Italy's AI Implementing Decrees: Training at the Centre, but the Italy–EU Dual Track Remains Open

On 10 June 2026, the Council of Ministers approved, by preliminary review, two draft legislative decrees implementing Law no. 132/2025, the Italian law on artificial intelligence. The first concerns the powers of the national Authorities and the use of AI in training and education; the second concerns the deployment of AI systems in policing activities and the profiles of civil and criminal liability. It must be stressed from the outset that these are non-final texts: they will be subject to the scrutiny of the parliamentary Committees, the Conference of the Regions and the competent Authorities, and the final content may undergo even significant changes.

The Government presents the package as the completion of the first comprehensive national regulatory framework on AI in Europe, declaredly consistent with the AI Act (EU Regulation 2024/1689). The press release insists on one point: the decrees do not introduce a regime alternative to the European framework but ensure its implementation within the national legal order, the result of constant dialogue with the European Commission — informal, through meetings and exchanges, and formal, within the notification procedure for certain parts of the AI law — and respectful of the completeness of the EU regime on the fundamental requirements of systems, "without overlaps". This is an important premise because — as we shall see — it is precisely the relationship between national rules and the European regulation that determines whether the entire framework holds together. The declared defining feature is its human-centric approach: technology cannot become the measure of the human nor substitute itself for the responsibility and discernment of the person.

Training as an "Enabling Condition": What Really Changes

The most distinctive element of the first decree is the central role attributed to training, defined not as mere technical instruction but as critical literacy: awareness of risks, the ability to interpret outputs, responsibility in the use of the tools. It is, in essence, the national articulation of the AI literacy obligation set out in Article 4 of the AI Act, applicable from February 2025, which the decree translates into concrete sectoral measures. The approach is progressive: competences spread "from school education onwards, and therefore on the basis of the specific sectors of belonging".

Schools

Within the school system, AI enters educational pathways on a stable basis, both as content to be known and as a tool to innovate teaching — not to chase technology, the press release specifies, but to strengthen the educational mission of the school. The main measures provide for the updating of the national guidelines for the second cycle in order to integrate advanced technologies and generative AI into study pathways, the inclusion of AI in civic education with attention to ethical profiles and digital citizenship, the strengthening of STEAM competences and educational guidance, and stable teacher training on how the systems work, the risks of error and bias, data protection and responsible use.

Three elements deserve particular attention from an operational standpoint:

• The territorial technical-ethical committees, organised as a network in support of schools, with functions of pedagogical guidance, support for teaching experimentation, protection of fundamental rights and data protection. They will also contribute to the updating of school regulations, to make the use of AI "safe and verifiable". This is an unprecedented governance structure for the school world, the effectiveness of which will depend entirely on the quality of its composition and on coordination with already existing roles (the institutions' DPOs, digital animators, USR). The risk of overlapping competences is real.
• The 100-million-euro plan for teacher training, expressly linked to the educational emergency connected with the abuse of social media, digital platforms and AI: the aim is to strengthen the school system's capacity to prevent risks, digital dependencies, algorithmic opacity and forms of conditioning of minors, with the involvement of families for the integral well-being of the person within the digital space. It is the first time that the Italian legislator has expressly linked AI training, the protection of minors and digital well-being with a dedicated financial endowment, qualifying the school as a "prevention safeguard".
• The extension to adults: structured pathways of literacy and training on AI, with recognition of competences already acquired and support for upskilling, professional reskilling and reintegration into the labour market. Training on AI is therefore treated also as an instrument of active labour market policy, not merely as a school subject.

Universities, AFAM, ITS Academy and Research

The approach is one of cross-cutting application: by reason of its horizontal impact, AI enters all areas of tertiary education, not only specialist pathways. Universities and AFAM institutions integrate training activities on the safe and informed use of the systems, including through laboratory-based and interdisciplinary methods; the minimum content covers how the systems work, the interpretation of outputs (predictions, content, recommendations or decisions), legal profiles, cybersecurity risks and impact on rights.

Of particular interest is the logic of the cross-fertilisation of competences expressly provided for: integration of ethical and legal profiles into science-oriented courses, and of technical profiles into courses with a predominantly economic or legal focus. It is a choice that recognises a fact of reality: the jurist who does not understand how an AI system works and the engineer who is unaware of its legal constraints are both, today, incomplete professionals.

The picture is completed by: ANVUR monitoring of the quality of the training offer, also for the purposes of incentive policies; the recognition of dissemination and literacy activities by lecturers and researchers for the purposes of evaluation and career progression; and the recognition of the ITS Academy as a strategic segment of higher tertiary education, in order to train figures capable of operating in technology-intensive contexts.

Public Administration

Within public administration the decree goes beyond training in the strict sense: administrations will have to introduce AI systems into their recruitment, training and organisational innovation policies, with the aim of simplifying administrative action and accelerating procedures. The Minister for Public Administration assumes a function of guidance and coordination — identification of common needs, definition of training priorities, homogeneity between central and territorial administrations — precisely in order to prevent a fragmented or unequal adoption of AI in the public sector.

