What the Council of Ministers Approved
On June 10, 2026, Italy's Council of Ministers gave preliminary approval to two legislative decrees implementing Law 132/2025, the national AI statute that took effect on October 10, 2025 (Gazzetta Ufficiale, Law 132/2025). One decree addresses AI governance; the other bundles law-enforcement use of AI with new civil and criminal liability rules for AI-caused harm. Undersecretary Alfredo Mantovano told reporters the "novelty of civil liability" was the standout feature of the package (ANSA). Neither decree is law yet — both now go to parliamentary committees, the State-Regions Conference, and the competent authorities for opinions before returning to the Council for final adoption, with a statutory deadline of October 10, 2026, twelve months after the parent law's entry into force (noze.it).
Four Levers for Claimants
According to the Council of Ministers' own press release, the civil liability decree gives someone harmed by an AI system four tools they did not clearly have before (Comunicato stampa CdM n. 177):
- Access to technical documentation — claimants can compel disclosure of a system's technical records to establish how it behaved.
- A rebuttable presumption of causality — where harm stems from a breach of AI Act obligations, the causal link between the system and the injury is presumed, subject to contrary proof by the defendant.
- Home-court venue — an injured natural person may sue in the court of their own residence rather than wherever the developer or deployer is domiciled.
- Direct action against the insurer — claimants can proceed straight against the liable party's insurance carrier rather than establishing the defendant's liability first.
The government describes this as procedural, not substantive: it does not create new duties for developers or deployers, and compliance with the AI Act does not by itself extinguish liability. The stated goal is to correct the "informational asymmetry" between a claimant who cannot see inside a model and a defendant who holds all the technical evidence.
The Case for It
The strongest argument for this approach is that traditional negligence doctrine assumes both sides can access the relevant facts, and that assumption collapses with opaque AI systems. A patient harmed by a diagnostic model, or a worker denied a loan by an automated scoring tool, has no way to reconstruct which inputs produced the output, let alone whether the system breached an AI Act obligation. Absent a presumption, the claimant effectively cannot prove causation regardless of the merits of their case. This is precisely the failure mode the European Commission itself identified when it proposed an EU-wide AI Liability Directive in 2022, and Italy's mechanism — presumption tied to a documented AI Act breach, rebuttable by the defendant — is a fairly narrow, evidence-anchored version of that logic rather than a leap to strict liability.
The Case Against Going It Alone
But the Commission withdrew that 2022 proposal. It signaled the move in its February 2025 work programme and formally withdrew the directive on October 6, 2025, citing an absence of "foreseeable agreement" among member states and a broader push toward regulatory simplification (IAPP). That leaves the EU with no harmonized fault-liability regime for AI at exactly the moment Italy is building its own. A developer selling the same model across the single market may now face a causality presumption, a home-court rule, and direct insurer exposure in Italy that does not exist in Germany, France, or Poland — the opposite of the predictability the AI Act's uniform risk classifications were supposed to deliver. Compounding this, Italy's decree ties the presumption to "breach of AI Act obligations" without narrowing it to high-risk systems specifically, leaving real ambiguity about how broadly courts will read it once the decree is finalized.
The combination of easier venue, a claimant-friendly evidentiary shortcut, and a straight line to the insurer's pocket is also, practically, an invitation to volume litigation — generative systems produce enough plausible-sounding errors that plaintiffs' firms have every incentive to test the presumption's edges before the ink on the final decree is even dry. None of that means claimants harmed by defective AI should be left with no remedy; it means the remedy should be built with the same rigor the AI Act itself applied to risk classification, rather than bolted on through a domestic decree that gets ahead of, and diverges from, the framework it is nominally implementing.
What to Watch
The draft still faces parliamentary and regional review before the October 10 deadline, giving industry a real window to narrow the presumption's scope — particularly whether it attaches only to high-risk systems under the AI Act's own taxonomy. Whether Brussels responds with a fresh EU-level liability proposal, or leaves member states to fill the vacuum individually, will determine whether Italy's decree becomes a template or an outlier.