Rome Moves First, But Not Cleanly
On August 4, 2026, Italy's Council of Ministers gave final approval to two legislative decrees implementing the EU AI Act (Regulation (EU) 2024/1689) into national law, completing the framework set out under enabling Law 132/2025. The package makes Italy the first EU member state to finish a full national implementation architecture for the Act, ahead of several peers still drafting their own governance statutes.
The core achievement is institutional clarity. One decree names the Agency for Digital Italy (AgID) as the national notifying authority, responsible for accrediting the conformity-assessment bodies that certify high-risk AI systems before they reach the market. The National Cybersecurity Agency (ACN) becomes the market surveillance authority and Italy's single point of contact with Brussels. Sector regulators keep their existing turf: Bank of Italy, Consob and IVASS oversee high-risk AI in financial services, while the data protection authority, the Garante, retains jurisdiction over AI used in law enforcement, border management, justice and elections. For a regulation as sprawling as the AI Act, assigning one clearly empowered notifying authority and one clearly empowered surveillance authority — rather than leaving compliance diffused across a committee — is exactly the kind of legal certainty that lets AI vendors and deployers plan investment instead of guessing who they answer to.
The second decree governs AI in policing and introduces liability rules that apply economy-wide. Employers cannot base hiring, discipline or dismissal decisions solely on automated processing; a dismissal that violates this rule is void. Italy's Criminal Code gains a new offense, Article 437-bis, punishing failure to adopt required security measures in high-risk AI systems, with corporate liability extending under Legislative Decree 231/2001. On the civil side, victims harmed by an AI Act violation get a rebuttable presumption of causal link between the violation and the damage, the right to sue the liable party's insurer directly, and the option to file suit at their own place of residence rather than the defendant's. That is a genuinely useful reform: it gives people harmed by AI systems a workable path to compensation without forcing every dispute into an open-ended negligence inquiry, while keeping the presumption rebuttable rather than a strict-liability trap that would deter deployment.
The Part Brussels Isn't Sure About
The police provisions are where the decree gets contested, and the tension is legitimate on both sides. Italian officials point to real incidents — including violent clashes with police during protests in Bologna and in the Susa Valley — as the reason law enforcement needs modern identification tools. That is a fair ask: the AI Act does not ban forensic, after-the-fact facial recognition, and no serious civil-liberties framework requires police to investigate assaults on officers with one hand tied behind their back.
Where the decree strains is the mechanism it chose. It authorizes police to record and retain camera footage from "sensitive" locations — squares, stadiums, protest routes — for seven days, comparing faces against a purpose-built database only if a crime is later reported, and deleting everything if none is. Real-time activation and any comparison run require prior judicial authorization, sought within 24 hours of the system going live. The decree separately bans building biometric databases by scraping the open web in bulk — a clear rebuke of the Clearview AI model that has already drawn fines from data-protection authorities elsewhere in Europe.
But bounding retrospective comparison with judicial sign-off doesn't resolve the objection Italy's own Garante and the European Commission have raised: the underlying data — everyone's face, at every protest and stadium, whether or not they ever become a suspect — is captured before any judicial authorization exists at all. The AI Act's Article 5 bans real-time remote biometric identification in public spaces except in narrowly listed situations — missing persons, imminent terrorist threats, or suspects in serious crimes — precisely because indiscriminate capture chills lawful assembly regardless of what happens to the footage afterward. Rome's position, that collection and identification are legally distinct stages, may hold up domestically; a European Commission spokesperson has so far declined to confirm or deny the prohibition applies here, an ambiguity Rome should not mistake for cover.
What Should Happen Next
The rest of this package deserves to be judged on its own merits, and mostly it is well-built: a single notifying authority, a single market-surveillance authority, real liability teeth for victims, and guardrails against solely-automated employment decisions are the kind of proportionate, evidence-based regulation this publication favors. The facial-recognition provision doesn't need to be scrapped to fix this — it needs the same judicial gate applied at the moment of capture that already applies at the moment of comparison. Absent that fix, Italy risks becoming the test case that hands Brussels its first AI Act enforcement dispute against a member state, a distraction the rest of a genuinely solid framework doesn't deserve.