Training is structured on three levels: basic literacy for all public employees; professional reskilling for specialist pathways (administrative procedures, citizen services, document management, recruitment, data analysis); advanced training for managers, digital transition officers and personnel offices. Operational coordination is entrusted to the National School of Administration (SNA), which will translate the ministerial guidance into common modules, differentiated by function and level of responsibility. The underlying principle is that of effective human oversight: AI may assist administrative action, but responsibility for the decision must remain clear, verifiable and attributable to competent persons.

Healthcare Professionals

In healthcare, AI training is included as a mandatory requirement, with a specific percentage, in the ECM programme, with content covering the operational use of the tools but also the deontological, ethical and legal profiles, so that the professional retains full clinical responsibility. Training also enters the managerial pathways of healthcare managers, with a view to organisational efficiency (managing waiting lists, rationalising waste). The reference to the "MIA" platform, funded by the PNRR and under trial at Agenas, signals a preference for reliable and certified institutional tools for the national health service (SSN).

Regulated Professions

For the professions, AI literacy enters initial and continuing training on three planes: technical (how the systems work, their potential and limits), legal (the European regulation and national rules) and deontological — the latter indicated as the most significant profile, since it concerns the professional's responsibility in the use of AI, the disclosure obligations towards the client and compliance with the human-centric principle of Law 132/2025. The professional bodies will have to adapt their regulations within six months, with the involvement of the supervisory authority where provided for.

Of particular relevance for professional firms is the connection with fair compensation (equo compenso): the decrees setting the parameters — including legal fee scales — will be supplemented within twelve months, with parameters commensurate with the risk classification of the system used, the actual professional contribution and the level of responsibility connected with the use of AI. The declared intent is to prevent automation from devaluing intellectual work; the translation of this principle into objective and transparent fee parameters, however, promises to be technically complex, and will be one of the testing grounds of the implementation.

The Judiciary and Police Operators

In the justice sector — a high-risk activity under the AI Act — training is qualified as a condition of reliability: it must ensure that the decision belongs to the magistrate and not to the machine. The pathways, entrusted to the Higher School of the Judiciary on the basis of programmatic guidelines from the Ministry of Justice and the CSM, are structured on three planes (technical, including questioning techniques and cybersecurity; legal; organisational and value-based), are differentiated by function and by the risk level of the systems, and are aimed in particular at human oversight over high-risk systems. The cardinal principle remains that AI does not replace the act of adjudication (ius dicere): the magistrate's discretion in interpreting and applying the law remains intact.

The second decree too, dedicated to policing activity, includes among its key points the specific training of operators: decisions remain human and overseen by personnel trained in the correct use of the tools, including in the two regulated biometric deployments (real-time identification with judicial authorisation and ex-post facial recognition). Training, in short, is the connecting thread of the entire package.

Governance and Sanctions: AgID, ACN and the Sectoral Authorities

The first decree also defines the configuration of the national authorities: AgID as the notifying authority and ACN as the market surveillance authority and single point of contact with the EU institutions, flanked by the sectoral authorities by reason of the areas of risk. Banca d'Italia, CONSOB and IVASS supervise the high-risk systems of financial intermediaries directly connected with the provision of financial services; the Data Protection Authority intervenes, for the profiles within its competence, on high-risk systems in law enforcement, borders, justice and democracy activities. Provision is made for instruments of informational and operational cooperation among the authorities and for an annual report to the Presidency of the Council of Ministers, through the Coordination Committee at the Department for Digital Transformation, also with a view to assessing any regulatory revisions.

On the sanctioning front, Italy avails itself of the option provided by the AI Act to introduce maximum limits lower than the European ones, calibrating the sanctions to the degree of responsibility of the parties along the supply chain. It is a choice favourable to operators, but also a first example of how national implementation may produce sanctioning geometries that differ from State to State.

The second decree is completed by two liability safeguards that directly concern those who develop or use high-risk systems: on the civil plane, procedural instruments in favour of the injured party (access to technical documentation, presumption of the causal link, alternative venue at the residence of the injured natural person, direct action against the insurer), declaredly designed to fill the gap left by the withdrawal of the European proposal on AI liability without imposing new substantive obligations on businesses; on the criminal plane, the new Article 437-bis of the Criminal Code, which sanctions the failure to adopt or the alteration of safety measures in high-risk systems where this gives rise to a concrete danger to life, public safety or the security of the State, with extension of liability to the entity under Legislative Decree 231/2001 and punishability anchored, for the negligent form, to gross negligence.

The Risks: Regulating a Fast-Moving Technology, Within an Already Regulated Framework

Alongside the positive elements — and the choice to focus on training as an enabling infrastructure is one of them — the package raises questions of method that those working in compliance cannot ignore.

