The two decrees complete the institutional framework envisaged by Regulation (EU) 2024/1689 (the AI Act). They address key aspects including the organization of the competent national authorities, the penalties regime, regulatory sandboxes and AI literacy measures, as well as specific rules governing the use of artificial intelligence systems in public security activities and the civil and criminal liability arising from their use.
Decree 179: authorities, training, penalties, employment and trade secrets
The competent national authorities
The authorities responsible for implementing and overseeing the application of the AI Act in Italy have been designated, building on the framework already established by Article 20 of the Law.
Specifically:
- Agenzia per l’Italia Digitale (AgID) is designated as the national notifying authority, responsible for establishing and implementing the procedures required for the assessment, designation and notification of conformity assessment bodies, as well as for monitoring those bodies.
- Agenzia per la Cybersicurezza Nazionale (ACN) is designated as the market surveillance authority, single point of contact and competent authority for the national registration of the high-risk AI systems referred to in Annex III, point 2, of the AI Act (i.e., systems used as safety components in the management and operation of critical infrastructures).
The Decree also confirms the supervisory powers of the relevant sectoral authorities—Banca d'Italia, CONSOB, IVASS and the Italian Data Protection Authority—within their respective areas of competence. The overall framework is based on institutional cooperation and is intended to facilitate coordination and the exchange of data between authorities through a dedicated Coordination Committee.
The penalties regime
Decree 179 makes use of the discretion granted to member states under the AI Act to define their national penalties regime, aligning the applicable maximum penalties with those laid down in the Regulation while calibrating the amounts according to the seriousness of the infringement and the role of the relevant operator within the AI system value chain.
Consistently with the approach taken under the AI Act, the Decree establishes a graduated system of penalties. These range from EUR1 million (or 0.5% of total worldwide annual turnover, if higher) for failure to comply with the obligations imposed on deployers under Articles 27 and 86 of the AI Act, to EUR7.5m (or 1% of total worldwide annual turnover, if higher) for supplying incorrect, incomplete or misleading information to the competent authorities.
Penalties may increase to EUR15m (or 3% of total worldwide annual turnover, if higher) for breaches including, among others, the obligations applicable to deployers of high-risk AI systems under Article 26 and the transparency obligations under Article 50 of the AI Act. The same threshold also applies to breaches of obligations imposed on providers, authorized representatives, importers, distributors and notified bodies, as well as of the requirements relating to the reporting of serious incidents.
The highest tier of penalties—up to EUR35m (or 7% of total worldwide annual turnover, if higher)—applies to infringements of the prohibitions on AI practices set out in Article 5 of the AI Act.
AI literacy
Another key pillar of Decree 179 concerns measures aimed at improving AI literacy among the various groups affected by the increasing use of artificial intelligence systems. To this end, the provision promotes training initiatives aimed, among others, at students, teachers, public-sector employees, members of regulated professions, healthcare professionals and members of the judiciary.
Employment law
Specific rules governing the use of artificial intelligence systems in the employment context have been introduced, consistently with the principles already set out in Article 22 of Regulation (EU) 2016/679 (GDPR) and Article 11 of the Law .
In particular, where an employer uses artificial intelligence systems in decision-making processes relating to the employment relationship, it must ensure that final decisions concerning the establishment, modification or termination of that relationship—including disciplinary measures and performance assessments—are ultimately taken by a natural person exercising genuine and independent decision-making authority. An employee’s consent does not constitute an exception to this prohibition.
By contrast, candidate search and selection activities (so-called “CV screening”) fall outside the scope of the aforementioned rules governing final decisions, including where they result in a candidate not progressing to subsequent stages of the recruitment process.
In such cases, the employer must have provided the relevant information in advance pursuant to Article 1-bis of Legislative Decree No. 152/1997. The individual concerned retains the right, upon request and with the involvement of a natural person, to obtain an intelligible explanation of the decision affecting them, including an explanation of how the artificial intelligence system influenced the decision-making process and of the main parameters considered.
Any dismissal imposed in breach of the requirements outlined above is expressly deemed null and void by law, as are any other decisions taken by the employer in breach of those requirements.
The Decree also addresses occupational health and safety. Employers are required, as part of their Risk Assessment Document (Documento di Valutazione dei Rischi—DVR), to assess the effects of AI systems that may affect work organization, the pace of work, the manner in which work is performed or decision-making processes relevant to occupational safety.
