AI for International Arbitration: Tools & Risk Framework

International arbitration sits at the harder end of the AI-in-legal-practice spectrum. The same case may touch privilege rules from three jurisdictions, data-residency requirements from a fourth, and institutional norms set in a fifth. AI tools that work fine in a single-jurisdiction domestic context can create real risk when deployed across borders.

This guide is a framework for counsel and tribunal secretaries thinking about AI in international arbitration in 2026. We built the AAAi Chat Book for the American Arbitration Association and operate the same architectural pattern for adjacent international clients; the perspective below is informed by that operating experience. Where claims about specific institutional rules or survey statistics are too time-sensitive to commit to here, we point at the authoritative source rather than reproduce numbers that will be out of date by next quarter.

The 2025 IA Survey results: AI adoption across institutions

The 2025 Queen Mary–White & Case International Arbitration Survey, based on 2,402 questionnaire responses and 117 interviews, shows 90% of practitioners expect to use AI for research, data analytics and document review, with 54% citing time savings as the main driver and 51% citing AI errors and bias as the main obstacle. Regionally, 55% of Asia Pacific respondents — versus 30% in Europe — believe AI can accelerate arbitration. The Survey does not break AI adoption down by arbitral institution; institutional positioning has to be read from each institution’s own published guidance, rules, and partnerships.

These are expectation and driver percentages, not measured adoption rates — practitioners reporting what they plan to do, not what they currently do. The directional pattern across the IBA’s parallel work and earlier QMUL/White & Case editions is consistent with this:

• AI use for research, summarisation, and document review is now mainstream for counsel.
• AI use for first-draft generation (briefs, statements, witness preparation) is rising but more contested.
• AI use for substantive merits decisions by tribunals remains rare and largely off-record.
• Concern about hallucinations, citation accuracy, and confidentiality is the single biggest brake on broader adoption.

The fourth point is what makes citation-grounded, source-cited tooling — see Citation-Grounded LLMs — the natural fit for serious arbitration deployment. The technology that addresses the dominant concern is the technology that gets adopted at scale.

Institutional rule differences: ICC vs LCIA vs SIAC vs HKIAC vs AAA-ICDR

The five major institutional rule sets have approached AI from different starting positions. All five remain silent on AI in their procedural rules — tribunals’ general procedural discretion governs in the absence of explicit provisions. The table below captures the published guidance and most recent AI development at each institution as of mid-2026.

ICC. The International Chamber of Commerce has no standalone AI guidance note, and the 2021 Rules contain no explicit AI provisions. On 14 September 2024 the ICC Commission on Arbitration and ADR announced a Task Force on Artificial Intelligence in International Dispute Resolution; the Task Force is still being formed and its scope is still being defined.

LCIA. The London Court of International Arbitration has not published AI guidance. The 2020 Rules contain no explicit AI provisions. LCIA has facilitated internal discussions on AI in the context of broader rule-review but has not issued formal guidance.

SIAC. The Singapore International Arbitration Centre’s 2025 Rules (effective 1 January 2025) introduced procedural changes — ex parte emergency relief, third-party-funding rules, preliminary determination, streamlined procedure — but no explicit AI provisions. SIAC has publicly stated it “remains open to support the responsible and informed adoption of AI and technology in arbitration.”

HKIAC. The Hong Kong International Arbitration Centre’s 2024 Rules contain no explicit AI provisions. On 11 April 2025 HKIAC announced a partnership with Jus Mundi; under the partnership, free access to the HKIAC Case Digest went live on 1 July 2025, with Jus Mundi’s “Jus AI” GenAI technology preparing abstracts of HKIAC procedural decisions. HKIAC Secretary-General Joanne Lau described the partnership as a step that will “ultimately help guide users and improve overall efficiency and transparency.”

