Data protection authorities have applied the GDPR to AI for years, and the delay to AI Act deadlines under the 2026 Digital Omnibus changes none of that. According to enforcement trackers, regulators in Italy, Greece, France and the Netherlands fined facial-recognition provider Clearview AI between 2022 and 2024, and Italy's Garante has acted against OpenAI, Replika and DeepSeek. In December 2024 the European Data Protection Board (EDPB) added a formal opinion on AI models.
The record is not one-directional. In March 2026 a Rome court annulled the Garante's EUR 15 million fine against OpenAI, reportedly on jurisdiction rather than on the merits. This article draws out what the EDPB opinion and the enforcement record, imperfect as it is, tell businesses that build or buy AI today. It is general information, not legal advice, and it complements our earlier piece on GDPR fines for AI mistakes and our guide to how the GDPR and the AI Act overlap.
EDPB Opinion 28/2024: the rulebook for AI models
The EDPB adopted the opinion in December 2024 (sources give 17 or 18 December) at the request of the Irish Data Protection Commission under Article 64(2) GDPR. It answers four questions, summarised by Hunton and CMS.
- Anonymity. A model is anonymous only if two conditions hold: the likelihood of extracting training subjects' personal data directly, including probabilistically, is insignificant, and so is the likelihood of obtaining it through queries. Claims need documentation such as DPIAs and technical safeguards, per William Fry.
- Legitimate interest. It is not a default basis. Controllers must pass a three-step test of interest, necessity and balancing. Examples that may qualify include a conversational agent that assists users, fraud detection and threat detection in information systems.
- Reasonable expectations. Relevant factors include whether the data was public, the relationship with the controller, the nature of the service, the context and source of collection, further uses of the model, and whether people know their data is online.
- Mitigating measures. These include pseudonymisation, synthetic data, transparency beyond Articles 13 and 14, and an opt-out offered from the start.
- Unlawful development. The effect on later deployment is assessed case by case. Corrective measures can include a fine, a temporary limitation, or erasure of the unlawfully processed data and, where that is impossible, of the whole dataset or the model, subject to proportionality, as Baker McKenzie reports.
Two points deserve emphasis. If a model is genuinely anonymous, the GDPR does not apply to its later operation, and earlier unlawfulness should not affect it. And deployers cannot assume the developer got it right: they must assess their own compliance, with depth proportionate to risk.
The enforcement record
The table collects the main actions reported by enforcement trackers and law-firm commentary. Figures and dates come from secondary sources, and the primary decisions were not reviewed.
| Authority | Date | Target | Action | Status |
|---|---|---|---|---|
| Garante (Italy) | Feb 2022 | Clearview AI | EUR 20m fine; biometric processing ban; deletion of Italian residents' data | Appeal status not confirmed |
| HDPA (Greece) | Jul 2022 | Clearview AI | EUR 20m fine | Appeal status not confirmed |
| CNIL (France) | Oct 2022 | Clearview AI | EUR 20m fine | Appeal status not confirmed |
| AP (Netherlands) | Sep 2024 | Clearview AI | EUR 30.5m fine, plus a possible further penalty of up to EUR 5.1m | Appeal status not confirmed |
| Garante (Italy) | Late 2024 (sources say Nov or Dec) | OpenAI (ChatGPT) | EUR 15m fine and media-campaign order: no legal basis for training, transparency, age verification, breach notification | Annulled by the Court of Rome on 18 March 2026; a possible Garante appeal was reported as under consideration in June 2026 |
| Garante (Italy) | Jan 2025 | DeepSeek | Order blocking the app in Italy; no fine | Single source |
| Garante (Italy) | May 2025 | Luka Inc. (Replika) | EUR 5m fine: no valid legal basis, no working age verification | Single source; outcome unknown |
Sources for the table: privacyterms, gdprfine, Blocksize and Wilson Sonsini. The last of these is written by counsel for OpenAI and should be read as a partisan account.