Employers must also ensure that workers receive adequate information and training on the specific risks associated with the use of such systems and on the prevention and protection measures put in place.
Protection of training data as trade secrets
Lastly, the Decree recognizes the strategic value of the resources used in the development of artificial intelligence systems. It clarifies that training data, algorithms and mathematical methods may qualify for protection as trade secrets under Article 98 of the Italian Industrial Property Code, provided that the applicable requirements are met.
Decree 160: public security, civil and criminal liability
The provision addresses two specific areas: on the one hand, it governs the use of AI systems in law enforcement activities; on the other, it regulates certain aspects of civil and criminal liability and of the administrative liability of legal entities.
Use of AI systems in law enforcement activities
The specific circumstances in which public security authorities may use artificial intelligence for biometric surveillance have been defined. It does so in a manner consistent with the AI Act and with the principles governing the protection of fundamental rights.
Real-time remote biometric identification systems may be used only where specific public-interest needs exist. These include preventing serious and concrete threats to public security, searching for missing persons, and investigating certain criminal offenses.
The use of such systems requires authorization from the competent judicial authority. It is also subject to strict limits as to duration (no more than 15 days, unless extended by a reasoned decision), territorial scope and the persons concerned.
Similar safeguards apply to post-remote (ex post) facial recognition technologies. These may be used for the targeted search of a person suspected of having committed an offense, subject to authorization by the judge for preliminary investigations upon application by the public prosecutor. No authorization is required where the recognition technologies are used solely after the offense has been committed, for the sole purpose of initial identification. Specific procedural safeguards also apply in this case. They include prior impact assessments (Data Protection Impact Assessments—DPIAs), consultation with the Italian Data Protection Authority, data retention limits, and a prohibition on generalized use, meaning any use unconnected to specific criminal proceedings or concrete investigative needs.
Criminal liability and corporate criminal liability
Decree 160 introduces a new criminal offense covering the failure to adopt security measures in artificial intelligence systems and the unlawful tampering with such systems.
The new Article 437-bis of the Italian Criminal Code, entitled “Failure to adopt security measures in artificial intelligence systems and unlawful tampering with such systems”, criminalizes three types of conduct:
(1) Failing to adopt technical security measures capable of preventing malfunctions or alterations of the system's operation in the design, training, production or placing on the market of high-risk AI systems, or failing to adopt the necessary human oversight measures. The offense is committed where the failure endangers life or public or individual safety. The penalty is increased where the security of the state is endangered.
(2) Willful tampering with high-risk artificial intelligence systems.
(3) Willful failure by a professional user (deployer) of high-risk artificial intelligence systems to adopt human oversight measures.
Whether a system is “high-risk” must be determined system by system, under Article 6 and Annexes I and III of the AI Act, taking into account the system’s intended purpose and the concrete manner in which it is used.
Summary—applicable penalties
- Failure to adopt security or human oversight measures: imprisonment of one to five years where the conduct endangers life or public or individual safety; two to eight years where it endangers the security of the state.
- Unlawful tampering with high-risk AI systems: imprisonment of two to six years where the conduct endangers life or public or individual safety; three to ten years where it endangers the security of the state.
- Professional user: anyone who intentionally fails to adopt human oversight measures for a high-risk AI system is punishable by imprisonment of one to five years where the failure endangers life or public or individual safety, and two toeight years where it endangers the security of the state.
As regards the quasi-criminal liability of legal entities under Legislative Decree No. 231/2001, the Decree introduces Article 25-vicies, entitled “Offences committed through the use of artificial intelligence systems”. It adds two predicate [change to “basic” or “preliminary”?] offenses: the new Article 437-bis of the Criminal Code, and the offenses of unlawful dissemination of content generated or altered by artificial intelligence systems under Article 612-quater of the Criminal Code, which was already introduced by the Law.
The entity may also be held liable under Article 8 of Legislative Decree No. 231/2001. That provision establishes that the entity's liability is autonomous and applies even where the perpetrator has not been identified or cannot be held criminally liable. This is particularly significant along the chain of parties that develop, supply, configure, deploy and oversee AI, because the way activities are spread across those parties may make it harder to identify the contribution of any individual natural person.
