Anonymization glossary: the terms worth knowing before handing a document to an AI

Alias

A replacement value that software substitutes for a piece of personal data in a document. Biffer produces stable, numbered aliases such as PERSONNE_001 or EMAIL_002: the same person always gets the same alias throughout the file, which keeps the text readable and consistent for an AI. The alias fully hides the original value; only the mapping table, if you keep it, can bring it back.

Anonymization

Processing that makes it impossible, irreversibly and by reasonable means, to identify a person from a document. Truly anonymized data falls outside the scope of the GDPR. Anonymization is demanding: beyond names, the context itself must keep people from being recognized. Biffer comes close to it when you work without a mapping table; an analysis of your context lets you reach a conclusion. See anonymize a document.

Biffage (striking out)

The act of "biffer", French for striking out or crossing out a passage of a text to make it illegible or remove it. The word comes from legal and administrative usage, where a clerk strikes out an entry on a deed. We chose it as the name of the software because it describes the operation exactly: removing from a document what must be kept from a third party, here an artificial intelligence, while keeping the rest.

CNIL

Commission nationale de l'informatique et des libertés, the French data protection authority. It publishes recommendations, practical guides and fact sheets, including on the use of artificial intelligence and on the difference between anonymization and pseudonymization (in French). It can audit organizations and impose penalties. For French organizations, its publications are the first source to consult before handing documents to an AI tool.

Confidentiality

The property of information that only authorized people can access. For business documents, confidentiality is a contractual, ethical or legal obligation depending on the profession. Local processing, on your own computer, preserves confidentiality because the document stays within the perimeter you control, before anything is sent to an online AI service.

Data breach

The disclosure, deliberate or not, of data to people who should not have access to it. Under the GDPR, a personal data breach that presents a risk must be notified to the supervisory authority (the CNIL in France) within 72 hours. Pasting a confidential document into an online AI assistant can amount to such a disclosure, even without any attack; preparing it before sending protects you from that.

Data controller

The person or organization that determines the purposes and means of processing personal data. A firm, a company or a self-employed professional is the controller of the processing it carries out on its clients' documents, including when it uses an AI tool. That responsibility remains its own, including toward the tool's publisher: hence the value of controlling what leaves the organization.

Data minimization

A GDPR principle (Article 5) under which only data that is adequate, relevant and limited to what is necessary for the purpose may be processed. Applied to AI, it means sending an assistant only what it needs to answer: the structure of a contract rather than the identity of the parties. Anonymizing a document before submitting it is a direct application of this principle.

EDPB

The European Data Protection Board, the body that brings together the supervisory authorities of the European Union, including the CNIL. It publishes guidelines and opinions that harmonize how the GDPR is applied. Its Opinion 28/2024 deals with the status of personal data in AI models and the conditions under which a model can be considered anonymous.

Encryption

Transforming data into a form that cannot be read without the matching key. Biffer offers to encrypt the mapping table with AES-256-GCM, an authenticated encryption algorithm, with the key derived from your password by Argon2id, a function designed to resist brute-force attacks. Your password is then the only way back to the original values, even if the file is copied.

False positive

An item wrongly detected as personal data. A common noun taken for a surname, a product reference taken for a case number, or a date unrelated to any person are examples. Biffer shows every detection in a highlighted preview; you simply untick the false positive and it is kept as is. The opposite, a missed item, is fixed by adding it manually.

GDPR

The General Data Protection Regulation, Regulation (EU) 2016/679, in application since 25 May 2018. It defines personal data (Article 4), requires data protection by design (Article 25) and security of processing (Article 32), and governs the use of processors (Article 28). It applies to any use of AI on documents containing data about people located in the European Union.

Generative AI

A family of artificial intelligence systems that can produce text, images or code from a prompt. ChatGPT, Claude and Gemini are the most widely used examples. These tools are useful for summarizing, rewriting or structuring a document, and most of them run on remote servers: preparing the document before sending it lets you benefit from them while keeping personal data on your computer.

IBAN

International Bank Account Number, the standardized identifier of a bank account. A French IBAN starts with FR, followed by two check digits and then the national account identifier. The two check digits are verified with a modulo 97 calculation, which lets Biffer recognize a genuine IBAN and tell bank details apart from an ordinary string of digits.

Identifying data

Data that, alone or combined, leads back to a specific person. There are direct identifiers (name, social security number, email address) and indirect identifiers (job title, employer, date, place) that identify by cross-referencing. Biffer detects direct identifiers and the fields specific to each profession; indirect identifiers are for you to review, because whether they identify someone depends on the context.

License key

The code that allows Biffer to run on a computer. The key is tied to your email address and to your computer's ID. It is issued in the software through the one-time email sign-in at first launch, either free (5 files per month) or after buying Biffer (€10 excl. VAT (€12 incl. VAT), lifetime for that version, unlimited files), and is verified locally: once received, the software runs fully offline. One license covers one computer.

Local processing

Running a program on the user's own computer, as opposed to processing on a remote server. Local processing ensures data stays on the machine, away from any network and any third party. This is how Biffer works: analysis, replacement with aliases, rewriting the file and license verification all happen on your machine, which settles the processor question from the outset.

Luhn check

A checksum formula used to verify that a number is consistent, including French company numbers (SIREN and SIRET) and payment card numbers. The last digit is calculated from the previous ones using a simple algorithm. Biffer runs this check before treating a string of nine or fourteen digits as a SIREN or SIRET, which sharply reduces false positives on amounts and references.

