Personal data and AI: what to remove before using an assistant
Under Article 4 of the GDPR, personal data is any information relating to an identified or identifiable natural person, directly or through cross-referencing. Personal data and AI can coexist as long as the document is prepared before it leaves your computer. A contract, meeting minutes, an insurance claim file or a payslip dropped into an online assistant becomes data processed by a third party. The GDPR applies to that upload just as it does to any other.
Assistants are genuinely useful; on this page we help you know what counts as personal data, what the CNIL (the French data protection authority) and the European Data Protection Board say about it, and how to remove this information from a document before sending it, with Biffer or otherwise. The goal is concrete: reduce what goes out at the source, the measure most directly within your reach.
What is personal data in a document?
Key points
- Personal data is any information relating to an identified or identifiable person, directly or through cross-referencing (GDPR, Article 4).
- Sending a document to an AI assistant is processing: it needs a legal basis, a purpose and safeguards.
- The CNIL and the EDPB accept the use of AI models, provided the data transmitted is limited and choices are documented.
- Pseudonymizing or masking data before sending, on your computer, reduces what goes out without stopping the assistant from working.
Article 4 of the GDPR (Regulation (EU) 2016/679) defines personal data as any information relating to an identified or identifiable natural person. The key word is "identifiable": the notion goes far beyond a name. An address, a date of birth, a case number, a license plate, a job title in a small organization, or a combination of several of these, is enough. The text is available on EUR-Lex (Article 4 of the GDPR). The American term PII (personally identifiable information) covers roughly the same idea, the European notion being broader.
In a business document, you mainly come across:
- Names: last name, first name, title, signature, the name of a relative mentioned in the text.
- Contact details: postal address, email, phone number.
- Identifiers: social security number, IBAN, French company number (SIREN or SIRET) of a sole trader, license plate, case, policy, claim or contract number, employee ID, lawyer's bar locker number (toque), RPPS, CRPCEN.
- Special categories of data under Article 9: health, ethnic origin, opinions, trade union membership, sexual orientation, biometric data, whose processing is prohibited in principle, with exceptions.
- Confidential information that is not personal: amounts, clauses, trade secrets, which fall under professional or contractual confidentiality rather than the GDPR.
GDPR and ChatGPT: what happens when you upload a file
When you drop a document into ChatGPT, Claude, Gemini or another online assistant, three things happen from the regulation's point of view. First, a transfer: the file is sent to the provider's servers, which may be outside the European Union, bringing in the rules of Chapter V of the GDPR. Second, processing by a third party: the provider becomes a processor, or a joint controller depending on the terms of use, and you need a suitable contractual framework (Article 28). Finally, depending on the plan and settings, retention and sometimes reuse of the content to improve the service.
These three points make the use possible, under precise conditions. You need to be able to answer the following questions:
- What is the purpose of the upload (summarizing, translating, drafting, analyzing), and what is its legal basis?
- Have the people concerned been told that their data may be processed by an AI provider?
- Does the provider offer the required contractual guarantees, and where are its servers located?
- Does the document sent contain more personal data than the purpose requires?
The last question is the one that depends most on you, and it is the one a tool like Biffer helps with. To summarize a contract or rephrase meeting minutes, an alias (PERSONNE_001) gives the assistant everything it needs for a coherent answer; Marie Dupont's name and address stay on your computer. That is data minimization (Article 5 of the regulation) put into practice. We describe this use on the pages anonymize before ChatGPT, anonymize before Claude and anonymize before Gemini.
What the CNIL and the EDPB say
Since 2023, data protection authorities have published a set of landmarks on generative AI.
The CNIL has put online a section devoted to artificial intelligence, with recommendations for developing AI systems and practical guidance sheets, notably on building training databases and informing people (the CNIL's artificial intelligence section, in French). It points out that GDPR principles apply at every stage, from collecting training data to using the model, and that pseudonymizing and anonymizing datasets are among the measures expected. The criteria for effective anonymization are set out in the reference Article 29 Working Party opinion on anonymisation techniques, which the CNIL's own guide follows.