The new AI offenses may therefore trigger the entity’s quasi-criminal liability in addition to the liability of the natural person who committed the offense.
Article 612-quater of the Criminal Code is an intentional offense requiring a specific harmful result, punishable by imprisonment of one to five years. It requires unjust harm to a person, and the transfer, publication or other dissemination, without consent, of images, videos or voices that have been falsified or altered by means of AI and are capable of misleading as to their authenticity. It is not limited to high-risk systems. It may therefore affect a wider range of businesses, including those operating in marketing, communications and digital channel management. Transfer to even a single third party is sufficient, provided the other elements of the offense are met.
Summary—applicable penalties
- Article 437-bis of the Criminal Code: pecuniary sanction from a minimum of 600 to a maximum of 1,000 quotas.
- Article 612-quater of the Criminal Code: pecuniary sanction from a minimum of 200 to a maximum of 700 quotas.
- Disqualification sanctions: including suspension or revocation of authorizations and licenses; a ban on contracting with the public administration; exclusion from benefits, financing or grants; and a ban on advertising goods and services.
These changes call for a specific risk assessment to determine the entity’s actual exposure to the new offenses and whether its control safeguards are adequate.
Civil liability: damage caused by artificial intelligence
Decree 160 addresses civil litigation, with the aim of making it easier for people who suffer damage attributable to an artificial intelligence system to obtain compensation. The changes focus on two aspects: access to evidence and allocation of the burden of proof.
The Decree also introduces a specific evidence disclosure mechanism, designed to make it easier to establish the facts in proceedings involving AI systems. Upon application by the injured party, the court may order the defendant or a third party to disclose documentation relevant to understanding how the AI system operates. This includes system logs, technical documentation, human oversight and risk management measures, and documentation prepared for the purpose of AI Act compliance.
A disclosure order may be issued only where the claim has sufficient prima facie merit (fumus boni iuris) and must comply with the principles of necessity and proportionality and with the protection of trade secrets. An unjustified refusal to comply with a disclosure order may have significant evidentiary consequences. In particular, where the documents not produced concern obligations under the AI Act, the court may draw adverse inferences from the conduct of the non-compliant party. In the cases provided for by the Decree, it may also deem the facts alleged by the claimant to be proven.
Decree 160 further introduces a rebuttable presumption of causation in damages actions based on breach of obligations under the AI Act. Where the damage is linked to a breach of one or more AI Act obligations, the breach is presumed to have caused the damage. The defendant may nonetheless rebut this presumption with evidence to the contrary.
Finally, an injured party who is a consumer may bring proceedings before the court of their place of residence or domicile. Where the relevant conditions are met, they may also bring a direct action against the insurer with which the allegedly liable party has taken out a policy covering damage arising from the use of AI.
Concluding remarks
With the publication of the two legislative decrees in the Official Gazette, the legislator completes the national framework for implementing the AI Act. The decrees set out the institutional structure responsible for ensuring the AI Act’s application in Italy. They also introduce provisions that will have a concrete impact on AI governance, on supervisory mechanisms and on liability arising from the use of AI. These measures cut across several areas of law, including criminal, civil, employment and intellectual property law.
For businesses, this calls for a comprehensive compliance effort. The starting point is mapping the AI systems currently in use, classifying them according to the AI Act’s risk classifications and identifying the role the business actually plays (provider, deployer, importer or distributor). From there, companies will need an internal governance framework with clearly assigned responsibilities, policies on the adoption and use of AI systems, and safeguards ensuring effective human oversight.
On that basis, they will need to meet the technical and documentation requirements of the AI Act where applicable. In doing so, they should take into account their exposure to the national penalty regime and keep logs and technical documentation in good order, as these may be acquired or produced as evidence in court proceedings.
On the employment side, businesses will need a process for providing, on request, a clear explanation of AI-assisted decisions, and must ensure that the final decision rests with a natural person who has genuine decision-making authority. They will also need to update the information notice required under Article 1-bis of Legislative Decree No. 152/1997 and the risk assessment document (DVR) and provide employees with targeted AI literacy training.
Finally, the 231 Model must be updated to reflect the new predicate offenses, with specific protocols and dedicated reporting lines to the Supervisory Body (OdV). The entire framework should be reviewed regularly and kept under constant review as legislation and industry guidelines evolve.
