Patents and ChatGPT: avoiding disclosure of an unfiled invention

What kind of disclosure destroys novelty?

An invention is new if it doesn't form part of the state of the art. Under article L. 611-11 of the French Intellectual Property Code (Code de la propriété intellectuelle, CPI), the state of the art comprises “everything made available to the public before the filing date of the patent application”, by “a written or oral description, by use, or in any other way” (our translation). Article 54 of the European Patent Convention (EPC) uses the same definition.

• The date. It is the filing date. For a European application that claims a priority, the priority date takes its place (art. 89 EPC).
• The exceptions. The EPC provides only two. A disclosure made within the six months preceding filing is not taken into consideration if it results from an evident abuse in relation to the applicant, or from an official exhibition (art. 55 EPC).
• The criterion. According to the EPO Guidelines for Examination, a document is available to the public if members of the public could gain knowledge of it, with no confidentiality obligation restricting its use or dissemination (G-IV, 1). Conversely, subject matter is not made available by use or in any other way if secrecy was agreed, expressly or tacitly, and has not been breached (G-IV, 7.2.2).

Is describing an invention to ChatGPT a disclosure?

The texts don't say. Neither article L. 611-11 of the CPI, nor article 54 of the EPC, nor the passages of the Guidelines cited above mention prompts sent to an AI. The question isn't settled by the texts; caution is called for.

The Guidelines' criterion only indicates what would need to be examined: who receives the text, and under what confidentiality obligation. Yet, according to the CNCPI's guide, the protection of data sent to an online AI relies heavily on trust in the provider, which alone sets its contractual terms. Information communicated may be used for training, then returned, directly or indirectly, to other users (p. 6). The guide speaks of a risk of disclosure in the confidentiality sense; it doesn't rule on novelty.

In practice, before filing: no description of the invention and no draft claims in an online AI. This isn't a written rule, but the cautious consequence of an open question.

Can a French patent attorney use ChatGPT despite professional secrecy?

Under article L. 422-11 of the CPI, “in all matters […], the conseil en propriété industrielle observes professional secrecy” (our translation). This secrecy covers consultations, correspondence with the client, a colleague or a lawyer (except correspondence marked “officielle”), meeting notes and “all documents in the file”. A draft application or a note on filing strategy falls within it.

Filing doesn't make the application public. The EPO in principle publishes the European application as soon as possible after eighteen months from the filing or priority date, or earlier at the applicant's request (art. 93 EPC).

What does the CNCPI's Guide to good practices for the use of generative AI (July 2026) recommend?

The Compagnie nationale des conseils en propriété industrielle (CNCPI), the French national association of patent and trademark attorneys, published in July 2026 a guide to good practices for the use of generative AI (Guide des bonnes pratiques pour l'usage des IA génératives). These are suggestions: each patent attorney can go further and remains individually responsible before the law (p. 1). What it recommends:

• No confidential information in prompts, because of the risk of disclosure, especially in a consumer tool (pp. 7 and 8).
• Pseudonyms and codes. Replace proper names with pseudonyms and “industrial property title numbers with codes” (our translation); regularly clean up the prompt history (p. 7).
• Inform the client and obtain their prior consent, in particular at the start of the relationship, and agree with them on the tasks concerned: drafting, official letters, translations (p. 4). Consent is sought again if needed, for example when the contact person changes or the tool evolves. Even with clear consent, the patent attorney remains vigilant about the risk of inadvertently disclosing confidential information (p. 10).
• Leave sensitive subjects aside. No generative AI for subjects classified as national defense or restricted, nor more broadly for sensitive subjects (p. 7). Data exported outside France may also breach national defense rules as long as authorization to disclose hasn't been obtained (p. 3).

Consumer AI or professional plan: what to check

According to the CNCPI's guide, free consumer generative AI services may, by default, record prompts and train on them, and the provider holds a copy of the information communicated (p. 6). The guide recommends checking and configuring the settings, checking whether the provider has the right to store, monitor and review prompts, and preferring a private cloud, ideally in France or in Europe (p. 7).

We don't summarize any provider's terms: they vary with the plan and the settings, and only those of your plan count. Check five points in them: training on your data, retention period, access by the provider's staff, location and subprocessors. A professional plan changes the contract, not the route: with an online service, the text leaves the firm.

What pseudonymization covers, and what it doesn't

Pseudonymizing means replacing each identity with a stable label, such as [ORGANISATION_1], while keeping the mapping at the firm: this is the approach the CNCPI suggests. It isn't anonymization (see the difference between anonymization and pseudonymization under the GDPR).

No detection is perfect: each document is reviewed next to the original.

What it doesn't cover: the technical content. A described invention stays the same, whether the applicant appears under their real name or as [ORGANISATION_1]. A technical passage masked in one click becomes invisible to the AI: useful for a letter, useless for working on claims. Decide, task by task, what the AI should see. Before filing, caution means not showing it the invention.

Watch out for cross-referencing. A product name, an earlier application, a rare combination of features or a priority date can be enough to find a file. The CNCPI's guide also notes that the weight given to cited prior art may reflect the client's instructions, which are confidential (p. 10).

The method with AnonymFiles

AnonymFiles is a desktop app, coming soon for Apple silicon Macs and Windows, that prepares these copies on the firm's computer.

1. Drop in the file. Word (.docx), text or scanned PDF (50 MB at most), emails, spreadsheets, presentations, images, .txt or .md. Scans are read by local OCR. The copy of a Word file is a redacted .docx that keeps its layout, without comments, tracked changes or images.
2. Let detection do its work. Names of people, organizations, addresses, emails, phone numbers (French format), IBANs (nine European countries), private dates and ID numbers become stable labels, identical across all documents.
3. Review and complete. Each document is shown next to the original. Select application numbers, amounts and technical passages: each selection becomes a label, applied to the whole file. A label can be renamed, for example [PERSONNE_CLIENTE].
4. Approve, then export. Nothing is exported without approval. The export creates a new folder: copies with labels (for a text PDF, a PDF with the same layout, with labeled black blocks; for a Word file, a redacted .docx with the same layout), masked content truly removed, metadata, annotations, forms and links removed, neutral file names, an index of exported and excluded documents. The mapping table stays on the computer, never in the export.
5. Work with the AI, then put the names back. “Copier la consigne IA” (copy the AI instruction) gives the instructions to paste into the conversation: keep the labels, recall the [DOSSIER_…] reference. “Réinjecter un résultat” (re-inject a result) puts the real names back into the AI's answer, as text, Word or PDF, on the computer, with a preview before saving.

The app works with no account, no telemetry and no online license check. On Mac, the detection models are included: no connection at all. On Windows, they are downloaded once, on first use, without any document text (see security and network perimeter). For details, see how it works for patent attorneys and document anonymization software for patent attorneys.

Three prompts on a pseudonymized copy

These examples assume a copy that has been prepared and reviewed, the client's consent for this type of task, and the AnonymFiles instruction pasted at the start of the conversation. None of them requires showing the invention. General method: how to anonymize a document before ChatGPT or Claude.

Rephrasing a letter to the client

Preparing the reply to an office action

An office action often quotes the claims. If the application isn't published, these passages remain confidential: mask them, or only send the useful part.

Translating a chosen passage

Only paste the passage to translate. An online translator may also rely on generative AI: the CNCPI's guide cites DeepL among the tools that integrate it without the user necessarily knowing (p. 2).

Proofread each answer: if you can't control, check and correct the results, the guide advises refraining (p. 10). Then put the real names back on your computer.