Legislative Decree No. 160 of 9 September 2026 was published in the Gazzetta Ufficiale (general series no. 214) on 15 September and took effect on 30 September, according to Noze's summary of the Gazette text. It is the first implementing decree under Law 132/2025, Italy's national AI law, and it runs to 22 articles in three titles. One of them regulates how police may use AI biometric identification. The Italian government wrote it while the European Commission was publicly restating that facial recognition in publicly accessible spaces is banned in principle.
The case for suspicion
The sceptics' argument is strong and deserves a fair statement. The AI Act's recitals call real-time remote biometric identification in public spaces "particularly intrusive" because it can create a feeling of constant surveillance. They allow it only in "exhaustively listed and narrowly defined situations" (Regulation (EU) 2024/1689). A national decree that operationalises the tool, and gives police a procedure for switching it on, can look like a loophole being turned into a doorway. On 30 July a Commission spokesperson said the AI Act bans facial recognition in accessible public space. The spokesperson added that the Commission had no specific information on the Italian text and would not comment on it, ANSA reported. Sources in the prime minister's office replied that Italy complies with EU law.
What the decree actually does
The worry is overstated for the final text, at least as secondary reporting describes it. I could not load the Gazzetta text itself, so the details below come from commentary written from it.
The AI Act already lets member states choose whether to permit real-time use under national law, and to restrict it to some of the permitted objectives. It requires a fundamental-rights impact assessment and an authorisation for each use. Decree 160 appears to use that room narrowly. According to CASRAI's guide, Article 8 allows real-time identification for prevention only to find missing persons or victims of abduction, trafficking or sexual exploitation. A prosecutor must authorise each use. The authorisation is capped at 15 days, renewable in 15-day blocks. The authorising decree must fix the geographic area and name the persons sought.
The criminal-procedure track is stricter. A new Article 359-ter of the code of criminal procedure lets a preliminary-investigations judge (GIP) authorise real-time identification of a suspect. It requires sufficient indicia of an offence carrying a maximum sentence of at least four years. The guide also reports that:
- a fundamental-rights impact assessment and a data-protection impact assessment are required before any real-time use;
- each use is logged in immutable files kept for five years;
- the Garante is notified after use, once the judicial authority gives clearance, and that notice can be deferred by up to three months for secrecy reasons;
- reference databases are built for each specific use, deleted when the authorisation expires, and cannot be accumulated across authorisations;
- databases built wholly or partly by untargeted scraping are prohibited, and a use that breaks the rules must be stopped, its data deleted and its results excluded from evidence.
This tracks the AI Act's own logic. The Act's recitals require an express authorisation from a judicial or independent administrative authority for each use, and bar any decision with an adverse legal effect resting solely on the system's output. The decree's scraping ban matches the Act's prohibition on building facial databases through untargeted scraping of internet or CCTV images.
Where the real risk sits
The weaker points are narrower than the headline suggests. Post-event searches are handled in Article 10. Targeted searches for a suspect need GIP authorisation, which police must request within 48 hours of activation and which the judge must decide within the next 48 hours. CASRAI also reports an exception. Where the system is used only after an offence, solely to make an initial identification from objective elements tied to the offence, no judicial authorisation is needed, and a designated officer takes responsibility. That is a sensible operational carve-out, since waiting for a warrant to look at a photo from a crime scene is impractical. It is also the place where a safeguard can quietly turn into a routine.
Two further gaps matter. First, the Interior Ministry decree that is meant to set reliability, accuracy and bias-monitoring standards had not been issued as of 25 September, per the same guide. Until it exists, the accuracy floor is undefined. Second, reports differ on transition. The guide says police have until 30 September 2027 to align with the general police-AI rules in Title I, Chapter II (Articles 3 to 6), and that the biometric provisions get no grace period. Other commentary describes a year-long adjustment for police AI systems already in use. Readers should treat the transitional scope as unsettled until the official text is checked.
One more detail shows the drafting is not careless. Italia Oggi reports that municipal CCTV may carry biometric components, but this gives local police no power to use them. Where such systems are lent to the Questura, it holds full and exclusive control, and the Interior Ministry is the data controller. Municipalities must keep a register of any biometric functions, even disabled ones. Clear lines of accountability are what make later audits possible.
Our view
A blanket ban that police ignore is worse than a narrow, reviewable permission. A prohibition that exists only on paper leaves forces to use vendor tools with no logs, no assessment and no judge. Italy's approach pushes the decision toward a magistrate, requires a written record, and forbids the data-harvesting practices that made the technology controversial. That is proportionate regulation, and a pro-innovation reading of it should welcome clear rules for lawful use.
But text is not enforcement. The decree's value depends on whether prosecutors and GIPs scrutinise requests or approve them in bulk, whether the Garante gets timely notice, and whether the Commission decides that a national permission fits Article 5 of the AI Act. Rights groups will be right to demand published statistics on requests, approvals and false matches. Italy should publish them annually. If it does, the decree could be a model. If it does not, the safeguards will be only as strong as the least curious magistrate.