Back to the blog
Italy's Decree 160/2026 on AI: new offence, entity liability, evidence
AI & automation

Italy's Decree 160/2026 on AI: new offence, entity liability, evidence

Since 30 September 2026, Italy's Decree 160/2026 adds Article 437-bis to the criminal code, brings two offences into the liability of entities and gives judges new powers over evidence. It concerns companies that develop, integrate or use AI systems.

8 October 2026· 7 min·by NaCode Studios
Share
In short

Since 30 September 2026, Italy's Decree 160/2026 adds Article 437-bis to the criminal code, brings two offences into the liability of entities and gives judges new powers over evidence. It concerns companies that develop, integrate or use AI systems.

Legislative Decree no. 160 of 9 September 2026 came into force on 30 September 2026 (Official Gazette no. 214 of 15 September) and implements law 132/2025. It creates the offence of failing to secure high-risk AI systems (Article 437-bis of the criminal code), brings that offence and the unlawful dissemination of AI-altered content among those for which a company also answers (Decree 231/2001 on the liability of entities), and gives claimants new tools of proof. If you develop, integrate or use AI systems, first establish who oversees them and where their logs are kept.

Updated 8 October 2026. This text rests on the decree in the Official Gazette, on law 132/2025 and on commentary by Italian legal writers. It is general information and not legal advice: consult a lawyer about your own case.

What the decree provides

  • Article 437-bis of the criminal code (Article 12 of the decree) punishes with one to five years' imprisonment anyone who omits the technical security measures of high-risk AI systems or the human oversight measures, if danger to life or to public or individual safety results. If state security is endangered the sentence is two to eight years. For gross negligence it is reduced by one third to one sixth. Anyone who alters a high-risk system, if danger results, faces two to six years, or three to ten where state security is endangered. A professional user who deliberately omits human oversight is punished like anyone who omits the measures.
  • Article 15 inserts Article 25-vicies into Decree 231/2001. An entity (a company or association) answers for offences committed in its interest or to its advantage (Article 5 of Decree 231/2001) and here faces a fine of 600 to 1,000 quotas for Article 437-bis and 200 to 700 quotas for Article 612-quater, plus disqualifying sanctions, including a ban on contracting with public authorities and the suspension of authorisations; disqualification from carrying on the business is excluded. Article 612-quater, in force since 10 October 2025 under law 132/2025, punishes with one to five years' imprisonment anyone who causes unjust harm by disseminating, without consent, images, video or voices altered with AI and capable of deceiving.
  • In damages claims for harm caused by the use of an AI system, the judge can order the production of evidence on how it works if the claimant makes the claim credible (Article 17): logs, risk management documentation, technical documentation, information on human oversight. If a party fails to comply without justified reason, the judge may draw evidentiary inferences and, for those documents, after weighing all other evidence, treats the facts alleged by the claimant as admitted. A non-compliant third party pays EUR 1,500 to EUR 10,000.
  • If the damage results from breach of obligations under the European regulation, causation is presumed unless proven otherwise (Article 18). A system's conformity, even if certified, does not of itself exclude the defendant's liability (Article 19). The injured party can sue the insurer directly, within the policy limit (Article 20).
A quota is worth from 500,000 to 3,000,000 lire (Article 10 of Decree 231/2001), roughly EUR 258 to EUR 1,549: 600 to 1,000 quotas come to about EUR 155,000 to about EUR 1.5 million for the Article 437-bis offence [illustrative estimate, to be confirmed against your own project].

Who it concerns

The Article 437-bis offence applies only to high-risk systems: certain areas of use in Annex III of the regulation, such as recruitment or access to essential services, and safety components of regulated products. According to BibLus, a company chatbot used to write emails or summarise documents does not become high-risk just because it is used at work. The rules on evidence apply to any AI system (Article 16).

The decree and the European deadlines

Regulation (EU) 2026/1744, in force since 27 July 2026, moved the rules on Annex III high-risk systems to 2 December 2027 and those on Annex I to 2 August 2028. Article 437-bis applies from 30 September 2026. According to BibLus, this does not make the postponed requirements mandatory at once: an omission is established against the measures actually due when the act was committed.

The European postponement does not move the order to produce evidence in Article 17, in force since 30 September. According to La legge per tutti, many companies do not have their logs at hand: they sit on the supplier's servers, often outside the European Union, and standard contracts often say nothing about logs and documentation.

What to do now

The decree imposes neither a register nor a checklist, as Altalex notes: these are the checks we start from.

  1. List the AI systems in use, including tools the company has not approved, with the supplier of each.
  2. Flag those that fall in the Annex III areas or are safety components of a product.
  3. For each one, decide who approves it, who exercises human oversight and which controls leave a trace.
  4. Ask suppliers in writing which logs and documentation you can access, for how long, and with what cooperation if a judge orders them.
  5. If you have a 231 model, update its risk mapping with the supervisory body and check your liability insurance.

Where this touches our work

We integrate language models into clients' existing workflows, with control over costs and answer quality: that is our AI automation and chatbot service. We also build management systems and internal tools starting from an analysis of the processes already in use, in our custom software development service. The decree touches this work: logs, human oversight and documentation are decided when the software is designed.

If you ask us to look at an AI feature already in use, we start from four points: which decision the system makes and who confirms it; which events it records and who can hand them over; what technical documentation exists; and what the contract with the model supplier says about logs, incidents and access. For a new project we propose planning, from the analysis stage, an event log the client can export and a step where a person confirms sensitive decisions: this is our own proposal, to be assessed case by case.

We do not give legal opinions: classifying a system and assessing liability is for a lawyer.

What we do not know yet

  • How the offence will be read against the postponed European obligations: for now there is only BibLus's reading.
  • Where danger to life or to safety stops: the text does not define it and the enabling law speaks of concrete danger.
  • How the new rules apply to facts before 30 September: the transitional regime in Article 21 covers only police systems.
  • For how long and in what form logs must be kept: the decree sets no periods or formats.
  • When the second decree of 4 August, on authorities, administrative sanctions and training, will appear: it is not in the Official Gazette up to 7 October and, according to the Government, it will set maximum penalties below European ones.

If you want us to look at an AI feature in your software, starting from logs and human oversight, see our AI automation and chatbot service or get in touch.

Sources

Share

Got a project in mind?

Let us build it together

We can help you turn an idea into a digital product that is solid, fast and ready to grow.

Related articles