This is a TechCred essay drawing on a forthcoming SSRN working paper, After ARIHQ: The Non-Delegable Arbitration Mandate in the Age of AI. It is the second piece in the TechCred series on the ARIHQ decision.

When the Quebec Superior Court annulled the arbitral award in ARIHQ v. Santé Québec on April 22, 2026, much of the international arbitration community treated the decision as new law. It is not. The doctrine that decided ARIHQ — that an arbitrator cannot delegate the substantive reasoning of an award to a third party, whether human, institutional, or algorithmic — has been the rule since 1989. ARIHQ applied an old rule to a new delegate.

That distinction matters. If ARIHQ announced a new doctrine, arbitrators would be entitled to wait for further case law before adjusting their practice. If ARIHQ applied an existing doctrine to AI, arbitrators are already bound by it everywhere the doctrine runs — which is everywhere arbitration runs.

This piece traces the lineage. The forthcoming SSRN working paper, After ARIHQ: The Non-Delegable Arbitration Mandate in the Age of AI, develops the argument in full and situates the doctrine in a wider three-layer framework of arbitral governance (lex arbitri, lex data, lex AI). What follows is the doctrinal spine.

The Italian Origin: Sacheri v. Robotti (1989)

The earliest articulation of the personal-mandate doctrine in modern arbitration came not from a common-law court but from the Italian Court of Cassation. In Sacheri v. Robotti, decided in 1989, the Court held that an arbitrator who relied on a non-arbitrator expert to perform the substantive analysis on which the award rested had improperly delegated the adjudicative function. The remedy was annulment.

The reasoning has aged remarkably well. The Court did not say that arbitrators may never consult outside expertise. It said that when the outside party performs the reasoning that produces the result, the arbitrator has substituted a different decision-maker than the one to whom the parties consented. Party autonomy is the engine of arbitration. What the parties bargain for is the personal judgment of the arbitrator they selected. When that judgment is supplied by another, the bargain has not been honoured.

Sacheri established what comparative arbitration scholars would later call the rule of intuitus personae — the arbitrator’s mandate is personal to the arbitrator. It is found in Gary Born’s International Commercial Arbitration, in Gaillard and Savage’s Fouchard Gaillard Goldman, in Redfern and Hunter, and in every major treatise written in the last thirty years. The principle is not contested. The only contest is over its application.

The English Codification: P v. Q (2017)

Twenty-eight years after Sacheri, the question reached the English Commercial Court in a different form. Tribunal secretaries had become a normal feature of LCIA, ICC, and ad-hoc arbitrations. Some parties suspected that the line between administrative assistance and substantive delegation had been crossed. In P v. Q [2017] EWHC 194 (Comm), Mr Justice Popplewell took up the question directly.

Popplewell J. acknowledged what he called “considerable and understandable anxiety in the international arbitration community” about tribunal secretaries becoming “fourth arbitrators.” His ruling drew the line where Sacheri had drawn it. A secretary may perform research, prepare summaries, manage logistics, and even prepare drafts for the tribunal’s consideration. What a secretary may not do is “express a view on the substantive merits of an issue or application.” The reasoning of the award must be the tribunal’s reasoning.

The English judgment matters because it confirmed that the personal-mandate doctrine survives translation across legal traditions. Sacheri had been an Italian civil-law decision construing Italian arbitration law. P v. Q applied the same logic under section 24 of the English Arbitration Act 1996. The convergence is not coincidence. It reflects the deep structural feature of arbitration: the arbitrator’s authority is a personal delegation from the parties, and personal delegations do not sub-delegate by their own force.

Institutional codification followed. The ICC’s Note to Parties and Arbitral Tribunals on the Conduct of the Arbitration at paragraphs 184 to 194 limits the role of administrative secretaries. The Young ICCA Guide on Arbitral Secretaries (2014) makes the same distinctions in practitioner-facing form. France’s Code de procédure civile Article 1450 goes further still: “La mission d’arbitre ne peut être exercée que par une personne physique.” The arbitrator’s mandate may be exercised only by a natural person. Germany’s ZPO § 1035(5), Scotland’s Arbitration Act 2010 Schedule 1 Rule 3, and the case law of Canada, Australia, and the Netherlands all converge on the same point.

By the time ARIHQ arrived in Quebec, the doctrine had a forty-year lineage and a consolidated foundation in the leading treatises and statutes.

The Quebec Application: ARIHQ (2026)

ARIHQ v. Santé Québec did not invent the personal-mandate doctrine. It applied it to a new delegate. The arbitrator in ARIHQ had delegated the substantive reasoning of the award to a generative AI system. The reasoning rested on five citations. All five were fabricated by the AI. The award was annulled.

The four doctrinal pillars Justice Sheehan invoked — personal mandate, reasoned award, the secret of deliberation, and delegatus non potest delegare — were not Quebec inventions. They were the framework that Sacheri, P v. Q, the ICC Note, the Young ICCA Guide, and Article 1450 had together built. The first piece in this TechCred series traced the four doctrines in detail. What that piece left for this one is the genealogy.

What changed with ARIHQ is the delegate, not the rule. Tribunal secretaries, paralegal researchers, librarians, drafting assistants, and language interpreters have accompanied arbitrators for decades. None of them ever became a fourth arbitrator, because the doctrine drew the line at substantive delegation. Generative AI is a new tool with a different risk profile — its outputs are persuasive even when wrong, its hallucinations are difficult to detect, and its reasoning style mimics the reasoning style of arbitral awards. Those properties make verification more important, not less. They do not change the doctrine. They change the discipline required to apply it.

Why the Lineage Matters for Arbitrators Now

If the personal-mandate doctrine were new, arbitrators could reasonably wait for institutional guidance and further case law before adapting their practice. Because the doctrine is forty years old, the obligation is already in force. ARIHQ did not impose a new duty. It made visible a duty that has always been there.

Three operational consequences follow:

First, an arbitrator who uses a generative AI tool to draft any portion of an award must independently verify every statement of fact, every citation, and every line of reasoning before that draft becomes part of the award. The verification cannot be delegated to the AI itself. Asking a model to check its own work is not verification — it is recursion within the same delegate. The doctrine requires verification by the arbitrator personally.

Second, the arbitrator’s reasoning must be the arbitrator’s reasoning. Reasoning generated by a machine and adopted without independent judgment is not the arbitrator’s. The award may carry the arbitrator’s signature, but the personal-mandate doctrine looks at the substance of the deliberation, not the format of the signature. Sacheri and P v. Q are both clear on this. ARIHQ applied the principle to AI without difficulty because the principle was already in place.

