Privacy × AI A weekly EU privacy and AI brief, with primary sources and practical context
Home Archive Add weekly reminder

Trained AI models and data subject rights

Issue 10 · 4 November 2026 · Covers 28 October to 3 November 2026 · About 9 min read
The weekly brief in video
1

Executive Summary

2

In Focus

AI in public administration Italy Opinion of the supervisory authority 7 October 2026, published 29 October 2026

The Garante makes its opinion on Trento's AI rules conditional on rights that can be exercised against models already trained

The Autonomous Province of Trento is preparing a regulation on how its administration develops, trains and uses AI. The rules implement Article 40-ter.1 of Provincial Law no. 23 of 1992, a provision added in August 2026 that promotes AI as a tool to support administrative activity. The province consulted the Garante on the draft. In opinion no. 760 of 7 October 2026, the Garante gave a favourable opinion, subject to two conditions.

1. What the draft regulation provides

The draft has 16 articles and covers the whole life of an AI system, from development and training to use, monitoring and updating. According to the Garante's account of the draft:

  1. AI systems may not be used to adopt administrative decisions based solely on automated processing.
  2. Each system or area of use is described in a specific technical sheet, which cannot introduce new purposes.
  3. The data used come from the archives of the province and its public bodies, provincial information systems, national databases and data collected directly from data subjects. They must be stored on systems in the EU, and transfers outside the EU are restricted.
  4. To train systems on criminal offence data, the province must use synthetic data or, failing that, pseudonymised or masked real data.

The first application is the province's document management and registry system, P.I.Tre, for which the draft includes a technical sheet and a data protection impact assessment.

2. The first condition: the legal basis for sensitive data

The draft leaves the details of processing special category and criminal offence data to the technical sheets. The Garante noted that each sheet is drawn up and approved unilaterally by the administrative unit concerned, and found this arrangement not fully consistent with Articles 2-sexies and 2-octies of the Italian Data Protection Code, which govern the processing of special category data and of criminal offence data respectively. In the Garante's view, a technical sheet cannot take the place of the act that sets the essential elements of the processing and its safeguards. It therefore asked for the provisions concerned to be amended so that they refer to an act of at least regulatory rank or, for special category data only, to a general administrative act of the province, rather than to a sheet approved by a single unit.

3. The second condition: rights against a trained model

The province's own impact assessment for P.I.Tre identified a risk that rectification or erasure "may not be reflected in the algorithmic parameters of the model already structured", and that only the source document would be corrected, so that the correction would take effect only in future training cycles. The Garante noted that the assessment made this point about a specific case, but described the risk as "potentially systemic and generalised". It asked the province to add measures so that rectification and erasure under Articles 16 and 17 GDPR, and all the other rights in Articles 15 to 22, can be exercised with certainty and effectively whatever AI system or model is used, in line with the EDPB's Opinion 28/2024 on AI models.

4. The national framework

The Garante found the draft broadly consistent with the national framework, including Article 14 of Italian Law no. 132 of 2025, which allows public administrations to use AI only as an instrument and support for administrative decisions, respecting human autonomy and decision-making power. The opinion does not refer to the AI Act.

Why it matters

A supervisory authority has made the effective exercise of rights against an already trained model a condition of a public body's AI programme, in a formal opinion rather than in general guidance. The province's own impact assessment had identified, for a specific case, that correcting the source data might not change the model, and the Garante asked for measures to overcome that limit. The opinion is advisory and concerns one draft regulation, but it indicates a question the Garante may ask of other public bodies that train models on personal data: how a request to correct or erase reaches the model, not only the database.

Practical impact

For a controller that trains or fine-tunes models on personal data: set out in the impact assessment how a rectification or erasure request is carried into the model itself, and when. The opinion does not prescribe a technique. Possible approaches include retraining at defined intervals and filtering of inputs and outputs. Do not treat correction of the source data alone as the answer. For a public body in Italy: make sure the processing of special category and criminal offence data rests on an act of at least regulatory rank or, for special category data only, a general administrative act, not on internal technical documents.

Copies a clean, cited sentence for use in a note.
3

Key Developments

Platforms Germany / EU Reference for a preliminary ruling 27 October 2026

Germany's Federal Court of Justice asks the EU Court of Justice whether a social network is a controller for users' posts and whether the hosting exemptions bar damages under Article 82 GDPR

On 27 October 2026 the Bundesgerichtshof (BGH), Germany's Federal Court of Justice, referred six questions to the Court of Justice of the EU. In the case, VI ZR 64/24, a member of the Bundestag is suing the operator of Facebook over a meme that attributed to her a quotation she never made, and that spread in many copies and variants. She seeks an order covering all identical or similar memes without having to supply their web addresses, and damages of at least 10 000 euro. The operator argues that, as a neutral host, it need only remove posts that are specifically notified.

