AI Regulation Tracker  /  Regulator guidance

France's CNIL Says an AI Model Trained on Personal Data Is Usually Covered by the GDPR

France's data regulator finalized its recommendations on building AI systems under the GDPR, warning that a trained model is not automatically anonymous and often qualifies as personal data itself. Any firm training or deploying AI on data from people in France should map the guidance.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

France's data-protection regulator has closed off one of the most convenient assumptions in AI development: that once a model is trained, the personal data used to build it disappears. In recommendations finalized around January 5, 2026, the Commission Nationale de l'Informatique et des Libertés (CNIL) states that AI models trained on personal data must, in most cases, be treated as subject to the GDPR. A model, in other words, is not automatically anonymous, and the burden is on the controller to prove otherwise.

The recommendations apply the existing General Data Protection Regulation to the practical steps of building and using AI systems. They are guidance, not a new law. But because the GDPR is binding and the CNIL enforces it, the recommendations set out how the regulator expects controllers to behave, and they preview how other European authorities are likely to treat AI models.

Purpose comes first, and it limits the data

The CNIL frames purpose as the anchor for everything else. Controllers must define what it calls "a well-determined, explicit and legitimate objective" before collecting training data. That objective then caps how much personal data is fair to gather, because the GDPR requires data to be adequate, relevant, and limited to what is necessary. The regulator acknowledges general-purpose systems, where the exact use is broad, but still expects controllers to specify the system type and the functionalities they can foresee, rather than collecting broadly and deciding later.

In practice, the CNIL points to pilot studies, data cleaning, and regular monitoring of dataset integrity as ways to keep collection proportionate. The message is that "we are building AI" is not a purpose.

The model itself may be personal data

The most consequential part of the guidance concerns the status of the trained model. The CNIL says controllers must assess whether their model contains personal data, and that they cannot assume it is anonymous. Models can memorize training examples, which means individuals whose data went into training may still be identifiable from the model.

To reach anonymity, the regulator expects controllers to run re-identification attack tests on the model and to document the result. Where a model is found to carry personal data, the GDPR continues to apply to it, which affects rights requests, retention, and security long after training ends. A controller who ships a model without that test has not shown it is anonymous; it has simply assumed so, and the CNIL is signaling that the assumption will not hold.

This is the "what it does not do" beat worth stressing: the guidance does not declare all models to be personal data, and it does not ban training on personal data. It requires controllers to test and prove the point rather than assert anonymity. Nor does it invent new obligations. Every duty it describes flows from the GDPR that already governs these firms.

Security across the lifecycle, and telling people

The CNIL asks for security controls spanning the full AI lifecycle, including encrypted communications, access controls, dataset versioning, and security audits. It strongly recommends a Data Protection Impact Assessment where sensitive data, large-scale collection, or vulnerable groups are involved.

On transparency, controllers must give clear, accessible information to people whose data may be used, including where that data could be memorized by a model, and must preserve rights such as access and deletion. The 2026 agenda extends this work to non-anonymous AI models and to sector guidance, signaling that the CNIL sees this as the start of a program, not a one-off publication.

Why a US reader should care

For a US firm, the CNIL is not a US regulator and these recommendations are French guidance. But any company that trains or deploys AI on the data of people in France falls within the GDPR's reach, and the CNIL is one of Europe's most active authorities. The practical value is directional: mapping these recommendations now shows how EU data authorities are starting to treat AI models as potential personal data, a stance that tends to spread across the bloc and inform how regulators elsewhere frame the same question. A firm that already tests its models for re-identification and documents a clear training purpose will be better placed when the next authority, in France or beyond, asks the same thing.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.