AI in the judiciary and justice: high-risk under Annex III

Short answer: AI systems intended to assist a judicial authority in researching and interpreting facts and the law, or in applying the law to a concrete set of facts, are high-risk under Annex III of the AI Act. Purely ancillary, administrative tasks fall outside it. The core stays: the judge decides. AI may prepare and advise, but not adjudicate.

What the AI Act means by "administration of justice"

Annex III, point 8, classifies as high-risk: AI intended to assist a judicial authority (or on its behalf) in researching and interpreting facts and law, or in applying the law to a concrete set of facts. Think of case analysis, outcome prediction or automated file summaries that touch the substantive assessment. The high-risk label brings the full set of high-risk obligations: risk management, data quality, logging and human oversight.

What falls outside

Not all AI in the justice chain is high-risk. Purely administrative or organisational support — scheduling, anonymisation, translation, document management with no influence on the administration of justice — falls outside point 8. The line is whether the system touches the substantive assessment. A tool that only speeds up logistics is different from a model that weighs into the judgment.

Judicial independence as a limit

Judicial independence sets a hard limit on automation. Human oversight (Art. 14 AI Act) is no formality here: the judge must be able to grasp how the system works, weigh the outcome and deviate without friction. Automation bias — the tendency to trust a system output — is especially dangerous in adjudication, because it hollows out the duty to reason and to weigh. The output must never replace the judgment.

Transparency and the defence

A defendant or party must be able to understand whether and how AI contributed to a decision, in order to contest it. This touches fair trial and equality of arms. Logging and explainability are therefore not just compliance requirements but procedural safeguards — see explainability of government algorithms. For the broader public-sector context, see AI in the public sector.

What to do

• Classify per use: does the system touch the substantive administration of justice (high-risk) or is it purely administrative (outside)?
• Secure effective human oversight: the judge must be able to understand, weigh and deviate — guard against automation bias.
• Keep AI advisory: never the decision-maker, always supporting a reasoned judgment.
• Make it traceable: logging and explanation so parties can contest AI's contribution.
• Run a fundamental rights assessment before use — see FRIA.

In the judiciary, AI is at most an instrument. The judgment — and the responsibility for it — stays with the human in the robe.