Artificial intelligence can already locate case law, organise thousands of pages in a case file and suggest connections that a human reviewer might need days to verify. The crucial question, however, is not whether such a tool is fast. It is who checks the sources, who explains the legal reasoning and who accepts responsibility when the result affects a person's liberty, property or family life.

The European framework does not prohibit every use of AI in the administration of justice. It does, however, draw a substantive line between technical support and the exercise of judicial judgment. The decision must remain human, reasoned and open to scrutiny. Software output cannot operate as an invisible authority that the judge, a party or counsel is unable to examine.

1. Which systems are classified as high-risk

Annex III, point 8(a), of the European Union Artificial Intelligence Act classifies as high-risk those systems intended to be used by, or on behalf of, a judicial authority to assist in researching and interpreting facts and the law and in applying the law to a concrete set of facts. Certain uses in alternative dispute resolution fall within the same approach.

The classification does not cover every program used in a court. An electronic filing, diary-management or anonymisation system may perform a purely ancillary administrative task without affecting judicial decision-making. The boundary depends on what the system actually does, not on the commercial name given to it.

If a tool selects the case law a judge will see, assesses the credibility of a contention, proposes the legal provision to be applied or produces a draft statement of reasons, its influence is substantive. It is not enough to call it an "office assistant" merely because a human still provides the final signature.

2. What must remain a human decision

A recital to the Regulation explains that AI may support the judicial power but should not replace ultimate decision-making, which must remain a human-driven activity. That wording guides the interpretation of the framework; by itself, it is not a simple and absolute ban on every algorithmic function.

In practice, a human decision requires more than a final click. A judge must be able to:

  • check the factual and legal sources on which the output was based,
  • identify omissions, bias or incorrect citations,
  • reject or alter the proposal without technical or institutional pressure,
  • give independent reasons that address the parties' arguments,
  • take personal judicial responsibility for the resulting decision.

Merely accepting automatically generated text does not become meaningful human oversight because a signature has been added. Judicial independence requires a genuine ability to understand the proposal and to depart from it.

3. What applies after the change to the timetable

Regulation (EU) 2026/1744 postponed until 2 December 2027 the application of specific parts of Chapter III concerning the high-risk systems listed in Annex III. The postponement concerns the dedicated package of regulatory obligations; it does not turn the intervening period into a legal vacuum.

Before those specific provisions become fully applicable, other rules continue to apply where relevant: the General Data Protection Regulation, the European Convention on Human Rights, the right to a fair trial, the principle of judicial independence, national procedural rules and the duty to give reasons. The delay is not permission to conduct unsupervised experiments on real cases.

Any pilot deployment needs a clearly defined purpose, restricted access to data, controlled sources, a record of use, a risk assessment and a procedure for immediate suspension when an error or unequal treatment is detected.

4. When a right to an explanation exists

Article 86 of the AI Act provides, subject to specific conditions, a right for an affected person to obtain clear and meaningful explanations of the role of a high-risk system in the decision-making procedure and of the main elements of the decision. It concerns an individual decision taken by the deployer on the basis of the output of a particular Annex III system where that decision produces legal effects or significantly affects health, safety or a fundamental right.

The right does not automatically amount to access to source code, nor does it mean that every use of software invalidates a decision. It does mean that a person should not be left with the vague answer that "the system produced it". The person must be able to understand whether and how the output influenced the result in order to exercise their rights effectively.

Article 86 became applicable on 2 August 2026 and is not included in the specific postponement introduced by Regulation 2026/1744. Whether it is triggered must always be assessed against the factual and legal conditions of the individual case.

5. Minimum safeguards for a fair trial

When a system plays a substantive part in processing a case, safety is not limited to cybersecurity. Procedural safeguards are required so that a party can know about, examine and challenge its use.

