AI arbitration case law: what the courts have decided, case by case

The short version

• In the cases we found, one court has annulled an arbitral award over AI use: the Superior Court of Québec, in April 2026, because the authorities the award relied on did not exist. This page is the courts sibling of the pillar, AI in arbitration.
• A US challenge on similar grounds, LaPaglia v. Valve, was dismissed in December 2025 for lack of jurisdiction, so no US court has ruled on whether AI use can vacate an award.
• Most AI-and-arbitration rulings so far are not about the award at all. They are courts dealing with non-existent citations in litigation about arbitration: motions to compel arbitration, an application to lift a stay, a removal under the New York Convention.
• No court we found has ruled on an AI arbitrator or on enforcing an award drafted with AI.
• The grounds a challenge uses already exist: exceeding powers, serious irregularity, public policy. From 1 January 2027, California also bars arbitrators from delegating any decision-making to generative AI.^{1,4,6,7,24,25,27,28}

Has any court set aside an award because the arbitrator used AI?

Yes, once, in the cases we found. In ARIHQ c. Santé Québec, 2026 QCCS 1360 (22 April 2026) the Superior Court of Québec annulled a sole arbitrator’s award: the doctrine and case law behind it were either invented or did not support the result, and the arbitrator had delegated part of the decision itself ([102] to [116]). The court set the award aside under articles 646 and 648 of Québec’s Code of Civil Procedure, not under the New York Convention.^1,2,25

The court drew limits too. Using AI is not itself a ground for annulment under article 646 ([69]), and a dependable tool is not a defect on its own ([94]); light use, or use on a side issue, could point the other way ([117] to [118]). The full reasoning, those limits and what is still open on appeal are in our explainer on awards challenged over AI use.¹

Was the award in LaPaglia v. Valve vacated?

No. The petition was dismissed before the court looked at AI. In LaPaglia v. Valve Corporation a consumer asked a federal court in California to vacate an award under the Federal Arbitration Act, arguing the arbitrator had outsourced the decision to AI (9 U.S.C. section 10(a)(3) and (a)(4); petition p. 2). The claims that the arbitrator used a chatbot and that the award carried the hallmarks of AI drafting were the petitioner’s, not findings by the court.^3,24

On 9 December 2025 the court dismissed the amended petition for lack of subject-matter jurisdiction (Doc. 18). There was no federal question, because the FAA does not by itself give a federal court jurisdiction, and a zero-dollar award did not meet the amount in controversy for diversity, so the court never reached the AI argument (pp. 7 to 12). Pages that still describe this case as pending pre-date the dismissal. The same facts, and why jurisdiction ended it, are in our set-aside explainer.^4,5

Which decisions are there, and what did each one decide?

This dated index of AI arbitration case law is ordered by what the court was deciding: the award challenged over AI use, non-existent citations in arbitration litigation, privilege and confidentiality of AI material, the AI arbitrator, and context cases that are not about arbitration.

Most entries are non-existent-citation rulings in litigation about arbitration. A few are award challenges, and none is about an AI arbitrator. The two set-aside cases, Québec and LaPaglia, are argued in full in our set-aside explainer.^1,4

What happens when a lawyer or party cites fake AI cases in arbitration?

Courts have refused the filing, remanded, or imposed costs, but in the cases we found they did so on motions about arbitration, not on the award itself. Two orders show the range.

• In Chenco Engineering v. Do-Fluoride (D. Idaho, 22 August 2025), a New York Convention removal dispute, a party’s motion for leave to file a surreply contained at least five case quotations that “do not exist”. The court denied leave to file the surreply and remanded (pp. 9 to 11). It made no finding that AI was used, so this is a citations-that-do-not-exist ruling, not an AI ruling.
• In Tercero v. Sacramento Logistics (E.D. Cal., 9 September 2025), a motion to reconsider an order compelling arbitration cited cases that do not exist and others that did not support the point. Counsel said AI was not used. The court still imposed a US$1,500 personal sanction under Rule 11 and directed the clerk to serve the order on the State Bar of California (pp. 23 to 24).
• In Deleman v. HighLevel (M.D. Pa., 5 January 2026), a motion to lift a stay pending arbitration was denied where the exhibits were what the court called an AI-generated statement of the terms and an AI-generated cost estimate, which could not prove the arbitration agreement was unconscionable (p. 10 n.5, p. 11).
• In Blue Cross Blue Shield of Georgia v. HaloMD (N.D. Ga., 10 July 2026), an insurer alleged that the filing company used AI to flood the federal surprise-billing process with ineligible disputes. The court held that review of a determination is limited to the FAA section 10(a) grounds and dismissed with prejudice, never deciding the AI allegation (pp. 20, 56 to 69).^6,7,8,9

Several self-represented parties opposing motions to compel arbitration have also filed non-existent citations in 2025 and 2026; the courts noted them and still decided the arbitration question. How courts sanction lawyers for AI-invented citations outside arbitration is in our guide to AI in law. The duty to verify every authority is the same one the Québec court applied to a decision-maker.^1,13,14,15

What have arbitral tribunals themselves done about AI-invented authorities?