AAA-ICDR. The American Arbitration Association — International Centre for Dispute Resolution has the most developed institutional position. Published documents include the “AAA-ICDR Principles Supporting the Use of AI in Alternative Dispute Resolution” (1 November 2023); “Guidance on Arbitrators’ Use of AI Tools” (March 2025); and the AAAi Standards framework launched in 2024. Products in production include the AI Arbitrator (November 2025), ClauseBuilder AI Beta (June 2024), AAAi Panelist Search (October 2024), and the AAAi Chat Book (January 2025). The March 2025 Guidance asks arbitrators to retain complete control over decision-making and to disclose AI use that “materially impacts” the arbitration process or reasoning.

Cross-cutting observations from this landscape:

• No major institution prohibits AI use, and none expressly permits it either — all five rule sets are silent. Tribunals exercise general procedural discretion within the institution’s broader conduct framework.
• Disclosure expectations are converging. Where AI use materially affects an output (a brief, a witness statement summary, a procedural position), disclosure is increasingly the default expectation even where rules do not explicitly require it.
• The arbitrator’s responsibility is unchanged. AI tools accelerate the work but do not transfer responsibility. The arbitrator (or counsel) remains responsible for the output.
• Institutional governance is the strongest steer. Where institutions deploy AI tooling themselves — AAA-ICDR being the most active example — the governance framework around that deployment sets norms beyond the specific institution.

For tribunal secretaries and counsel preparing for cross-institutional work, the practical posture is: assume AI use needs to be disclosed where material, document the human oversight applied to AI outputs, and avoid AI tools that cannot show their work (no citation provenance, no audit trail). The institutional landscape is also moving quickly — particularly the ICC Task Force output — so positions should be re-verified against the institution’s own published materials at the time of deployment.

Where AI is being used: research, drafting, document review, hearings

The four main work areas where AI is in routine use across institutions.

Research and authority synthesis. Surveying case law, institutional precedents, doctrinal commentary. Source-cited RAG tools — that retrieve from a curated corpus and answer with citations — are the dominant pattern. Generic LLMs without grounding are recognised as unreliable for this work in serious practices.

Drafting first drafts. Briefs, witness statements, procedural correspondence. AI-assisted drafting is now common at major firms. The practice norm is human revision of the AI output before submission, with the AI assistance treated as accelerator rather than author.

Document review. Classification, summarisation, timeline extraction. See AI for Arbitration Document Review & E-Discovery for the per-task detail. The technology category here is mature (technology-assisted review has been in use for over a decade) and AI assistance is well-understood by tribunals.

Hearings. Real-time transcription with AI assistance is standard. AI-assisted summary of testimony for tribunal deliberation is emerging but more cautiously deployed. Live AI-assisted advocacy (real-time research, real-time citation) is technically possible but uncommon in serious cases — the disclosure and reliability questions are still being worked through.

In each of these, the AI is a productivity tool, not an authority. The disclosure and oversight expectations reflect this. Tools that try to be authorities (an AI that delivers a substantive merits position, an AI brief without human revision) generate the most caution and the most institutional pushback.

What the Ciarb Guideline (and equivalents) requires

The Chartered Institute of Arbitrators (Ciarb) published its Guideline on the Use of AI in Arbitration originally in March 2025 and issued an updated version on 5 September 2025. It is soft law — non-binding unless parties or the tribunal adopt it — organised into four Parts (context; general recommendations; arbitrators’ powers over parties’ AI use; arbitrators’ own AI use) plus appendices including a Template Agreement on the Use of AI and a Template Procedural Order. The Drafting Committee was chaired by Claire Morel de Westgaver, and the Guideline adopts the OECD definition of AI.

The Guideline anchors three obligations in particular: participants “remain fully responsible for their actions and submissions, regardless of whether AI tools were involved” (¶3.4); arbitrators “should not relinquish their decision-making powers to AI” and must ensure “independent judgement” (Part IV); and disclosure of AI use may be required where it materially affects the evidence, the outcome, or other procedural elements.