  • Verification of sources: every judgment, statutory provision or factual item suggested by the system must be checked against the authentic text or record.
  • Logging: a trace should be retained of the tool, its version, the input data, the time of use and the critical interventions made by a human.
  • Disclosure: where the use materially influenced the judicial assessment, concealing it may undermine adversarial proceedings and an effective challenge.
  • Equality of arms: no party should face a technical conclusion that only the opposing party or the court can examine.
  • Bias testing: the data, indicators and error rates should be tested in relation to groups that may be affected differently.
  • Opportunity to object: there must be a practical means of correcting inaccurate data and challenging the method or the result.

A data protection impact assessment and a fundamental rights impact assessment are not automatically required for every simple application. They must, however, be considered before a system is used in real cases whenever the conditions set by the GDPR and the AI Act are met.

6. What the Greek framework provides

Greek Law No. 5321/2026 establishes the national framework for applying the AI Act and provides, among other measures, for a register of artificial intelligence systems used in the public sector. The Hellenic Data Protection Authority has taken on a role within the national supervisory structure in matters falling within its competence.

The register increases transparency, but it does not replace a legality review. Registration alone does not prove that a system is accurate, unbiased or suitable for every judicial use. Equally, this article does not claim that every digital tool currently used by a Greek court is a high-risk system. Its real function must be assessed.

7. What is binding law and what is guidance

The AI Act, data protection rules and national procedural safeguards form the binding legal framework. In parallel, the Council of Europe's European Commission for the Efficiency of Justice, CEPEJ, has issued the 2018 European Ethical Charter, a 2023 assessment tool and 2025 guidelines on generative AI. These texts are important standards for design and review, but they are not law in themselves.

The Council of Europe Framework Convention on Artificial Intelligence, CETS 225, had not yet entered into force when the sources were reviewed, and Greece was not listed in the official table as having signed or ratified it. In addition, the page of the Consultative Council of European Judges concerning Opinion No. 29 described work in preparation; it must not be presented as an opinion that has already been adopted.

8. What a citizen or lawyer can ask

Where there are specific indications that an AI tool materially influenced an individual determination, the useful questions are practical:

  1. Which system was used, for what function and by which authority?
  2. Was its output merely administrative support, or did it influence factual or legal conclusions?
  3. Which sources and which personal data were used?
  4. Who verified the result, and what changes did that person make?
  5. Is there a record of the system version and of the particular use?
  6. Was there an opportunity to correct the data and mount an effective challenge?
  7. Does Article 86 or another right to information, access or reasons apply?

The answer depends on the role of the tool, the type of proceedings and the applicable law. Suspicion that AI was used does not, by itself, invalidate a judicial act. There must be a connection between the use, the particular procedural safeguard and the actual effect on the case.

9. What the new framework does not say

  • It does not prohibit a judge from using a lawful search or organisational tool.
  • It does not permit the final judgment to be handed over to an opaque model.
  • It does not guarantee that a certified or registered system will not make a mistake.
  • It does not give every person unlimited access to code, trade secrets or third-party data.
  • It does not displace national rules on evidence, admissibility, reasons and remedies.

10. Frequently asked questions

Can AI draft a judicial decision?

A tool can technically produce a draft, but responsibility for the decision remains with the judge. The more the text influences the selection of facts, rules and reasons, the stricter the review, logging and opportunity to challenge must be.

Do I always have a right to know whether AI was used?

There is no unlimited, uniform right covering every ancillary use. Depending on the system's function and impact, however, Article 86 of the AI Act, rights under the GDPR, procedural disclosure or the duty to give reasons may apply.

Can an AI error be a ground of appeal or review?

It may be material if it affected the decision, introduced inaccurate data or prevented adversarial examination. The appropriate remedy and consequence depend on the procedure, the applicable time limit and the specific rule that was breached.

Does the postponement until 2027 mean that there are no rules?

No. Specific obligations for high-risk systems were postponed. Fundamental rights, data protection, judicial independence and procedural legality continue to bind the relevant bodies.

11. Official sources

Legal information: The official sources were reviewed on 30 August 2026. This article explains the general regulatory framework and does not assert that a particular Greek court uses an AI system in a particular case. It is not personalised legal advice.