AI regulations and confidential data: GDPR, professional secrecy, AI Act
AI regulations and confidential data come down to three sets of texts once a business document is handed to an AI tool: the General Data Protection Regulation (GDPR, Regulation (EU) 2016/679), professional secrecy, which French law enforces through Article 226-13 of the Criminal Code, and the EU regulation on artificial intelligence (AI Act, Regulation (EU) 2024/1689). The French data protection authority, the CNIL, adds recommendations on the use of generative AI systems.
Across all of them, the same measure keeps coming back: reduce the data you send to what is necessary, and pseudonymize or anonymize whatever can be. Every statement below links to the official text on EUR-Lex, Légifrance or the CNIL website. This summary is meant to guide your reading of the texts; for a specific situation, you still need legal advice.
The GDPR applied to a document handed to an AI
The GDPR applies as soon as a document contains personal data, which Article 4(1) defines as "any information relating to an identified or identifiable natural person". A name, an email address, a client file number or a social security number all qualify. Uploading that document to an AI tool is processing within the meaning of Article 4(2). Six articles shape the resulting obligations.
| GDPR article | What it provides | Application to a document handed to an AI |
|---|---|---|
| Article 4 | Definitions: personal data (point 1), processing (point 2), pseudonymization (point 5) | A redacted document whose table is kept separately and protected meets the definition of pseudonymization |
| Article 5 | Principles, including data minimization: data that is "adequate, relevant and limited to what is necessary" | Summarizing a contract or rewriting a letter does not require the identity of the parties |
| Article 6 | Lawfulness: processing rests on one of six legal bases; paragraph 4 lists encryption and pseudonymization among appropriate safeguards | Pseudonymization counts when assessing whether further processing is compatible with its original purpose |
| Article 9 | Special categories (health, racial or ethnic origin, opinions, beliefs, trade union membership, genetic and biometric data, sex life and sexual orientation): processing prohibited except in specific cases | A medical report or a social services report calls for extra care before anything is sent |
| Article 28 | Processors: only providers offering "sufficient guarantees" may be used, under a contract | The provider of an online AI that processes data on your behalf must be bound by contract |
| Article 32 | Security: appropriate measures, "including inter alia as appropriate: (a) the pseudonymisation and encryption of personal data" | Biffer applies both: aliases in the document, a table encrypted with AES-256-GCM on your computer |
Article 25 completes the picture: it cites pseudonymization as an example of data protection by design. Biffer applies these measures on your computer, before anything is sent, following the steps described on the how Biffer works page.
Pseudonymization and anonymization: Recital 26
The GDPR distinguishes two outcomes. Pseudonymization, defined in Article 4(5), means data can no longer be attributed to a specific person "without the use of additional information", provided that information is "kept separately" and protected. Recital 26 draws the consequence: pseudonymized data that could be attributed to a person by using that additional information is still information about an identifiable person. The GDPR therefore continues to apply.
The same recital states that data protection principles do not apply to anonymous information, meaning data "rendered anonymous in such a manner that the data subject is not or no longer identifiable". To decide, account must be taken of "all the means reasonably likely to be used" to identify the person, considering cost, the time required and the technology available.
The European data protection authorities set out three criteria in Opinion 05/2014 on anonymisation techniques (Article 29 Working Party, the predecessor of the EDPB), which the CNIL also applies: it must be impossible to single out an individual, to link records relating to the same person, and to infer new information about them.
- Redacted document with the table kept: the data is pseudonymized. The table, encrypted and stored on your computer, is the additional information kept separately.
- Redacted document without a table: the aliases no longer point to any value. The result comes close to anonymization, provided the remaining text (a unique job title, a dated event) does not allow the person to be found; the on-screen review is there to check this.
- Images and scanned PDFs: the pixels are blacked out and the information is gone from the file.
These concepts are covered in more detail on the anonymize a document and personal data and AI pages.
Professional secrecy: Article 226-13 of the French Criminal Code
Article 226-13 of the French Criminal Code (in French) punishes with one year's imprisonment and a €15,000 fine the disclosure of secret information by a person who holds it by virtue of their position or profession, or because of a temporary function or assignment. This text applies on top of the GDPR, whether the information concerns an individual or a company.
Profession-specific texts define its scope. For lawyers, Article 66-5 of Law No. 71-1130 of 31 December 1971 covers advice sent to the client, correspondence with the client and between lawyers, meeting notes and every document in the file. For health professionals, Article L1110-4 of the Public Health Code guarantees everyone in care respect for their private life and the confidentiality of information about them.
Sending the identifying content of a file to an AI service hosted by a third party amounts to disclosing it to that third party. Biffer removes the identity of clients, patients and parties before anything is sent: the AI works on PERSONNE_001 and RG_001, and the answer is unredacted on your computer. The lawyer, healthcare, notary and accountant pages list the identifiers recognized for each profession.
