ARIHQ v Santé Québec: When Must an AI-Drafted Award Be Set Aside?

Artificial intelligence (“AI”) is no longer just confined to improving the efficiency of arbitration proceedings: it is increasingly becoming part of arbitral decision-making (see here). But can an arbitral award survive judicial scrutiny when an arbitrator relies on authorities generated by AI that do not exist? In a recent Canadian case, ARIHQ v Santé Québec, the Québec Superior Court first confronted this issue directly and set aside an award after finding that the key authorities cited by the arbitrator were hallucinated. This decision also appears to be the first time a state court has set aside an AI-drafted award, which is thus significant and provides guidance to legal practitioners.

Facts of ARIHQ v Santé Québec

ARIHQ v Santé Québec involved two applicants: a non-profit legal entity named the Association des ressources intermédiaires d'hébergement du Québec (“ARIHQ”) and an adult accommodation service provider named Centre de Santé Osman (“Osman”). Osman provided services to the respondent, Santé Québec, a Canadian public health and social services authority.

In 2018, ARIHQ, as a recognized representative organization, entered into a National Agreement (“Agreement”) with the Minister of Health and Social Services on behalf of Osman. The Agreement contained a multi-tier dispute resolution procedure requiring that if any disagreement could not be resolved amicably, Osman or ARIHQ must submit a written notice of disagreement to the representative designated by Santé Québec within a peremptory period of 90 days of the event or the date Osman or ARIHQ became aware of it. If still unresolved, ARIHQ shall submit the dispute to arbitration. A year later, Osman and Santé Québec reached two further agreements that were incorporated into the Agreement.

In 2021, a disagreement arose over retroactive payments, with Osman requesting payment and receiving a further refusal from Santé Québec. Nearly three years later, ARIHQ submitted a written notice of disagreement to Santé Québec seeking approximately CAD 1,225,000. Santé Québec rejected the claim and argued the 90-day deadline had long expired. ARIHQ then commenced the arbitration. The arbitrator ultimately dismissed the dispute, holding that the 90-day deadline was legally valid and did not violate public policy.

ARIHQ and Osman then sought to annul the arbitral award before the Superior Court, advancing two main grounds. First, they argued that the award was contrary to public order because it gave effect to a contractual limitation period shorter than the three-year period provided by law. Second, they contended that the arbitral procedure was not respected, as there was strong evidence that the award was drafted by AI. This post focuses on the AI issue alone.

The Superior Court’s Reasoning on the AI Issue

The AI issue arose in the context of the application to annul the award under 3° of Article 646 of the Code of Civil Procedure (“CCP”), which permits annulment where the procedure for the appointment of the arbitrator or the applicable arbitral procedure was not observed. The Superior Court emphasized that the violation of the arbitration procedure agreed by the parties must be significant and must thereby undermine the integrity of the arbitral process. By contrast, an insignificant procedural irregularity causing no prejudice, injustice, or effect on the award or the outcome of the dispute does not warrant judicial intervention. Once a significant violation is established, the court must assess whether the nature of the violation is likely to undermine procedural integrity and how it might affect the award.

No Delegation of Decision-Making Power

The Superior Court first stressed that a key advantage of arbitration is the parties’ autonomy to choose their own decision-maker. Accordingly, parties who opt for arbitration are entitled to expect that the selected arbitrator will personally render a reasoned decision (see discussion on necessity of reasoned awards in this post). Additionally, together with the duties of confidentiality and maintaining public confidence, these considerations all necessarily mean that the arbitrator cannot delegate their core decision-making functions to a third party.

The AI Issue

The Superior Court acknowledged AI’s potential benefits but further identified key risks, namely:

(1) AI may create misleading hallucinations;

(2) AI may not be able to exercise discretionary power;

(3) AI may reproduce hidden biases that are difficult to detect or correct;

(4) AI use may threaten confidentiality; and

(5) AI may undermine public confidence in justice.