On 17 December 2024, the European Data Protection Board (EDPB) adopted its Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models. It addresses three questions: when an AI model trained on personal data can be considered anonymous, how legitimate interest can serve as a legal basis for development and deployment, and what consequences unlawful training has on the later use of the model. The Board states that a model's anonymity is assessed case by case, and that anonymized or pseudonymized training data is among the measures that weigh in that assessment.
For someone working with documents, the practical consequence is twofold: choose an AI provider whose terms are compatible with data protection, and reduce at the source the personal data you send, the measure most directly within your reach.
Removing personal data before sending: the method
Biffer is desktop software, for Mac and PC (Windows), designed by Romain Rissoan, a trainer and consultant based in Lyon, France. It works offline: the document is analyzed, reviewed and rewritten on your computer, and your files and every fragment of text stay on your machine. This choice is explained on the offline software page.
1. Identify
The software highlights people, email addresses, phone numbers, addresses, French company numbers (SIREN and SIRET), French social security numbers, IBANs, license plates, dates and profession-specific fields. You untick false positives and add whatever is missing, for all occurrences at once.
2. Decide
For each data type: pseudonymize (a stable alias such as PERSONNE_001), mask ([PERSONNE]) or keep. Create a mapping table or not, encrypted (Argon2id and AES-256-GCM) or in plain text.
3. Rewrite
The text is replaced in place in the PDF, Word, Excel, CSV or text file; formatting is kept and metadata (author, company, comment authors) is cleared.
4. Send, then restore
The pseudonymized document is submitted to the assistant. If you kept the table, the original values are restored in the answer, on your computer.
What this method guarantees: the data you removed stays on your computer, and the file you send no longer contains it. What remains in your hands: reviewing for cross-referencing, choosing a provider that offers guarantees, and informing the people concerned. The distinction between pseudonymization and anonymization, and what it means under the GDPR, is covered on the page anonymize a document.
Frequently asked questions
Can I use ChatGPT with personal data and stay GDPR compliant?
Yes, provided you treat this use like any other processing: a defined purpose, a legal basis, informing the people concerned, a provider offering sufficient contractual guarantees, and minimizing the data sent. In practice, the simplest lever is to send only what is necessary: a summary or a rephrasing works very well with aliases in place of names, addresses, social security numbers or IBANs. Pseudonymizing the document before sending, then restoring the values in the answer, lets you work with the assistant while greatly reducing the identifying share of what leaves your computer.
What is the difference between personal data, sensitive data and confidential information?
Personal data is any information relating to an identifiable natural person: name, contact details, identifiers, but also indirect elements such as a date tied to a place. Sensitive data, or special categories of data, is a subset defined in Article 9 of the GDPR: health, ethnic origin, political or religious opinions, trade union membership, sex life, genetic and biometric data. Processing it is prohibited in principle, except in limited cases. Confidential information, for its part, can concern something other than a person: amounts, clauses, processes, trade secrets. It falls under professional confidentiality or contract law rather than the GDPR.
Is anonymized training data enough to make an AI model anonymous?
It is one of the measures that count, among others. EDPB Opinion 28/2024 states that a model's anonymity is assessed case by case, by checking that the model resists, with reasonable means, the extraction of personal data and queries aimed at obtaining it. Anonymizing or pseudonymizing training data is one of the measures taken into account, along with the training method, output filters and documentation. For most professionals, the question is more modest: it is about using an assistant rather than training a model. Reducing personal data in the documents submitted then remains the most effective measure and the easiest to demonstrate.
Does Biffer guarantee that my use of AI is GDPR compliant?
Biffer contributes directly, and you stay in control of the whole. It removes from the document, on your computer, the personal data it detects and you approve, while keeping it readable thanks to consistent aliases. This is a pseudonymization measure, encouraged by Articles 25 and 32 of the GDPR, and a concrete application of data minimization. It comes on top of analyzing the legal basis, informing the people concerned, choosing an AI provider that offers guarantees and reviewing the document produced, which remain in your hands: you are the data controller, and Biffer gives you the means to act as one.
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