AI in arbitration: Germany's missing rulebook is nothing to worry about
If you are looking for specific legislation or institutional guidance to tell you whether artificial intelligence (AI) belongs in your arbitration in Germany, you can stop waiting. Although the DIS has not yet published such guidance and the German Code of Civil Procedure (ZPO) does not expressly address AI, German law already answers the most important questions. Additional requirements apply, and will increasingly apply, under the EU Artificial Intelligence Act (Regulation (EU) 2024/1689; AI Act). This post explains what already binds tribunals and counsel in arbitration proceedings conducted in Germany and why the most important AI rules in your next case will be the ones you write yourself.
What other institutions have published and where Germany stands
The institutional soft law on the use of AI in arbitration has developed quickly. The SVAMC issued its AI Guidelines in April 2024. The SCC followed with a guide in October 2024, CIArb with a guideline and a template procedural order in 2025, and VIAC and CIETAC with their own notes the same year. The ICC has convened a task force on AI in dispute resolution, though it has not yet published any guidance.
The instruments agree on a handful of principles: arbitral tribunals must not delegate their decision-making powers to AI, counsel remain responsible for the accuracy of factual and legal submissions, and all participants must ensure that AI tools comply with applicable confidentiality obligations. The disclosure of AI use is handled flexibly.
The 2018 DIS Rules are technology-friendly (e.g. electronic communication is the default, article 4.1; the use of information technology is among the efficiency measures the tribunal is to discuss with the parties, article 27.4(i) and Annex 3). The DIS has engaged with the debate on AI through events and through a cooperation with the SVAMC announced in December 2025. It also recently made a joint submission with 15 other arbitral institutions on the European Commission’s draft guidelines on the classification of high-risk AI systems under the AI Act. The submission asks the Commission to distinguish between AI used by tribunals in adjudicative functions and AI used by arbitral institutions for administrative case management. However, the DIS has not, to date, published any specific AI rules or practice note.
Nor will the German legislator provide AI-specific rules any time soon. On 10 June 2026, the Federal Cabinet adopted a government bill (Gesetzentwurf der Bundesregierung) to reform the 10th book of the ZPO. This bill would modernise German arbitration law in several respects, among them video hearings and electronic awards, but it does not address the use of AI tools. That is consistent with the position in state-court litigation, where the ZPO likewise contains no provisions specifically regulating AI.
Seen from abroad, these absences might look like Germany falling behind. We would suggest a different reading.
What German procedural and professional law already requires
For arbitration proceedings with a German seat, many of the key principles found in the international AI guidelines already follow from mandatory German law, in particular the ZPO and German professional law, as well as from EU regulation.
Prohibition on delegation of the decision to AI
The prohibition on delegating the decision to AI follows from the way the ZPO structures the arbitrator’s mandate. The statute does not say expressly that an arbitrator must be human, but sections 1025 et seq. are built around a personally appointed and personally responsible human decision-maker. Section 1036(1) ZPO, for example, speaks of the “person” to whom the position of arbitrator is offered. Section 1054 ZPO requires the award to bear the arbitrators’ signatures, a formal act that only a natural person can perform.
That premise matters at the set-aside stage. In its decision of 12 January 2023 (I ZB 41/22), the Federal Court of Justice held that an award may be set aside under section 1059(2) no. 1(d) ZPO where it is made without the participation of an appointed arbitrator and the conditions for deciding without that arbitrator are not met. The case concerned an arbitrator who refused to take part in the vote, not an arbitrator outsourcing a decision to AI. Its rationale nevertheless extends to that scenario. The provision protects the integrity of the tribunal’s deliberation by requiring the decision to be reached by those the parties appointed. An arbitrator who delegates the substance of the decision to an AI tool no longer participates in the act of deciding. Such a defect will typically be taken to have affected the award, making it vulnerable to a set-aside application.
Delegation of this kind may also engage the procedural public policy ground under section 1059(2) no. 2(b) ZPO. Written submissions can, technically, be uploaded to and processed by a large language model. However, the right to be heard is not satisfied merely by processing a party’s case. It requires the appointed decision-maker to take note of, understand, and consider the party’s submissions in reaching the decision. A large language model does not understand a party’s case in that sense. It generates text by predicting likely continuations based on statistical patterns learned from its training data.
Counsel’s responsibility for the accuracy of factual and legal submissions
When it comes to AI-generated submissions, including fabricated case law, professional law is particularly relevant. In every professional capacity, including arbitration, a German Rechtsanwalt is required to practise conscientiously (section 43 of the Federal Lawyers’ Act (BRAO)) and must not knowingly disseminate untruths in the course of professional practice (section 43a(3) sentence 2 BRAO). Knowingly presenting a fabricated authority engages that duty. Knowingly advancing false factual submissions does so even more clearly. A lawyer who breaches these professional duties will not necessarily lose the case for the client but may be personally sanctioned with a reprimand or fine.