Third, disclosure is the operational corollary of the doctrine. Where AI use materially affects the integrity of the decision-making process, the parties must be told. They cannot exercise informed consent — to the arbitrator they selected, to the procedures they agreed to, and to the award they receive — unless they know how the reasoning was produced. The disclosure obligation is not a separate rule grafted onto arbitration. It is what the personal-mandate doctrine has always required, surfaced by the new delegate’s distinctive risk profile.

The Institutional Framework Catches Up

In the year before ARIHQ, the major arbitration institutions issued AI guidance that, in retrospect, anticipated the doctrinal application Quebec would make:

• The SVAMC Guidelines on the Use of AI in Arbitration (April 30, 2024), Guideline 6: “Arbitral decision-making cannot be delegated to AI.”

• The CIArb Guideline on the Use of AI in Arbitration (March 2025, updated September 2025): “arbitral functions should not be delegated to AI.”

• The VIAC Note on the Use of AI in Arbitration Proceedings (April 2025): AI may not substitute for the arbitrator’s reasoning.

• The SCC Guide to the Use of AI (October 16, 2024): AI use raises questions about the arbitrator’s independent judgment.

• The CIETAC Guidelines on Generative AI in Arbitration (effective July 18, 2025): first AI-specific rules from a major Asia-Pacific institution; tribunals must not delegate the adjudicative function.

• The AAA-ICDR Guidance on Arbitrators’ Use of AI Tools (January 2026): disclosure required where AI materially affects the process or reasoning.

These instruments did not announce a new prohibition. They translated Sacheri and P v. Q into operational language for the AI use case. ARIHQ then translated the same principle into a judicial holding.

What This Means for the TechCred Practitioner

The personal-mandate doctrine has three implications that every TechCred-trained arbitrator should be operating against today.

The first is genealogical confidence. When a party suggests in a procedural conference that the arbitrator should be free to use AI as the arbitrator sees fit because no court has yet ruled definitively on AI delegation, the answer is that Sacheri, P v. Q, the ICC Note, and Article 1450 have all ruled, and ARIHQ has confirmed that the rule applies to AI. The doctrinal foundation does not need to be invented.

The second is disclosure-readiness. A procedural-order template that includes an AI disclosure clause, a verification protocol, and a record-of-tools provision is not a defensive document. It is a transparent application of the personal-mandate doctrine that Sacheri established. Practitioners drafting procedural orders for tribunals they sit on or appear before should treat AI disclosure language as part of the standard set, not as an optional or experimental add-on.

The third is institutional alignment. The institutional guidance has moved faster than the case law because the institutions saw the doctrinal point earlier. The SVAMC and CIArb Guidelines, the VIAC and SCC Notes, the CIETAC Guidelines, and the AAA-ICDR Guidance are not soft-law speculation. They are the operational form of an old hard-law principle. Practitioners working under those institutional rules are working with the doctrine, not around it.

Looking Ahead

The forthcoming SSRN working paper, After ARIHQ: The Non-Delegable Arbitration Mandate in the Age of AI, develops the doctrine at length and traces its consequences for enforcement under the New York Convention and the Federal Arbitration Act. It situates the personal-mandate doctrine within a wider tripartite framework of arbitral governance: lex arbitri governs procedural validity, lex data governs informational sovereignty, and lex AI governs the digital tribunal. The companion working papers on each layer will be available on SSRN in parallel.

For arbitrators reading this in the present moment, the operational message is short. The doctrine you are accountable to is forty years old. The application to AI is new. The discipline of disclosure, verification, and human judgment is what the doctrine has always required. ARIHQ did not change the rule. It made the rule unmissable.

Where TechCred Comes In

Bridging old doctrine to new practice is the curricular mission of the ABA TechCred Program, the first internationally accredited certification in AI, digital innovation, and technology dispute resolution. Sponsored by the ABA International Law Section and credentialed by New York Law School, TechCred trains arbitrators to operate the personal-mandate doctrine under the conditions that generative AI has now created — verification protocols, AI disclosure clauses, procedural-order templates, hallucination-detection workflows, and the data-sovereignty implications of using cloud-based AI tools in cross-border proceedings.

The Fall 2026 program is now open for registration.

REGISTER NOW

Further Reading

Primary sources

• ARIHQ v. Santé Québec, 2026 QCCS 1360 (Que. Sup. Ct. Apr. 22, 2026) (Sheehan J.C.S.) — canlii.ca/t/kkjtm

• P v. Q [2017] EWHC 194 (Comm) (Popplewell J.) — bailii.org/ew/cases/EWHC/Comm/2017/194.html

• Sacheri v. Robotti, Italian Court of Cassation (1989) — discussed in Born, International Commercial Arbitration (3d ed. 2021)

• French Code de procédure civile, Article 1450

• German ZPO § 1035(5)

• Scotland Arbitration Act 2010, Schedule 1, Rule 3

Institutional guidance

• SVAMC AI Guidelines (April 2024)

• CIArb Guideline on the Use of AI in Arbitration (March 2025, updated September 2025)

• VIAC Note on the Use of AI in Arbitration Proceedings (April 2025)

• SCC Guide to the Use of AI (October 2024)

• AAA-ICDR Guidance on Arbitrators’ Use of AI Tools (January 2026)

Empirical

• Damien Charlotin, AI Hallucination Cases Database (1,455 entries as of May 17, 2026)

Forthcoming

• Barry Appleton, After ARIHQ: The Non-Delegable Arbitration Mandate in the Age of AI (SSRN, May 2026)

• Companion working papers: The Lex AI, Lex Data, Arbitration Hacked, AI and International Arbitration: A Comprehensive Listing

Earlier in this series

• ARIHQ v. Santé Québec: The Case That Changed Arbitrator AI Liability

Prof. Barry Appleton, FCIArb, is the Faculty Director of the ABA TechCred Program, Co-Director and Distinguished Senior Fellow of the Center for International Law at New York Law School, and a Fellow of the Balsillie School of International Affairs. He is a dual-licensed (Canadian and American) lawyer with three decades of experience in North American trade, commerce, and investment law and a continuing scholarly focus on the intersection of AI governance and international arbitration.

Prof. Barry Appleton, FCIArb TechCred
Faculty Director, ABA TechCred Program, Co-Director and Distinguished Senior Fellow, New York Law School

This is the third piece in an ongoing series.

