2026-10-12 - Facial recognition and the AI Act: the Italian choice in Legislative Decree 160/2026 deserves careful reading
This contribution continues the series devoted to Italian
Legislative Decree no. 160 of 9 September 2026
(in Italian), which began with an
. The focus here is on Chapter III of Title I, devoted to biometric systems: the part in which the national legislator exercised the most sensitive margin that
grants to Member States.
The margin left by the AI Act
Article 5(1)(h) of the AI Act prohibits the use of real time remote biometric identification systems in publicly accessible spaces for law enforcement, unless such use is strictly necessary for one of three exhaustively listed objectives: the targeted search for specific victims of abduction, trafficking or sexual exploitation and for missing persons; the prevention of a specific, substantial and imminent threat to the life or physical safety of persons or of a genuine and present or foreseeable threat of a terrorist attack; the localisation or identification of persons suspected of having committed the offences listed in Annex II, punishable by a maximum sentence of at least four years.
The prohibition is the rule; the exceptions are subject to cumulative conditions. Paragraphs 2 to 7 of Article 5 require, among other things, a fundamental rights impact assessment and registration of the system, prior authorization granted by a judicial authority or by an independent administrative authority whose decision is binding, and notification of each use to the market surveillance authority and to the national data protection authority. Above all, paragraph 5 makes such use possible only where a Member State provides for it through detailed national rules.
Legislative Decree no. 160/2026 therefore reflects Italy’s decision to use that option. Alongside it, the decree regulates retrospective facial recognition, which the AI Act does not prohibit but, in Article 26(10), subjects to a dedicated authorization when used to target a person suspected or convicted of a criminal offense.
A dual track for real-time use
The legislator has distributed the three objectives of Article 5 across two separate tracks.
The prevention track is governed by Article 8 of the decree and covers preventing threats and searching for missing persons or victims of specific offenses. Use is permitted exclusively to confirm the identity of the persons specifically concerned, including dynamically and with their localization, and requires the authorization of the public prosecutor at the court of the district capital, upon request of the chief of police (questore) or of the provincial commanders of the Carabinieri and the Guardia di Finanza, or of the heads of the central inter-force services. The authorization specifies the event or the strictly necessary period, in any case not exceeding fifteen days and renewable, the delimited territorial area and the persons sought.
The criminal proceedings track is governed by the new Article 359-ter of the Code of Criminal Procedure, introduced by Article 13 of the decree, and covers identifying and locating persons suspected of Annex II offenses, searching for fugitives, and conducting targeted searches for victims within criminal proceedings. The judge for preliminary investigations grants authorization on the public prosecutor’s request.
Both tracks provide for an urgency regime: use may start before authorization, with strict time limits for the request and the decision, consistent with the twenty-four-hour limit laid down in Article 5(3) of the Regulation. Where authorization is missing or refused, use stops, data and outputs are deleted, and the results cannot be used, without prejudice, in the prevention track to evidence lawfully acquired on another legal basis and, in criminal proceedings, to establish the corpus delicti.
AspectPrevention (Art. 8)Criminal proceedings (Art. 359-ter CCP)Retrospective (Art. 10)
AI Act referenceArt. 5(1)(h)(i) and (ii)Art. 5(1)(h)(iii); point (i) for the search for victimsArt. 26(10)
Authorising authorityPublic prosecutorJudge for preliminary investigationsJudge for preliminary investigations
DurationUp to 15 days, renewableUp to 15 days, renewableLinked to the targeted search
UrgencyStart ordered by the questore or the commanders; request to the prosecutor within 24 hours, decision within the following 24 hoursProsecutor's decree notified to the judge within 24 hours, validation within the following 48 hours; if the judicial police start use, request to the prosecutor within 12 hours and request for validation to the judge within 24 hours of start, validation within the following 48 hoursRequest to the judge within 48 hours of start, decision within the following 48 hours
Without authorisationStop, deletion, results unusable, save for evidence acquired on another legal basisStop, deletion, results unusable, save for the corpus delictiStop, deletion, save for the corpus delicti
Which authority authorizes
In the
contribution on the preliminary examination
I flagged, based on the draft version then circulating, the attribution to the public prosecutor of the power to authorize use for preventive purposes. The final text confirms that choice, and the table shows the resulting asymmetry: for the same technology, with the same degree of interference with rights, authorization lies with a judge in criminal proceedings and in retrospective recognition, and with the public prosecutor in prevention.
Article 5(3) of the AI Act requires a judicial authority or an independent administrative authority whose decision is binding. In the Italian legal order, public prosecutors belong to the judiciary and enjoy constitutional guarantees of independence, which is the strongest argument for this choice.
The benchmark, however, is a concept of Union law, to be interpreted autonomously. In a related field, access by authorities to traffic and location data, the Court of Justice ruled out that a public prosecutor responsible for directing the investigation and bringing the prosecution can carry out the independent prior review required by Union law, since it is not in a position of neutrality with respect to the interests at stake (
judgment of 2 March 2021, Case C-746/18, Prokuratuur
, ECLI:EU:C:2021:152). The Italian legislator followed up on that ruling with
Decree-Law no. 132 of 30 September 2021
(in Italian), which entrusted a judge with authorizing the acquisition of traffic data.
