Can an award made with AI, or by an AI arbitrator, be enforced?

The short version

• No court we found has refused to enforce, or enforced, an arbitral award because AI was used. The two AI cases so far are a Québec annulment and a United States petition thrown out on jurisdiction.
• The New York Convention, the 1958 treaty on recognising and enforcing foreign awards, sets no rule that an arbitrator be human and none that an award be signed. A challenge has to fit Article V.
• The UNCITRAL Secretariat has written that an AI tool that filtered out essential evidence, or a decision handed to AI, could ground refusal under Article V(1)(b) and (d).
• Some seats need a natural-person arbitrator, among them France for domestic cases, the Netherlands, Spain, Portugal and Scotland; Brazil requires a capable person and voids an award from someone who could not be an arbitrator. An AI-only tribunal risks Article V(1)(d) at a natural-person seat.
• AAA’s AI arbitrator has a human review and sign the award. That human-signed, AI-assisted award is the lower-risk path, but no court has tested one.
• From 2 December 2027 the EU AI Act can treat AI that helps decide a dispute with legal effects as high-risk, a duty on providers and deployers, not a rule that an award is invalid.^{1,5,8,10,12,15,16,17,18,19,29,34}

Has a court ever refused to enforce an award because AI was used?

No. In the cases we found, no court has refused or granted recognition because AI was used. The two AI cases are a seat set-aside and a vacatur petition dismissed before the court reached AI. Neither tests enforceability under the New York Convention; there is no ruling yet. Every AI decision, by category, is in AI arbitration case law.^8,10

In ARIHQ c. Santé Québec, 2026 QCCS 1360, the Superior Court of Québec annulled a sole arbitrator’s award after finding that the authorities it relied on did not exist or did not support it, and that the arbitrator had delegated the decision ([114] to [116], [122]). That was a domestic annulment, decided under articles 646 and 648 of Québec’s Code of Civil Procedure; the New York Convention, which governs foreign awards, did not apply.^8,9

In LaPaglia v. Valve Corporation the losing party asked a federal court to vacate an AAA award, arguing the arbitrator had handed his role to AI. On 9 December 2025 the court dismissed the petition for want of subject-matter jurisdiction and never reached the AI question (Order, pp. 7 to 12). The order records the respondent’s statement that a Washington state court had confirmed the award (p. 3, fn. 3); it does not say AI was argued there.¹⁰

Does the New York Convention require a human arbitrator?

No. The Convention never says an arbitrator must be a natural person (a human being, not a company). It refers to awards made by “arbitrators appointed for each case” and by “permanent arbitral bodies” (art. I(2)). There is no human-arbitrator or signature rule in the text. So whether an AI-rendered arbitral award is enforceable turns on national law at the seat and at the enforcing state, funnelled through Article V: the seat decides whether the tribunal was validly composed, and the enforcing state decides recognition (exequatur) without re-running the merits.¹

States are moving slowly. Several States have told the Secretariat that final decisions must be taken by natural persons (A/CN.9/1271 paras 12 to 14, 20 to 21), and Working Group II’s colloquium took stock of what AI means for recognition and enforcement (A/CN.9/1237); UNCITRAL has so far kept the subject at an exploratory stage (A/81/17 paras 170, 172). For whether AI can be an arbitrator at all, including a human in the loop and non-delegation of the mandate, see our pillar guide to AI in arbitration.^4,6,7

Tribunal composition: can an AI arbitrator satisfy Article V(1)(d)?

Article V(1)(d) lets a court refuse where the composition of the tribunal or the procedure was not as the parties agreed or, failing agreement, not in line with the law of the seat. That is the most direct route against an AI-only tribunal seated where the law needs a human. The resisting party must raise and prove it, and courts have sometimes read the ground restrictively (NYC Guide 2016, art. V(1)(d)).^1,2

What is settled: at a seat that requires a natural person, an AI sitting alone as arbitrator is not a validly composed tribunal. What is argument: whether a human-signed, AI-assisted award is composed “not as agreed” where the parties expected human arbitrators. The UNCITRAL Secretariat frames heavy reliance on AI in deciding as a possible Article V(1)(d) ground (A/CN.9/1215 para 42), though no court has so held.⁵

Seat law and enforcement can come apart. Article 1450 of France’s Code de procédure civile allows only a natural person to act as arbitrator, but article 1506 does not apply it to international arbitration. A March 2025 Ministry of Justice reform report would keep that rule for arbitrators in France, “surtout à l’heure où l’on voit apparaître des projets d’arbitrages entièrement générés par l’Intelligence Artificielle” (our translation: especially now that wholly AI-generated arbitrations are appearing, Proposition 17), while still allowing recognition in France of foreign awards made or signed by legal persons (personnes morales), even where natural persons in fact decided.^12,14

For a seat not in the table, see the relevant guide, such as arbitration in London, Singapore, India, the Netherlands or Hong Kong.