Five lessons from the record
- Where you are established decides who regulates you. According to secondary reports, the Rome court focused on jurisdiction under the one-stop-shop. OpenAI had set up OpenAI Ireland Ltd in February 2024, which made the Irish authority the lead regulator, and the court did not decide whether the GDPR was breached. The outcome is a procedural win, not a clean bill of health, and the Irish Data Protection Commission remains free to act.
- Children and age verification recur. The Garante cited missing age checks in both the OpenAI and Replika cases. Any AI service that minors could reach should be able to show how it assesses age and limits what it does with children's data.
- The legal basis for training is the weak point. Authorities repeatedly allege that no valid basis exists for training data. The EDPB's three-step test gives a route, but only to controllers who can document it.
- Territorial reach is broad. In the Clearview matter the Garante relied on the GDPR's extraterritorial scope and required an EU representative, according to privacyterms. A provider outside the EU is not outside the regime if it monitors people in the EU.
- Remedies go beyond fines. Suspensions, blocking orders and deletion appear throughout the record. The Garante's provisional block of ChatGPT in 2023 was lifted after changes, as gdprfine reports, and the EDPB opinion contemplates erasure of datasets and, in the end, models.
The common thread is that regulators ask basic questions: what data went in, on what legal basis, who can be affected, and what people were told. Businesses that can answer them in writing are in a much stronger position than those who cannot, whatever the final outcome of any single case.
What businesses should do now
The lessons translate into a short list of actions for any organisation that builds, fine-tunes or buys AI.
- Document the legal basis. For training and fine-tuning, record the lawful basis and, if relying on legitimate interest, a written three-step assessment with the mitigating measures chosen.
- Test anonymity claims. If you or a vendor describe a model as anonymous, ask for the evidence against both EDPB conditions, not a marketing statement.
- Handle age and vulnerable users. Where minors may use a service, put age assurance and tailored defaults in place and keep proof of them.
- Make transparency specific. Go beyond a generic notice: say what sources and categories of data feed the model and how people can object.
- Prepare for rights requests. Decide how you will handle access, objection and erasure when personal data may sit in training sets or outputs.
- Map your lead authority. Know where you are established, since that determines which regulator leads.
- Question your vendors. The EDPB expects deployers to assess their own compliance, so collect training-data provenance and legal-basis statements from suppliers.
The bottom line
The enforcement record is uneven, and appeals show that procedure matters. Yet the direction is consistent. Regulators expect AI providers and deployers to explain, in documents they can show, why personal data is used, how people are protected and what happens when something goes wrong. Waiting for the AI Act's high-risk dates is not a reason to postpone that work.
To see how your own AI tools measure up on these basic questions, try GDPRGard's audit tool.
Sources
- CMS: EDPB Opinion 28/2024, key takeaways
- Hunton: EDPB publishes opinion on AI models
- William Fry: unwrapping the EDPB's guidance on AI models
- Baker McKenzie: EDPB opinion on AI models
- Wilson Sonsini: OpenAI prevails in Italian enforcement case (counsel for OpenAI)
- gdprfine: Italian Garante enforcement record
- privacyterms: AI privacy fines tracker
- Blocksize: GDPR and generative AI 2026
- SecurityWall: GDPR fines tracker 2026
Enforcement figures come from trackers and commentary that often repeat each other, so agreement between them is not independent confirmation. Dates conflict in places (the OpenAI fine is dated November or December 2024), several items rest on a single source, and no primary decision or judgment text was reviewed. Verify against the Garante, CNIL, HDPA, AP and court records before relying on any figure.
Also read:
- Forget the AI Act — GDPR is already fining companies for AI mistakes
- Is Your “Anonymous” Data Really Anonymous? The EDPB's New Draft Changes the Test
- The "It Was Public" Excuse Is Over: GDPR Now Formally Covers AI Training Data