Germany has a growing number of cases in which courts have “caught” lawyers relying on AI-fabricated authorities. The Local Court of Cologne (2 July 2025 – 312 F 130/25) appears to have been the first to call out such conduct expressly, cataloguing a series of invented decisions and non-existent commentary references in a lawyer's brief. The court warned that such submissions impeded its work, misled non-expert readers and harmed the reputation of the legal profession. It added that knowingly false submissions about the content and meaning of statutes and court decisions fall within section 43a(3) BRAO. More recently, the Higher Administrative Court of Baden-Württemberg (12 March 2026 – 12 S 470/26) held that, in proceedings subject to mandatory legal representation, that requirement makes counsel responsible for independently examining, reviewing and legally analysing the subject matter of the dispute. Where AI is used to draft a submission in such proceedings, counsel must review the generated text independently and may not adopt it unexamined if the procedural act is to be valid.
Other courts have been more restrained in their reactions. For instance, the Regional Court of Frankfurt am Main (25 September 2025 – 2-13 S 56/24) found three case citations to be entirely fabricated and expressed the hope that they had not been invented by counsel himself, but rather “hallucinated” by a chatbot. The Higher Regional Court of Celle (29 April 2025 – 5 U 1/25) and the Higher Administrative Court of Baden-Württemberg (8 July 2025 – 3 S 1012/25) merely noted that the cited decisions could not be found.
All these examples come from state-court litigation. No German arbitration decision dealing with AI-hallucinated submissions is publicly known to date.
None of this means that the legal profession discourages the use of AI. Sections 5 and 11(1) of the Rules of Professional Practice (BORA) require lawyers to maintain the resources, staffing and organisational arrangements necessary for their practice and to handle mandates within a reasonable time. In large-scale matters in particular that may point towards the use of AI.
Confidentiality duties
Confidentiality is the least novel aspect of the AI debate. Section 43a(2) BRAO requires a German Rechtsanwalt to keep confidential any information learned in professional practice, a duty further specified in section 2 BORA. Unauthorised disclosure of a professional secret may also be punishable by imprisonment or a fine under section 203(1) no. 3 of the German Criminal Code.
Section 43e BRAO nonetheless permits lawyers to give service providers access to such confidential information, but only where necessary and subject to safeguards. These include careful selection of the provider, passing on the confidentiality obligations, ensuring comparable protection abroad and obtaining client consent for services relating to a specific mandate. Uploading confidential mandate material to an ordinary consumer AI tool will fail that test, in particular where the provider may access, retain or reuse the material for training or product improvement. Enterprise-grade tools may offer stronger safeguards, but they too require review against these obligations before mandate material is uploaded.
For arbitrators not admitted to the German bar, the applicable confidentiality regime usually derives from institutional rules and will typically lead to the same practical result. Article 44 of the 2018 DIS Rules, for example, subjects the parties, their counsel, the arbitrators and the DIS employees and any other persons associated with the DIS who are involved in the arbitration to a duty of confidentiality, unless the parties agree otherwise. This obligation extends to the confidentiality of data introduced into IT systems and has corresponding implications for cyber security.
What the EU AI Act means for arbitration in Germany
As an EU regulation, the AI Act is directly applicable in Germany. One obligation that already applies is the requirement for providers and deployers to take measures to ensure AI literacy among their staff and other persons using AI systems on their behalf (article 4 of the AI Act). This does not determine how AI may be used in arbitration proceedings, but it is intended to promote the informed use of such systems.
Looking ahead, certain uses of AI by arbitral tribunals fall within the regulatory regime for high-risk AI systems. This includes AI systems intended to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, as well as systems intended to be used in a similar manner in alternative dispute resolution (article 6(2) of the AI Act in conjunction with Annex III no. 8(a)). Not every use of AI by an arbitral tribunal is therefore high-risk. The classification targets adjudicative tasks, and article 6(3) excludes certain systems from that classification where they do not materially influence the outcome of decision-making, provided the further conditions of that provision are met. A provider relying on this exception must document its assessment before the system is placed on the market or put into service and must still register itself and the system in the EU database (articles 6(4) and 49(2) of the AI Act).
The specific requirements for these high-risk systems will apply from 2 December 2027, following the deferral agreed in the Digital Omnibus on AI (Regulation (EU) 2026/1744), which moved the date back from 2 August 2026. For deployers, these include using the systems in accordance with the provider’s instructions, monitoring their operation, retaining system logs and assigning human oversight to individuals with the necessary competence, training and authority (see article 26 of the AI Act).
Drawing the line in Procedural Order No. 1
German law already provides a workable framework for the use of AI in arbitration proceedings. Procedural Order No. 1 can translate that framework into case-specific rules, reducing the risk of challenges based on AI use in the arbitration.
In LaPaglia v. Valve, the losing party alleged that the arbitrator had effectively outsourced the decision to ChatGPT, relying among other things on the speed with which a lengthy, reasoned award followed the hearing. The motion failed on procedural grounds, leaving the substantive question open.
Nevertheless, the case points to a risk likely to recur, namely that parties will increasingly raise the tribunal's – alleged – AI use as a ground to challenge the award. It is therefore sensible to address the matter early, discussing with the parties at the first case management conference which uses of AI by the tribunal are permissible and recording the outcome in Procedural Order No. 1. The order can identify which uses, if carried out as agreed, should not, without more, support a later challenge to the award. It can also specify which uses must be disclosed by the tribunal.
Procedural Order No. 1 can also address the data side of AI use. It can set specific confidentiality and data conditions for any tool used with case materials, including access, retention, training, deletion and onward disclosure.
***
