15 legal AI tools in the French bar's AI grid: 9 declare nothing on pseudonymisation
France's National Bar Council (CNB) summed up fifteen legal AI tools in a ten-field form. We read the fifteen forms: 9 of 15 declare nothing on pseudonymisation or anonymisation, and 14 of 15 name a US company among those hosting the language model.
- 9 of 15 legal AI tools summed up in the CNB's AI grid declare nothing on pseudonymisation or anonymisation: 8 forms read "N/C" (not communicated) and 1 "N/A".
- 5 of 15 declare a mechanism: 3 pseudonymisation, 2 anonymisation. One more answers "no", with short retention instead.
- 14 of 15 name a US company among those hosting the LLM; the fifteenth runs on the user's machine.
- Source: the CNB's Grille de lecture, June 2025, v1.2, summary forms on pp. 29–43. The count is ours (24–25 October 2026), first published in our report State of AI confidentiality 2026. It is not a ranking.
Disclosure: this article is published by Tardigrade SARL, which sells Hidden Tokens, a pseudonymisation tool. The figures below are a count of what the CNB published, not a test of the tools. It is not legal advice.
The findings
The CNB's AI working group published its Grille de lecture (1st edition, June 2025, v1.2): the criteria a firm should use to choose a legal AI tool, and a ten-field summary form ("fiche de synthèse") for each of fifteen tools it consulted. We read the fifteen forms and transcribed two fields: the company hosting the LLM, and pseudonymisation or anonymisation.
- 9 of 15 forms declare nothing on pseudonymisation or anonymisation. 8 read "N/C" (not communicated) and 1 "N/A" (not applicable). #
- 5 of 15 declare a mechanism (3 pseudonymisation, 2 anonymisation), and 1 answers "no" with short retention instead (LexisNexis). #
- 14 of 15 forms name a US company among those hosting the LLM. 11 name only US companies; 3 name a US company alongside a French model or an on-premises installation (Jimini, Lexbase, Ordalie). The fifteenth, Cedie, runs on the user's machine. #
- The grid excludes US hosting as a sovereignty criterion. Its second and fourth criteria exclude any company whose national law allows its authorities to requisition data stored on servers owned by companies of that nationality, "such as US companies under their current legislation". #
- Every tool says it does not train on client data; the grid says that cannot be checked. All tools audited say they apply the no-reuse prerequisite, but the grid notes that such a statement from a proprietary model is "in general neither provable nor verifiable". #
- The bar now recommends pseudonymising first. The CNB's ethics guide on AI (April 2026) recommends to "systematically pseudonymise or anonymise data before any use of a generative AI tool", while warning that pseudonymisation does not fit every task and that people can be re-identified by cross-referencing. #
Tool by tool: what the fifteen forms declare
The wording is the CNB's, translated; "N/C" means not communicated, "N/A" not applicable. Tools are listed in the grid's order.
| Tool | Company hosting the LLM | Pseudonymisation / anonymisation |
|---|---|---|
| Doctrine | US (Microsoft) | Pseudonymisation |
| Lefebvre Dalloz | US (Microsoft) | N/C |
| Jimini | French (Mistral) and US (Microsoft) | Pseudonymisation |
| Juri'Predis | US (Microsoft) | Pseudonymisation |
| Jus Mundi | US (Microsoft) | Anonymisation |
| Lamy Liaisons | US (Microsoft) | N/C |
| Legora | US (Microsoft, AWS and Amazon Bedrock) | N/C |
| Lexbase | US (Microsoft) and on premises at Lexbase | Anonymisation |
| LexisNexis | US (Microsoft, AWS and Amazon Bedrock) | No, but deletion after a few minutes or days |
| Notae | US (Google) | N/C |
| Ordalie | US (Microsoft) and on premises at Ordalie | N/C |
| Septeo | US (Microsoft) | N/C |
| Cedie | N/A (runs on the user's machine) | N/A |
| Predictice | US | N/C (may be specified in its Trust Center) |
| Haiku | US | N/C |
Source: CNB, Grille de lecture, June 2025, v1.2, summary forms (pp. 29–43 of the PDF). Our transcription and count, 24–25 October 2026. Same table as in section 2 of State of AI confidentiality 2026.
