AI in the justice system: where courts are trialling AI

Assistants, not robot judges

At the Higher Regional Court of Stuttgart, software has been pre-sorting diesel claims since the end of 2022. At the Local Court of Frankfurt, a programme assembles draft judgments in air passenger cases from standard text modules. In Lower Saxony, AI reads the courts' electronic mail, and Bavaria and North Rhine-Westphalia are training a language model of their own. Artificial intelligence has arrived in the German judiciary, and it has arrived as an assistant: the programmes structure case files, suggest text modules and take over routine work. The decision is still made by a human being.

The federal government is paying part of the bill. Through its digitalisation initiative for the judiciary, it is providing up to 200 million euros from 2023 to 2026. From 2027, a new pact for the rule of law is intended to follow. In April 2025, the federal and state governments adopted a joint AI strategy, followed in June 2025 by a joint declaration of the justice ministers. The key sentence of that declaration: the final decision remains an activity controlled by humans.

Mass proceedings: diesel and air passengers

It began with proceedings that repeat themselves in series. In 2022, more than 13,000 appeals arising from the diesel scandal were pending at the Higher Regional Court of Stuttgart. Together with the Baden-Württemberg Ministry of Justice and IBM, the court introduced the assistant OLGA. The software reads the judgment, the grounds of appeal and the response, looks for features such as vehicle type, engine and date of purchase, and assigns the case to a case group. The result is a pre-sorting; the assessment remains with the panel.

The volume is even greater for flight delays. In 2024, the German Association of Judges counted around 131,000 air passenger claims at the local courts in the twenty largest airport locations. At the Local Court of Cologne alone there were more than 41,000, roughly three quarters of its civil cases. For the Local Court of Frankfurt, Hesse developed the judgment configurator FraUKe. It evaluates pleadings, reads metadata and assembles a draft judgment from modules which the judge can then amend step by step.

Lower Saxony built the mass proceedings assistant MAKI, which has been in live use at several courts in the state since the end of 2024. The system summarises case files, categorises them and suggests decision patterns. Hesse, Brandenburg and North Rhine-Westphalia are taking part. NRW intends to deploy the system first at local courts in whose districts large airports are located.

Thick files and asylum proceedings

The second field of use is proceedings running to thousands of pages. Since 2023, the Regional Court of Frankfurt has been trialling the Codefy tool in two civil chambers and one commercial criminal chamber. It captures extensive case files, creates tables of contents and offers review assistants for preparing decisions. The Higher Regional Court of Koblenz is testing the same programme in four civil panels, primarily in construction law, and Bavaria has introduced it at the regional courts of Ingolstadt and Augsburg.

Baden-Württemberg is going a step further with AKIRA at the Social Court of Ulm. There, AI is to summarise the content of social court files including the administrative file and give them a preliminary legal structure. Such individual solutions are to be developed into StruKI, a structuring tool for all courts.

Administrative courts are using AI in asylum proceedings. EMIL, the country-information assistant from Lower Saxony, searches around 1.5 million documents on countries of origin, summarises passages and translates foreign-language sources. Ten further states, the Federal Administrative Court and the Federal Office for Migration and Refugees are trialling it. The facts assistant TabeA drafts the statement of facts, that is the account of the case in the judgment, at four administrative courts in Lower Saxony.

Mail room, translation, language model

Less spectacular, but noticeable in everyday work, is AI in the registry. Since 2023, Bavaria has been trialling the SMART/IMJ system at the Regional Court of Ingolstadt, which reads electronic submissions using text recognition and transfers the case data automatically. In 2026, Lower Saxony is piloting the INDATA software together with Bavaria and Rhineland-Palatinate. The programme names incoming documents, separates multi-part submissions and extracts claimants, defendants and addresses. After retraining, the ministry says the accuracy rate rose from around 80 to over 99 per cent.

Baden-Württemberg is building a translation platform for at least 40 languages for the federal and state governments. And since March 2026, Bavaria has been using the AI tool ALeKS to anonymise judgments in order to publish 50,000 decisions.

The most ambitious project is called GSJ, the judiciary's generative language model. Bavaria and North Rhine-Westphalia have commissioned the Technical University of Munich and the University of Cologne to develop it. It is being trained on anonymised judicial data from both states, funded from the digitalisation initiative. By the end of 2026, the model is to answer questions about long documents, organise the parties' submissions and summarise decisions in testing. Behind this lies the desire for digital sovereignty: a model that the judiciary controls itself and may release as open software after the trial phase.

Open questions for judges and lawyers

The justice ministers stress that humans decide. The question is how free that decision remains once the draft is already on the screen. Anyone dealing with air passenger cases in series using pre-formulated modules must also be willing to disagree with the suggestion. This is where the German Association of Judges draws the line: IT must neither standardise nor steer decision-making processes; it should support the independent judge's own responsible work. At the same time, the association soberly notes that usable AI tools for routine use across the board are still lacking.

The legal framework has become tighter. The EU AI Act classifies AI systems that assist courts in establishing and interpreting facts and legal provisions as high-risk systems. This entails obligations relating to risk management, documentation, logging and human oversight, and these apply to the judiciary as a deployer as well. What remains open is transparency towards the parties: are they told that software has pre-sorted their case? The joint declaration names transparency as a principle, but it does not say what that looks like in an individual case.

For law firms, this has two sides. Anyone bringing a claim must expect that a machine will read the pleading first and search it for particular features. Clearly stated facts then help the programme as much as the judge. Conversely, courts complain about hundred-page, sometimes faulty submissions produced by AI. The CSU's working group of lawyers has therefore been calling for changes to procedural law since September 2026, and the Federal Ministry of Justice has announced further development within six months.

Conclusion

AI is already at work in German courts day to day, in clearly defined niches: in diesel claims and air passenger rights, in thick case files, in asylum proceedings and in the mail room. In all of these it is an assistant, and the federal and state governments have put it in writing that the decision remains with a human being.

The difficult questions come with use across the board: the judge's independence from the machine's suggestion, the information given to the parties and the obligations arising from the AI Act. The tools are already running; the answers are still outstanding.