Open data of court decisions: how the Cour de cassation pseudonymises

France has committed to publishing all of its court decisions for free, millions of documents full of names, addresses and family situations. To do so, the Cour de cassation and the Conseil d'État had to solve, at national scale, the problem every law firm now faces with AI: removing what identifies people without destroying what makes the text useful. Here is how they do it, and what it teaches about your own files.

1. What the law requires

The principle of open court decisions comes from the 2016 Digital Republic Act. Its redaction rules were rewritten by article 33 of Law no. 2019-222 of 23 March 2019, which amends article L. 10 of the Administrative Justice Code and article L. 111-13 of the Code of Judicial Organisation in identical terms. They work on two levels.

The same article forbids reusing the identity of judges and court staff "to evaluate, analyse, compare or predict their actual or supposed professional practices" (our translation), on pain of criminal penalties.

Decree no. 2020-797 of 29 June 2020 organises the implementation, and an order of 28 April 2021 sets the timetable court by court. According to the Ministry of Justice, administrative decisions, those of the Cour de cassation and civil decisions of the courts of appeal are already available; the timetable runs to 31 December 2027 for minor offences.

2. Who does it, and how

Redaction is carried out by the Cour de cassation, which publishes judicial decisions on Judilibre, and by the Conseil d'État, which publishes administrative decisions on its open data portal. The ministry puts it this way:

The Cour de cassation documents its approach publicly in its GitHub repository. Until December 2019 it used a rule-based engine whose error rate "was close to 5%" and which required "a significant amount of manual correction" (our translation). Through the Entrepreneurs d'intérêt général programme, it built a machine-learning engine that finds the data to pseudonymise. Decisions pre-annotated by that engine are then proofread by staff in LABEL, an annotation tool the Court has released as open source.

3. The debate is not closed

A report submitted to the Minister of Justice on 11 July 2025 by Daniel Ludet, honorary judge at the Cour de cassation, proposes extending systematic redaction to company names, addresses and localities, dates relating to personal status and directly identifying character strings (Next, 16 July 2025). Whether or not these proposals are adopted, they say what experience says: names alone are not enough.

4. What a law firm can take from it

• Names are the floor, not the ceiling. The law itself targets "any element that would identify". An address, a date of birth, a company name or a rare job title is often enough to find someone. That is the subject of our article on re-identification risk.
• Automation plus human review. The State's own model is a detection engine followed by a human check. That is also how Hexagone AI works: every detection appears in the preview pane and can be corrected before export.
• Rules engines plateau. The Cour de cassation moved from rules to machine learning because rules could not keep up with the volume and variety of documents. A list of names to search and replace has the same limit in a firm.
• Publication and AI work are different goals. A decision published as open data is not meant to be re-identified. A file sent to an AI is: you want the answer back with the real names. Hence reversible pseudonymisation, with the mapping kept in the firm. The difference is explained in anonymisation or pseudonymisation.