The Higher Regional Court in Frankfurt granted the injunction and dismissed the damages claim, and both sides appealed. The BGH waited for the Court of Justice's judgment of 2 December 2025 in C-492/23 Russmedia before referring.

The six questions cover four issues:

  1. Whether an operator that only stores user content and makes it accessible, without reviewing it in advance or controlling it editorially, is fully a controller under Article 4(7) GDPR.
  2. Whether such an operator benefits from the liability exemptions in Articles 12 to 15 of the E-Commerce Directive, and when it counts as a hosting service under the Digital Services Act (Article 3(g)(iii) and (i), and Article 6). Both questions ask whether it matters that the operator selects, orders, presents or recommends content by algorithm beyond basic categorisation and indexing.
  3. Whether Article 15 of the E-Commerce Directive or Article 8 of the Digital Services Act precludes an obligation to carry out an independent, manual and substantive check of content that may be equivalent to content already found unlawful.
  4. Whether the hosting exemptions in Article 14(1) of the E-Commerce Directive, Article 6(1) of the Digital Services Act, or both, apply to a claim for damages under Article 82 GDPR.

This account is based on the BGH's press releases. The Court of Justice has not yet given the case a number.

Why it mattersIn Russmedia the Court of Justice held that the operator of an online marketplace can be a controller for personal data in users' advertisements, and that it cannot rely on Articles 12 to 15 of the E-Commerce Directive in respect of infringements of its obligations under Articles 5(2), 24 to 26 and 32 GDPR. The BGH now asks whether that reasoning extends to a social network that hosts without editing, and whether the hosting exemptions can shield a platform from GDPR damages at all. The answers could clarify the liability of social network operators for personal data in content their users post. Until the Court rules, the questions remain open.
Practical impactFor platforms: document what the service does with user content beyond storing it, including ranking and recommendation, since the reference treats that as a possible factor in the questions on the hosting exemptions. For parties to pending damages claims over user content: expect the reference to be raised.
Copies a clean, cited sentence for use in a note.
AI agents Spain Guidance for users 27 October 2026

The AEPD sets out what users should check before delegating tasks to AI agents, and what it expects providers to offer

On 27 October 2026 the AEPD published a decalogue for people who use AI assistants and agents. Its ten recommendations include granting each agent only the access a task needs; deciding what it may remember; requiring confirmation before it sends, shares, pays or contracts; being wary of hidden instructions in websites, emails and documents; and protecting the data of other people that an agent can reach through email, calendars and photos.

The document also states that responsibility for AI assistants and agents "cannot rest only on the user", while reminding users that they may be responsible for the actions they have allowed an agent to carry out. The document then lists seven controls that, in the AEPD's view, providers should offer:

  1. Understandable permissions that distinguish consulting from acting.
  2. Configurable confirmation for the actions with the greatest impact.
  3. A history of what the agent did on the user's behalf.
  4. An immediate way to stop the agent.
  5. Management and deletion of the agent's memory.
  6. Clear information on who the data are sent to and whether they are used to train AI.
  7. A simple way to exercise rights.
Why it mattersThe decalogue is addressed to users and creates no obligation. Its list for providers gathers in one place the controls the AEPD expects AI agent products to build in. That makes it a practical benchmark for data protection by design and by default in such products.
Practical impactFor providers of AI assistants and agents in the EU: check the product against the seven controls. For an organisation whose staff use AI assistants: build into the workplace AI policy the three rules the AEPD gives users, namely to use only the assistants the organisation has authorised, not to install agents on a work device themselves, and not to connect work email or documents to a personal AI account.
Copies a clean, cited sentence for use in a note.
4

Enforcement Watch

Freedom of expression Sweden Decisions, may be appealed 3 November 2026

IMY disapplies Sweden's constitutional protection for people-search services and finds breaches of Articles 10 and 17 GDPR

IMY, the Swedish data protection authority, has issued decisions against four online services that let anyone search for information about private individuals: Lexbase, Krimfup, MrKoll and Upplysning. All four hold a voluntary publication certificate, which gives a database the constitutional protection of the Swedish Fundamental Law on Freedom of Expression. Under Chapter 1, Section 7, first paragraph, of the Swedish Data Protection Act, the GDPR does not apply where it would conflict with that law or with the Freedom of the Press Act.

IMY relied on the Court of Justice's judgment of 9 July 2026 in C-199/24 Legal Newsdesk Sweden, a reference from a Swedish district court involving the company behind Lexbase. According to the Court's press release, the Court held that Member States may not provide exemptions under Article 85 for processing carried out for purposes other than those the article lists. On the Court's test, processing is journalistic where it aims to disseminate information, opinions or ideas to the public, in line with press ethics, with editorial work or at least an editorial line, and with the facts checked. The Court added that a paid online database of criminal convictions does not appear to meet those conditions, subject to the national court's assessment.