In the cases we could verify against an official or institutional source, none. Every ruling so far has come from a court, not a tribunal. The arbitration-related orders above, such as Chenco and Tercero, were made by courts reviewing or supervising arbitration, not by the tribunals.^6,7

Arbitral opinions refusing motions built on fictitious citations have been reported, at the AAA and at JAMS, but we have not been able to confirm them against an official copy, so we do not set them out here. What the guidance requires is clearer: the Silicon Valley Arbitration & Mediation Center keeps the decision and the duty to verify with the arbitrator (SVAMC Guideline 6), and the Chartered Institute of Arbitrators does the same (Ciarb).^30,31

Are conversations with an AI tool confidential or privileged?

US courts that have looked at it say chats with a public AI tool are generally not privileged and not automatically work product, the US shield for material prepared for litigation. That is a direct confidentiality warning for anyone putting case material into a tool.¹⁰

• In United States v. Heppner (S.D.N.Y., 17 February 2026), a defendant’s chats with a public AI platform, run on his own initiative, were neither privileged, because the tool is not a lawyer and the chats were not confidential given the provider’s data terms, nor work product, because they were not made by or at the direction of counsel (pp. 4 to 12). The court left open that counsel-directed use might differ (p. 7).
• In Morgan v. V2X (D. Colo., 30 March 2026), a self-represented party’s AI work could be work product, and AI use does not automatically waive it, but he had to name the AI tool used with confidential material. The protective order was amended to bar AI tools for confidential material unless the provider is contractually barred from using inputs to train its model and from passing them to a third party (pp. 5 to 16).
• In Fontainebleau v. Botach (S.D. Fla., 21 October 2026), the court held that AI chats are not categorically work product and that AI use “does not automatically establish either discoverability or protection”; it declined to compel consent to subpoenas aimed at the AI providers (pp. 16 to 23).
• Clearly labelled, not an arbitration case: a UK tribunal held that putting confidential documents into an open-source AI tool breaches confidentiality and places them in the public domain, while closed tools can summarise without those risks ([2026] UKUT 81 (IAC), headnote 4 and [21]).^10,11,12,23

What this means for confidentiality, privilege and data protection when AI touches case material is in confidentiality and AI in arbitration, and how to vet a tool before it touches case data is in AI legal tech.

Has any court ruled on an AI arbitrator, or on enforcing an award made with AI?

No, in the cases we found. The only US challenge to an award over alleged AI use, LaPaglia, ended on jurisdiction, and no court has ruled on an award drafted by an AI or made by an AI arbitrator. There are two different things here: an AI-assisted award, where a person decides and signs, and an AI-rendered award, where the decision is handed to a machine. The cases so far reach neither.⁴

The grounds a challenge would use exist, but are untested here. On enforcement abroad under the New York Convention, a court may refuse where a party was unable to present its case (art. V(1)(b)), where the procedure departed from the parties’ agreement (art. V(1)(d)), or on public policy (art. V(2)(b)); the UNCITRAL Model Law’s grounds to set aside mirror them (art. 34(2)(a)(ii), (iv) and (b)(ii)).^25,26

How the AAA’s opt-in AI arbitrator works is in AAA and AI in arbitration (tracker entry); whether a human-signed, AI-assisted award can be recognised and enforced is in AI and award enforcement.³²

Which non-arbitration decisions set the direction?

Several widely cited rulings are not arbitration cases, but they are what tribunals and supervising courts will reason from. Each is marked.

• DJP v DJO [2025] SGCA(I) 2, upholding DJO v DJP [2024] SGHC(I) 24, set an ICC award aside for breach of natural justice where at least 212 of the award’s 451 paragraphs were copied from parallel awards ([2]); the Singapore Court of Appeal dismissed the appeal ([89]). Neither judgment mentions AI, but it is the closest analogue to a copied-reasoning or delegated-drafting challenge. For how Singapore’s courts set aside and enforce awards, see arbitration in Singapore.
• Ayinde v Haringey [2025] EWHC 1383 (Admin) and Mata v. Avianca (S.D.N.Y., 22 June 2023): both held the advocate, not the tool, responsible for every citation. Mata imposed a US$5,000 sanction on the lawyers and their firm for standing by opinions that did not exist, though the court saw no fault in a dependable tool used with care.
• Kupetz v Terraform Labs [2026] SGHC(I) 9 (Singapore International Commercial Court, 29 June 2026): AI may assemble the form of a witness statement but must not supply its substance ([30] to [31]).
• Pooja Ramesh Singh v. J&K Bank, 2026 INSC 668 (Supreme Court of India, 2 July 2026): tribunal orders resting on hallucinated precedents were set aside; such a decision, the court said, is “no decision in the eyes of the law” (para 7).
• Jackson Federation of Teachers v. Fitch (5th Cir., 22 October 2026): a trial judge’s own orders carried AI-style errors and a case that does not exist; the US Court of Appeals for the Fifth Circuit vacated the injunction and directed that another judge take the case (pp. 2, 4, 27).^{16,17,18,19,20,21,22}

Each turns on hallucination: a filing or a decision built on authorities that do not exist. The lawyer-sanction and judge’s-own-use line is tracked in our guide to AI in law.

What is still open?

• How much AI use is too much. Québec left minimal use open, and we found no US ruling on the point.
• Whether a US court would vacate an award for AI use. The one challenge ended on jurisdiction.
• Enforcement under the New York Convention where the decision was handed to AI. We found no reported decision.
• Whether an award drafted mainly by AI but signed by a human is the arbitrator’s own reasoned award (Model Law art. 31).
• What counts as delegating “any part” of decision-making under California’s section 1282.1, in force 1 January 2027.
• Whether any arbitral tribunal’s ruling on AI-invented authorities will be published and citable.^1,4,25,26,28