Keeping up with the machines: how arbitral institutions are responding to AI (Part 2)

The AAA

The AAA has certainly gone the furthest of the three institutions. In November 2025, the AAA launched an AI arbitrator for documents-only construction disputes – a model trained on over 1,500 annotated AAA construction awards. This product operates under its own dedicated rulebook, the AI Led Arbitration rules, which came into effect on 18 February 2026. The key features of the framework include:

• Both parties having to opt in – if the respondent declines, the case reverts to standard AAA administration.
  
  
• The parties' consent is to a preliminary AI decision, which a human arbitrator then reviews, revises, validates and issues as the award.
  
  
• The AI selects the human arbitrator from the AAA's national roster.
  
  
• There is no hearing. Parties exchange written submissions on ten-business-day cycles. The AI summarises the claims, evidence and authorities, and the parties are given the opportunity to correct those summaries.

Separately, the AAA-ICDR published general guidance in March 2025 governing arbitrators' use of AI tools. This requires arbitrators to cross-check AI outputs against primary sources, retain complete control over decision-making, keep confidential material out of insecure tools, and disclose the use of generative AI where it materially affects the process or the reasoning.

Most recently, the AAA this month (September 2026) launched its AI Ambassador program. The AI Ambassadors (arbitrators and mediators) are organised into four focus groups and will collaborate to identify emerging legal and professional issues arising from AI and develop practical resources for those in the ADR space. The four working groups will examine: (i) AI evidence and arbitration procedure; (ii) emerging AI disputes; (iii) AI-driven commerce and automated transactions; and (iv) digital assets and algorithmic finance. The topics of these focus groups indicate the issues that are already emerging in practice and how the rules of the AAA in relation to AI may develop in the future.

CIArb

CIArb published its Guideline on the Use of AI in Arbitration in 2025. This contains four parts and two appendices: a template agreement on AI use and a template procedural order.

The substantive provisions can be summarised under three themes:

1. Accountability: Using an AI tool does not reduce the parties' or the tribunal's responsibility for the output.
   
   
2. Procedural power: The tribunal has the ability to direct AI use as part of its general procedural powers. The tribunal may require a party to disclose the way in which it has used an AI tool where such use may affect the evidence, the outcome of the arbitration, or the discharge of an express duty. However, arbitrators may not regulate the private use of AI by parties so long as the above conditions are met and such use is generally allowed in litigation in the relevant domestic courts. The Guide does not elaborate on what is meant by "private use", though we can infer that a party using AI to manage internal workload, documents and budgets, with appropriate safeguards and supervision, would be beyond the tribunal's reach.
   
   
3. Decision-making: Arbitrators must not relinquish their decision-making responsibilities, must verify AI outputs independently, and must remain responsible for every aspect of the award. Notably, where the tribunal wishes to use AI and the parties disagree, the Guideline provides that the arbitrator should refrain from such use.

Further, the Guideline goes on to define "High-Risk AI Use", which includes use of an AI tool in the proceedings entailing risks such as breach of privacy and confidentiality, the potential materially to undermine the procedural integrity of the arbitration, or the potential to assert a non-human influence on the award. Any engagement in High-Risk AI use is treated with gravity and must promptly be disclosed to the tribunal. Before the tribunal itself undertakes a high-risk use, it must give notice and allow the parties to brief the point.

As discussed in the first article in this series, the CIArb template procedural order may be a useful starting point for parties and tribunals seeking to incorporate AI governance into their proceedings, including in institutions that do not yet have their own AI-specific provisions.

The SCC

The SCC was an early adopter of AI-specific regulations, having published its Guide in October 2024. The Guide provides flexible guidance without imposing specific obligations.

The SCC's approach is distinguished by its connection to legislation. The Guide notes that the EU AI Act classifies AI used by an arbitral tribunal "in researching and interpreting facts and the law and in applying the law to a concrete set of facts" as high-risk. On that basis, the Guide makes four points. The first of these is addressed to all participants, while the remainder appear to be directed to tribunals.

1. Confidentiality: Participants should understand how their input data is used.
   
   
2. Quality: AI can perpetuate bias and produce false or manipulated evidence. Tribunals must apply an appropriate level of review before relying on any output.
   
   
3. Integrity: Tribunals are encouraged to disclose their AI use, which helps protect the right to be heard and guards against exceeding the mandate.
   
