AI in the Justice Administration: what the 2024 use policy allows and forbids

Artificial intelligence is already part of the ordinary running of the Justice Administration. According to a presentation by the Directorate-General for Digital Transformation of the Justice Administration, of the Ministry of Justice, dated May 2025, the Artificial Intelligence Use Policy was approved in this area in June 2024, with the aim of ensuring a responsible, lawful and ethical use of these technologies.

The same document sums up the approach: the tools do not seek to replace people or make decisions that could affect citizens' rights and freedoms. For the legal profession, which litigates every day before courts that use these tools, it is worth knowing what the policy allows and forbids.

Where to find the text

The Council of Europe publishes on its website a preliminary English version of the policy, warning that it may contain translation errors. That text says its adoption must be ratified by the General Council of the Judiciary, the State Attorney General's Office, the autonomous communities with powers in Justice and the Ministry. The data in this article about the policy come from that version, so it is wise to check them against the Spanish text when it is available.

Scope

The policy applies to systems and services fed by judicial data, from the first filing to the final decision, including publications and notifications. It reaches the staff of the Justice Administration, that of suppliers and other institutional actors with access to the systems. It distinguishes between uses that may affect judicial independence and those that do not.

Three levels of use

Generating computer code also requires approval from the technology area, with peer review and security validation.

What is forbidden

According to the version consulted, the policy prohibits:

• Using protected data to support decisions affecting rights, except with specific authorisation.
• Generating final decisions that are issued automatically.
• Reporting automatically on the status of proceedings without human supervision.
• Predicting people's future behaviour to justify judicial actions, such as pre-trial detention or parole.
• Using generative systems to replace attendance at meetings or committees.

Safeguards

The policy requires that every generative AI output be identified as such, and clearly marked in materials addressed to the outside. What is generated must be reviewed by a person whenever it affects, directly or indirectly, rights or judicial activity, and final responsibility lies with judges and magistrates. Citizens must know whether an AI output is binding and keep the right to be heard by a court.

On data protection, it asks for impact assessments before use and warns that data entered into public generative AI tools may persist or reappear. It recommends publishing FAT sheets for each service, covering data, algorithms and possible biases.

The tools the Ministry describes

The May 2025 presentation lists several tools:

• Recording Transcription: turns audio and video into text, in production.
• Document anonymiser: recognises data that identifies people.
• Document Classifier: automatically catalogues the documentation arriving at the courts.
• Similarity analysis: locates the same person in different parts of a document.
• Easy Read: in proof of concept, to adapt legal texts.
• Delfos: an intelligent search engine in legal documents, with a pilot planned.
• Draft assistants for payment orders and air passenger claims, developed with CENDOJ, which generate a draft decision for review and final editing by the judge or magistrate.

Several services are offered through the serviciosia.justicia.es portal. The document also mentions the Data Manifesto, which treats data as a public good.

What it means for lawyers

1. Recordings may have automatic transcription. It is useful for preparing an appeal, but it is wise to check it against the recording before citing it.
2. Assisted drafts are limited to very defined matters and always go through the judge. If a decision looks standardised, the route is the usual appeal.
3. The criteria for judges are completed by the CGPJ instruction we analyse in the review of the CGPJ instruction on AI for judges.
4. The firm needs its own rules, in the same spirit of human supervision, as we proposed in the editorial on AI policy at the law firm. Courts already sanction fake citations, as we reported when covering the sanctions for fake case law.

More analysis in Legal AI.