AI Law - International Review of Artificial Intelligence LawCC BY-NC-SA Commercial Licence ISSN 3035-5451
G. Giappichelli Editore

17/09/2026 - Court of Cassation trials AI to map tax-case files, magistrate-supervised Copilot pilot (Italy)

argument: Notizie/News - AI in Judicial Activities

Source: Famiglia Cristiana

The Court of Cassation is testing artificial intelligence to assist in the initial review ("spoglio") of case files in the tax chamber, using an AI agent running on Copilot that was trained by magistrate Gianluca Grasso to filter acts and produce a procedural "map" of the case. The agent’s scope is limited to the documents uploaded by the human operator; it does not search prior cases or external databases, and it returns a map which must be compared and supervised by a magistrate who then manually enters the verified outcome into a platform inaccessible to the agent. The experiment ran for two months and is intended to free resources and accelerate long processes, with consideration of extending it beyond taxation.

The trial is conducted within a legal framework that the piece describes as including the EU AI Act (Regulation 2024/1689/UE), the CSM resolution of 8 October 2025 permitting Copilot for classification, summarization and analysis of acts, and article 15(1) of law 132/2025 which reserves all interpretative and decisional functions to judges. The dossier notes that on 16 September 2026 the Decreto Legislativo n.160 of 9 September 2026 was published in the Gazzetta Ufficiale to implement the EU AI Act in Italy, stressing human supervision and banning solely automated decisions that may cause prejudicial effects. That decree adds criminal provision art. 437‑bis punishing omitted security measures and illicit tampering of high‑risk AI, and inserts art. 359‑ter in the criminal procedure code authorizing, in extraordinary, time‑limited circumstances with judicial authorization, real‑time remote biometric identification for serious crimes listed in Annex II while forbidding indiscriminate biometric archives from web‑harvested data. The article also cites a 2026 Siracusa judgment that convicted a lawyer for using AI‑generated, unverified fictitious citations and a Brazilian labor decision that found a hidden AI command in a pleading, and it highlights the pervasive challenge of distinguishing genuine from falsified evidence (deepfakes) in court.