Making Crowds Searchable: Italy’s New Rules on Police Facial Recognition*

Arianna Vedaschi & Chiara Graziani

Arianna Vedaschi is Full Professor of Comparative Public Law at Bocconi University, Milan, Italy

Chiara Graziani is Assistant Professor of Comparative Public Law at Bocconi University, Milan, Italy

When does a face become a biometric identifier, searchable by the police before any offence has been committed? And who should oversee such intrusive power? These questions lie at the heart of the current debate on police use of facial recognition in Italy and provide the starting point for this blog post, which explores how the Italian legal framework regulates the use of facial recognition technologies by law enforcement authorities, while offering some comparative insights.

On 4 August 2026, the Council of Ministers gave final approval to a legislative decree adapting national law to the EU Artificial Intelligence Act (AI Act), including police use of biometrics. Although the final decree has not yet been officially published, the draft submitted to the Houses of Parliament and the institutional opinions issued on it highlight the legal risks of making crowds searchable in the absence of any criminal offence.

Delegated Powers and the Real-Time/Post-Remote Divide

The instrument chosen is a legislative decree, in which the Government retains legislative discretion, although within limits set by Parliament. Specifically, under Article 76 of the Italian Constitution, the Legislature may delegate the exercise of the legislative power to the Government (precisely, to the Council of Ministers, i.e. its collegial body) for a defined subject and period, according to guiding principles and criteria. The resulting decree has the force of an ordinary statute. In this case, Article 24 of Law No. 132/2025 delegated the adaptation of national law to the AI Act, including rules for police use of AI. A draft was approved on 10 June; the Italian Data Protection Authority (DPA) delivered its opinion on 14 July; parliamentary committees followed on 30 July; and approval of the – still unpublished – final version came in early August.

On the substance, “remote biometric identification” is an umbrella concept. The AI Act distinguishes real-time identification – in which the capture of facial images, their comparison against a reference database and identification occur simultaneously – from post-remote identification, where previously recorded images are analysed later to identify suspects.

Article 5 of the AI Act prohibits real-time use in publicly accessible spaces for law enforcement, subject to narrow exceptions concerning specified victims or missing persons, imminent threats (like terrorist attacks) and suspects in serious offences. In any case, such deployment must be authorised by a judicial or at least independent administrative authority.

Article 26(10) of the AI Act confines post-remote use to targeted searches for people suspected or convicted of an offence. It requires the deployer to seek authorisation from a judicial or administrative authority, either before use or within forty-eight hours of the moment of use. Notably, authorisation is not required when a suspect’s identification is based on objective facts directly linked to the offence.

The Italian draft replicates the distinction between real-time and post-remote identification. More specifically, two contexts govern real-time use. First, Article 8 regulates its use to prevent a specific, substantial and imminent threat to life or physical safety, or a genuine and present or foreseeable terrorist attack. It also covers searches for missing persons and victims of kidnapping, human trafficking or sexual exploitation. In these circumstances, senior police authorities must seek authorisation from the chief public prosecutor, specifying the target, purpose, area, duration, database and type of technology to be used. Authorisation lasts fifteen days and may be renewed. In urgent cases, notifying the prosecutor, even orally, is enough to allow use to begin, but a formal request must follow within twenty-four hours, and the prosecutor must decide whether to validate the use within the next twenty-four hours.

Second, Article 13 proposes a new Article 359-ter of the Code of Criminal Procedure for real-time use during preliminary investigations. It permits targeted searches for suspects in serious offences listed by the AI Act, fugitives, and specific victims. Here, the public prosecutor in charge of the case must obtain authorisation from the judge for preliminary investigations (GIP).

Article 10 instead governs post-remote identification, with a two-stage system. Upon entering certain venues or events, each person’s face is converted into a biometric template and stored locally for seven days, together with ticket information where available. If an offence subsequently occurs, police may compare the image of a person already suspected on objective and verifiable grounds with those templates to establish that person’s identity. Of greatest concern, the draft allows the comparison to take place without prior judicial authorisation. Although identification is retrospective, the biometric database on which it relies is created in advance by everyone present. This generates the draft’s main tension with the AI Act’s requirement that post-remote identification must remain targeted.