The recurring themes across the Guideline and equivalent professional-body work:

• Understanding the tool (¶3.1). Users are encouraged to make reasonable efforts to understand the technology, functionality, and data behind any AI tool used in arbitration.
• Risk assessment (¶3.2). Users should assess potential risks and benefits of AI use, especially in relation to due process, the rule of law, environmental impact, and the credibility of arbitration.
• Human responsibility (¶3.4). The arbitrator or counsel remains responsible for any output the AI produces. AI use does not transfer professional responsibility for the result.
• Transparency (¶9.1–9.2). Arbitrators are encouraged to consult parties before using AI tools and to avoid use where objections are raised.
• Disclosure to the parties where material. If the use of AI materially affects an output — the evidence, the outcome, or other procedural elements — the parties should know.
• Confidentiality. AI tools that process confidential matter — and that includes most arbitration material — should have an appropriate confidentiality posture. Vendor contractual protections, in-perimeter deployment, no-training-on-content guarantees.
• Citation and verification. AI outputs that cite authorities should cite verifiably. AI outputs that cannot be traced to sources should not be relied on as authority.

These themes map closely onto the architectural choices that make a tool deployable in arbitration: source-cited retrieval, refusal on out-of-corpus questions, audit logging, in-perimeter deployment for the most sensitive matters. The technology and the professional standards are converging on the same answer.

AI clauses in Terms of Reference — drafting guidance

Several institutions and tribunals have begun including AI-use clauses in Terms of Reference (or equivalent procedural orders early in the case). The drafting practice is still evolving; a few patterns are emerging.

Disclosure clauses. Parties (or counsel) commit to disclosing material use of AI in producing submissions, witness statements, or other case materials. The threshold for “material” varies; the trend is toward broader disclosure rather than narrower.

Permitted-use clauses. Tribunals specify which AI uses are presumptively acceptable (research, summarisation, transcription) and which require additional notice or process (first-draft generation of party submissions, AI-assisted witness preparation).

Confidentiality clauses. Specific provisions on the AI tools’ confidentiality posture — whether content can be sent to SaaS tools, whether vendor processing is acceptable, whether on-prem deployment is required for specific document categories.

Verification clauses. Where AI is used to identify authorities (case citations, statutory references, expert opinions), the relying party commits to having verified those authorities directly. This addresses the common failure mode of AI citing real-looking but fabricated authorities.

For counsel preparing TOR drafts in 2026, the safe posture is to address AI use explicitly rather than leaving it unspoken. The unspoken default — that any AI use is permitted as long as it does not violate the institutional rules — increasingly looks insufficient as adoption rises and the failure modes become better understood.

Cross-border privilege and data-residency considerations

This is the area where international cases differ most sharply from domestic ones, and where SaaS AI tools most often run into trouble.

Privilege rules vary by jurisdiction. What is privileged in one jurisdiction may not be in another. AI processing of materials whose privilege status is jurisdiction-dependent creates non-trivial complications. The conservative posture is to evaluate the privilege rules of every jurisdiction whose laws might apply to the matter and to deploy AI tools that are compatible with the most restrictive of them.

Data-residency rules vary similarly. GDPR is the most-discussed, but several other regimes (Swiss FINMA, German national, Singaporean PDPA in certain contexts) have specific positions on cross-border data processing. SaaS AI tools whose control plane sits in one jurisdiction may not satisfy data-residency requirements in another, regardless of where the underlying servers are located.

The cross-border combination is what hurts. A case with English-law contracts, Singapore-seat arbitration, German parties, and Swiss confidential information runs into privilege-rule fragmentation and data-residency fragmentation simultaneously. The deployment models that handle this cleanly — typically in-perimeter or VPC-in-jurisdiction — are different from the default SaaS setup most generic AI tools ship with.

For specifics on in-perimeter deployment patterns that address these requirements, see On-Premise RAG: Deployment Guide for Regulated Sectors.

Why source-verified RAG matters more in international cases

The general case for citation-grounded RAG — citation provenance, refusal on out-of-corpus questions, audit logging — applies to any regulated AI deployment. In international arbitration the case is sharper for three reasons.

Verification matters more across borders. A counsel relying on AI-derived authority for a Singapore-seated arbitration based on English law cannot easily verify the citation through casual reading. Page-level or claim-level citations let counsel verify quickly across legal traditions they may not be intimately familiar with.

