Anonymize a document before sending it to ChatGPT or an AI

Why anonymize a document before ChatGPT?

A contract, a résumé, a medical report or a client file contains names, addresses, phone numbers and sometimes trade secrets. As soon as you paste this text into a general-purpose AI such as ChatGPT, Claude, Copilot or Gemini, this information leaves your computer and passes through a third party’s servers. You no longer know exactly where it is stored, for how long, or who at the provider may be able to access it.

Anonymizing or pseudonymizing the document before sending it (replacing “Marie Dupont” with “[PERSONNE_1]”, an address with “[ADRESSE_1]”) lets you use the AI to summarize, rephrase or analyze a text without exposing the identity of the people concerned. It is a simple precaution, it costs nothing, and it applies to personal use as much as to professional use subject to the GDPR.

What happens to your documents in ChatGPT, Claude or Copilot?

Policies vary by vendor and above all by the plan used. On free or paid consumer versions, conversations may often be used to improve the models, unless you turn this option off in the privacy settings. On business plans (API, ChatGPT Enterprise, Team with the right configuration, Copilot in a work context), providers generally say they do not use the data for training.

In both cases, the document has left your machine. It has been read, processed, and potentially logged for security or technical support purposes. If the provider suffers a breach, or if an employee accesses the logs for debugging, your confidential data is exposed. This is not an accusation against any particular vendor: it is the very nature of an online service, AI or not.

What should you remove from a document before ChatGPT?

The list changes with the type of document, but the same families of information always come up:

• People’s identity: last names, first names, nicknames, identifying initials.
• Contact details: postal addresses, email addresses, phone numbers.
• Unique identifiers: social security number, file number, customer number, IBAN.
• Organizations: the name of the company or department, sometimes even the city if it points to a specific site.
• Amounts: prices, salaries, damages, when they are enough to recognize a matter. Automatic tools do not always catch them: in AnonymFiles, you mask them by hand during review.
• Confidential content: financial clauses, medical information, industrial secrets (for example inventions not yet filed), passages covered by a non-disclosure agreement.

A good practice is to replace each occurrence with a stable placeholder rather than a simple “[confidential]”: if “Marie Dupont” consistently becomes “[PERSONNE_1]” throughout the document, the AI can still understand who does what in the text, which keeps the analysis useful while removing the real identity.

How do you anonymize a document for free, by hand?

If the document is short or you do not have a dedicated tool yet, manual anonymization is still possible:

1. Make a copy of the original file; never work on the only source version.
2. Find each piece of sensitive data with the search function of your word processor or PDF reader (Ctrl/Cmd+F on the known names, then a full read-through for the rest).
3. Replace with consistent placeholders: “[PERSONNE_1]”, “[PERSONNE_2]”, “[ORGANISATION_1]”, keeping the same placeholder for the same entity throughout the text.
4. If the document is a PDF with passages to hide rather than replace, use true redaction, as explained in how to redact a PDF properly: Adobe Acrobat Pro (the Redact tool), Preview on a Mac or LibreOffice Draw (Tools › Redact), then check the final file with a copy and paste. A simple black box drawn over the text is not enough: in most cases the text can still be selected underneath.
5. Review the whole document before sending it to the AI, including the metadata (file properties) and comments or revision notes, which may contain names.

These free tools do the job properly if you take the time to review. Their limit: everything is manual, which becomes long and error-prone as soon as the document runs beyond a few pages or several files need processing.

How do you anonymize a whole case file before ChatGPT?

By hand, anonymizing a thirty-document case file without missing a variant of a name takes hours. For a firm, the right tool is local software that processes the whole case file at once. AnonymFiles, French document anonymization software, coming soon for Apple silicon Macs and Windows, automates this step. It would be ironic to go through an online service to prepare what will be sent to an AI: once installed, AnonymFiles works offline, with no account, no online license check, no telemetry and no cloud AI chat (the security page details what the app does not do). On a Mac, the detection models are included; on Windows, they are downloaded once, at the first processing run, without any document text. No document leaves your computer until you decide yourself to send a copy.