On June 24, 2019, a three-member arbitral tribunal in Tennant Energy LLC v. Government of Canada, dismissed the applicability of European data protection law to its own proceedings in a single paragraph. The reasoning was terse: the European Union was not a party to NAFTA, so the GDPR did not, “presumptively, come within the material scope” of the arbitration.1

That paragraph is now the textbook example of what happens when a tribunal treats data protection law as if it were optional. The arbitration was Toronto-seated. The presiding arbitrator was established in Singapore. One of the co-arbitrators was a United Kingdom-based practitioner who had publicly acknowledged acting as a data controller under the GDPR in his own data privacy notice.2 The Permanent Court of Arbitration secretariat processed case materials through European-based infrastructure.3 California-resident counsel and California-resident expert witnesses participated on the investor side, with their personal data and the personal data of California-resident witnesses moving through the proceeding throughout. The Government of Canada was the respondent. The proceeding processed personal data of witnesses, experts, and counsel located across multiple jurisdictions.

The tribunal addressed all of that with five words: “does not, presumptively, come within.”4 That answer was not merely insufficient. It was wrong. The author was lead counsel for the investor in the proceeding and is now publishing a working paper that walks through, regime by regime, why.5

Seven years later, on April 22, 2026, the Quebec Superior Court annulled an arbitral award in ARIHQ v. Santé Québec6 on grounds that included the arbitrator’s transfer of party data outside Quebec and Canada in violation of Quebec and Canadian data privacy law. Tennant and ARIHQ are now the two ends of the same arc. Tennant shows what happens when a tribunal declines to identify the applicable data protection regimes. ARIHQ shows what happens in court when those obligations are then breached. The intervening seven years have turned the lex data from an unaddressed analytical gap into an enforcement-determinative requirement.

What the Tribunal Actually Missed

The Tennant tribunal did not miss one data protection regime. It missed six. Each one attached to the proceeding through an independent jurisdictional trigger that no tribunal ruling could displace.

The GDPR. Article 3 of the General Data Protection Regulation establishes extraterritorial reach in three ways. It applies whenever personal data is processed in the Union, whenever a controller or processor is established in the Union, and whenever the targeting criteria are met. The UK-based co-arbitrator processed personal data on a personal device located in the United Kingdom (then still subject to the GDPR’s EU-law successor regime), and the European-based PCA secretariat processed case materials in The Hague. Either of those facts, standing alone, was sufficient to engage the Regulation.

The UK GDPR. Following the United Kingdom’s withdrawal from the European Union, the United Kingdom retained the GDPR through the Data Protection Act 2018 and the Data Protection, Privacy and Electronic Communications (Amendments etc.) (EU Exit) Regulations 2019.7 The Tennant tribunal’s UK-based co-arbitrator was personally subject to the UK GDPR, regardless of NAFTA’s identity.

California’s CCPA, as amended by the CPRA. The California Consumer Privacy Act, codified at Cal. Civ. Code § 1798.100 et seq. and amended by the California Privacy Rights Act effective January 1, 2023, applies to businesses that collect or process the personal information of California residents and meet statutory thresholds.8 The investor’s California-resident counsel and California-resident expert witnesses participated in the Tennant proceeding. Their personal data, and the personal data of California-resident witnesses whose information was processed in connection with the case record, fell within the CCPA/CPRA’s scope. The California regime is enforced by the California Privacy Protection Agency, which has rule-making and administrative-fining authority that does not depend on the parties’ choice of arbitral seat.9 The tribunal’s confidentiality order was silent on the regime entirely.

Canada’s PIPEDA. The Personal Information Protection and Electronic Documents Act applies to private-sector processing of personal data in the course of commercial activities anywhere in Canada.10 The investor’s counsel was based in Toronto. The arbitration was seated in Toronto. The investor itself was a private-sector commercial party. PIPEDA applied.

Canada’s Privacy Act. The Privacy Act governs federal government institutions in their handling of personal data. 11 The Government of Canada was the respondent. The respondent’s processing of investor and witness personal data in connection with the proceeding fell squarely within the Privacy Act’s scope.

Singapore’s PDPA. The Personal Data Protection Act 2012 applies to organizations that collect, use, or disclose personal data in Singapore. 12 The presiding arbitrator was established in Singapore and practiced there throughout the proceeding. The presiding arbitrator’s processing of case materials was governed by the PDPA.

A single proceeding. Six mandatory regimes. The tribunal addressed none of them.

Why “Mandatory” Means What It Says

International arbitration has long accommodated mandatory norms. Tribunals have applied mandatory tax laws despite contrary party stipulations, recognized legal privilege as inviolable irrespective of the procedural code, and invalidated agreements tainted by corruption on public policy grounds. The structural logic is settled: certain obligations attach regardless of party choice because the obligations protect interests the parties cannot waive.

Data protection regimes belong on that list. The GDPR, the UK GDPR, the CCPA/CPRA, PIPEDA, the Canadian Privacy Act, and the Singapore PDPA each protect a fundamental right of natural persons whose personal data is processed in the course of arbitration. None of those regimes asks for the parties’ permission. None can be displaced by an arbitration agreement. None turns on whether the European Union, the United Kingdom, the State of California, Canada, or Singapore is a party to the underlying treaty. 13

The lex data is the doctrinal name for this overlay. 14 It is the mandatory, extraterritorial data protection framework that attaches to international arbitral proceedings through five digital connecting factors: the seat, the location of arbitrators, counsel, and institutional secretariats, the location of data subjects, the location of data storage and processing, and the location where the proceeding’s outputs will be used. Each connecting factor is independent. Each can engage a different regime. Where two or more attach simultaneously, the proceeding becomes subject to multiple regimes operating in parallel. This is the transnational simultaneity problem, and it is now a feature of every digitally active international arbitration.

What ARIHQ Now Adds: Data Privacy Breach as an Enforcement Risk

For the seven years after the Tennant ruling, the consequence of ignoring the lex data was, in practice, deferred. Tribunals could decline to perform the analysis without an immediately visible cost. ARIHQ changed that.

In ARIHQ v. Santé Québec, Justice Sheehan of the Quebec Superior Court annulled an arbitral award rendered August 8, 2025, by sole arbitrator Michel Jeanniot. The award was annulled on multiple grounds, four of which addressed the arbitrator’s undisclosed use of generative AI to produce hallucinated legal authorities. A fifth ground, less commented on but equally important for present purposes, addressed data sovereignty directly. The court observed that entering party information into an AI platform likely transferred personal data to servers outside Quebec and Canada, and that such a transfer would violate Quebec’s and Canada’s data privacy laws. 15 In the ARIHQ court’s framing, this is not an abstract compliance problem. It is a jurisdictional problem that affects the validity of the arbitral process itself.

The doctrinal step is significant. The ARIHQ court did not treat data protection law as parallel to the arbitration. It treated compliance with data protection law as part of what makes the arbitral process valid. Breach of data privacy law in connection with the proceeding, in ARIHQ’s analysis, undermines the integrity of the procedure and the validity of any resulting award.

