legal liability remains governed by the national instruments of the Member States, as the AI Act is silent on the matter and the proposed directive was withdrawn; the gap is partly covered by Directive (EU) 2024/2853, while in criminal law negligence and duty of care, together with new endangerment offences, appear the most appropriate route. In the field of evidence, Article 86 of the AI Act provides only a partial answer through its requirement of explanation, and the paper therefore sets out a minimum standard of adversarial process derived from Article 6 ECHR, namely the logging of use, a meaningful explanation, defence access to logs, a narrow reading of exceptions and the verifiability of decisive evidence. The risks are thus institutional rather than technological, and AI must remain added rather than super intelligence.