Method
- Source. The CNB AI working group's Grille de lecture — synthèse des consultations…, 1st edition, June 2025, v1.2, as published on cnb.avocat.fr (French). The fifteen summary forms are on pages 29 to 43 of the PDF.
- What we did. We read each form and transcribed two of its ten fields: the company hosting the LLM, and pseudonymisation or anonymisation. We kept the CNB's wording, translated, and counted.
- Counting rule. A form "declares nothing" when the field reads "N/C" or "N/A". "No, but deletion after a few minutes or days" (LexisNexis) is counted as an answer, not as silence.
- Dates. CNB grid: June 2025. Our transcription and count: 24–25 October 2026. First publication: State of AI confidentiality 2026, edition 1, 25 October 2026. This article: 6 November 2026, with the same figures.
Limits
- This is not a ranking. The forms reflect what vendors declared to the working group in 2025; the grid says so itself: "subject to change and depending on what the companies declared, found and/or sent". Several vendors have published more since.
- A declaration is not a test. We did not test any tool. "N/C" says the form is silent, not that the tool sends names in clear.
- Our count, not the CNB's. The CNB published the forms, not these totals.
- Conflict of interest. We sell a pseudonymisation tool. Our own answers to the CNB's ten fields, with their limits, are in the CNB compliance file; they are not part of this count.
Products and vendor statements change. If a vendor's position has moved since the CNB's forms, write to info@hiddentokens.com and we will add a dated note here.
How to cite
Hidden Tokens, "15 legal AI tools in the French bar's AI grid: 9 declare nothing on pseudonymisation", 6 November 2026, https://hiddentokens.com/blog/french-bar-ai-grid-legal-ai-tools-pseudonymisation/ — data first published in Hidden Tokens, State of AI confidentiality 2026, edition 1, 25 October 2026, section 2, https://hiddentokens.com/state-of-ai-confidentiality-2026/#regulators
To cite one finding, link to its anchor, for example …/french-bar-ai-grid-legal-ai-tools-pseudonymisation/#finding-1. Please also cite the primary source: CNB, Grille de lecture, 1st edition, June 2025, v1.2. A French version of this article is available: 15 outils d'IA juridique passés à la grille du CNB.
More from the same report
The full report, State of AI confidentiality 2026, also covers what professional secrecy, the GDPR, Moroccan law 09-08 and the EU AI Act ask, how well generic detectors catch French and Moroccan identifiers (section 3), a market map, and a 10-point checklist, starting with: ask each vendor the CNB's ten questions, whatever your profession. Its key figures each have their own link.
Frequently asked questions
Did the CNB publish these figures?
No. The CNB published the fifteen summary forms; the counts (14 of 15, 9 of 15, 5 of 15) are our own, made from the CNB's PDF on 24–25 October 2026. Each count can be checked against the table on this page and against pages 29 to 43 of the CNB's grid.
Does "N/C" mean the tool does not pseudonymise?
No. "N/C" means not communicated: the form says nothing either way. The forms reflect what vendors declared to the CNB's working group in 2025, and several vendors have published more since. Ask the vendor directly, and date the answer you get.
Which tools declare pseudonymisation or anonymisation?
In the CNB's forms: Doctrine, Jimini and Juri'Predis declare pseudonymisation; Jus Mundi and Lexbase declare anonymisation. LexisNexis answers "no, but deletion after a few minutes or days".
Why does it matter who hosts the model?
The CNB's grid ranks four data-sovereignty criteria. Two of them exclude companies whose national law allows their authorities to requisition data stored on their servers abroad, "such as US companies under their current legislation". Pseudonymising data before the model sees it is one of the routes the CNB names to reduce that risk.
Can I reuse the table and the figures?
Yes, please cite them, with a link to this page or to the report. Each finding has its own anchor (#finding-1 to #finding-6) and the table is at #table-1. For the spreadsheet behind the count, write to info@hiddentokens.com.