Professional secrecy in the age of AI: where are the red lines?

Generative AI did not change the rules of professional secrecy. It just made them harder to keep. Here is what the codes say, what the bars have started ruling, and where a tool like Hexagone AI actually fits.

1. The rule has not moved

Article 226-13 of the French Penal Code is short and old: revealing information of a secret nature, by a person who holds it through their state or profession, is punishable by a year in prison and a €15,000 fine. The Bar's internal rules (RIN) go further: secrecy is general, absolute and unlimited in time. It covers everything, consultations, correspondence, file notes, meetings, the names of clients, and the very fact that you advise them.

No technological innovation has carved out an exception. Sending a memo to a paralegal, a colleague, or to ChatGPT, all three move information. Only one of those three is governed by a duty of secrecy.

2. What changed: the destination

A consumer LLM is, by default, a third party with no professional duty toward your client. The provider can read prompts to improve models, retain them for abuse detection, and host them outside the EU. Even when training is opt-out, the prompt is still processed on infrastructure you do not control.

The Paris Bar reminded its members in late 2024: the use of generative AI tools requires the same caution as outsourcing any task. A specific authorization from the client is not in itself enough, because the lawyer remains personally responsible for the conditions under which the data leaves the firm.

3. The red lines

From recent disciplinary discussions and the most explicit deontology notes, four red lines are emerging:

• Pasting raw client names, addresses or amounts into a consumer chat interface. Even for a quick reformulation. It transfers the information to a non-bound third party.
• Uploading a contract or pleading verbatim to summarize it. Same issue, magnified by volume. A single PDF can hold dozens of identifiers.
• Letting a generic assistant access your inbox without filtering. The reasoning chain ends up seeing everything: client correspondence, draft strategy, fee discussions.
• Sharing case files inside a shared team workspace on a US-hosted SaaS. Hosting localization matters. CLOUD Act exposure is a real and documented question.

4. What ethics actually allow

Used carefully, AI is not banned. The codes have always tolerated the use of third-party tools, as long as secrecy is preserved. Three guardrails make the practice defensible:

• The data leaving your machine no longer identifies anyone, directly or indirectly.
• The transformation is auditable: you can show, after the fact, what was sent and what was kept.
• The mapping between markers and real values stays local, under your control.

5. Where Hexagone AI fits

Hexagone AI is not a chatbot. It is a layer that sits between your files and any AI tool you already use. Before a prompt or a file reaches Claude, ChatGPT or Cursor, identifiers are replaced with stable markers. The mapping stays on your machine. The AI receives a protected version, reasons on it, and you get the answer back in clear.

6. The honest disclaimer

No tool exempts a professional from their own judgment. The duty of secrecy remains personal. A protected workflow makes the line easier to hold, it does not erase the line. We publish a security page, a list of substitution categories and a sample audit trail precisely so your CISO, your DPO and your batonnier can form their own opinion.