On 20 December 2024 Italy’s data protection authority, the Garante, fined OpenAI EUR 15 million, closing an investigation that began in 2023 into the use of personal data to train ChatGPT. The sum is modest by global standards; the reasoning is not. The decision lands on the question every model developer faces — can you train on personal data, and on what basis?
The substance of the decision
The Garante’s concern was not that a generative model exists. It was that personal data was used for training without an adequate legal basis, and that users were not told clearly enough what was happening to their data. The regulator also examined the handling of data subjects’ rights, including how requests to correct or delete data can be satisfied when that data has shaped a model.
That framing matters. It treats the training pipeline as an ordinary data-processing activity, subject to the ordinary GDPR questions: what is the lawful basis, what was disclosed, and can rights be exercised. A model is not a legal shield around the data that built it.
A pattern, not an outlier
This decision did not arrive alone. Our notes on the GEMA v OpenAI ruling in Munich cover the copyright side of training; the ruling on disclosure of OpenAI user logs covers the procedural side. Together with the EDPB’s coordinated enforcement on transparency, they show regulators converging on the same pressure points from different directions.
The EU AI Act adds a separate layer through its GPAI obligations, which address documentation and disclosure for general-purpose models. Data-protection enforcement and AI Act compliance are distinct tracks, and a company can be exposed on both.
Practical actions for AI companies
First, map your training data. A data map that names sources, categories and retention periods is the document every regulator asks for first, and the one most companies lack.
Second, document the legal basis before training begins, not after a complaint. Retrospective justification is the weakest form of defence.
Third, make transparency notices specific. Vague statements that “data may be used to improve services” do not satisfy the standard the Garante applied.
Fourth, treat vendor diligence as part of the same exercise. The same logic applies in China through the obligations of major personal information processors, so a single, well-maintained data map serves both markets.
This article is based on the Garante’s decision concerning OpenAI and public information and is for reference only; it does not constitute legal advice.