AI regulations and confidential data: what the AI Act provides
Regulation (EU) 2024/1689 classifies AI uses by level of risk and sets obligations for providers and deployers, meaning organizations that use an AI system in a professional setting. It adds to the GDPR without replacing it. The selector below shows the rules that apply to seven common situations.
The EU regulation on artificial intelligence (Regulation (EU) 2024/1689, known as the AI Act) entered into force on 1 August 2024 and applies in stages until 2027. It does not replace the GDPR: it classifies AI uses by level of risk and sets obligations for providers and for "deployers", meaning you, as soon as you use AI in a professional setting. The confidential documents you hand to a tool remain governed by the GDPR and by professional secrecy. Choose your situation below and the rules that apply are shown alongside. We are not lawyers: this overview points to the official texts and does not replace legal advice.
What you do
What the AI Act says
I summarize or rewrite a client document with ChatGPT, Claude or Copilot
Minimal risk- AI Actyou are a "deployer" of a general-purpose AI system. For this everyday use, the regulation imposes no specific formality, but since 2 February 2025 Article 4 requires you to ensure the AI literacy of your staff (understanding what the tool does with data, and its limits).
- GDPRthe AI provider becomes a recipient of the personal data in the document. You need a legal basis, compliance with data minimization (Article 5) and a contract governing the processor (Article 28); any transfer outside the European Union must be covered.
- Professional secrecyfor a lawyer, a doctor, a notary or an accountant in France, passing the identifying content of a file to a third party engages professional secrecy (Article 226-13 of the French Criminal Code).
- In practiceredact the document before uploading it; the AI works on aliases and the answer is restored on your computer. The issue disappears along with the data.
I screen résumés or assess candidates and employees with AI
High risk- AI Actemployment and worker management are listed in Annex III: recruitment, selection, evaluation, promotion and task allocation are high-risk uses. As a deployer, from 2 August 2026 Article 26 requires human oversight, use in line with the provider's instructions, relevant input data, retention of logs and informing employees and their representatives before the system is put into service.
- Prohibited(Article 5, since 2 February 2025): emotion recognition in the workplace, except for medical or safety reasons.
- GDPRno fully automated decision producing legal effects without safeguards (Article 22); candidates must be informed.
- In practicea redacted applicant spreadsheet (names, emails, phone numbers and application numbers replaced by aliases) can be analyzed without exposing anyone; the decision stays human.
I use AI on patient or beneficiary files
Sensitive data- AI Acta system that contributes to decisions on access to healthcare or essential social benefits falls under Annex III (high risk); a tool built into a medical device falls under Annex I. In both cases the provider has stricter obligations, and you must ensure human oversight and inform the people concerned (Article 26).
- GDPRhealth data is sensitive data (Article 9), processed only by way of exception; in France, hosting health data is also regulated.
- Medical secrecyArticle L1110-4 of the French Public Health Code.
- In practiceBiffer replaces patient identifiers, RPPS numbers (French health professional IDs), admission numbers, names and dates; a redacted report can be summarized without the tool being able to identify the patient.
I publish or distribute text, images or voice produced by AI
Transparency- AI ActArticle 50 (applicable from 2 August 2026) requires content generated or manipulated by AI to be disclosed: deepfakes (images, audio, video) must be labeled, and AI-generated text published to inform the public on matters of public interest must state its origin, unless it has undergone human review and editorial responsibility.
- Providerssystems must mark their outputs in a detectable way (machine-readable marking).
- Public releasebefore anything is published, the personal data in the source document must have been removed, this time under the GDPR.
- In practiceredacting the source document prevents a name, an address or a case number from ending up in the generated text, and then in print.
I train or fine-tune a model on my own files
Provider- AI Actyou become a provider. For a high-risk system, Article 10 requires governance of training data (relevance, representativeness, bias examination, provenance); for a general-purpose model, Article 53 requires technical documentation, a copyright compliance policy and a public summary of the training data (obligations applicable since 2 August 2025).
- GDPRthe French data protection authority (CNIL) recommends building training sets that are minimized, anonymized or pseudonymized, with a defined purpose; the EDPB (Opinion 28/2024) points out that whether a model is anonymous must be assessed case by case.
- In practicerun each file through Biffer before adding it to a training set; keep the mapping table separately, or do without it for permanent anonymization.
I have creditworthiness, an insurance risk or a public benefit application assessed
High risk- AI Actassessing the creditworthiness of individuals, pricing life and health insurance, and granting essential public benefits are high-risk uses (Annex III). Deployer obligations from 2 August 2026: human oversight, informing the people concerned, logs, and a fundamental rights impact assessment (Article 27) for public bodies and for banking and insurance institutions.
- GDPRthe right not to be subject to a fully automated decision (Article 22), and information about the logic involved.