The right to be heard: does AI use engage Article V(1)(b)?

Article V(1)(b) lets a court refuse where a party was “unable to present his case”, the same equality-of-arms idea as the Model Law’s “full opportunity of presenting his case” (art. 18). It is the route for AI that filters what the tribunal reads, or springs reasoning a party could not meet. Weighing evidence, the UNCITRAL Secretariat says, is “inherently human and cannot be delegated”, and a decision not based on the tribunal’s own assessment could ground refusal under Article V(1)(b) and (d) (A/CN.9/1215 paras 35, 41 and 42). That an opaque output offends the right to be heard is argument, not a decided case; how a tribunal may and may not draft with AI is in AI and award drafting.^1,3,5

Public policy: can an award made with AI be refused under Article V(2)(b)?

In principle yes, but the bar is high. Article V(2)(b) is one of the two grounds a court may raise on its own motion, the other being non-arbitrability under Article V(2)(a), and the Secretariat Guide calls it a “safety valve” for exceptional cases. The grounds are exhaustive with no review of the merits, so “the AI got the law wrong” is a merits attack the Convention does not allow; it must be reframed as procedural public policy, composition or the right to be heard.^1,2

The argument, and it is only argument, is that undisclosed or unexplained AI drafting offends procedural public policy. Informed consent has limits: parties can agree to an AI-assisted process and so waive the party-raised Article V(1) grounds, but consent cannot cure the court’s own-motion public-policy ground. An award must apply the law the parties chose (Model Law art. 28(1)) unless they authorised the tribunal to decide ex aequo et bono, that is by what it judges fair (art. 28(3)), and state reasons unless the parties agree otherwise (art. 31(2)).^1,3

The closest analogue is not an AI case. In DJP and others v DJO [2025] SGCA(I) 2, upholding DJO v DJP [2024] SGHC(I) 24, the Singapore courts set an ICC award aside for breach of natural justice where at least 212 of its 451 paragraphs had been copied from parallel awards. A reviewing court, it held, is concerned with the integrity of the process, not the correctness of the result ([50]). Neither judgment mentions AI, but it is the closest analogue to a delegated-drafting challenge.¹¹

Signature and form: is an AI-drafted award duly authenticated?

Only when a human signs it. The Convention sets no signature rule, but Article IV(1)(a) makes the party enforcing supply the “duly authenticated original award” or a certified copy, and the Guide treats authentication as confirming the arbitrators’ signatures are genuine (NYC Guide 2016, art. IV). The Model Law requires a written award signed by the arbitrator or arbitrators (art. 31(1)); its 2026 amendments let an award be in electronic form and bar refusal “solely on the ground that it is in electronic form” (new arts 2(g), 31(5), 35(3), A/81/17 annex IV), but the signature requirement is unchanged.^1,2,3,4

France’s Décret n° 2026-741, from 1 January 2027, lets an award be on paper or digital (new art. 1480-1); a digital award “est signée au moyen d’un procédé de signature électronique qualifiée” (new art. 1480-2 al. 3), though the circular says that qualified-signature rule does not reach international arbitration (art. 1506 carries over only art. 1480-1). Under the EU’s eIDAS Regulation a “signatory” is “a natural person who creates an electronic signature” (art. 3(9)), with “the equivalent legal effect of a handwritten signature” (art. 25(2)). So the signature points back to a human: an arbitrator who reviews and signs satisfies the form; an award signed only by an AI has no natural-person signatory to authenticate.^13,27

Set aside at the seat: does that block enforcement abroad? Article V(1)(e)