That linkage matters far beyond Quebec. New York Convention Article V(2)(b) permits a recognition court to refuse enforcement of an award where recognition would be contrary to the public policy of the enforcing forum. Schrems II identifies data protection as constitutional public policy in the European Union. 16 Awards rendered in proceedings that disregarded mandatory data protection obligations, or that involved transfers of personal data in violation of the GDPR, the UK GDPR, the CCPA/CPRA, PIPEDA, the Canadian Privacy Act, or the Singapore PDPA, become structurally vulnerable at the recognition stage. The Lex Data working paper develops the analysis for three enforcement jurisdictions and concludes that the risk is no longer hypothetical. 17

Tennant defined the failure mode at the front end of the proceeding: the tribunal declines to identify which data protection regimes apply. ARIHQ defined the failure mode at the back end: the arbitrator breaches a data protection regime in connection with the proceeding, and the resulting award does not survive judicial review. The two cases connect through a single proposition: data protection law is integral to the validity of the arbitral process, and its breach is no longer an off-balance-sheet liability.

What Proactive Lex Data Governance Looks Like

The tribunal that has identified the applicable regimes at the first procedural meeting, allocated controller and processor roles transparently, addressed cross-border transfer mechanisms, conducted a Data Protection Impact Assessment where Article 35 requires it, and documented compliance, has built the record that defeats Article V(2)(b) at the enforcement stage. When California-resident participants are involved, the tribunal has identified and addressed the CCPA/CPRA threshold question. Where AI tools are used in the proceedings, whether by the arbitrators, counsel, or experts, the tribunal has documented the platform’s data-handling characteristics and ensured that personal data is not transferred to servers in a manner that breaches applicable regimes. The cost is an hour at the first session. The benefit is enforceability.

Why the Cybersecurity Protocol Is Necessary But Not Sufficient

The leading institutional instrument for digital governance in international arbitration is the ICCA-NYC Bar-CPR Cybersecurity Protocol for International Arbitration (2020, revised 2022).18 The Protocol normalized information security planning as a component of the arbitral procedure. Its contribution is substantial.

It is also incomplete. The Protocol is a cybersecurity instrument. It addresses the protection of case information against unauthorized access, use, disclosure, modification, and destruction. It does not address the affirmative obligations that data protection law imposes regardless of any security threat: the lawful basis for processing under GDPR Article 6; the controller and processor mapping under Article 26 and Article 28; the Data Protection Impact Assessment under Article 35; the cross-border transfer mechanisms under Articles 44 through 49; the consumer-rights mechanisms under the CCPA/CPRA; and the post-quantum security dimension that NIST FIPS 203, 204, and 205 (published August 2024) now adds to the Article 32 “appropriate to the risk” standard.19

The Lex Data working paper proposes five targeted enhancements to close the gap: a data-role mapping appendix, a DPIA decision tree, a transfer mechanism appendix, a multi-regime coordination matrix, and a post-proceeding data governance protocol that incorporates a documented plan for post-quantum re-encryption of any materials retained over a long time horizon. The enhancements are designed to be adopted by the institutional drafters of the Protocol’s next edition, not to compete with the Protocol.

How Tennant and ARIHQ Define the Arc

Tennant Energy and ARIHQ v. Santé Québec: the front and back covers are of the same volume. Tennant is the case in which a tribunal dismissed a claim under mandatory data protection law on grounds that did not survive analysis. ARIHQ is the case decided seven years later in which a court annulled an arbitral award in part because an arbitrator’s data-handling practices breached Quebec and Canadian data privacy laws and undermined the procedural integrity of the proceeding.

Both decisions belong to the same arc. Both turn on the same insight: the lex arbitri is necessary but no longer sufficient to govern the digital conduct of international arbitration. The lex data governs informational sovereignty. The companion lex AI framework governs the digital tribunal.20 The Digital Security Framework governs the security of the infrastructure. 21 Together with the lex arbitri, they constitute the Digital Procedural Constitution.

The Tennant tribunal could have closed the analytical gap in 2019 with one additional paragraph. The ARIHQ arbitrator could have avoided annulment in 2025 through disclosure and a different platform choice. Neither did. Their successors have run out of room to repeat the omission.

The TechCred Program

The American Bar Association’s TechCred program certifies digital competency for arbitrators and counsel. The program addresses, in three sessions, the core obligations the Lex Data framework identifies: identifying applicable data protection regimes through the digital connecting factors; allocating controller and processor responsibilities; integrating data protection into Procedural Order No. 1; and managing the cybersecurity, AI, and breach-response dimensions that intersect with the lex data.

The Fall 2026 TechCred program is now open for registration: https://events.americanbar.org/event/ee369916-3182-40d6-a1e9-5347aefeda2f/summary. The program is delivered live online and credentialed by New York Law School.

REGISTER NOW

A Final Word

The Tennant tribunal’s single-paragraph ruling is a useful artifact. It marks, with unusual clarity, the moment when the digital governance gap in international arbitration became visible. The proceeding processed multimillion-dollar claims across six mandatory data protection regimes simultaneously. The tribunal addressed none of them. The result was a regulatory vacuum that no party had requested and no law required.

The post-Tennant, post-ARIHQ tribunal has no excuse. The applicable regimes are identifiable. The connecting factors are objective. The procedural responses are now in the institutional literature and will shortly be consolidated into a Digital Governance Protocol designed to be dropped directly into Procedural Order No. 1.

When the next tribunal is asked at its first procedural meeting whether the GDPR, the CCPA/CPRA, or any other applicable regime applies, the answer is no longer one paragraph. It is an analysis. The lex data is the framework for that analysis. Performing it is now the baseline.

References

Appleton, Barry. Lex Data: Reconceptualizing the Law of the Digital Seat in International Arbitration (Working Paper, April 30, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6610079.

Appleton, Barry. Lex AI: Mandatory AI Governance and the Digital Tribunal in International Arbitration (Working Paper, April 30, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6610019.

Appleton, Barry. Arbitration Hacked: Toward a Unified Cybersecurity and Data Breach Framework for International Arbitral Proceedings (Working Paper, April 21, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6625218.

Appleton, Barry. ARIHQ v. Santé Québec: The Case That Changed Arbitrator AI Liability, TechCred Substack (May 4, 2026).

Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c. Santé Québec — Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l’Île-de-Montréal, 2026 QCCS 1360 (Que. Sup. Ct. Apr. 22, 2026) (Sheehan, J.C.S.), https://canlii.ca/t/kkjtm.

California Consumer Privacy Act, Cal. Civ. Code §§ 1798.100–1798.199.100 (West 2024), as amended by the California Privacy Rights Act of 2020.

