Docs

Regulations

Italian AI Law

Italian AI Law — A practical guide for developers and businesses 🇮🇹

Italy’s AI law is not a parallel regulation. It is the glue between the EU AI Act and the national enforcement machine: it designates the authorities who will inspect and sanction, adds criminal liability that the AI Act deliberately left to member states, and sets the principles for sector adoption. No new obligations on the systems themselves, but serious consequences for those who ignore them.

What it is and why it matters 🇮🇹

Law 23 September 2025, No. 132 — “Provisions and delegations to the Government on artificial intelligence” — is Italy’s first organic AI law. Published in the Official Gazette No. 223 of 25 September 2025, it entered into force on 10 October 2025 (15 days after publication).

It is a national law that complements and implements Regulation (EU) 2024/1689 (the AI Act) without duplicating it. Article 3.5 states plainly:

This law creates no new obligations compared to those provided by Regulation (EU) 2024/1689 for artificial intelligence systems and general-purpose AI models.

In other words: if you already comply with the AI Act, the Italian law adds no extra technical requirements on the systems themselves. But it does impose other kinds of obligations — governance, designated authorities, criminal penalties — that no “AI Act-compliant” system can safely ignore.

It completes the regulatory picture: GDPR for data, NIS 2 for organisations, CRA for products, the AI Act for high-risk systems, and now this law for national enforcement on Italian soil.

The golden principle: AI Act conformity and “human oversight” 🎯

Article 1.2 establishes that the law’s provisions shall be interpreted and applied in conformity with Regulation (EU) 2024/1689. Article 3 sets the general principles:

respect fundamental rights and freedoms of the Constitution, EU law, and the principles of transparency, proportionality, security, data protection, accuracy, non-discrimination, gender equality and sustainability;
develop AI models on data and processes ensuring correctness, reliability, security, quality, appropriateness and transparency, according to the principle of proportionality;
ensure human autonomy and decision-making power, harm prevention, knowability, explainability, human monitoring and intervention;
guarantee cybersecurity throughout the entire lifecycle of AI systems and general-purpose models, based on a risk-based approach, with specific security and resilience controls against attempted alterations;
guarantee persons with disabilities full access to AI systems, without discrimination.

For developers: “human oversight” is not just a high-risk AI Act requirement. The Italian law elevates it to a universal principle, alongside cybersecurity throughout the lifecycle — the same language as the CRA, now applied to AI systems.

Who enforces: the national AI authorities 🏛️

Article 20 designates the National Authorities for Artificial Intelligence:

AgID (Agency for Digital Italy): promotes innovation and AI development; defines procedures, functions and tasks for notification, evaluation, accreditation and monitoring of conformity-assessment bodies; is designated the notification authority under Article 70 AI Act;
ACN (National Cybersecurity Agency): responsible for market surveillance — including inspection and sanctioning activities — of AI systems under national and European law; is designated the market-surveillance authority and single contact point with EU institutions under Article 70 AI Act;
Bank of Italy, CONSOB and IVASS remain market-surveillance authorities for financial sectors (Art. 74.6 AI Act);
AgID and ACN jointly manage AI regulatory sandboxes for compliant systems.

The Garante (Data Protection Authority) and AGCOM retain their competences (Art. 20.4): the Garante on personal data, AGCOM as Digital Services Coordinator under the DSA.

For businesses: you have two authorities, not one. AgID for innovation and notified bodies; ACN for enforcement and penalties. If your system is inspected, the agency knocking on your door is the same one that enforces national cybersecurity.

Government delegations: where the future is being drafted 🧾

Article 24 delegates to the Government (within 12 months of entry into force, i.e. by 10 October 2026) the power to adopt legislative decrees to:

transpose the AI Act into national law: grant AgID/ACN the supervisory, inspection and sanctioning powers of the regulation, modify sector-specific legislation, define the sanctioning framework (including enforcement of Article 99 AI Act penalties: up to €35M or 7%, €15M or 3%, €7.5M or 1.5%);
regulate cases of illicit creation and use of AI systems and introduce autonomous criminal offences for the circulation and professional use of AI systems where security measures have been omitted (posing a concrete danger to life, public safety or national security);
regulate the use of AI in preliminary investigations, respecting defence rights and personal data;
establish literacy and training paths on AI for professionals and sectors; promote STEM competences in school curricula;
valorise research and technology transfer on AI in universities, AFAM institutions, ITS Academies and public research bodies;
define the market-surveillance authority’s powers over providers (remote or on-site inspections, information requests, real-world testing controls);
regulate administrative liability of legal entities for offences related to AI systems.