An award set aside or suspended at its seat is a ground to refuse enforcement elsewhere under Article V(1)(e). But refusal is discretionary: Article V says recognition “may be refused”, so a court that finds a ground proven still has a discretion to enforce, and Article VII lets a state apply a more favourable regime (NYC Guide 2016, art. V discretion). An annulled award is not automatically dead abroad.^1,2

If the annulled Québec award were presented for enforcement abroad, Article V(1)(e) is the ground a resisting party would reach for, and a vacatur at the seat of an international award would work the same way; that the Québec reasoning carries across is our inference, not a holding. The set-aside cases are argued in full in awards challenged over AI use. An ICSID Convention award is a separate track, enforced under its own article 54, not the New York Convention (arts 53, 54).^8,26

Is an award from an AI arbitrator, like AAA’s, enforceable?

AAA’s AI-led process is built so that a human arbitrator reviews, may rewrite and signs the award, which is designed to keep it enforceable, though no court we found has yet confirmed or refused one. Under the AI Led Arbitration Rules, effective 18 February 2026, both parties consent to this process (R-2(c)), which runs on written submissions with no hearing (R-6; AAA FAQ); the human arbitrator “can approve, edit, or rewrite” the AI’s draft and then “finalizes and signs the award” (R-8); and judgment may be entered in any federal or state court (R-12(a)).^34,35

In the US the confirm route is statutory: within three years a court “shall confirm the award unless it finds one of the grounds for refusal or deferral” in the Convention (9 U.S.C. 207), but whether a court will confirm an award from an AI arbitrator is untested. The rules, fees and reported results are in AAA and AI in arbitration; the tracker has an AAA AI arbitrator entry.²¹

Does the EU AI Act change whether an award is enforceable?

Not directly: the Act is product-safety law aimed at providers and deployers, not a rule on whether an award is valid. Annex III point 8(a) can class AI used “in a similar way in alternative dispute resolution” as high-risk; recital 61 limits that to ADR outcomes with legal effects for the parties; and article 14 requires oversight by natural persons. It never says an award is invalid, so its only link to enforcement runs through the Article V public-policy and procedure arguments.²⁸

The timing has moved: under Regulation (EU) 2026/1744 the high-risk obligations apply from 2 December 2027, not the 2 August 2026 date often quoted. Mapping a breach of the Act onto Article V(1)(d) or V(2)(b) is argument only; no reported decision we found has refused an award on that basis, and a product-safety breach does not by itself invalidate an award. The full role-and-duty read is in the EU AI Act and dispute resolution.^28,29

How to protect an AI-assisted award’s enforceability

The safer path is simple: the human decides, AI only assists, and that use is recorded and disclosed to the parties. Every institution text we read keeps the decision with the arbitrator, from SVAMC and Ciarb to the SCC and VIAC texts.^30,31,32,33

• Keep the decision with the arbitrator and do not delegate any part of it (SVAMC Guideline 6; Ciarb paras 8.2 to 8.4; and California’s Code of Civil Procedure section 1282.1(b)(1), from 1 January 2027, which bars delegating “any part of their decisionmaking process” to generative AI).
• Raise AI at the first case management conference, record what was agreed, and consult the parties before the tribunal uses its own AI tools (Ciarb paras 5.3, 4.6 and 9.1, and its template agreement and procedural order).
• Check every authority against an authoritative source before it enters the award (Ciarb para 8.3). Invented or hallucinated authorities annulled the Québec award (2026 QCCS 1360).
• Have a human arbitrator review and sign, and keep a record (AAA FAQ; AAA Rules R-8; AAA-ICDR Guidance, January 2026, Considerations 1 and 3).
• Disclosure helps: noting the tools used is good practice, not a requirement in any rule we read.^{8,30,31,34,35,36,37}

Open questions, all unresolved for enforcement: whether an award drafted mainly by AI but signed by a human still counts as the arbitrator’s own reasoned award; whether Article V(1)(d) or V(2)(b) bites where the deciding itself was handed to AI; and, as the Québec court left open, how much AI use is too much. The pillar guide sets out the wider list.

England is silent: the Arbitration Act 1996 sets the award’s form and mirrors the Article V grounds (ss. 52, 103), the Arbitration Act 2025 says nothing on AI, and the Law Commission’s 2023 report (Law Com No 413, 5 September 2023) did not take it forward.^22,23,24,25