The two situations do not coincide. The regulatory context differs, and in prevention the prosecutor does not act as a party to proceedings already underway. In my view, however, the question remains open and deserves further attention, bearing in mind that the ruling concerns access to traffic and location data and does not interpret the AI Act. A restrictive reading of judicial authority could also affect the usability of the results.
Comparison databases
The decree precisely addresses an aspect that the AI Act leaves in the background: what the detected face is compared against. The reference database must be adequate for each purpose and specifically formed for each use, with only the relevant biometric data; it may be derived from databases used by the police forces or held by other administrations, or from images lawfully acquired in the course of police activities; it must be deleted when the authorization expires, so that the comparison set is specific and cannot be incrementally fed with respect to previous authorizations.
The use of biometric databases fed, in whole or in part, through untargeted scraping, or created in breach of data protection law, is prohibited in all cases. The provision ties in with the prohibition in Article 5(1)(e) of the AI Act, which concerns the creation or expansion of facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage, but draws a further consequence from it: it is not enough to refrain from creating such databases; they must not be used either.
Retrospective facial recognition
Article 10 allows AI components to be integrated into video surveillance systems installed under specific statutory provisions. Facial recognition components may be activated only after images have been acquired, with a time interval that does not amount to real-time identification.
The wording deserves attention. The AI Act defines a real-time system as one in which identification occurs without significant delay, including not only instant identification but also limited short delays, specifically to prevent circumvention (Article 3, point 42). The decree sets no minimum interval and relies on a substantive criterion; in practice, that is where the boundary between the two regimes will be tested.
For a targeted search of a suspect, the public prosecutor requests authorization from the judge for preliminary investigations without delay and, in any case, within forty-eight hours of the start of use, and the judge decides within the following forty-eight hours. In line with Article 26(10) of the Regulation, no authorization is required where the technologies are used exclusively after the commission of an offense solely for the initial identification of a potential suspect, based on objective and verifiable facts directly linked to the offense; in that case, use takes place under the direct and exclusive responsibility of the public security officer designated by the questore or, outside that case, of the judicial police officer in charge.
Three safeguards complete the picture: no decision producing adverse legal effects may be based solely on the results of facial recognition; the systems may not be used in an untargeted way, without any link to an offense or proceedings, nor for generalized or indiscriminate identification; access and operations are recorded in logs that cannot be altered, kept for five years.
In terms of timing, Article 26(10) of the AI Act will apply from 2 December 2027, under the timetable amended by
. It will therefore be necessary to verify how that date is coordinated with the provisions of Article 10 of the decree, in light of the reference to the European dates contained in
(in Italian).
Public venues and events: storing without processing
The most original solution concerns venues and events with particular public order and security requirements, identified by decree of the Minister of the Interior. In these contexts, systems automatically process facial images of people entering, without processing their biometric data. The images are stored locally, along with personal details and, where applicable, the assigned seat, obtained by electronically scanning access tickets, and are kept for seven days, after which they are automatically deleted.
Facial recognition is activated only after an offense has been committed, for retrospective comparison with the suspect’s face. The controller is the Ministry of the Interior, Department of Public Security, which conducts a prior impact assessment and consults the Garante; venue operators and event organizers may install and maintain the equipment, which the questura is lent free of charge.
The model rests on the distinction, established in Union data protection law, between a facial image, which is personal data, and biometric data, which presupposes specific technical processing aimed at unique identification. The distinction is legally sound, but its effectiveness will depend on the technical details set out in the ministerial decree: the systematic, named collection of all participants’ faces at an event remains, in itself, processing of considerable breadth that calls for proportionate safeguards.
Biometric categorization
Article 7 concludes the chapter by regulating the labeling, filtering, and categorization of lawfully acquired biometric datasets, which Article 5(1)(g) of the AI Act excludes from the prohibition on biometric categorization in law enforcement. The decree adds conditions: these operations may not be aimed at inferring the characteristics protected by the European provision, must serve comparison or search activities or further activities compliant with the applicable rules and exclusively for police purposes, may not constitute the sole basis for decisions producing legal effects, and the controller must prevent the reuse of data for incompatible purposes.
Concluding remarks
Chapter III translates principles into verifiable conditions of use: authorizations, time limits, databases formed for each use, deletion, unusability of results, logging. This is the direction the Regulation requires, and the direction I hoped for in the June contribution.
In my view, three aspects deserve to be followed in practice: the authority entrusted with authorizing use for preventive purposes, measured against the European concept of judicial authority; the boundary between real-time and retrospective use, entrusted to a substantive criterion; and the model for public venues and events, whose proportionality will largely depend on the implementing ministerial decree.
The next contribution in the series focuses on the obligations accompanying these systems: fundamental rights impact assessment, data protection impact assessment, logging, and documented human review.
For those wishing to navigate the regulatory framework referred to, the
of this site are available: the
, with the state of application of the AI Act, the map of Annex III and operational checklists; the
, devoted to roles, obligations and the chain of responsibility; and the
, on the operational intersections between the two regulations.
The reference sources are available on the institutional websites:
Legislative Decree no. 160 of 9 September 2026
in the Official Gazette (in Italian),
on EUR-Lex, the
judgment of the Court of Justice in Case C-746/18
on Curia and
Decree-Law no. 132 of 30 September 2021
on Normattiva (in Italian), and
on EUR-Lex.
New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.