Measure 4.02 · progressive deployment

Legal File 4.02: Deploying AI Across 164 Courts

The number of courts must never turn a pilot into a rushed nationwide rollout.

Legal File 4.02: Deploying AI Across 164 Courts
Legal File 4.02: Deploying AI Across 164 Courts
Scope of the proposal

What measure 4.02 actually proposes

Measure 4.02 proposes deploying artificial-intelligence tools across the 164 courts identified by the Plan de Rupture. That figure belongs to the Plan’s own implementation scenario. It must not be interpreted as an instruction to purchase one identical product and install it everywhere at once.

Courts differ in jurisdiction, caseload, staffing, digital maturity and existing software. A credible programme therefore begins with clearly defined uses, limited pilots and measurable conditions for moving from one wave to the next.

Why deployment must proceed in waves

🧭 Different uses

Document search, filing assistance, administrative triage and legal analysis do not create the same risks.

🏛️ Different courts

Volumes, subject matter, teams and technical environments vary substantially.

🧪 Evidence before scaling

A pilot makes it possible to measure errors, time saved and unintended effects before wider use.

↩️ Reversibility

Each wave must be capable of being stopped without preventing the court from operating.

👥 Training

Users need time, competence and genuine authority to challenge or reject the system’s output.

📊 Public evaluation

Indicators must be defined before deployment, not invented after the decision to generalise.

Recommended implementation path

  1. Inventory. Map current tools, data, workflows and needs.
  2. Classification. Determine the legal and operational risk level of each proposed use.
  3. Pilots. Select a limited number of courts and narrowly defined tasks.
  4. Acceptance testing. Test performance, rights, security, human oversight and fallback arrangements.
  5. Controlled waves. Extend deployment only when published thresholds have been met.
  6. Review. Suspend, correct or withdraw a system when evidence no longer supports its use.
Legal position in August 2026. Under the European AI Act, certain systems used by or on behalf of judicial authorities to assist in researching and interpreting facts and law may fall within the high-risk category. The European Commission published draft classification guidelines in May 2026; those guidelines remained a draft following the consultation that closed in July 2026.

Decision gate before the next wave

QuestionEvidence required
Does the tool produce a real benefit?Comparison with a documented reference procedure.
Are error rates acceptable?Tests by subject matter, court type and model version.
Is human oversight genuine?Competence, available time, authority to refuse and a duty to give reasons.
Are affected persons protected?Impact assessments, information, remedies and security controls.
Can the court return to a safe mode?A tested continuity and reversibility plan.

Working resources

Main official sources

  1. Regulation (EU) 2024/1689 — Artificial Intelligence Act.
  2. European Commission — draft guidelines on high-risk AI systems.
  3. CNIL — carrying out a data-protection impact assessment for an AI system.