The sheets cover development (design, dataset creation, training), not deployment, and the GDPR and the AI Act both apply where personal data is involved.
This page reads them as of 2026-11-17 for a company that supplies or acquires a training dataset, within AI training data licensing. No CNIL sanction concerning AI training data is recorded on the pages read.
Every statement below cites a CNIL page read on 2026-11-17, and the method is set out in how we source, date and check every figure.
The CNIL's AI How-To Sheets: What They Are and What Status They Have
The CNIL's English summary page, dated 2026-01-05, lists how-to sheets 0 to 13 and a checklist, and the sheets are guidance with no legal force. The table lists the sixteen entries as of 2026-11-17, with the date shown on each English page and what the entry covers for a data supplier or acquirer.
| Sheet | Title (CNIL English) | Date shown | What it covers for a data supplier or acquirer | Source |
|---|---|---|---|---|
| 0 | What is the scope of the AI how-to sheets? | 2024-06-07 | development phase only, not deployment | cnil.fr |
| 1 | Determining the applicable legal regime | 2024-06-07 | GDPR applies to all processing of personal data | cnil.fr |
| 2 | Defining a purpose | 2024-06-07 | explicit, legitimate purpose; general-purpose not too broad | cnil.fr |
| 3 | Determining the legal qualification of AI system providers | 2024-06-07 | controller, joint controller, processor | cnil.fr |
| 4 (1/2) | Ensuring the lawfulness of the data processing: defining a legal basis | 2024-06-07 | legal bases compared; consent often impractical for data collected online | cnil.fr |
| 4 (2/2) | In case of re-use of data, carrying out the necessary additional tests and verifications | 2024-06-07 | compatibility test; checks on public and third-party datasets | cnil.fr |
| 5 | Carrying out a data protection impact assessment when necessary | 2024-06-07 | covered on the DPIA page | cnil.fr |
| 6 | Taking into account data protection when designing the system | 2024-06-07 | not summarized here | cnil.fr |
| 7 | Taking data protection into account in data collection and management | 2024-07-02 | minimization, cleaning, dataset documentation, retention | cnil.fr |
| 8 | Relying on the legal basis of legitimate interests to develop an AI system | EN page 2026-01-05 | three-step test, examples, safeguards | cnil.fr |
| 8 bis | The legal basis of legitimate interest: focus sheet on web scraping | EN page 2026-01-05 | scraping safeguards, publisher signals | cnil.fr |
| 9 | Informing data subjects | EN page 2026-01-05 | individual and general notice | cnil.fr |
| 10 | Ensuring and facilitating the exercise of data subjects' rights | EN page 2026-01-05 | datasets, models, timeframes | cnil.fr |
| 11 | Annotating data | EN page 2026-01-05 | minimization, accuracy, sensitive labels | cnil.fr |
| 12 | Ensuring the security of an AI system's development | EN page 2026-01-05 | datasets, supply chain, models, access | cnil.fr |
| 13 | Analysing the status of an AI model with regard to the GDPR | EN page 2026-01-05 | anonymity tests, documentation | cnil.fr |
Sheets 0 to 6, including both parts of sheet 4, are dated 2024-06-07 and sheet 7 is dated 2024-07-02 on the CNIL's own pages, while the English pages of sheets 8 to 13 show 2026-01-05. The English pages are courtesy translations, and the French text prevails.
The summary page numbers its "Steps" differently, so this page uses sheet numbers. Sheet 5 presumes an impact assessment is required for AI development when at least two criteria apply, and the full treatment is on the page about the DPIA for an AI data licensing deal.
On 2026-08-26 the CNIL updated Genmod, a demonstrator that maps the genealogy of open-weights models from public models and datasets on Hugging Face.
Checks on a Dataset Acquired From a Third Party (Sheet 4)
Sheet 4 (re-use), dated 2024-06-07, says a developer must not re-use a dataset whose creation or transmission was manifestly unlawful, and that its own processing must comply with the GDPR whatever the dataset's origin.
The sheet separates three cases: a provider's own data, a publicly accessible dataset, and a dataset from a third party, for which the CNIL's own example is a data broker (the sheet's example, not an endorsement of any intermediary).
For the third party sharing the data, the sheet says the duties depend on why the data was collected. If it was collected to be shared for training, the sharer ensures the legal basis and the information given to people.
If it was collected for another purpose, the sharer first runs the compatibility test, which is owned by the page on purpose compatibility (Article 6(4)).