ICCA, New York City Bar Association & CPR Institute, Cybersecurity Protocol for International Arbitration (2022 ed., first published 2020), https://drs.cpradr.org/rules/protocols-guidelines/icca-nyc-bar-cybersecurities.

ICCA-IBA Joint Task Force on Data Protection in International Arbitration, Roadmap to Data Protection in International Arbitration, ICCA Reports No. 7 (2022), https://www.arbitration-icca.org/icca-reports-no-7-icca-iba-roadmap-data-protection-international-arbitration.

National Institute of Standards and Technology, Federal Information Processing Standards Publications 203, 204, and 205 (August 2024), https://csrc.nist.gov/projects/post-quantum-cryptography.

Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation), 2016 O.J. (L 119) 1.

Prof. Barry Appleton, FCIArb, TechCred Faculty Director, ABA TechCred Program,
Co-Director and Distinguished Senior Fellow, New York Law School

This is the second piece in an ongoing series.

On November 3, 2025, the American Arbitration Association (AAA) and its international division, the International Centre for Dispute Resolution (ICDR), launched the AAA-ICDR AI Arbitrator. It is the first deployment of artificial intelligence in the adjudicative function of arbitration by a major international institution.1

The product is currently available for two-party, documents-only construction cases. It generates draft awards from case submissions, is trained on more than 1,500 prior construction awards, and is refined through expert-labeled examples and structured legal prompt libraries. The system was developed in collaboration with QuantumBlack, AI by McKinsey, and tested across more than 1,000 simulated cases.2 The AAA has announced plans to expand into insurance, commercial disputes, multi-party proceedings, and cross-border matters in subsequent phases.

The AI Arbitrator does not issue awards on its own. A human arbitrator, specifically trained for AI-led case reviews, revises if necessary, confirms, and signs. The Integra Ledger blockchain partnership provides tamper-proof authentication of the entire documentary record.3 In structural terms, the system is designed to satisfy the “human oversight” requirement of EU AI Act Article 14 and the non-delegation principles articulated in ARIHQ v. Santé Québec.4

Whether the architecture in fact does satisfy those obligations is the question this post examines. The answer matters because the AAA-ICDR AI Arbitrator is the first concrete test of whether mandatory AI governance can be reconciled with institutional efficiency. If it can, it sets the template for every other major institution. If it cannot, it sets the limit.

Why This Is a Lex AI Question

International arbitration is now subject to two direct AI governance instruments and a growing body of indirect regulatory architecture. The two direct instruments are the EU AI Act5 and ABA Formal Opinion 512.6 The indirect architecture includes the California Transparency in Frontier Artificial Intelligence Act (SB 53)7 and the New York Responsible AI Safety and Education Act (RAISE Act).8 None of these instruments is principally about arbitration. All of them now apply to it.

The Lex AI working paper proposes that this overlay constitutes a distinct mandatory framework on international arbitral proceedings, analogous to the lex data but coequal and independent.9 It is the law of the digital tribunal. Like the lex data, it attaches through objective connecting factors and cannot be displaced by party agreement.

The AAA-ICDR AI Arbitrator engages the Lex AI on every available connecting factor. It is a regulated AI system under the EU AI Act. It performs an arbitral function classified as high risk under Annex III, point 8(a). It is deployed by an institution whose case work routinely produces awards used in the European Union, which engages Article 2(1)(c)’s extraterritorial reach. It is used in connection with proceedings in which one or more arbitrators may be EU-established, which engages Article 2(1)(b). And its use by US-barred counsel is subject to ABA Formal Opinion 512.

The High-Risk Classification Is Not an Aspiration

The EU AI Act classifies AI systems intended for use by competent authorities in the administration of justice and democratic processes, including alternative dispute resolution, as high-risk. Recital 61 makes the legislative intent explicit. Annex III, point 8(a), is the operative provision.

The Stockholm Chamber of Commerce Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules (October 2024) acknowledges that the EU AI Act “qualifies as high-risk AI systems used by an Arbitral Tribunal ‘in researching and interpreting facts and the law and in applying the law to a concrete set of facts’”, citing Annex III, point 8(a) verbatim. 10 The Chartered Institute of Arbitrators Guideline on the Use of AI in Arbitration (March 2025, updated September 2025) similarly identifies AI use in arbitration as presenting risks to “enforceability of arbitral awards and other adverse implications for due process rights, the rule of law, the administration of justice, the credibility and legitimacy of arbitration”. 11 Both institutions thus acknowledge the high-risk character of AI in adjudicative contexts.

Neither institutional instrument provides the mandatory implementation mechanism that high-risk classification requires. The Lex AI framework supplies it through three mechanisms applicable at the first procedural meeting:

1. Risk-tier classification. Each AI system in the proceeding is characterized by EU AI Act tier: unacceptable-risk (prohibited), high-risk (mandatory compliance), limited-risk (transparency required), or minimal-risk. The characterization is documented in the procedural record.

2. Disclosure on a three-tier framework. Tier 1 administrative use (translation, format conversion) requires no disclosure. Tier 2 research assistance requires disclosure. Tier 3 substantive reasoning, including award drafting, requires disclosure, party consent, and verification. The framework is described in the Lex AI working paper and tested against the ARIHQ facts. 12

3. Platform qualification. A platform may be used in a proceeding only if it satisfies enterprise licensing terms that prohibit training on submitted data, that satisfy GDPR Article 28 processing-agreement requirements where applicable, and that offer the post-quantum cryptographic agility required by the lex data framework’s reading of Article 32. 13

The AAA-ICDR AI Arbitrator’s documented architecture suggests that it is designed with these requirements in mind: enterprise-grade infrastructure, four-step human-in-the-loop workflow, tamper-proof documentary authentication, and confined application to two-party documents-only construction cases. The next several phases of expansion will test how the architecture performs as scope grows.

Equality of Arms and Agentic AI

The EU AI Act’s transparency requirements reflect a deeper procedural justice concern that applies across the Lex AI framework: the equality of arms. 14 Agentic AI systems deployed to monitor proceedings, identify inconsistencies in opposing arguments, and suggest real-time procedural responses are now in use by certain well-resourced parties in high-value international arbitrations. A party whose counsel deploys such a system gains a real-time analytical advantage over a party whose counsel does not.

The Silicon Valley Arbitration and Mediation Center Guidelines on the Use of Artificial Intelligence in Arbitration (April 30, 2024) specifically recommend that tribunals consider whether asymmetric access to AI tools compromises procedural fairness and address the issue in Procedural Order No. 1. 15 Lex AI extends that recommendation into a mandatory framework: where the asymmetry is material, the tribunal’s duty to ensure procedural fairness now includes a duty to consider the AI architecture each party deploys, including agentic systems.