Post-Remote Identification and the Limits of Targeting

Despite the targeting requirement provided by Article 26(10) of the AI Act, Article 10 of the draft decree builds the database beforehand, converting every entrant’s face into a biometric template before any offence, thereby failing to satisfy the targeting requirement.

The Italian DPA’s opinion addresses this point by requesting a reversal of the sequence. Cameras may retain ordinary footage, but biometric analysis should begin only after an offence creates a concrete need. The investigation would then define the target before biometric data are extracted, implementing purpose limitation, data minimisation and proportionality. This approach reduces the risk that searchable attendance deters movement, assembly and expression protected by Articles 16, 17 and 21 of the Italian Constitution.

If targeting must precede biometric processing, independent scrutiny should verify the search’s justification and scope. Article 10 leaves the decision to the police; to increase independence, the parliamentary committees’ opinion proposes prior judicial authorisation. The committees recommend human oversight and prohibit rights-affecting measures based solely on a biometric match.

Conclusion: Drawing the Line on Biometric Search

At the heart of the Italian debate lie three distinct and sensitive stages: creating biometric data, authorising a search and acting on a match. This sequence provides the basis for a brief comparison. French law bars real-time facial recognition, while allowing retrospective identification. Law No. 2023-380, amended in 2026, permits algorithms to detect predefined security threats in live video, without biometric processing or personal data linkage.

Separately, French police may conduct post-remote identification through the Traitement d’antécédents judiciaires (TAJ) – a national database containing photographs from criminal investigations – by searching an investigative image against stored photographs. Dutch CATCH, under the Police Data Act, follows the same post-remote model: it searches an investigative image against a pre-existing criminal-justice database; experts review suspects, and Dutch police stated that it is never used live. Therefore, neither the French nor the Dutch framework creates biometric templates from all people attending an event before any offence. Article 10, at least in its draft version, does: it makes the crowd searchable first and searches for a suspect later. The comparative analysis confirms that recording images, extracting biometric data, and matching identities are distinct exercises of public power, each requiring its own legal basis and safeguards. As we argued in our recent book, rule of law scrutiny should cover the entire technological process.

Publication of the final decree will reveal which safeguards survived. The key challenge is striking a balance: security remains a crucial interest, but classifying a search as post-remote cannot justify prior biometric processing of an entire crowd. In other words, a democratic society need not promise invisibility in public and has a duty to safeguard its citizens; it must nevertheless allow participation in public life without making one’s face a police search term by default. Facial recognition must hence remain targeted, necessary, proportionate and independently scrutinised.

Arianna Vedaschi is Full Professor of Comparative Public Law at Bocconi University, Milan, Italy 

Chiara Graziani is Assistant Professor of Comparative Public Law at Bocconi University, Milan, Italy

*Note: After this post was written, Legislative Decree No. 160 of 9 September 2026 was published in the Official Gazette on 15 September 2026. The final text addresses some of the central concerns raised in this post regarding the draft. Most importantly, Article 10 now provides that, before an offence occurs, facial images may be stored without processing the corresponding biometric data; biometric processing may take place only subsequently, following the commission of an offence, for the purposes of retrospective comparison. The final text further explicitly prohibits generalised, untargeted or indiscriminate biometric identification unrelated to a specific offence or criminal proceeding. Moreover, taking into account an issue that had also been raised in this post, it introduces judicial oversight for targeted searches of persons already suspected of an offence: the public prosecutor must seek authorisation from the judge for preliminary investigations within forty-eight hours of the system being activated. The final decree therefore addresses a significant part of the concerns discussed in this post, while leaving the initial identification of a potential suspect outside the requirement of prior judicial authorisation.

Suggested Citation: Arianna Vedaschi and Chiara Graziani, ‘Making Crowds Searchable: Italy’s New Rules on Police Facial Recognition’, IACL-AIDC Blog (01 October 2026) Making Crowds Searchable: Italy’s New Rules on Police Facial Recognition