For a public dataset, the sheet lists four checks:
- the source is named in the dataset description;
- the dataset is not manifestly the result of a crime, and no public sanction requires its deletion;
- there is no clear doubt about the lawfulness of the collection;
- there are no special category or offence data, or extra checks are made.
For a re-user, the sheet recommends a contract covering the source, the collection context, the legal basis, an impact assessment where needed, the information given to people and guarantees of lawful sharing and compatibility.
Consent obtained by a third party must be provable, and a clause obliging the other party to obtain consent is not enough alone. Knowingly re-using a manifestly illegal dataset may amount to concealment under Article 321-1 of the French Criminal Code.
Web Scraping Safeguards in Sheet 8 bis
Sheet 8 bis, the CNIL's focus sheet on web scraping, dated 2026-01-05 on the English page, says scraping is "not prohibited per se" and must be analyzed case by case, and that a controller must exclude sites that clearly oppose the scraping of their content.
Before collection, the sheet asks for defined criteria and a default exclusion list covering intrusive or sensitive sites and sites used mainly by minors.
Collection is limited to freely accessible content, irrelevant Article 9 data is excluded automatically, and incidental sensitive data is deleted once known. After collection, irrelevant data is deleted at once and the rest is anonymized or pseudonymized.
The table lists the publisher signals the sheet names, as of 2026-11-17, and how the sheet presents each one.
| Signal or measure | Type in the sheet | What the sheet says the controller does | Source |
|---|---|---|---|
| CAPTCHAs | blocking | exclude sites that clearly oppose scraping | cnil.fr |
| IP analysis, rate monitoring, bot IP and VPN registers | blocking | same | cnil.fr |
| robots.txt | listed with the non-blocking measures; the sheet also says publishers should use it or CAPTCHAs | exclude sites that clearly oppose scraping | cnil.fr |
| ai.txt, TDMRep | non-blocking | listed as publisher protocols; no duty beyond the exclusion rule is stated (editorial reading) | cnil.fr |
| "noai" meta tags, dynamic page loading | non-blocking | listed as publisher measures | cnil.fr |
| do-not-train registries | non-blocking | listed as publisher measures | cnil.fr |
| Objection lists | controller measure | discretionary and prior right to object; skip the data of objecting people | cnil.fr |
The sheet treats copyright and the text and data mining reservation as separate limits. Sheet 8 gives the three-step test with examples, and the full test is on the page about legitimate interest for AI training data.
Informing People: What Sheet 9 Accepts as a General Notice
Sheet 9 says people whose data was collected indirectly must be told as soon as possible and at the latest at first contact, in any case within one month, but accepts a general notice where individual notice is disproportionate and documented.
Article 14(5)(b) turns on factors the sheet lists: missing or old contact details, the number of people, how identifiable they are, the sensitivity of the data and the safeguards in place. Lawfully scraped, indirectly identifying data often qualifies, volume alone does not, and a held email address should be used.
A general notice is published on a freely accessible site and says that individuals cannot be identified and what extra information would help. Where sources are many, they are given by category, and for a re-used dataset the original controller is named.
The sheet pairs the notice with compensating measures: an impact assessment even if not required, pseudonymization, less data, shorter retention and security. The Article 14 rules themselves sit on the page about GDPR notices before a data sale.
People's Rights Over Datasets and Models (Sheet 10)
Sheet 10, dated 2026-01-05 on the English page, says access, rectification, erasure, objection, restriction and portability apply to training datasets and to models that are not anonymous, within one month, extendable by two months, so up to three months in total.
For datasets, a controller that cannot identify a person and can show it says so, and the sheet states no duty to keep or collect identification data, with blurred faces as its example. Where legitimate interest is the basis, objection applies, and opt-out and rejection lists are encouraged.
For models, the sheet describes query and attack tests and a documented procedure for generative providers. Retraining is the most effective method where the training data is held, and filters on outputs are a fallback where retraining is disproportionate, provided they are robust and not easily bypassed.
The summary page adds that retraining "may take up to three months". Erasure across datasets and models is owned by the page on removing content from AI training datasets.
For re-used datasets, the sheet recalls the Article 19 notice to recipients unless impossible or disproportionate, and recommends documenting dataset versions. It also suggests a contract clause in the re-use license requiring re-users to pass on objection, rectification and erasure requests.
Annotation Under Sheet 11
Sheet 11 says annotations and the data being annotated must be limited to what training requires, and that annotation can create special category data even when the source data is not sensitive. Sensitive annotation is prohibited in principle under Article 9, and the organization annotating must identify a lawful exception.