- In practiceredacted supporting documents (statements, pay slips, tax notices) keep what the analysis needs (amounts, dates, categories) without the identity or account numbers.
Uses the AI Act prohibits, whatever the sector
Prohibited- Since 2 February 2025 (Article 5), prohibited practices include: subliminal manipulation or exploitation of vulnerabilities, general social scoring, assessing the risk that a person will commit an offense based solely on profiling, building facial recognition databases through untargeted scraping of images, emotion recognition in the workplace and in education, biometric categorization inferring sensitive data, and real-time remote biometric identification by law enforcement (with narrow exceptions).
- Penaltiesup to €35 million or 7% of worldwide annual turnover for prohibited practices.
- In practiceno anonymization makes a prohibited use lawful; Biffer serves permitted uses by removing what has no reason to circulate.
Sources: Regulation (EU) 2024/1689 on EUR-Lex, European Commission page on the AI Act. Timeline: prohibitions and AI literacy since 2 February 2025, general-purpose models since 2 August 2025, Annex III high-risk systems from 2 August 2026, Annex I regulated products in 2027.
The CNIL's recommendations on generative AI
In its Q&A on using a generative AI system (in French) published on 18 July 2024, the CNIL makes several recommendations to organizations.
- Deployment model: for personal data, an on-premises deployment is considered more appropriate and more secure; a cloud service requires a processing contract that sets out responsibilities and access.
- Usage policy: the organization defines permitted and prohibited uses and, depending on the deployment model, forbids entering certain confidential data.
- Training: users are trained on how these systems work and on their limits.
- Review: outputs are never reused as is, and their accuracy is checked.
For organizations that train or fine-tune a model, the CNIL's AI how-to sheets cover data selection and minimization (sheet 7) and secure development (sheet 12), among other topics. The CNIL's artificial intelligence section brings all these publications together.
Biffer fits these recommendations: processing happens on your computer, with no service provider; the document sent to the AI no longer contains the data your usage policy forbids; on-screen review and unredaction keep the final check in your hands. Biffer is published by OPCADIA, in Lyon, France, under the direction of Romain Rissoan, trainer and consultant. Practical questions are answered in the FAQ.
Frequently asked questions
Is a pseudonymized document still subject to the GDPR?
Yes. Under Recital 26 of the GDPR, pseudonymized data that could be attributed to a person by using additional information remains information about an identifiable person. With Biffer, the mapping table plays the role of that additional information: as long as you keep it, the redacted document stays within the scope of the GDPR. Pseudonymization is still a security measure explicitly cited in Articles 25 and 32: it sharply reduces what is exposed at the recipient's end, since the recipient only receives aliases.
What is the difference between pseudonymization and anonymization?
Pseudonymization replaces identifiers with aliases and keeps the means of returning to the original values separately, under protection: it is reversible for whoever holds the table. Anonymization makes the person unidentifiable by any means reasonably likely to be used, irreversibly; the GDPR then no longer applies. European authorities assess anonymization against three criteria: singling out, linkability and inference. A document redacted without a table, whose remaining text does not allow the person to be found, comes close to it; redacted images, whose pixels are blacked out, no longer hold the information.
Does professional secrecy prohibit using AI?
Article 226-13 of the French Criminal Code punishes the disclosure of secret information by the person entrusted with it; it does not target any particular tool. The question is therefore what content is sent. Passing the identifying content of a file to an AI service hosted by a third party amounts to disclosing it to that third party. A document from which the identity of clients, patients and parties has been removed, with the mapping table kept on the professional's own computer, lets you use AI for legal, accounting or medical reasoning without sending anything covered by secrecy.
Does the AI Act require anything from a ChatGPT or Claude user?
For everyday use, such as summarizing or rewriting a document, the organization using the tool is a deployer. Since 2 February 2025, Article 4 of Regulation (EU) 2024/1689 requires it to ensure a sufficient level of AI literacy among its staff. Obligations become heavier for the high-risk uses listed in Annex III, such as recruitment or creditworthiness assessment, from 2 August 2026. Personal data in the documents remains governed by the GDPR, and information covered by professional secrecy by criminal law.
Does Biffer guarantee GDPR compliance?
Biffer applies two measures the GDPR cites explicitly, pseudonymization and encryption, and does so on your computer, with no additional service provider. Whether a processing operation complies also depends on factors specific to your organization: legal basis, information given to the people concerned, choice of AI provider and the contract with it, retention period. Biffer handles document preparation, and the on-screen review lets you check every mask before sending. For an analysis of your situation, your data protection officer or a lawyer remains the right person to ask.
Your business documents anonymized, ready to be published or handed to an AI tool.
Biffer runs on your computer without sending a single file. Free for five files a month, unlimited for €10 excl. VAT per computer.
100% offline · Free: 5 files a month · No file sent