Applying that test, IMY found that the Swedish exemption is not necessary to reconcile data protection with freedom of expression for services of this kind, and so is not compatible with Article 85(2) GDPR. It found that the Swedish rules cannot be read in conformity with EU law, and that EU law therefore takes precedence over the constitutional rules and Section 7. It also found that none of the four services processes data for journalistic purposes, and that the use of a database by journalists does not make the service journalistic.

The findings differ by service:

  1. Lexbase, run by Legal Newsdesk Sweden AB, and Krimfup, run by Fuplex AB: processing of data on criminal convictions and offences in breach of Article 10 GDPR. IMY reprimanded each and ordered it, within two months of the decision becoming final, to change the service so that a search by name, personal identity number or address no longer shows such data about the person searched.
  2. MrKoll, run by Nusvar AB, and Upplysning, run by Upplysning Checknode AB: failure to comply with erasure requests under Article 17(1) GDPR. In the case of MrKoll, requests were dealt with only after IMY intervened, and the company showed no ground for refusing them. IMY reprimanded each.

IMY, which had received about 3 800 complaints against the companies, imposed no fines. It said the scale of the processing would normally justify one, but took into account the constitutional protection the companies held and the absence of case law until recently. The decisions can be appealed to the Administrative Court in Stockholm, through IMY, within three weeks.

Why it mattersA supervisory authority has set aside its own country's constitutional protection for these services and applied the GDPR, relying on the primacy of EU law and on a judgment that concerned one of the same companies. In IMY's reading of C-199/24, a national media exemption covers only processing for the purposes Article 85 lists, and a searchable database of individuals does not become journalism through a publication certificate. The decisions are not final.
Practical impactFor any service that publishes or sells searchable data about individuals in reliance on a national media or press exemption: test the service against the C-199/24 criteria, which apply to every Member State's exemption under Article 85, rather than against the national label. If it falls outside them, apply Article 10 to conviction data and handle erasure requests on their merits under Article 17.
Copies a clean, cited sentence for use in a note.
Cookies and trackers France Fine 22 October 2026, published 29 October 2026

The CNIL fines Fnac 7 million euro over an app tracker that ran without consent and fed its recommendation algorithm

By a decision of 22 October 2026, the CNIL's restricted committee fined FNAC DARTY PARTICIPATIONS ET SERVICES 6 million euro and FNAC DIRECT 1 million euro for breach of Article 82 of the French Data Protection Act, the provision on cookies and other trackers, which transposes Article 5(3) of the ePrivacy Directive.

The decision concerns the Fnac mobile app only. The app sent requests to a third-party domain, and those requests contained an identifier assigned to the user by Fnac and information on their activity, such as the products viewed. The data were used to show personalised recommendations of complementary and similar products, and to train the recommendation algorithm. The requests were sent even after the user had chosen to continue without accepting. The CNIL held that neither exemption in Article 82 applied: the tracker's sole purpose was not to enable or facilitate a communication by electronic means, and it was not strictly necessary for an online communication service the user had expressly requested. Prior consent was therefore required.

In setting the amount, the CNIL took into account the tracker's disregard of users' refusal, the large number of the app's users, and the CNIL's settled position on trackers used for personalisation and commercial targeting. The companies must show, within one month of notification of the decision, that the app now obtains consent where required, under a penalty of 10 000 euro per day of delay. The case began with a complaint by noyb, a campaigning organisation, in 2023; the CNIL inspected the app in March 2024.

Why it mattersThe decision confirms that the consent rule for trackers covers data collected inside an app and sent to a third party, and that using those data to train a recommendation model does not bring them within an exemption. In setting the fine, the CNIL weighed not only the missing consent but the fact that the tracker overrode users' refusal.
Practical impactFor app publishers: list the requests the app sends to third-party domains, check which carry an identifier or activity data, and confirm that each runs only after consent unless it is strictly necessary for a feature the user asked for. Where activity data feed a recommendation or other model, include that use in the consent request.
Copies a clean, cited sentence for use in a note.
Identity documents and consent Poland Court judgments 6 and 23 October 2026

Polish courts uphold UODO on a bank's blanket scanning of identity documents and on proof of explicit consent to health data

The ING Bank Śląski fine

On 23 October 2026 the Voivodeship Administrative Court in Warsaw dismissed ING Bank Śląski's complaint against a UODO decision fining it 18 416 400 zloty for breach of Article 5(1)(a), (b) and (c) and Article 6(1) GDPR. According to UODO, after amendments to the Polish anti-money laundering act took effect on 13 July 2018, the bank's procedures required the identity documents of customers and prospective customers to be scanned, without an individual risk assessment, in every case the procedures listed. Scanning was often a condition of carrying out an operation for the customer. According to UODO, scanning also took place in situations unrelated to anti-money laundering duties, such as a complaint about an ATM. In UODO's view, which the court upheld, scanning is lawful under the anti-money laundering act only where it is necessary to apply the required security measures on the basis of an individual risk assessment. The judgment is not final and can be appealed to the Supreme Administrative Court.

