AI in arbitration: what it does today and where the rules draw the line

The short version

• Most AI use in arbitration is support work, mainly research, document review and data analytics.
• Institutional guidance agrees that the arbitrator decides, and differs on when AI use must be disclosed.
• From 1 January 2027, California bars arbitrators from delegating any part of their decision-making to generative AI.
• France (for domestic cases), Portugal and Scotland allow only a human arbitrator, and AAA’s AI-led process keeps a human who signs the award.
• One court we found has annulled an award challenged over AI use (Québec, April 2026).
• From 2 December 2027, the EU AI Act can treat AI that helps decide an arbitration with legal effects as high-risk.^{1,2,3,4,5,7,8,9,10,11,33,35}

What is AI in arbitration?

Artificial intelligence (AI) in arbitration is the use of AI tools in a case by the parties and their lawyers, the arbitrators or the institution running it. The Chartered Institute of Arbitrators (Ciarb) adopts the OECD’s definition: a computer system that works out from its input how to produce predictions, content, recommendations or decisions.^3,44

The American Arbitration Association (AAA) separates AI-assisted arbitration, where AI tools support a traditional case that a person decides (a human in the loop), from AI-led arbitration, where the AI makes a “preliminary decision” that a human arbitrator reviews and finalises (R-2(c)). No institution we checked offers a fully automated award.^8,44

How is AI used in arbitration today?

AI in arbitration is used mostly for research and document work. In the 2025 Queen Mary University of London survey (2,402 responses, 117 interviews), 64% had used AI for factual and legal research in the past five years and 91% expect to in the next five (printed pp. 27, 37). The main reasons were saving parties’ and lawyers’ time (54%), cutting costs (44%) and reducing human error (39%) (Chart 19).¹

• Document review: a 2022 International Chamber of Commerce (ICC) Commission report lists time and cost savings and better accuracy among the benefits of predictive coding, machine learning that finds relevant documents (section 5.2).
• The record: a tribunal at the International Centre for Settlement of Investment Disputes (ICSID) said in its first procedural order that it may use AI for legal research and to summarise, search and organise the record, without delegating any decision (ARB/24/46, section 28).
• Transcription: AAA offers AI transcription with two levels of human review and says it reaches 99% word accuracy. The ICC report warned that free live transcription may not be accurate enough for the record (section 6.1.4).
• Choosing arbitrators: AAA says its AI panellist search cut case managers’ time preparing arbitrator lists by 55%, and it offers an AI clause builder.
• Tribunal tasks: assuming the tribunal oversees and reviews the output, 77% of respondents think it appropriate for arbitrators to use AI to help calculate damages and costs, 66% to summarise submissions or evidence and 60% to draft procedural orders and procedural parts of awards, but only 23% to help draft an award’s reasoning (Chart 21).^{1,12,13,14,15}

The Dubai International Arbitration Centre (DIAC) gives its case management team an AI research and workflow tool, and the Hong Kong International Arbitration Centre (HKIAC) uses AI to summarise procedural decisions for its Case Digest.^17,18

What are the risks of AI in arbitration?

The main risks are undetected errors, bias and leaks of confidential material. In the Queen Mary survey, 51% cited undetected errors and bias, 47% confidentiality, 44% lack of experience and 38% missing rules, and 28% feared due-process challenges (Chart 20). The State of AI in U.S. Arbitration 2026, a study AAA co-published, put inaccuracy or hallucination first (78%).^1,16

• Invented authorities: tools can cite sources, legislation or evidence that do not exist, known as hallucination (Ciarb para 2.6). Every authority in the Québec award either did not exist or did not support the point ([102] to [113]).
• Confidentiality: WIPO says to treat AI tools as potentially disclosing any input unless clear, verifiable safeguards are in place (section 4).
• The right to be heard: a tool that filters what the arbitrator reads can undercut a party’s right to present its case (Ciarb para 2.5).
• Bias: training data and system design can carry bias into outputs (Ciarb para 2.4; WIPO section 2).
• AI-generated evidence: before admitting it, a tribunal can ask how material in the record links to the output (Ciarb para 6.6).
• Reasons: under the Model Law of the United Nations Commission on International Trade Law (UNCITRAL), an award must state reasons unless the parties agree otherwise (art. 31(2)). “Black box” outputs make that harder, and the arbitrator stays responsible for the whole award (Ciarb paras 2.6, 8.4).
• Delegation: the power to decide is personal and cannot be delegated (Guideline 6 of the Silicon Valley Arbitration & Mediation Center, SVAMC; Québec judgment [81] to [86]).^2,3,9,19,21

What rules and guidelines govern AI in arbitration?

