Part of the AI Regulation News hub.
France's Cour de cassation published an internal AI charter that bars automated judicial decisions and sets verification and data rules for its own staff
This is a court setting limits for itself. The charter is useful to outside lawyers mainly as a statement of how France's highest civil and criminal court treats AI output, legal references and confidential data.
Bottom line: An internal ethical charter. It is addressed to the court's own magistrates, civil servants and contractors. On the sources we opened it is not a statute, a decree or a rule of procedure, and it does not bind advocates, litigants or other French courts. Its legal force against the court's own staff, such as any disciplinary consequence, is not stated in the material we read.
Who this affects: Magistrates, civil servants and contractors of the Cour de cassation who use AI in their functions. Indirectly, French and foreign litigators who appear before the court, and legal technology vendors that supply or hope to supply the court.
Issue date: The charter says it was settled ("arrêtée") on 30 September 2026, and its cover reads September 2026. The Cour de cassation published it with a news item dated 22 October 2026. The charter says it takes effect on its date of diffusion, which we read as the publication date, as inference.
What changed: The court now has a written AI framework with six principles. No judicial decision may be produced automatically, even under a magistrate's control. AI in judicial work is limited to support tasks. No generative AI tool can be imposed on a magistrate. Outputs are to be verified, legal references especially. Personal data and deliberation secrets stay out of external tools. The charter is to be revised at least every two years.
Analysis: Two points for outside lawyers, both our inference. First, the court does not require its magistrates to tell the parties when they used an AI tool, so counsel should not expect a disclosure from the bench. Second, the verification standard the charter sets for staff covers whether a cited decision exists, whether it is relevant, where it sits in the hierarchy and its exact scope. The charter does not say it is a standard for filings, but it shows what the court itself treats as the checking task.
Primary sources: IA : la Cour se dote d'une charte éthique, Cour de cassation, 22 October 2026 · Charte éthique relative à l'utilisation des systèmes d'IA à l'usage des agents de la Cour de cassation (PDF, 27 pages)
- Instrument (EN)
- Ethical charter on the use of artificial intelligence systems, for all agents of the Cour de cassation
- Instrument (FR)
- Charte éthique relative à l'utilisation des systèmes d'intelligence artificielle, à l'usage de l'ensemble des agents de la Cour de cassation
- Authority
- Cour de cassation (France), prepared by the court's AI reflection committee
- Jurisdiction
- France
- Status
- Adopted and published by the court. 27-page PDF with six principles and three annexes
- Bindingness
- Internal ethical charter for the court's own personnel. No effect on advocates or other courts is stated. Any sanction for breach is not stated in the material we read
- Issue date / next deadline
- Settled 30 September 2026 per the charter; published 22 October 2026. Revision at least every two years
- Scope
- Magistrates, civil servants and contractors using AI in their functions, across the system life cycle from design to withdrawal from service, covering external and internal systems
- Primary source
- https://www.courdecassation.fr/toutes-les-actualites/2026/10/22/ia-la-cour-se-dote-dune-charte-ethique
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
What did the Cour de cassation adopt, and for whom?
The court's news item says it has adopted an ethical charter on AI systems, following an April 2025 report that called for an ethical framework. It applies to everyone working within the court, magistrates, civil servants and contractors, when they use an AI system in their functions.
The charter sorts systems into external ones, built by third parties and reached over the internet, and internal ones, built by the court or supplied by the Ministry of Justice. It says external systems account for most use today and call for particular care on confidentiality.
It describes itself as deliberately more demanding than the law. It notes that the EU AI Act classes judicial activities as high risk, then says the Act alone is not enough and that respecting the law is a minimum requirement for magistrates. The charter's stated aim is to protect the judge's office where positive law does not yet reach.
No automated decisions, and support tasks only
Principle 1 opens with the sharpest line: "Aucun système d'IA ne doit se voir confier de responsabilité décisionnelle", and no judicial decision may be produced automatically even under a magistrate's control.
The charter then separates tasks. Some can be delegated entirely to AI with a simple after-the-fact check, and its example is formatting of decisions. Others cannot be fully delegated in any case, including interpreting a rule of law and characterising the facts. At the Cour de cassation, summarising the grounds of appeal is also ruled out.
Annex 3 turns this into a table. Drafting a draft judgment, order, opinion or report, legal characterisation of facts, constitutionality review, interpretation of a legal text and summarising the parties' grounds are all marked for exclusion on both internal and external systems.
Magistrates keep a freedom to decline. The charter says no generative AI system can be made mandatory for a magistrate in judicial work, and that neither using it, following its recommendations nor even receiving its suggestions can be imposed where they feed the reasoning behind a decision.