Aware of these risks, the Superior Court held that decision-makers using AI bear stricter duties than counsel, notwithstanding counsel’s duty to verify the sources on which they rely. This is because the legitimacy of a judicial decision rests on the sufficient reasoning provided by the decision-maker rather than external rationalization. Consistently, the Canadian Judicial Council has prohibited judges from delegating their decision-making power to any third party, including AI.

Finally, the Superior Court found that the arbitrator had relied on hallucinated case law and non-existent doctrinal sources, and that these fictitious authorities were central to his reasoning. As a result, he effectively delegated his authority and failed to verify the outcome. For this reason, the Superior Court set aside the award and ordered the parties to choose a new arbitrator within 60 days of the judgment.

Beyond Québec: Situating ARIHQ v Santé Québec within the NYC and Model Law Framework

The Superior Court’s grounds to set aside the award in ARIHQ v Santé Québec on the AI issue were based on 3° of Article 646 of the CCP, which closely mirrors Article V(1)(d) of the New York Convention (“NYC”) and Article 34(2)(a)(iv) of the UNCITRAL Model Law (“Model Law”). Given that the amendment of the CCP in 1986 incorporated the principles of the NYC and was largely inspired by the Model Law (see Dell Computer Corp. v Union des consommateurs, [44]), ARIHQ v Santé Québec might offer valuable guidance to courts in Model Law jurisdictions and NYC Contracting States.

While Article V(1)(d) of the NYC does not specify the threshold of procedural irregularities, most courts require a substantial defect in the arbitration proceeding or a causal nexus between the defect and the award. The approaches mainly include assessing whether the alleged procedural irregularity caused substantial injustice to one of the parties, or requiring the party opposing enforcement to demonstrate that the outcome would have been different had the irregularity not occurred (see UNCITRAL Secretariat Guide on Article V(1)(d)). Similarly, although the travaux préparatoires of the Model Law suggested that courts retain discretion to set aside an award regardless of the materiality of the defect (see A/40/17, [303]), courts in Model Law jurisdictions generally assess on a case-by-case basis whether procedural errors regarding Article 34(2)(a)(iv) reach a certain degree of seriousness (see here and here).

In ARIHQ v Santé Québec, the arbitrator’s use of AI amounted to a serious procedural irregularity. The evidence overwhelmingly showed that all the cases and doctrinal authorities were hallucinated. Because these authorities formed a central part of the award, the arbitrator’s failure to verify those authorities infected the very process by which the award was reasoned and rendered. As a result, there was a direct causal nexus between the procedural irregularity and the final outcome. In this sense, the applicants could reasonably have believed that, had those authorities been verified, the arbitrator might have reached a different conclusion.

Observations

While ARIHQ v Santé Québec signals that efficiency of arbitration cannot come at the expense of authenticity, it should not be understood as establishing a blanket prohibition on the arbitrator’s use of AI. As the Superior Court noted, not every award containing erroneous citations or involving AI assistance must be annulled. If the arbitrator had independently verified the authenticity of the cited authorities, or if the hallucinated authorities had constituted only minor procedural irregularities, the arbitrator’s use of AI in drafting the award would not, in itself, have warranted the setting-aside of the award. It therefore remains to be seen how courts in other jurisdictions will deal with similar issues involving AI in the future, depending on the nature and seriousness of the arbitrator’s use of AI in each case.

Ultimately, ARIHQ v Santé Québec serves as a reminder that AI may assist but cannot replace arbitrators. As the Superior Court emphasized, the parties to an arbitration are entitled to expect that their disputes will be resolved by the arbitrator they select. Thus, the responsibility to render a reasoned award must remain with the arbitrator. This expectation is fundamental to the parties’ confidence in both the outcome and the arbitration process itself. Where an award largely relies on unverified AI-generated authorities, it can no longer be said to reflect the arbitrator’s genuine exercise of discretion — a function AI is inherently incapable of performing. In essence, arbitral integrity depends on a fair outcome reached by an independent and impartial human decision-maker who has weighed the facts, assessed the evidence, considered the parties’ competing arguments, and arrived at a reasoned decision.