1. The risk of regulatory layering. The AI Act is a regulation, directly applicable and designed to ensure harmonisation: every national intervention, however declaredly implementing, adds a further layer of interpretation. The Government states that the decrees respect the completeness of the EU regime on the fundamental requirements of systems, without overlaps, and that they represent the "fulfilment" of the regulation as regards the regulatory techniques falling within State competence. It is a claim to be verified against the final texts, article by article: the experience of the GDPR teaches that even legitimate spaces for national intervention can generate, multiplied by 27 Member States, precisely the fragmentation that the regulation sought to avoid. The very choice — albeit legitimate — of national sanctions with maxima lower than the European ones confirms that State implementation inevitably produces differentiation. For businesses operating across several EU markets, every national specificity is an additional compliance cost.

2. The risk of rapid obsolescence. Detailed rules are being set — down to the ECM percentages, the six months for the professional bodies' regulations and the twelve months for the fair-compensation parameters — for a technology that evolves over cycles of months, not years. The national school guidelines updated today on generative AI risk describing, three years from now, systems that no longer exist in the form considered. The mechanism of the authorities' annual report to the Presidency of the Council of Ministers, also aimed at assessing regulatory revision measures, is a welcome corrective, but it remains to be seen whether the timescales of legislative revision can ever be aligned with those of technological evolution.

3. The European context is itself in motion. The framework onto which the decrees are grafted is not static: in Brussels the Digital Omnibus package is under discussion which — in the proposals currently on the table, and with timings and content still to be defined — could affect deadlines and obligations of the AI Act itself, in particular for high-risk systems. Building today a detailed national implementation on a regulation whose application timelines could be revised exposes one to the concrete risk of having to reopen the decrees even before their full operability. The press release itself, moreover, indirectly acknowledges the instability of the EU framework when it justifies the intervention on civil liability with the withdrawal of the European proposal on the matter (AILD): Italy fills a gap, but if the Commission were to return to the subject with a new harmonised instrument, the national rules would have to be recalibrated once again.

4. Implementation is the real unknown. Territorial technical-ethical committees, SNA modules, ECM percentages, adaptation of the professional bodies' regulations within six months, integration of the fee parameters within twelve: the decree designs an imposing training architecture, the realisation of which depends on ministerial decrees, guidelines, administrative capacity and the resources actually available. The 100-million endowment for schools is significant, but the recent history of digital training in the Italian public administration suggests caution as to the speed of execution. And a training architecture that remains on paper does not even satisfy the AI literacy obligation of Article 4 of the AI Act, which is already applicable.

The Industrial Dimension

The press release closes by linking regulation and industrial policy: up to 1 billion euro from the Venture Capital Support Fund for the national AI ecosystem (Article 23, Law 132/2025), more than 300 million already allocated by CDP Venture Capital in support of more than 150 start-ups and around 20 third-party SGR funds, more than 500 million of new investments planned over the three-year period and, from 2026, the national technology transfer hub "SophIA" dedicated to AI and cybersecurity, with an endowment of around 30 million. The Italian AI market — 1.8 billion in 2025, +50% on the previous year — is presented as the base on which to build priority supply chains (humanoid robotics, autonomous driving, quantum, photonics for HPC, vertical AI). It is the declared attempt to make AI not only an object of regulation but an axis of growth: the consistency between regulatory ambition and industrial ambition will be measured, here too, in execution.

What Organisations Should Do Now

Although these are drafts under preliminary review, the direction is clear and it is advisable to move ahead:

• Schools and universities: map the initiatives on AI already under way (school regulations, use policies, teacher training) and prepare to align them with the future territorial technical-ethical committees and with the minimum training content that will be defined;
• Public administrations: launch the assessment of training needs on the three provided levels (basic, specialist, managerial) and take stock of the AI systems already in use, also in light of the human oversight obligations and the future guidance of the Minister for Public Administration;
• Healthcare facilities and regulated professionals: monitor the integration of the ECM programmes and the adaptation of the professional bodies' regulations, which will have to take place within tight timescales (six months), and follow the evolution of the fair compensation parameters linked to the use of AI;
• Financial intermediaries: take note that supervision of high-risk AI systems connected with financial services will fall to Banca d'Italia, CONSOB and IVASS, and integrate accordingly the mapping of the systems into existing compliance frameworks;
• Businesses that develop or use AI systems: oversee the dual track of the AI Act and national legislation, remembering that the literacy obligation of Article 4 of the AI Act is already applicable and that the Italian decrees will specify its sectoral arrangements; and treat the safety measures of high-risk systems as an effective organisational safeguard — not as a formal compliance step — in light of the new Article 437-bis of the Criminal Code and the extension of liability under Legislative Decree 231/2001, which make it advisable to update organisational models.

Training is the right choice: no regulatory framework on AI works without people capable of understanding and applying it. But the quality of this implementation will be measured by the capacity to remain light, updatable and genuinely complementary to European law — not by the quantity of compliance obligations it manages to generate.

Sources

• Press release of the Council of Ministers no. 177, 10 June 2026 — governo.it
• Law no. 132 of 23 September 2025 — Provisions and delegations on artificial intelligence
• Regulation (EU) 2024/1689 (AI Act), in particular Article 4 (AI literacy)
• MIM — 100-million-euro teacher training plan on AI — mim.gov.it