Delegations are not paper tigers. By October 2026, decrees could arrive that truly fix Italy’s AI Act sanctioning framework and new criminal offences. Anyone working in AI must track legislative decrees the way they track release notes of a critical dependency.

Criminal and civil penalties: what can actually happen 🚨

The most concrete and novel part of the law is the amendments to the Criminal Code (Capo V):

Art. 61(11-undecies) C.p. — new general aggravating circumstance: committing an offence using AI systems when they constitute an insidious means, obstruct public or private defence, or aggravate the consequences of the crime;
Art. 294 C.p. (coercion) — aggravating circumstance: imprisonment from 2 to 6 years if the coercion is carried out using AI;
Art. 612-quater C.p. — new offence of illicit dissemination of content generated or altered by AI systems (deepfakes): anyone who, without consent, distributes images, videos or voices falsified or altered by AI and capable of inducing a mistake as to their genuineness, faces imprisonment from 1 to 5 years. Private prosecution; public prosecution where connected to a public-offence crime or committed against incapacitated persons or public authorities;
Art. 2637 C.c. (market manipulation) — aggravating circumstance: imprisonment from 2 to 7 years if committed using AI;
Art. 185 TUF (Consolidated Law on Finance) — aggravating circumstance for market manipulation using AI: imprisonment from 2 to 7 years and a fine from €25,000 to €6,000,000.

In civil law, Article 1 assigns to the specialised enterprise-law chambers jurisdiction over disputes concerning the operation of an AI system.

Deepfakes are the immediate use case: before, it was a reputation or GDPR issue; since October 2025 it is a crime carrying one to five years. If your app can generate altered video or voice of people, consent is no longer just a checkbox — it is protected by criminal law.

Copyright protection: the “human factor” ✍️

Article 25 amends the Copyright Law (Law 633/1941):

works of authorship are protected when attributable to human ingenuity, “including where created with the aid of AI tools, provided they constitute the result of the author’s intellectual work”;
new Art. 70-septies: reproductions and extraction of text and data (text-and-data mining) from online works or databases for training AI models, including generative ones, are permitted in accordance with Articles 70-ter and 70-quater (implementing the TDM exception of Directive 2019/790);
new criminal offence Art. 171 a-ter: reproducing or extracting text or data in violation of Art. 70-ter and 70-quater, including through AI, is a criminal offence.

For companies training models: mining content downloaded or extracted without respecting rightholders’ reservations is no longer (only) a civil matter. A criminal sanction now backs TDM exception violations.

Sector impact 🏥

The law’s Annotations cover the main application domains:

Healthcare (Arts. 7-10) 🩺

AI in healthcare is supportive, not substitutive: the decision always rests with the doctor;
no selection or conditioning of access to healthcare based on discriminatory criteria;
the patient has the right to be informed about the use of AI technologies;
scientific research and experimentation for healthcare AI (prevention, diagnosis, treatment, drug development, prosthetics) is declared of relevant public interest, with simplified personal-data processing (including secondary use of data stripped of direct identifiers, subject to Garante consultation);
creation of a national AI platform for healthcare (managed by AGENAS), with data strictly limited to what is necessary and AGENAS as data controller;
measures for the electronic health record and digital health.

Employment (Arts. 11-12) 💼

AI is used to improve working conditions and productivity, without violating human dignity or personal-data privacy;
the employer or client must inform the worker of AI use (Art. 1-bis of Legislative Decree 152/1997);
AI must not discriminate on grounds of sex, age, ethnicity, religion, sexual orientation, political opinions or personal conditions;
an Observatory on AI adoption in the world of work is established at the Ministry of Labour.