Most arbitral institutions publish AI guidance rather than rules, and that guidance agrees the arbitrator decides. AAA’s AI Led Arbitration Rules govern its opt-in AI cases, and the 2026 rules of the Korean Commercial Arbitration Board (KCAB) encourage tribunals to discuss AI tools with the parties at the first case management conference or later (art. 16.5).^2,3,8,19,22

California’s SB 574, approved on 30 September 2026, adds section 1282.1 to the Code of Civil Procedure. An arbitrator “shall not delegate any part of their decisionmaking process to any generative artificial intelligence tool” ((b)(1)), nor rely on generative AI information outside the record without first telling the parties and, as far as practical, letting them comment ((b)(2)). It takes effect on 1 January 2027 under California’s constitution, and its wording is close to SVAMC Guidelines 6 and 7.^2,4,5

The texts differ most on disclosure. SVAMC says disclosure is not generally necessary and should be decided case by case, for parties and arbitrators alike (Guideline 3). Ciarb lets the tribunal require parties to disclose (paras 4.4, 7.1 to 7.7). The AAA-ICDR guidance (from AAA and its International Centre for Dispute Resolution) asks arbitrators to disclose AI use that “materially impacts” the process or the reasoning.^2,3,25,26

The London Court of International Arbitration (LCIA) rules consultation covers AI disclosure, and JAMS’s AI rules govern disputes about AI systems rather than a neutral’s own use of AI. The AI rules tracker summarises each text and links to it.^31,32

Can AI be an arbitrator?

Not where the law requires a human being. France (for domestic arbitration), Portugal and Scotland allow only a natural person (a human, not a company) or an individual to act as arbitrator. Elsewhere the law is silent or written with people in mind, and no court we found has ruled on a tribunal that is not human.^{6,7,21,33,34,35,36}

No rule we checked lets AI replace the arbitrator. At its 2026 session UNCITRAL kept its AI work exploratory, after it was stressed that the subject was “not yet sufficiently mature to warrant legislative work” (A/81/17, paras 170, 172).⁴⁰

How does AAA’s AI arbitrator work?

AAA’s AI arbitrator is an opt-in, documents-only process: an AI drafts the award and a human arbitrator reviews it, may rewrite it and signs it. AAA says it was trained on more than 1,500 construction awards, and it opened on 3 November 2025 for two-party construction cases, the scope AAA’s pages still give (tracker entry). The Rules apply to “all two-party disputes” on AAA’s platform (Introduction), but AAA says its current process does not suit complex, multi-party or high-value cases, or cases needing a hearing.^{8,41,42,43,44}

• Consent: both parties opt in (R-2(c)). If the respondent declines and the parties had agreed to AAA arbitration, the case moves to AAA’s usual process (R-4(b)(iii)); either way the claimant’s fee is refunded (Arbitration Cost; Refund Schedule).
• Procedure: written stages of 10 business days each (R-6(a)), on the parties’ submissions only (R-6(c)); parties can comment on the AI’s summaries (R-6(g)).
• Appointment: the AI picks the human arbitrator from AAA’s National Roster, and an objection leads to a replacement (R-7).
• Award: the arbitrator can approve, edit or rewrite the AI’s draft before signing, within five business days of the feedback deadline (R-8); parties do not see the draft (AAA FAQ).
• Cost and terms: US$2,500 to file claims and the same for counterclaims. Judgment on the award may be entered in any federal or state court (R-12(a)) unless the parties agree at the start to a non-binding award (R-13(a)); AAA and the arbitrator cannot be held liable or called to testify (R-12(c), (d)).
• Limits: AAA’s Release 1.6 notes say the platform takes only monetary claims and binding arbitration is not available in California.^8,43,45

AAA says case data and outputs are kept out of model training and its controls align with the SOC 2 Type 2 security audit standard. It reports “early results” of resolution 20 to 25% faster at a cost at least 35% lower in two-party, documents-only construction cases, with no published method, sample or case count. A separate Resolution Simulator gives one party a non-binding simulated decision, with no human arbitrator and no award.^43,46,47,48

Can an award made with AI be enforced?