For health research the sheet cites Article 44-3 of the French Data Protection Act and Article 9(2)(j) GDPR, and it prefers synthetic data. It recommends in-house or local annotation, encryption, logging, task-based access and an annotation protocol with ongoing checks.
On roles, a provider building a dataset for a client is generally a processor and must ensure only relevant annotations are made. A provider of ready-made datasets should keep annotation categories separable so that minimization can be respected.
Notices name the annotating organization and any transfer outside the EU. The Article 9 rules are covered on the page about special category data (Article 9) in AI training.
Security and Model Status (Sheets 12 and 13)
Sheet 12 asks developers to version and log datasets, vet libraries and pre-trained models, and restrict access, and sheet 13 says a model trained on personal data is subject to the GDPR unless re-identification is shown to be insignificant.
Sheet 12, dated 2026-01-05 on the English page, also covers safer file formats such as safetensors over pickle, hashing or signatures on trained models, a memorization assessment before release and encrypted backups.
Sheet 13 treats anonymity case by case and builds on EDPB Opinion 28/2024. Its indicators include rare or duplicated data, overfitting and generative use, and its tests include regurgitation, membership inference and inversion.
The documentation must be presentable to the CNIL, and re-identification test results are needed in all cases for systems and in most for models. The opinion is owned by the page on when an AI model is anonymous under EDPB Opinion 28/2024.
What the CNIL Can Require and What Is Recorded as Enforcement
Sheet 13 says that where the GDPR applies to a model and personal data can be extracted, the CNIL may require the model to be retrained or removed, but no CNIL sanction concerning AI training data is recorded on this page.
The sheet adds that extraction can be a personal data breach, with notification to the CNIL "within 72 hours" where risks to people are likely.
The CNIL's AI topic page and the sheets read on 2026-11-17 list no sanction or formal notice on training data, which describes those pages and does not claim that none exists.
Genmod is a lineage tool for studying memorization across open-weights models, not an enforcement step. The CNIL's 2022 note on selling customer files belongs to the French parent page.
Using the Sheets Before a Deal: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Following the CNIL's sheets costs documentation and review time before any license fee arrives, but they are non-binding, and licensing less or nothing remain valid choices.
Benefits
- The sheets give a regulator's own checklist of what to record.
- Documentation of the kind sheet 13 describes is what a buyer or authority may ask for.
- Sheets 4 and 10 supply a clause set for a re-use license.
Risks
- Ignoring a sheet is not itself an offense, but the sheets describe how the CNIL reads binding GDPR duties.
- Knowingly re-using a manifestly illegal dataset risks Article 321-1 of the French Criminal Code, as sheet 4 says.
Downsides
- Sheet numbering and dates differ across the CNIL's pages.
- Some sheets are English courtesy translations, and the French text prevails.
- No CNIL sheet is written for a seller of a dataset.
Alternatives
- License only anonymous or non-personal data.
- Use synthetic data, which sheet 11 prefers.
- Narrow the set to what a buyer's purpose needs.
- Not license at all, as set out in the alternatives to selling data to AI labs, including not selling.
Beyond the CNIL: The Other French and EU Rules for a Data Deal
The CNIL's sheets sit under the GDPR. The French parent page sets them beside the text and data mining objection and the works council rules: France: rules for licensing data for AI training.
- AI training data law by jurisdiction
- GDPR and AI training data
- Web scraping for AI training
- robots.txt for AI crawlers
- EU text and data mining exception
- EU AI Act Article 53
- Due diligence in an AI data deal
- De-identification before licensing
- Data licensing agreement clauses
- France: text and data mining opt-out
- France: CSE consultation
- French liquidation and data sales
How This Page Is Sourced and Dated
Every statement cites a CNIL page with the date it was read, and the GDPR and EDPB texts sit on their own pages. This page is general information about CNIL guidance, not legal advice for any specific dataset or deal.
Frequently Asked Questions
Are the CNIL's AI how-to sheets binding?
No, the sheets are guidance. The CNIL's English page dated 2026-01-05 presents them as its reading of the GDPR, which does bind.
What are the CNIL's "fiches pratiques IA"?
They are the CNIL's numbered how-to sheets on applying the GDPR to developing AI systems. They run from sheet 0 (scope) to sheet 13 (model status).
Does a company that sells a dataset have to follow the sheets?
A seller is a controller for its own disclosure, so the GDPR applies to it directly. The sheets describe re-user and provider duties, and none is addressed to a seller.
Has the CNIL fined anyone for AI training data?
No CNIL sanction concerning AI training data is recorded on the pages read on 2026-11-17. That describes the pages read and does not claim that none exists.