   
4. Non-delegation: Tribunals cannot delegate the decision, or the reasoning leading to it, to "anyone or anything".

All three frameworks appear to converge on the same fundamental principle: whilst AI assists, humans ultimately decide. However, the frameworks differ in important respects, and parties should be aware of the following considerations.

The AAA is the only institution that has operationalised its framework. Its design raises three points that parties should consider carefully:

• First, a point in respect of sequencing. AI is the real mover, drafting the award that is later reviewed by a human arbitrator. As a matter of practice, an initial draft has a tendency to set the tone and substance of the final output; this may therefore limit the extent to which the human arbitrator departs from the AI's analysis.
  
  
• Second, there is a question of transparency. Whilst parties are given the opportunity to validate summaries of their own case, they do not see the AI's issue analysis or the draft award.
  
  
• Finally, parties should have a careful eye to the appointment process. A fundamental feature of arbitration, which commercial parties tend to value significantly, is autonomy in selecting an arbitrator. However, here, the AI selects the human arbitrator on unpublished criteria.

The CIArb Guideline is the most detailed drafting aid currently available and provides a practical framework that parties and tribunals can adopt. However, parties should be aware that the long-form procedural order introduces substantive procedural machinery: disclosure obligations, comment rounds, and a notice and briefing requirement before the tribunal can use an AI tool. Whilst these safeguards are valuable, in the hands of a party seeking to delay proceedings, they could become a source of satellite disputes that increase both cost and time. Further, the Guideline's carve-out for "private" AI use is not defined in any detail. In practice, the boundary between private use (for example, counsel using AI to assist with internal work) and use that affects the proceedings (for example, counsel using AI to draft a submission) may be difficult to draw. Importantly, the Guideline is not binding unless adopted by the parties or the tribunal.

The SCC Guide is the shortest of the three but is notable for being the only framework anchored in legislation rather than best practice alone. However, the Guide encourages AI use rather than requiring compliance. As noted above, the deferral of the EU AI Act's high-risk obligations to December 2027 means that the legislative framework to which the Guide refers is not yet in force.

Significantly, there are gaps that are common to all three frameworks. Each framework regulates the tribunal's use of AI more tightly than counsel's (perhaps unsurprisingly), whilst none addresses the institution's own use of AI for case administration or appointments, even though institutions are now deploying AI for these purposes. The enforcement risk associated with AI use in arbitration also remains untested. As discussed in the first article, in LaPaglia v Valve Corporation [No. 3:25-cv-00833 (S.D. Cal. Apr. 8, 2025)], a losing party sought to vacate an AAA award on the ground that the arbitrator had outsourced his adjudicative role to AI. The court dismissed the petition in December 2025 on jurisdictional grounds without needing to address the substantive point. The argument is likely to return and may ultimately be tested under Article V of the New York Convention.

In light of the frameworks discussed in this article and the first article in the series, parties may wish to consider the following steps:

• Agree AI use at the drafting stage: CIArb's Appendix A template agreement can sit alongside an arbitration clause and settles the question of AI governance before any dispute arises.
  
  
• Ask the tribunal about its own use of AI: Under the CIArb Guideline, if the parties disagree with the tribunal's proposed AI use, the arbitrator should refrain from such use. Any concerns should be raised during the arbitration and not reserved for enforcement – this will create a record of any objection parties may have raised and may prevent costly disputes about AI usage.
  
  
• Address costs promptly: Where the tribunal does use AI in the proceedings, parties should address how such use may be reflected in the tribunals' fees.
  
  
• Interrogate confidentiality: Parties should enquire into the specific type of AI used by the tribunal, particularly in terms of confidentiality and data protection – asking questions such as what the AI tool retains, whether the model is closed, and whether inputs are used to train the model. Institutions that use their own, secure AI platforms may be preferred by parties on this basis.

The AAA, CIArb and the SCC have each contributed to a developing framework for AI in arbitration. These frameworks provide useful tools for parties and tribunals, particularly in the context of the institutional rules examined in the first article, which remain largely silent on AI. However, gaps remain. Parties who engage with these frameworks early and incorporate appropriate governance into their arbitration agreements and procedural orders will be better placed to manage the risks and benefits as this area continues to develop.