For institutional AI products like the AAA-ICDR AI Arbitrator, the equality-of-arms analysis takes a different form. When the institution itself supplies the AI, the question is not whether one party has access. It is whether the system has been disclosed, whether the parties have meaningfully consented to its use, and whether the human arbitrator’s review of the AI-generated draft is substantive rather than perfunctory.

The Hallucination Liability Chain

Five years. Five cases. Three jurisdictions. One unmistakable pattern. 16 In 2023, a New York federal court sanctioned counsel for submitting AI-generated, fabricated citations and then doubling down when challenged. 17 In 2024, the Second Circuit entered a disciplinary referral against an attorney who submitted a nonexistent ChatGPT-generated authority.18 In November 2025, the Qatar Financial Centre Civil and Commercial Court in Doha issued a landmark judgment addressing the same conduct. 19 In early 2026, the New York Appellate Division sanctioned counsel who continued filing AI-fabricated citations after being warned. 20In April 2025, a petitioner argued before a United States federal court that an arbitrator had outsourced the reasoning behind a twenty-nine-page award to artificial intelligence, presenting the first judicial test of whether AI-assisted award drafting constitutes grounds for vacatur under the Federal Arbitration Act. 21

In April 2026, the chain reached the developer. Nippon Life Insurance Company of America v. OpenAI Foundation, filed March 4, 2026, in the Northern District of Illinois, is the first tort action naming an AI developer as a defendant in litigation-conduct claims. 22In April 2026, the Quebec Superior Court annulled an arbitral award in ARIHQ v. Santé Québec because the arbitrator’s undisclosed AI use produced hallucinated authorities. 23

The chain runs counsel, arbitrator, and developer. Mata, Park, Sheppard, Deutsche Bank, LaPaglia, ARIHQ, Nippon Life. Six years. Each case extends the Lex AI framework from where the prior case left off. Each case clarifies that the duty of verification is now a hard rule.

Why the Voluntary Framework Cannot Carry the Load

The SCC AI Guide is candid about its scope: it provides “flexible guidance to participants in an SCC case without imposing specific obligations”. 24 The CIArb AI Guideline is equally explicit: it is “intended for use in conjunction with, and does not supersede, any applicable laws, regulations or policies”, and where it is “incompatible with any agreement of the parties, and or applicable mandatory rules and regulations, the latter shall of course prevail”. 25

Both institutional instruments thus subordinate themselves to mandatory law. Neither constitutes the mandatory framework it acknowledges. Where the EU AI Act’s high-risk obligations apply, the obligations apply whether or not a tribunal has consulted the SCC Guide or the CIArb Guideline. Where ABA Formal Opinion 512 applies to a US-barred lawyer, the Opinion applies whether or not counsel has consulted any guideline at all.

The Lex AI working paper therefore proposes an enhanced model procedural clause designed to operationalize the mandatory obligations at the first procedural meeting. The clause is structured around three pillars: applicable regimes (mapped through connecting factors); risk-tier characterization (with documentation); and disclosure, verification, and oversight obligations (calibrated to the tier). 26

A tribunal that adopts the CIArb Appendix B Option 1 short-form order, which simply directs participants to “be guided by the CIArb AI Guideline in the conduct of the proceedings”, has not discharged its EU AI Act compliance obligations for high-risk AI systems. A tribunal that adopts the Lex AI model clause has.

The TechCred Program

The American Bar Association’s TechCred program is the first internationally accredited certification in AI, digital innovation, and technology dispute resolution. The curriculum addresses the Lex AI framework directly: the EU AI Act risk-tier system, ABA Formal Opinion 512 obligations, the equality-of-arms problem, hallucination prevention, and the integration of AI governance into Procedural Order No. 1.

The Fall 2026 TechCred program is open for registration: https://events.americanbar.org/event/ee369916-3182-40d6-a1e9-5347aefeda2f/summary. The program is delivered live online, accredited by New York Law School, and is the practitioner companion to the working papers in this series.

REGISTER NOW

A Final Word

The AAA-ICDR AI Arbitrator is the institutional product that the arbitration community has been preparing for since the first AI hallucination case reached a court. It is also the most carefully structured deployment of AI in the adjudicative function to date. Whether its architecture survives the first New York Convention enforcement challenge, the first ABA disciplinary inquiry, and the first EU Member State court asked to recognize an AAA-ICDR AI-assisted award is the question the next two years will answer.

What is no longer in doubt is the framework within which the answer will be assessed. The EU AI Act applies. ABA Formal Opinion 512 applies. The Lex AI connecting factors will be tested in real cases. The ARIHQ line is now the precedent that gives all of this judicial flesh.

When the next AAA-ICDR AI Arbitrator award is challenged, the case file will not contain a single paragraph dismissing the relevant AI governance regime. It will contain a record built to show that every regime was identified, every obligation was performed, and every disclosure was made. That record is the institutional product, more than any algorithm.

References

Appleton, Barry. Lex AI: Mandatory AI Governance and the Digital Tribunal in International Arbitration (Working Paper, April 30, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6610019.

Appleton, Barry. Lex Data: Reconceptualizing the Law of the Digital Seat in International Arbitration (Working Paper, April 3, 2026), https://ssrn.com/abstract=6610079.

Appleton, Barry. Artificial Intelligence and International Arbitration: A Comprehensive Listing of Governing Laws, Institutional Rules, Professional Responsibility Standards, and Guidance (Working Paper, May 4, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6631182.

American Arbitration Association — International Centre for Dispute Resolution, Guidance on Arbitrators’ Use of AI Tools (March 2025), https://go.adr.org/rs/294-SFS-516/images/2025_AAAICDR%20Guidance%20on%20Arbitrators%20Use%20of%20AI%20Tools.pdf?version=0.

Chartered Institute of Arbitrators, Guideline on the Use of AI in Arbitration (March 2025, updated September 2025), https://www.ciarb.org/media/bpndtcgu/guideline-on-the-use-of-ai-in-arbitration_updated-sept-2025.pdf.

Stockholm Chamber of Commerce Arbitration Institute, Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules (October 2024).

Silicon Valley Arbitration and Mediation Center, Guidelines on the Use of Artificial Intelligence in Arbitration (April 30, 2024), https://svamc.org/wp-content/uploads/SVAMC-AI-Guidelines-First-Edition.pdf.

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act).

American Bar Association Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512: Generative Artificial Intelligence Tools (July 29, 2024).