No court we found has ruled on enforcing an AI-drafted award under the New York Convention, the 1958 treaty on foreign awards, but the grounds a challenge would use exist. A court may refuse where a party was “unable to present his case” (art. V(1)(b)), where the tribunal or procedure departed from the parties’ agreement or, failing one, the law of the place of arbitration (art. V(1)(d)), or on public policy grounds (art. V(2)(b)).⁴⁹

The Convention has no express signature or reasons rule, but the applicant must supply the “duly authenticated original award” or a certified copy (art. IV(1)(a)). The Model Law requires a written award signed by the arbitrators (a majority can suffice), with reasons unless the parties agree otherwise or it records a settlement (art. 31(1), (2)). Its grounds to set aside an award mirror the Convention’s (art. 34(2)(a)(ii), (iv)).^21,49

ICSID awards can be annulled only on the ICSID Convention’s grounds, including a serious departure from a fundamental rule of procedure or a failure to state reasons (arts 52(1)(d), (e), 53). English law allows a challenge for serious irregularity causing substantial injustice (Arbitration Act 1996, s 68(1), (2)), and a US court may vacate an award where the arbitrators “exceeded their powers” (9 U.S.C. section 10(a)(4)).^50,51,52

The Québec annulment was a challenge under articles 646 and 648 of Québec’s Code of Civil Procedure, not a New York Convention case ([3], [33] to [36], [114] to [116]).^9,53

Does the EU AI Act apply to arbitration?

It can, mainly where AI helps decide the case. The EU AI Act, Regulation (EU) 2024/1689, can classify such AI as high-risk: Annex III point 8(a) covers AI that helps a court research and interpret facts and law and apply the law to facts, “or to be used in a similar way in alternative dispute resolution” (ADR). Recital 61 limits this to ADR outcomes with legal effects for the parties and excludes purely ancillary administrative activities.^10,54

There is an exception. A system is not high-risk if it poses no significant risk of harm, for example because it does not materially influence the outcome, and it only does a narrow procedural task, improves work a person has finished, spots patterns in past decisions without replacing human review, or does a preparatory task (art. 6(3)). The exception never covers a system that profiles people. A provider relying on it must document its assessment and register the system (arts 6(4), 49(2)).¹⁰

The European Commission’s draft guidelines of 19 May 2026, not yet adopted, treat arbitration as ADR (para 426), say ADR bodies may include arbitral institutions (para 429) and say awards “generally have binding and final effect” (para 431). Search tools, even advanced ones, are generally outside; tools that interpret sources or infer legal reasoning are likely inside (paras 421, 422).⁵⁵

Duties fall on providers and deployers. A deployer must follow the instructions for use, assign human oversight, keep the logs it controls, generally for at least six months, and inform the people a system helps decide about (art. 26). Anyone who rebrands or substantially modifies a high-risk system, or repurposes a system so it becomes high-risk, is treated as its provider (art. 25(1)). Providers and deployers outside the EU are covered where the output is used in the EU (art. 2(1)(c)).¹⁰

Regulation (EU) 2026/1744 set 2 December 2027 for these obligations (art. 1 point (40)(b)). Systems already on the market by then are covered only after a significant design change, except those used by public authorities, which must comply by 2 August 2030 (art. 1 point (39)(a)).^11,56

On 23 July 2026, 16 arbitral institutions and bodies, including ICC, ICSID, AAA-ICDR and the LCIA, asked the Commission to clarify that the label covers AI used by or for the tribunal in deciding, not an institution’s administrative tools (LCIA entry).^57,58

Consumer cases have a separate rule. Directive (EU) 2025/2647 requires consumer ADR bodies that use automated decision-making to tell parties in advance and allow review by a natural person, under national rules applying from 20 September 2028 (art. 1 point (4)(b), art. 5).⁵⁹

What have courts decided about AI use?

