Skip to content
Atesto
Create account
All articles

Data protection

Italy sets GDPR erasure conditions for trained AI models

The Italian Data Protection Authority ruled that deleting source data is not enough if an AI model has already learned from it, setting new conditions.

On October 29, 2026, the Italian Data Protection Authority (opens in a new tab) (Garante Privacy) published Decision No. 760, establishing strict legal conditions for the deployment of artificial intelligence in administrative processes. The authority determined that deleting or correcting personal data in a source database does not automatically rectify or erase its residual impact on an already trained artificial intelligence model.

The decision responds to a draft regulation submitted by the Autonomous Province of Trento. The province sought to introduce artificial intelligence systems to support its daily administrative activities and organizational processes. Among the first planned applications was a system for document management and protocol registration. In a public administration, a protocol system collects daily requests from citizens, corporate communications, and documents containing names, addresses, economic situations, health information, and judicial records.

While the Garante Privacy issued a favorable opinion on the regulation, it conditioned its approval on specific modifications to elevate the protection of personal data. The regulator reviewed the data protection impact assessment attached to the draft, which highlighted a fundamental technical limit of machine learning (opens in a new tab). When an organization corrects or eliminates information in the source archive after the training phase, the modification affects the original document and future training cycles. However, the parameters of the already built model may continue to reflect the previous data.

During training, a model absorbs relationships and recurrences present in the data, translating them into numerical values distributed within its architecture. The Garante Privacy recognized this reality, noting that removing a file from the archive can leave the trace that the file impressed upon the model intact.

Who it applies to

The immediate subject of Decision No. 760 is the Autonomous Province of Trento and its administrative structures. The regulation governs the use of artificial intelligence by the province to increase the efficiency of administrative action and improve the quality of services offered to citizens and businesses.

However, the ruling establishes how the Italian regulator interprets the General Data Protection Regulation regarding machine learning models. The technical and organizational expectations outlined by the Garante Privacy signal how the authority views the obligations of any data controller that trains or deploys artificial intelligence systems using personal data.

The decision serves as a baseline for public administrations and private companies operating under the jurisdiction of the Garante Privacy. Executives, data protection officers, and chief information security officers must account for these conditions when bridging the gap between legal privacy requirements and engineering capabilities. IT service providers managing data infrastructure will also need to align their systems with these expectations.

What it requires

The Garante Privacy mandated that the province must implement technical and organizational measures to ensure that data subject rights remain effective against trained models. This specifically concerns the right to rectification under Article 16 and the right to erasure under Article 17 of the General Data Protection Regulation.

A data controller cannot rely solely on deleting or correcting the source file in the original database. The authority requires that the rights of rectification and erasure be guaranteed in a concrete manner, independently of the specific artificial intelligence system or model utilized by the organization. The burden is on the controller to adopt measures that address the residual memory of the model itself.

The regulator also set strict rules for processing sensitive data and information about criminal convictions. The Garante Privacy determined that an organization cannot govern the processing of special categories of data exclusively through internal technical specifications or operational IT sheets prepared by individual administrative structures.

The essential elements of these processing activities must be anchored in adequate normative sources. For the public administration in question, the authority required a reference to an act of at least regulatory rank or, for special categories of data, a general administrative act of the province. This ensures the essential aspects of the processing find their discipline in appropriate sources, guaranteeing citizens the concrete exercise of their rights (opens in a new tab).

The decision also sits within the broader framework of Italian artificial intelligence legislation. The opinion references Law No. 132 of September 23, 2025, which regulates the use of artificial intelligence in the public administration to increase efficiency while keeping final decisions attributed to responsible officials. The implementation of this law was completed with Legislative Decree No. 179 of October 7, 2026, which adapts the Italian legal system to the European Union Artificial Intelligence Act.

The territory of Trento had previously been the subject of an intervention by the authority. In January 2024, the Municipality of Trento, a distinct entity from the Province, was sanctioned with a 50,000 euro fine for research projects that analyzed data collected by cameras and microphones using artificial intelligence.

The dates

The Garante Privacy adopted Decision No. 760 on October 7, 2026. The authority formally published the favorable opinion and its compliance conditions in its newsletter on October 29, 2026.

The broader national rules referenced in the decision are already active. Italian Law No. 132 of 2025 entered into force on October 10, 2025. Legislative Decree No. 179 of 2026 entered into force on October 23, 2026.

What to check now

  • Map the flow of personal data from primary databases into any artificial intelligence models the company trains or fine tunes.
  • Review current data deletion and rectification workflows to identify where the technical process stops.
  • Verify with the engineering team how a specific individual's data influence is removed from a trained model after a deletion request.
  • Inventory all artificial intelligence systems that process sensitive data, health information, or criminal records.
  • Ensure that the use of sensitive data in artificial intelligence is governed by formal, high level documentation and explicit legal bases, rather than just technical deployment guides.
  • Update data protection impact assessments to explicitly address how the rights of rectification and erasure are handled within the model architecture.
  • Confirm that the final decisions produced by artificial intelligence systems in administrative processes remain subject to human supervision and responsibility.

The Italian decision highlights a technical reality of machine learning that directly intersects with privacy law. Data protection authorities expect the engineering capabilities of an organization to align with the legal reality of data subject rights, ensuring that a deleted record does not live on in the weights of an algorithm.

Sources

  1. Federprivacy (opens in a new tab) · www.federprivacy.org
  2. Byte.it (opens in a new tab) · www.byte.it
  3. Sardegnagol (opens in a new tab) · www.sardegnagol.eu

This article is general information, not legal advice. For your company's situation, talk to a qualified lawyer.

This article was written with AI and checked, sentence by sentence, against the sources linked above.