By Prof Barry Appleton, FCIArb, TechCred
Faculty Director, ABA TechCred Program, Co-Director and Distinguished Senior Fellow, New York Law School

This is the first piece in a series of three. Next on 05/11 and 05/12

On April 22, 2026, the Quebec Superior Court handed down a decision that rewrites the rulebook for artificial intelligence in arbitration. In Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c.Santé Québec - Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l’Île-de-Montréal, Justice Martin Sheehan became the first judge globally to annul an arbitral award based on an arbitrator’s undisclosed use of AI to generate substantive reasoning supported by hallucinated authorities.

The decision answers a question that has haunted the international arbitration community since LaPaglia v. Valve Corp., No. 3:25-cv-00833-RBM-DDL (S.D. Cal., petition filed Apr. 8, 2025; dismissed Dec. 9, 2025 for lack of subject matter jurisdiction), which was filed last year: what happens when an arbitrator delegates award drafting to AI? LaPaglia never reached the merits. ARIHQ did.

What happened in the ARIHQ case?

The arbitration concerned a contractual dispute between Quebec’s health authority and an association representing residential care facilities. The sole arbitrator, Michel Jeanniot, issued an award on August 8, 2025, addressing limitation periods under Quebec law. The award cited five authorities. All five were fabricated by artificial intelligence.

These “authorities” do not exist:

• An article purportedly authored by Professor Frédéric Bachand titled “Prescription et déchéance: frontières mouvantes et enjeux pratiques,” supposedly published in a 2016 Barreau du Québec collection. Professor Bachand of McGill University is a real and respected scholar. He did not write this article.

• A purported Quebec Court of Appeal decision cited as Ville de Montréal c. Syndicat des cols bleus, 2005 QCCA 591.

• A purported Quebec Superior Court decision cited as Groleau et Groupe Pages Jaunes Cie, 2011 QCCS 5386.

• A purported Quebec Court of Appeal decision cited as Tremblay c. Commission scolaire de la Jonquière, 2002 CanLII 24357 (QCCA).

• A purported arbitral precedent cited as Arbitrage CHU Ste-Justine (D.T.E. 2018-30). SOQUIJ, Quebec’s official legal information service, confirmed this precedent does not exist.

The neutral citations for the court “decisions” lead to entirely different, unrelated cases. No warning appears here without cause: arbitrators who accidentally cite these fabricated authorities risk perpetuating AI-generated hallucinations.

Why the Court Annulled the Award

Justice Sheehan analyzed the problem through four interlocking doctrines of Quebec arbitration law:

• The personal mandate doctrine. An arbitrator’s authority derives from party consent. That consent is to the arbitrator’s personal exercise of judgment, not to a machine’s. Delegation to AI substitutes a different decision-maker than the one the parties selected.

• The reasoned award doctrine. Quebec law requires arbitrators to provide reasoning. Reasoning copied from AI output without independent verification is not the arbitrator’s reasoning. The court invoked the Supreme Court of Canada’s observation in Baker v. Canada (Minister of Citizenship and Immigration), [1999] 2 SCR 817, that the process of writing reasons can in itself guarantee a better decision.

• The secret of deliberation. Quebec arbitrators have an affirmative duty not to communicate with third parties about a matter under advisement. Entering case facts into an AI platform is communication with a third party. The AI system becomes a consultant on the merits, which is prohibited during deliberations.

• The non-delegation doctrine. The Supreme Court of Canada established in Therrien (Re), 2001 SCC 35, that “the body to which the exercise of a power is attributed under its enabling statute must exercise it itself.” An arbitrator cannot delegate adjudicative judgment to a machine.

The court also addressed data sovereignty. Inputting party information into an AI platform likely transfers personal data to servers outside Quebec and Canada. This violates Quebec’s and Canada’s data privacy laws. For arbitrators, this is not an abstract compliance problem. It is a jurisdictional problem that affects the validity of the arbitral process itself.

Why This Matters Beyond Quebec

The ARIHQ decision applies everywhere, not just in Quebec. Here’s why.

In the United States. The Federal Arbitration Act allows vacatur of awards procured through “evident partiality,” “misbehavior,” or where “the arbitrators exceeded their powers.” The four doctrines that Justice Sheehan applied have plausible analogs in FAA section 10(a)(4) (arbitrators exceeded their powers), although the merits question remains open after LaPaglia’s dismissal on subject matter jurisdiction grounds. The AAA-ICDR Guidance on Arbitrators’ Use of AI Tools (March 2025) already requires arbitrators to disclose any use of AI that materially affects the arbitration process or reasoning. ARIHQ now provides a judicial precedent showing what happens when arbitrators fail to do so.

In Commonwealth jurisdictions. The English Arbitration Act 2025, the Singapore International Arbitration Act, the Hong Kong Arbitration Ordinance, and Canada’s provincial arbitration statutes all recognize duties of independence, impartiality, and procedural fairness. An arbitrator’s undisclosed use of AI to generate substantive reasoning violates these duties. The Chartered Institute of Arbitrators issued its Guideline on the Use of AI in Arbitration (March 2025, updated September 2025), stating that “arbitral functions should not be delegated to AI.”

Under risk-based AI regulation. The EU AI Act (Regulation (EU) 2024/1689) classifies AI systems used “in researching and interpreting facts and the law and in applying the law to a concrete set of facts” as high-risk (Annex III, point 8(a)). The high-risk obligations under Annex III take effect on August 2, 2026, subject to ongoing trilogue negotiations on the European Commission’s Digital Omnibus on AI proposal. California’s Transparency in Frontier Artificial Intelligence Act (TFAIA, SB 53) and New York’s Responsible AI Safety and Education Act (RAISE Act) regulate frontier AI developers. These regulatory frameworks establish that AI-assisted adjudication is an inherently high-stakes activity requiring disclosure, verification, and human oversight.

Through institutional arbitration rules. The ARIHQ decision aligns with emerging institutional guidance:

• The AAA-ICDR Guidance (March 2025) requires disclosure when AI use “materially impacts the arbitration process or the reasoning underlying their decisions.”

• The Stockholm Chamber of Commerce Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules (October 2024) warns that AI use raises questions about the independent judgment of the arbitrator.

• The Vienna International Arbitral Centre Note on the Use of Artificial Intelligence in Arbitration Proceedings (April 2025) provides that AI may not substitute for the arbitrator’s reasoning.

• The Silicon Valley Arbitration and Mediation Center Guidelines on the Use of Artificial Intelligence in Arbitration (April 2024) include Guideline 6: “Arbitral decision-making cannot be delegated to AI.”

• The China International Economic and Trade Arbitration Commission issued its Provisional Guidelines on the Use of Artificial Intelligence Technology in Arbitration (Trial) effective July 18, 2025, requiring disclosure and prohibiting reliance on unverified AI outputs.