Intellectual professions (Art. 13) 💼

AI use in professions is limited to instrumental and supporting activities, with intellectual work prevailing;
information on the AI systems used must be communicated to the client in clear, simple and exhaustive language (fiduciary relationship).

Public administration (Art. 14) 🏛️

PAs use AI for efficiency and quality of services, with transparency of operation and traceability of use;
AI is instrumental and supportive to administrative activity: the official remains the sole person responsible for decisions;
obligation of technical, organisational and training measures for responsible use.

Justice (Art. 15) ⚖️

every decision on interpretation and application of law, evaluation of facts and evidence, and adoption of measures remains reserved to the judge;
the Ministry of Justice regulates AI uses for organisation and simplification of judicial work; experimentation authorised until full implementation of the AI Act, following consultation with the national authorities;
compulsory digital training for judges and administrative staff.

National cybersecurity and defence (Arts. 6 and 18) 🛡️

Activities for national security, defence and policing (per Law 124/2007, D.L. 82/2021) are excluded from the law’s scope (constitutional principles and Art. 3.4 remain);
the ACN promotes every initiative, including public-private partnerships, to valorise AI as a resource for national cybersecurity;
investments authorised (up to €1 billion) in risk capital for innovative SMEs in AI, cybersecurity and quantum computing, via the Venture Capital Support Fund.

Impact for businesses 🏢

The checklist for anyone developing or using AI in Italy:

Map your AI system against the AI Act risk classification (if you have not done so already);
Know the competent authority: AgID for innovation and notifications, ACN for enforcement and penalties;
Verify human-overlook and cybersecurity processes across the lifecycle as a precondition of the law’s principles;
Audit deepfake / AI-generated content activities: consent, labelling, disclosure; the criminal risk is new and real;
Respect copyright for text-and-data mining: verify rightholders’ TDM reservations and the exceptions in Art. 70-ter/70-quater;
Track the government decrees (deadline October 2026): AI Act sanctioning framework, criminal offences, enforcement powers;
Financial sector: beware of aggravated penalties for market manipulation using AI (2-7 years);
If you operate with the PA or healthcare: transparency, traceability and certification prerequisites are now mandatory.

Golden rule for businesses: the law adds no AI Act compliance, but raises the stakes. European conformity protects you from administrative penalties; the new Italian criminal offences can land you in front of a criminal court if, with intent or negligence, you fail to adopt AI security measures.

Timeline and deadlines 📆

Date	What happens
25 September 2025	Published in Official Gazette No. 223
10 October 2025	Entry into force of the law
By 10 January 2026	Ministry of Health decree on data processing for healthcare research (120 days)
By early 2026	Ministry of Labour decree establishing the AI Observatory (90 days)
By 10 October 2026	Legislative decrees adapting national law to the AI Act (12 months)
By 10 October 2026	Legislative decrees regulating illicit uses of AI systems (12 months)
Biennially	Update of the national AI strategy
2025-2026	€300,000/year authorisation for MAECI experimental AI projects

Compliance checklist ✅

A concise checklist for developers and businesses.

System and model

AI system classified under the AI Act risk scale (unacceptable / high / limited / minimal)
Human oversight and human monitoring, even for non-high-risk systems
Lifecycle cybersecurity with security and resilience controls (resilience against attempted alterations)
Transparency, explainability, accuracy and non-discrimination documented

Governance and authorities

Awareness of national authorities (AgID, ACN) and the single contact point
If you are a provider: notification/accreditation procedures managed (AgID)
If you are in healthcare or PA: transparency, traceability, supportive-not-substitutive principle

Criminal risk

Audit of systems that generate/alter content (deepfakes): consent, labelling, prosecution rules
Verify that no process uses AI as an “insidious means” in illicit contexts
In finance: anti-market-manipulation controls using AI
Track government decrees (by October 2026)

Content and copyright

Verify TDM reservations of content used for training
Compliance with Art. 70-ter and 70-quater (text-and-data mining exception)
Policy on use of generative outputs in products (works of “human ingenuity”)

Employment and professional relationships

Disclosure to workers on AI use
No algorithmic discrimination in HR processes
Clear communication to professionals/clients about AI use