Machine ID

A code specific to a computer, shown in Biffer, used to link a license key to that machine. It only serves to recognize the computer; your documents and your identity have nothing to do with it. The software sends it with your email address during the initial sign-in, and the key received is tied to it; if you change computers, you sign in again on the new machine or enter the new ID in your Biffer account, and a new key is generated.

Mapping table

An optional file that links each alias to the original value it replaces. It lets you restore the real names and numbers in the document returned by the AI or by a third party. Biffer can encrypt it (Argon2id and AES-256-GCM) or save it unencrypted, as you prefer. Without a table, the processing is permanent: this is the point that tips pseudonymization over into anonymization. See how it works.

Masking

Replacing a piece of data with a generic label that shows its category without numbering it, for example [PERSONNE] or [EMAIL]. It is one of the three policies Biffer offers for each data type, alongside pseudonymization with aliases and keeping the data. Masking suits tasks where only the category of the data matters for the AI, such as a general summary.

Metadata

Information that describes a file and is stored outside its visible text: author, company, creation date, software used, authors of comments and tracked changes. It travels with a Word, Excel or PDF document and often reveals who wrote it. Biffer clears these fields during processing, so that anonymizing the body of the document extends to its properties.

NIR (French social security number)

The French national registration number for individuals, made up of thirteen digits and a two-digit check key. It encodes sex, year and month of birth, department and municipality. Its processing is strictly regulated. Biffer recognizes it by verifying the check key, which distinguishes a real NIR from any other fifteen-digit string.

OCR

Optical Character Recognition, a technique that converts an image of text into usable text. A scanned PDF is a series of images: OCR makes it possible to read a name in it. Biffer includes an OCR engine that runs on your computer, offline, for images and scanned PDFs; the data found is then redacted by blacking out the pixels, permanently.

Occultation

The French term, used in particular by courts and public authorities, for making part of a document invisible in published decisions and records. Removing the names of the parties from court decisions is the best-known example. In a digital file, reliable occultation means removing the text from the file's actual content, beyond how it is displayed on screen.

Offline

Describes software that works without an internet connection and keeps all data on the computer. Biffer is offline for all processing: it analyzes, modifies and saves a document with the network off, the license is verified locally and the software uses no telemetry. Its only connection is at first launch, to confirm your email and issue the key. You can check this by turning off the network afterward. See offline software.

Personal data

Any information relating to an identified or identifiable natural person, according to Article 4 of the GDPR. A name, an email address, a phone number, a license plate, a professional identifier or a combination of details that make someone recognizable all qualify. The notion is broad: a date of birth combined with a town can be enough to identify a person within a small group.

PII

Personally Identifiable Information, the term mostly used in the United States for information that can identify a person. It is close to the GDPR notion of personal data, but narrower in some uses: US practice focuses on direct identifiers, whereas European law includes everything that makes a person identifiable, including by cross-referencing.

Processor

A person or organization that processes personal data on behalf of the controller, within the meaning of Article 28 of the GDPR. An online anonymization service that receives your documents on its servers is a processor and must be bound by a contract. Biffer runs on your computer and keeps every piece of data on it: the processor question is settled upstream, since everything stays with you.

Professional secrecy

The legal obligation not to disclose information entrusted in the course of a profession. Lawyers, doctors, notaries, accountants and others are bound by it, with criminal penalties for breaches. Sending a file to an online service, including an AI assistant, can amount to disclosure. Processing the document locally, before anything is sent, is a precaution consistent with this obligation.

Pseudonymization

Processing that replaces identifying data with substitutes, so that the information can no longer be attributed to a person without additional data kept separately, according to Article 4 of the GDPR. This is exactly what Biffer does with its aliases and its mapping table. Pseudonymized data remains personal data as long as the table exists, but the risk is greatly reduced.

Queue

A feature for processing several files in a row. In Biffer, the queue runs in reviewed mode, where each document goes through the preview before being saved, or in automatic mode, where the policies you have set apply without intervention. Automatic mode, included even in the free plan, suits batches of similar documents, such as a series of invoices or agreements.

Redaction

Removing passages from a document, historically with a black box, to hide their content before release. Reliable redaction works on the actual content of the file, beyond simply covering it on screen. Biffer rewrites the file's content and replaces each piece of data with an alias or a label so that the original value truly disappears from the output document. See redact a document.

Re-identification

Recovering a person's identity from data that is supposed to no longer point to them. It can go through the mapping table, which is intended and controlled, or through cross-referencing clues left in the document, which is not. The risk of re-identification is the criterion that separates pseudonymization from anonymization, and it is assessed case by case.

Sensitive data (Article 9)

Special categories of data whose processing is prohibited in principle by Article 9 of the GDPR: racial or ethnic origin, political opinions, religious beliefs, trade union membership, genetic and biometric data, health data, sex life and sexual orientation. A medical file, an HR report or a litigation file often contains them. Their presence calls for extra care before anything is sent to an AI.

SIREN and SIRET

French company identifiers. The SIREN has nine digits and designates the legal entity; the SIRET, fourteen digits, adds five digits that designate a specific establishment. Both are verified with the Luhn check. They identify businesses; they also directly designate a sole trader and often make a client recognizable; Biffer detects them and handles them like other identifiers.

Training data

The set of texts, images or recordings used to build or improve an artificial intelligence model. Depending on the terms of service, content sent to an online assistant may end up in this data. Removing personal data before sending keeps your clients' names out of it, whatever the provider's policy.