The International Chamber of Commerce revised its Rules of Arbitration (effective June 1, 2026), but these revisions do not yet address arbitrators’ use of AI. The London Court of International Arbitration has not yet issued AI-specific guidance. Both arbitration entities have rules committees looking at these issues at this time. The gap matters. Institutional rules without AI provisions leave arbitrators, and the awards they render, exposed to the challenges ARIHQ now makes possible.

What ARIHQ Tells Arbitrators

The ARIHQ decision establishes five principles that apply to every arbitrator in every jurisdiction:

• Disclosure is foundational. Any use of AI that affects the integrity of the decision-making process must be disclosed to the parties. Disclosure gives parties the opportunity to consent, object, or require verification procedures. Without disclosure, parties cannot exercise informed consent.

• Verification is non-delegable. An arbitrator who uses AI to draft portions of an award must independently verify every statement of fact, every citation, and every element of reasoning. Verification cannot be delegated to the AI system itself. Asking the AI to “check its work” is not verification.

• Materiality has tiers. Not all AI use requires disclosure. AI use for administrative tasks (calendar management, email drafting, translation of submissions) does not materially affect adjudication. AI use for research assistance may require disclosure depending on how the results are used. AI used to draft substantive reasoning in the award always requires disclosure.

• Deliberation is personal. During deliberation, an arbitrator must not communicate with third parties about the merits of the case. Entering case facts into an AI platform is arguably communication with a third party. This prohibition applies during deliberations, even if the AI tool has been disclosed to the parties and they consented to its use.

• Some violations are non-waivable. Party consent does not cure all procedural defects. If an arbitrator delegates the decision-making function itself to AI, meaning the arbitrator reviews but does not independently reason through the AI’s output, the award may be annulled even if the parties consented to AI use. The arbitrator’s personal exercise of judgment is mandatory.

The TechCred Program

The American Bar Association’s TechCred program certifies digital competency for arbitrators. The first TechCred class graduated in 2025 and received the competency credential. The ARIHQ decision confirms what the TechCred program has been teaching since 2025: arbitrators must understand AI well enough to know when they are delegating judgment rather than using a tool. They must be able to evaluate AI outputs, detect hallucinations, and implement verification procedures. And they must know how to communicate the use of AI to parties in procedural orders and disclosures.

The next TechCred three-day international arbitration competency program takes place in October and November 2026. The program, delivered live but online, is the first internationally accredited certification in AI, digital innovation, and technology dispute resolution, presented by the ABA International Law Section. Registration opens at https://events.americanbar.org/event/ee369916-3182-40d6-a1e9-5347aefeda2f/summary

REGISTER NOW

A Final Word

The ARIHQ case is the first judicial decision globally to annul an arbitral award on the grounds of an arbitrator’s use of AI. It will not be the last. Damien Charlotin’s AI Hallucination Cases database catalogs 1,394 AI hallucination cases globally as of May 1, 2026. ARIHQ is the only case in that database addressing arbitrator use. But the pattern is clear: AI hallucination liability has moved from counsel (Mata v. Avianca, Park v. Kim, Sheppard v. Jillion LLC, Deutsche Bank v. LeTennier) to arbitrators (ARIHQ), and it will soon reach developers (Nippon Life Insurance Co. v. OpenAI Foundation, filed March 4, 2026, remains pending in the Northern District of Illinois).

For arbitrators, the lesson is simple: disclosure, verification, and human judgment are not optional. They are the baseline for lawful AI use in adjudication. The ARIHQ court did not hold that arbitrators cannot use AI. The court held that arbitrators cannot delegate judgment to AI, cannot hide AI use from the parties, and cannot rely on unverified AI outputs in their reasoning. Those principles apply in every seat of arbitration under every arbitration law.

References

Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) c. Santé Québec - Centre intégré universitaire de santé et de services sociaux du Centre-Sud-de-l’Île-de-Montréal, 2026 QCCS 1360 (Que. Sup. Ct. Apr. 22, 2026) (Sheehan, J.C.S.), https://canlii.ca/t/kkjtm

Therrien (Re), 2001 SCC 35, [2001] 2 SCR 3, https://canlii.ca/t/521j

Baker v. Canada (Minister of Citizenship and Immigration), [1999] 2 SCR 817, 1999 CanLII 699 (SCC), https://canlii.ca/t/1fqlk

LaPaglia v. Valve Corp., No. 3:25-cv-00833-RBM-DDL (S.D. Cal., petition filed Apr. 8, 2025; dismissed Dec. 9, 2025 for lack of subject matter jurisdiction), https://dockets.justia.com/docket/california/casdce/3:2025cv00833/810531

Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) (Castel, J.)

AAA-ICDR, Guidance on Arbitrators’ Use of AI Tools (March 2025), https://go.adr.org/rs/294-SFS-516/images/2025_AAA-ICDR%20Guidance%20on%20Arbitrators%20Use%20of%20AI%20Tools%20%282%29.pdf?version=0

Chartered Institute of Arbitrators, Guideline on the Use of AI in Arbitration (March 2025, updated September 2025), https://www.ciarb.org/media/bpndtcgu/guideline-on-the-use-of-ai-in-arbitration_updated-sept-2025.pdf

Stockholm Chamber of Commerce Arbitration Institute, Guide to the Use of Artificial Intelligence in Cases Administered Under the SCC Rules (October 2024), https://sccarbitrationinstitute.se/wp-content/uploads/2024/12/scc_guide_to_the_use_of_artificial_intelligence_in_cases_administered_under_the_scc_rules-1.pdf

Vienna International Arbitral Centre, Note on the Use of Artificial Intelligence in Arbitration Proceedings (April 2025), https://www.viac.eu/wp-content/uploads/2025/04/VIAC-Note-on-AI.pdf

Silicon Valley Arbitration and Mediation Center, Guidelines on the Use of Artificial Intelligence in Arbitration (April 30, 2024), https://svamc.org/wp-content/uploads/SVAMC-AI-Guidelines-First-Edition.pdf

China International Economic and Trade Arbitration Commission, Provisional Guidelines on the Use of Artificial Intelligence Technology in Arbitration (Trial) (effective July 18, 2025)

International Chamber of Commerce, ICC Rules of Arbitration (2026), effective June 1, 2026, https://iccwbo.org/dispute-resolution-services/arbitration/rules-of-arbitration

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Annex III, point 8(a) (administration of justice and democratic processes)

Damien Charlotin, AI Hallucination Cases, https://www.damiencharlotin.com/hallucinations/ (last visited May 4, 2026)

ABA TechCred Program, https://events.americanbar.org/event/ee369916-3182-40d6-a1e9-5347aefeda2f/summary