Verification, responsibility and training
The verification duty is framed as an obligation. The charter calls it especially pressing for legal references, which means checking not only that cited decisions exist but their relevance, hierarchy and exact scope. Its best-practice note says information produced by an AI system is to be checked by other means before it enters a professional document.
Responsibility stays with the author. The charter says responsibility for a document, a decision or preparatory work always and exclusively falls on the person who authored it, whatever tools were used. Reusing AI-generated content, even in part, means taking full responsibility for it.
Training is softer. Court staff are "incités" (encouraged, in the French wording) to take regular training, which reads as encouragement, not a requirement. The charter also asks staff to think independently before consulting an AI tool, as a defence against anchoring bias, and to report malfunctions through a dedicated mailbox.
Where court data may and may not go
Principle 4 is the practical core for anyone sending material into a model. For external systems, no document containing personal data is to be loaded, unless pseudonymised beforehand by the SDER. Data covered by a legally protected secret, notably the secrecy of deliberations, is never to go to an external AI, even pseudonymised.
Internal systems differ. If the AI system is under the responsibility of the court or the Ministry of Justice, in conditions of sovereignty and secure access, the charter says it is possible to load non-pseudonymised data from protected files. External tools remain open for tasks using only open data, such as summarising a published statute, provided the absence of personal or confidential data is checked with care.
The administrative-use precautions in the annex say they apply cumulatively, and that when there is doubt about the sensitivity of content, use of the internal system is mandatory.
What it does and does not change for lawyers outside the court
The charter does not regulate advocates. It says the court does not require a magistrate or agent to inform the parties of the use of an AI tool in handling their case, because AI use is limited to support tasks and the author answers fully for the output. Transparency is institutional, through publishing the charter.
That is a different design from the English CJC proposals on witness statements and expert reports, which target the parties' evidence. The two documents address different actors, so they should not be read as competing models. The comparison is our own.
For legal technology vendors, the charter says the court intends to prioritise internal systems and will watch server location, language model choice and any use of private providers. It does not name vendors or say what procurement will follow.
What we did not verify
We opened the court's news item and the 27-page charter PDF, and read the introduction, the six principles and Annexes 1 to 3. The extract file we saved for this article holds the passages we rely on rather than the whole PDF, which was read in full from the scrape.
We did not verify the charter's legal force against staff, any disciplinary consequence, or how it is enforced. We did not open the April 2025 report, the Ministry of Justice agreement of 10 July 2026 that the charter cites, or the other charters it references. The abbreviation SDER is not expanded in the material we read, and we do not expand it.
We have not established the exact date of diffusion beyond the 22 October 2026 news item, and the 30 September 2026 settlement date comes from the charter itself. French quotations keep their original diacritics, with typographic apostrophes straightened. We express no view on how any French court would treat AI use by advocates, and this is not legal advice.
A court has put in writing what it treats as off limits, what it will check and where its data stays. For outside counsel, the useful reading is the checking standard for legal references and the absence of a disclosure duty for the bench. The charter applies to the court's own personnel, and we did not establish its legal enforceability against them. It sets no rule for outside counsel.
Source File
Open the court's news item and confirm the date of 22/10/2026 and the six principles. Then open the charter PDF and confirm the 30 September 2026 settlement date in section 4 of the introduction, the wording of Principle 1 on decision-making responsibility, Principle 3b on not informing the parties, and Principle 4 on external systems.
Aucun système d'IA ne doit se voir confier de responsabilité décisionnelle, et aucune décision de justice ne saurait être produite de manière automatisée, même sous contrôle du magistrat. · Charte éthique de la Cour de cassation, Principe 1 a (French original; typographic apostrophes straightened), September 2026
FAQ
Does the Cour de cassation charter bind lawyers?
No. It is addressed to the court's own magistrates, civil servants and contractors who use AI in their functions. It sets no filing rule for advocates, and the sources we opened state no effect on other courts.
Does the court have to tell parties if a magistrate used AI?
Not under the charter. It says the court does not require a magistrate or agent to inform the parties of the use of an AI tool in handling their case, and puts transparency at the institutional level through publication of the charter.
Can AI draft a judgment at the Cour de cassation?
The charter's Annex 3 marks drafting a draft judgment, order, opinion or report as an exclusion for both internal and external systems. It also rules out summarising the parties' grounds of appeal.
What is the charter's date?
The charter states it was settled on 30 September 2026 and its cover reads September 2026. The court published it on 22 October 2026, and it takes effect on its date of diffusion.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.