Audit trails matter more across procedural cultures. Procedural cultures differ. What an AAA-ICDR tribunal expects as documentation of AI use may differ from what an LCIA tribunal expects. A defensible audit trail covers all reasonable institutional expectations.

Data-residency-compatible deployment matters more. A SaaS tool that works fine for a domestic US matter may run into hard problems for a cross-border matter with German confidential content. Tools deployable in-perimeter or in-jurisdiction handle this without making the case team re-architect their workflow.

The pattern that handles all three is the pattern this knowledge base argues for throughout: source-cited retrieval, configurable refusal behaviour, full audit logging, deployment-model flexibility. It is the same pattern whether the tool is the AAAi Chat Book or a private deployment for a specific case file.

Related reading

• AI for Arbitration Document Review & E-Discovery — the document-side workflow detail this article references.
• AI Arbitrator Tools Compared — institutional AI tooling for the AAA-ICDR specifically.
• On-Premise RAG: Deployment Guide for Regulated Sectors — for the in-perimeter deployments cross-border work often requires.
• Citation-Grounded LLMs — the category framing for why source-cited tooling is the right fit.

Frequently asked questions

Are AI-drafted submissions admissible in international arbitration?

Yes, generally — no major institution prohibits AI-assisted drafting. The submission’s admissibility depends on the same standards as any party submission: the responsible counsel adopts the submission, the content is accurate and properly cited, and any institutional disclosure expectations are met. The use of AI as a drafting tool is not in itself a basis for inadmissibility.

Do I need to disclose AI use to the tribunal?

Where AI use materially affects a submission, witness statement, or procedural position, yes — disclosure is the increasingly default expectation across major institutions, even where specific rules do not explicitly require it. The threshold for materiality is evolving; the prudent posture in 2026 is to err toward disclosure rather than silence.

Can I use SaaS AI tools for cross-border arbitration?

It depends on the case profile. For matters where data residency, privilege, or confidentiality rules of any relevant jurisdiction restrict cross-border processing, SaaS may not be appropriate. For matters without such restrictions, SaaS is usually acceptable provided the vendor’s contractual posture supports confidentiality (no training on content, appropriate data handling). The most conservative cross-border deployments use in-perimeter or VPC-in-jurisdiction tooling.

What does the Ciarb AI Guideline require?

The recurring themes across Ciarb and equivalent professional body guidance are: human responsibility for AI outputs is unchanged, disclosure where material is expected, competence to understand the tool used, confidentiality posture appropriate to the matter, and citation verifiability where authorities are cited. Specific provisions and current versions should be read from the published guideline directly.

How do AI clauses in Terms of Reference typically read in 2026?

The emerging patterns address four areas: disclosure of material AI use, permitted vs notice-requiring use cases, confidentiality posture of AI tools used (SaaS vs in-perimeter), and verification of AI-identified authorities. Drafting practice is still evolving and varies by institution and tribunal; addressing AI explicitly in the TOR is increasingly the safer default than leaving it unstated.

Which AI tools are accepted across all major institutions?

No tool has formal cross-institutional accreditation; institutions evaluate tools through their own governance processes. The architectural features that make a tool acceptable across institutions — citation provenance, refusal on out-of-corpus, audit logging, deployment-model flexibility — are well-defined. Tools with those features tend to satisfy institutional expectations broadly even without formal accreditation.

Can tribunals use AI for substantive decision-making?

Cautiously and within strict scope where institutions permit it; rarely for substantive merits decisions in commercial arbitration as of 2026. The pattern is procedural and administrative AI assistance under institutional governance, with substantive decisions remaining the tribunal’s. Specific institutional positions are evolving.

What about AI evidence (deepfakes, AI-generated documents)?

A distinct and serious issue: AI-generated content that ends up in evidence (witness statements drafted by AI without disclosure, deepfaked recordings, AI-generated documents purporting to be authentic) is a growing concern. Several institutions have begun addressing this through evidence-handling guidance. Tribunals should be prepared to test the authenticity of evidence in ways that did not require attention even five years ago.