How it works: you drop in an entire folder (text or scanned PDFs, Word .docx, emails, spreadsheets, presentations, images, .txt, .md); the app detects people, organizations, addresses, email addresses, phone numbers, IBANs, private dates and ID numbers, then replaces them with stable placeholders such as [PERSONNE_1] or [ORGANISATION_1], identical from one file to the next. The copy of a Word document is a redacted .docx that keeps its layout, without comments or tracked changes, along with a text file for the AI. You then compare each original with its copy side by side, redact by hand what automatic detection may have missed (amounts in particular), approve document by document, then export clean copies with neutral file names. The mapping table between placeholders and identities never leaves with the copies: it stays on your computer and is used to put the real names back into the AI’s answer, retrieved as text, Word or PDF. Scanned PDFs and images are read by local OCR; for a single scan, without the app, see our methods for scanned documents.

The exported copies are then ready to paste into ChatGPT, Claude or another assistant: the text keeps its structure and logic, without the masked identities. An MCP server is also announced for Claude and other agents: it will expose only the redacted copies to them. Still, review the copies before handing them to an AI: no automatic detection is perfect.

Confidential data and professional secrecy: what can you put into ChatGPT?

Confidential data should not be pasted into ChatGPT as is: not clients’ or employees’ personal data, not trade secrets, not documents covered by professional secrecy. You can, however, use an anonymized or pseudonymized copy in which nothing identifies a person or a matter anymore.

• Personal data: the GDPR requires sending only what the task needs (Article 5). To summarize a contract, the AI does not need the parties’ names. OpenAI’s plans are detailed in ChatGPT and GDPR: can you put client data into it?
• Professional secrecy: in France, lawyers, notaries, chartered accountants, commissaires de justice, insolvency practitioners and patent attorneys are bound by secrecy (article 226-13 of the French Criminal Code). The AI guide of the Conseil national des barreaux (CNB, the French national bar council) recommends anonymizing or pseudonymizing first. See AI and professional secrecy, profession by profession.
• Trade secrets and inventions: a contract under an NDA or an unfiled invention must not go out as is. For a contract, follow the checklist to anonymize a contract before ChatGPT.

For a firm, the right tool is local anonymization software, chosen by profession: document anonymization software for lawyers, for notaries, for chartered accountants, for in-house counsel and DPOs.

Special cases: code, spreadsheets, emails, contracts

A contract or a court document raises a specific problem: the parties’ names often come back dozens of times, including in indirect phrasings (“the landlord”, “the selling company represented by Mr. X”). Check that all variants of the same name are covered by the same placeholder. We explain how to anonymize a client file before AI as a lawyer, and the same logic applies to a notary’s office’s draft deeds and preliminary sale agreements.

A spreadsheet exported as text for an AI must be checked column by column: an entire “Client” or “Email” column can go unnoticed if you only review the first rows. This is typically the case for an accounting firm’s client files. AnonymFiles reads Excel and CSV files, one table per sheet, to be reviewed the same way.

An email pasted into an AI almost always contains a signature with name, position, phone number and sometimes an address: remember to remove it, not just the body of the message.

Source code can contain API keys, hard-coded passwords or internal domain names: these are not personal data under the GDPR, but they are secrets that should not end up in an AI provider’s logs either.

Which mistakes should you avoid?

• Do not ask the AI itself to anonymize the document: by then, the confidential data has already been sent to it, which defeats the purpose.
• Do not stop at the first page: lists, schedules and attachments often contain the most sensitive data.
• Do not reuse an “anonymized” file without checking the metadata (author, revision history), which may contain a real name even when the body of the text is clean.
• Never consider automatic processing 100% reliable: a human review is still needed before anything is sent.

To go further on the exact meaning of the terms used here, the difference between anonymization and pseudonymization under the GDPR is worth knowing before choosing your level of precaution. If the document is part of a court file, also see our guide to anonymizing a judgment or court documents. And for the general method for a PDF, see how to anonymize a PDF for free.