In the cases we found (how we search), one court has annulled an arbitral award over AI use: the Superior Court of Québec, in ARIHQ c. Santé Québec, 2026 QCCS 1360. It found that the award’s authorities did not exist or did not support it and that the arbitrator’s authority had been delegated ([102] to [116]), but said a reliable AI tool is not wrong in itself ([94]) and minimal use may not justify annulment ([117], [118]). Our explainer on awards challenged over AI use has the detail.⁹

In LaPaglia v. Valve Corporation, a petition to vacate an AAA award alleged that the arbitrator had outsourced the decision to AI. On 9 December 2025 the federal court dismissed the case for lack of jurisdiction over the subject matter, without ruling on AI (Order, pp. 7 to 12; why the court never reached AI).^60,61

In Ayinde, [2025] EWHC 1383 (Admin), a case about lawyers’ filings, the High Court of England and Wales held that lawyers must check AI research against authoritative sources ([7], [8]). In Mata v. Avianca (22 June 2023), a New York federal court found nothing “inherently improper” in using a reliable AI tool but fined the lawyers and their firm US$5,000 for submitting non-existent opinions and standing by them.^62,63

Courts’ guidance for their own judges, which does not cover arbitration, also keeps the decision human. In England and Wales, the judiciary’s Position Statement of 9 October 2026 says “AI tools must not determine the outcome of a case” (para 10). The Canadian Judicial Council says no judge may delegate decision-making to a computer program.^64,65

What can tribunals and institutions do about AI now?

• Raise AI at the first case management conference, the tribunal’s first procedural meeting with the parties (KCAB art. 16.5; Ciarb paras 5.3, 5.4; VIAC para 5.1).
• Record what was agreed in Procedural Order No. 1 (Ciarb para 4.6; VIAC para 5.3).
• Start from Ciarb’s template agreement and procedural order (Appendices A and B) or SVAMC’s model clause.
• Choose a standard for disclosing AI use and name its source, such as SVAMC Guideline 3 or Ciarb paras 7.1 to 7.7 (compared in our tracker).
• Check every authority against an authoritative source, such as an official database or a reputable legal publisher, before it goes into a submission, order or award (Ayinde [7], [8], on lawyers).
• Keep confidential material out of AI tools without verified safeguards (WIPO section 4).
• Tell the parties how the tribunal will use AI and let them comment before it relies on AI-generated information outside the record; if they disagree on a tool, Ciarb says not to use it (Ciarb para 9.1; SVAMC Guideline 7). Where California’s section 1282.1 applies, the comment step is a statutory duty for generative AI from 1 January 2027 ((b)(2)).
• Institutions: separate administrative and search uses from tools that interpret the law or draft decisions (Recital 61; draft guidelines paras 419, 421 to 423), and know whether you are the provider or the deployer of each tool before 2 December 2027.^{2,3,4,5,10,11,19,22,28,55,62}

Open questions on AI in arbitration

• Whether an award drafted mainly by AI but signed by a human is the arbitrator’s own reasoned award (Model Law art. 31).
• Enforcement under article V(1)(d) or V(2)(b) of the New York Convention where the decision was handed to AI. We found no reported decision.
• The status of a tribunal that is not human under laws that are silent, including in international arbitrations seated in France, where article 1450 does not apply.
• How much AI use is too much. The Québec court left minimal use open ([117], [118]), and we found no US ruling on the point.
• What counts as delegating “any part” of decision-making under California’s section 1282.1.
• How the AI Act’s deployer duties (arts 26(11), 27, 86) apply to a sole arbitrator or a private institution.
• What ICC (its AI task force), Ciarb and the LCIA will publish next.
• Whether AAA takes its AI arbitrator beyond construction. It said in 2025 that it would do so in 2026.^{4,9,10,21,30,31,33,41,49,55,61,66}