Explicit consent to health data

On 6 October 2026 the Supreme Administrative Court, in case III OSK 1943/24, dismissed the appeal of a firm that helps people claim compensation, mainly after road accidents. UODO had fined its partners 45 697 zloty and ordered them to stop processing prospective clients' personal data, including health data, without a legal basis, after the firm collected such data from public sources and through follow-up calls, before any contract was concluded. As reported by UODO, the court held that explicit consent to the processing of health data may be given orally, but that under Article 7(1) GDPR the controller bears the burden of proving it and cannot shift that burden to the authority or to the data subject. According to UODO's account, the court also held that silence, the absence of an objection, starting a conversation or sending medical documents does not by itself show explicit consent.

Why it mattersBoth judgments are known so far only from UODO's own account of them, and the ING judgment is not final. The ING ruling applies the data minimisation principle to the default scanning of identity documents that the bank justified by the Polish anti-money laundering act. The second ruling separates the form of explicit consent, which may be oral, from its proof, which the controller must be able to produce.
Practical impactFor firms subject to Polish anti-money laundering rules: check that identity documents are scanned only where an individual risk assessment requires it, and never for unrelated contacts such as complaints. For firms in other Member States: check what national anti-money laundering law requires before changing scanning practice, and whether the same minimisation reasoning applies to scans that law does not require. For any controller relying on explicit consent to special category data: record when, how and to what the person consented, including where consent is given by phone or in person.
Copies a clean, cited sentence for use in a note.
Data subject rights and security Italy / Romania Fines and orders 23 September and 30 October 2026

In brief: the Garante fines Barbuti over an erasure request left in a general mailbox, and ANSPDCP orders Auchan to restrict remote access after a breach

Italy: Barbuti

By a decision of 23 September 2026, reported in its newsletter of 29 October 2026, the Garante fined Barbuti S.p.A. 5 000 euro. A customer who could not unsubscribe from promotional text messages asked for erasure by writing to the general mailbox linked to the company's website contact form, which the company admitted no one was assigned to monitor. He received an automatic acknowledgement but no reply, and the messages continued. Relying on the EDPB's Guidelines 01/2022 on the right of access, the Garante noted that a request sent through a channel other than the preferred one is in general valid, and that a controller need not answer a request sent to a random address where it has provided an appropriate channel. It found that this request was clear and could easily have been redirected, and that the messages gave no contact point for exercising rights. It found breaches of Articles 5(2), 12 and 17 GDPR and of Article 130 of the Italian Data Protection Code. The company must take measures so that its promotional messages comply with Article 130(5) of the Code, which requires a contact point for exercising rights, and report to the Garante within 30 days.

Romania: Auchan

In a press release of 30 October 2026, ANSPDCP, the Romanian supervisory authority, reported that it had fined Auchan România S.A. 41 908 lei, about 8 000 euro, for breach of Article 32(1)(b) and (2) GDPR. An attacker compromised an administrator account on one of the company's security systems and accessed the usernames, names, email addresses and passwords of a significant number of internal users. The investigation, which followed the company's own breach notification, found that the company had not set specific requirements for secure access to its network storage equipment. ANSPDCP ordered it to ensure that its IT equipment is accessed only from the internal network and that any external access uses, at least, a VPN and multi-factor authentication together.

Why it mattersEach decision turns a general duty into a concrete expectation in the case at hand. For the Garante, a clear request that reaches a company mailbox the public can write to has to be redirected and answered. ANSPDCP required one retailer's IT equipment to be accessed from the internal network, with any external access through at least a VPN and multi-factor authentication together.
Practical impactCheck that every company mailbox the public can write to, not only the one named for privacy requests, is monitored, and that rights requests reaching it are forwarded to whoever handles them. Check that access to systems holding personal data is limited to the internal network or, where external access is needed, protected by VPN and multi-factor authentication together.
Copies a clean, cited sentence for use in a note.

Each story links to its supporting sources. Editorial analysis is clearly marked separately from factual reporting.

Sources and references

How the brief is prepared and checked

This newsletter is prepared with the help of AI and goes through several stages of review. Please check the original source before relying on the information presented here.