The Biggest AI Privacy Fines and Rulings So Far

Regulators and courts already penalize how AI uses personal data: in December 2024, Italy's authority fined OpenAI €15 million over ChatGPT's data processing; the CNIL and several European authorities hit Clearview AI with fines up to €30.5 million for scraping faces; and a U.S. federal court ordered OpenAI to retain conversation logs that users had asked to delete. The common thread across these rulings: penalizing the use of personal data without a legal basis or transparency. The most direct defense is to minimize and anonymize that data before entrusting it to an external AI system.

Italy: €15M against OpenAI, and the first ChatGPT suspension

On December 20, 2024, the Garante per la protezione dei dati personali (Italy's data protection authority) fined OpenAI €15 million after an investigation into ChatGPT's processing of personal data — one of the first major GDPR penalties against a large generative-AI player in Europe. The authority faulted the lack of an appropriate legal basis for training the model on personal data, a failure of transparency, and insufficient age-verification mechanisms to protect minors; OpenAI had also failed to notify a data breach that occurred on March 20, 2023. On top of the fine, OpenAI was required to run a six-month information campaign in Italian media. The company called the decision disproportionate and announced it would appeal.

That fine was not Italy's first intervention. On March 31, 2023, the Garante had already ordered a temporary suspension of ChatGPT in Italy — the first such block by a Western country — citing the lack of a legal basis for the mass collection of training data, a failure to inform users, and no age verification. The service was restored in late April 2023, about a month later, after OpenAI published an updated privacy policy and made an objection-to-processing form available to European users.

Clearview AI: penalized across Europe for scraping faces

Clearview AI, which built a biometric database by scraping face images from the web, was penalized repeatedly by European regulators. On October 20, 2022, the CNIL (Commission Nationale de l'Informatique et des Libertés) imposed the maximum €20 million fine for collecting and using biometric data without a legal basis, and ordered it to stop the collection and delete the data of people in France, under a €100,000-per-day penalty for delay. Comparable fines (around €20 million) followed in Italy and Greece.

In the Netherlands, the Autoriteit Persoonsgegevens (the Dutch data protection authority) announced a €30.5 million fine on September 3, 2024 — the largest imposed on the company — for the illegal build-out of a database of more than 30 billion photos converted into biometric codes, with an additional penalty of up to €5.1 million for non-compliance. In the United Kingdom, the Information Commissioner's Office (ICO) had penalized Clearview AI in 2022 (on the order of £7.5 million) for collecting and storing images of British citizens, with an order to stop the processing.

Courts and regulatory pressure: from frozen logs to Meta AI

Regulators aren't the only ones with leverage over data. In May 2025, in The New York Times v. OpenAI (a dispute originally over copyright), a U.S. federal judge ordered OpenAI to preserve and segregate ChatGPT output logs that would otherwise have been deleted — including ones users had asked to erase. In November 2025, the court upheld an order requiring OpenAI to produce roughly 20 million (de-identified) conversation logs to the plaintiffs, rejecting OpenAI's objection grounded in user privacy.

The lesson is concrete: a retention obligation imposed by a court can neutralize, for the duration of a lawsuit, your deletion policies and erasure requests. Conversations you believe are deleted can remain accessible and usable as part of a proceeding — a point we expand on in “Deleting a chat does not actually erase it.” Deleting therefore does not guarantee real erasure.

Meta AI: training on your data under scrutiny

Not every action ends in a fine: some take the form of regulatory pressure that reshapes practices upstream. In 2024-2025, Max Schrems's organization noyb (None of Your Business) challenged Meta's plan to train its AI on the public data of European Facebook and Instagram users starting May 27, 2025. noyb sent Meta a cease-and-desist and threatened a class action, contesting its reliance on “legitimate interest” rather than explicit (opt-in) consent for AI training.

After a December 2024 EDPB opinion and review by the Data Protection Commission (DPC, Ireland's authority), Meta revised its plan: updated transparency notices, longer notice periods, and working objection forms inside the app across European jurisdictions. On May 21, 2025, the DPC said it had made recommendations that Meta responded to — illustrating continued regulatory pressure on training AI from personal data, which ties into the framework we detail in “GDPR and generative AI: getting compliant.”

The timeline of penalties and rulings

Here are the main regulatory and judicial actions, in chronological order. Distinct from technical breaches, which we track separately in “The biggest AI data breaches”: these are penalties from authorities and court rulings, not security incidents.

The lesson: minimize and anonymize to cut legal exposure

A common denominator runs through these decisions: authorities and courts target the use of personal data without a legal basis, transparency or consent, and can impose data retention during litigation. You control neither an AI provider's jurisdiction, nor a court order, nor the legal basis it invokes. What you do control is what you transmit.

• Minimize: transmit only the data strictly necessary for the task.
• Anonymize or tokenize identities, identifiers and secrets before sending.
• Keep in mind that a “deletion” can be frozen by a legal obligation.

1. 1Map the personal data present in your prompts and files.
2. 2Remove or replace that data with tokens before any send to an external AI system.
3. 3Restore the real values when the response comes back, in the browser only.

That's exactly what ONYRI Sanitize is for: the engine replaces sensitive data with reversible tokens before sending; detection and the token↔value mapping stay in your browser, and only anonymized text reaches the model. Less personal data transmitted mechanically means less risk surface — whether a regulator investigates or a court demands log retention, it only finds tokens, not your real information.