How judges can use artificial intelligence: CGPJ Instruction 2/2026, point by point

The Plenary of the General Council of the Judiciary (CGPJ) approved on 28 January 2026 Instruction 2/2026, on the use of artificial intelligence systems in the exercise of judicial activity. It was published in the BOE on 30 January and has applied since that day.

The instruction is addressed to judges and magistrates, not to the legal profession. But anyone who litigates has an interest in knowing it: it defines what AI can do in the judge's chambers and, above all, what it cannot do.

The starting point: support, never replacement

The preamble relies on the European AI Act, on Spanish rules on the digitalisation of Justice and on the AI Use Policy approved by the State Technical Committee for Electronic Judicial Administration (CTEAJE). Its thesis is clear: AI is an aid to the exercise of the judicial function, not a substitute.

That approach translates into nine principles, set out in section four: effective human control, no replacement, judicial responsibility, independence, respect for fundamental rights, confidentiality and security, bias prevention, proportionality and limited use, and training.

Only authorised systems

Section five establishes that judges can only use AI systems provided by the administrations with powers in Justice or by the CGPJ itself, subject to their quality control and audit.

Section nine prohibits the use of systems that have not been provided. It admits a limited exception: preparation or study tasks that work only with open sources. In no case may judicial data be entered into those external systems.

Permitted uses

Under section six, AI may be used only as support to:

• Search for and locate legal information.
• Analyse, classify and structure the documents of the proceedings.
• Prepare outlines, summaries or internal working drafts with no decision-making character.
• Carry out organisational tasks.

Draft decisions

Section seven regulates the most delicate case. Draft decisions can only be generated with the systems provided, at the judge's will and in an editable form. They require a personal, complete and critical review and validation, and do not constitute automated decisions. Responsibility for the decision remains exclusively the judge's.

Prohibited uses

Section eight lists what cannot be done:

1. Replacing or delegating decision-making, the assessment of the facts or the evidence, or the interpretation and application of the law.
2. Using outputs that condition independence or freedom of judgement.
3. Incorporating AI-generated content into decisions without a critical and personal assessment.
4. Processing specially protected data or building profiles without express authorisation.
5. Any other use contrary to the instruction.

Data protection and oversight

Section ten prohibits the use of personal data in systems that have not been provided and limits its processing to what is strictly necessary and proportionate, without mass processing, with technical measures such as traceability. The CGPJ oversees this processing for judicial purposes, will offer training and warns that non-compliance may give rise to liability under the Organic Law of the Judiciary.

What the instruction does not say

The text does not expressly oblige the judge to inform the parties that AI was used in preparing a decision. It does make clear that responsibility is full and non-delegable and that human control must be real and effective.

What it means for lawyers

A reference standard

The instruction describes a model of responsible use: authorised tools, protected data, human verification and responsibility of whoever signs. It is the same standard the courts are applying to parties' filings. The cases in which lawyers have been sanctioned for citing non-existent judgments show it clearly, as we explain in our review of the sanctions for fake case law generated by AI.

Arguments for appeals

If a decision reflected a use of AI contrary to the instruction, for example an assessment of evidence without its own reasoning, the lawyer now has a written framework to refer to.

A parallel internal policy

A firm can draw on this structure for its own policy: a closed list of tools, a ban on entering sensitive data into services that have not been contracted, mandatory verification and training. The CCBE guide on generative AI for lawyers offers complementary recommendations from the legal profession's side.

The European timetable also matters: we review it in the new AI Act timetable after the digital Omnibus. And the organisation of the courts has changed with the Tribunales de Instancia (first-instance courts), about which we wrote in our editorial on the first-instance courts. More analysis in the Legal AI section.