THE AUTHOR:
Prof. Barry Appleton, New York Law School
This week the ICC convenes in New York for its 21st annual conference on international arbitration. AI is impossible to avoid at any professional gathering right now, and I expect arbitration institutions to be no exception. The real question is what, specifically, the ICC should be saying about it, and this piece is my answer.
The 2026 ICC Arbitration Rules never mention artificial intelligence, and they already govern it. That architecture is complete. This piece is about what belongs inside it, and why the reasoned award, the instrument arbitration has relied on for a century, is quietly failing at the one job it was built to do.
The Objection Worth Taking Seriously
The strongest reply to everything above is that no new rule is needed at all, because the arbitrator is already personally responsible. Whatever tool she used, the award is hers. She answers for it. Done.
That is true, and it misses the point. Personal responsibility is a liability rule, not a procedural rule. It tells you who bears the loss after an award has been annulled. It does not prevent the annulment. It gives the parties no information before the fact, nothing to object to, and no basis on which to calibrate their own conduct. It gives the reviewing court no record to review. Procedure exists to make responsibility operable before the loss happens, not merely nameable afterward.
Association des ressources intermédiaires d’hébergement du Québec (ARIHQ) v. Santé Québec [2026 QCCS 1360](“ARIHQ”), covered on Daily Jus when it was decided, is the proof. The liability rule functioned exactly as designed. The arbitrator was personally responsible, and the award was set aside anyway, and the parties lost the years. The award fell because the reasoning had been delegated, evidenced by fabricated authorities no arbitrator exercising personal judgment would have produced. That is why a rule which only assigns blame after the loss functions as an autopsy, not a safeguard. Nor is ARIHQ isolated: Damien Charlotin’s AI Hallucination Cases database had already cataloged well over a thousand fabricated outputs in legal proceedings by the time it was decided.
Six Things the ICC Should Consider Doing Now
None wait for the ICC’s next full Rules revision. Ranked from what needs no new text at all to what asks for one short addition.
- Publish a Model AI Clause for Procedural Order No. 1. The Note to Parties invites tribunals to use the ICC Model Data Protection Clause for Procedural Order No. 1. There is a model clause for the data. There is none for the reasoning. That gap should close, and closing it constrains no one. A model clause functions as a shared starting point, not a mandate. It lets a professional community circulate hard-won practice quickly, and any tribunal remains free to adapt it or depart from it entirely. That asymmetry, a model for the data and none for the reasoning, can close with no amendment and no party agreement. It should be first.
- Add authorities verification to award scrutiny. The ICC Award Checklist already guides arbitrators in drafting. Article 37(3) already directs the Court to consider enforceability and mandatory law. A committee already reads every draft award before it issues. Adding a verification step for cited authorities requires no rule change and targets the ARIHQ failure mode at the one chokepoint best placed to catch it before an award goes out, not after a court has to unwind it.
- Amend Article 24(2), the provision fixing the procedural timetable at the initial case management conference, so the initial conference must address AI use and information security. With the Terms of Reference gone, that conference is the only remaining moment at which the parties and the tribunal are guaranteed to agree on anything about AI before the case proceeds. Make the conversation compulsory. The tribunal remains free to decide that nothing further is needed.
- Extend Article 12(8), or use Article 23(3) deliberately. As drafted, an arbitrator who uploads the file to a public AI model breaches confidentiality, because Article 12(8) binds arbitrators alone. A party doing the identical thing breaches no ICC obligation, though her contract may bind her anyway. Article 23(3) lets the tribunal make confidentiality orders on a party’s request and protect trade secrets and confidential information, so the gap can close case by case. It should not have to. The asymmetry is the most obvious drafting flaw here, and the easiest to fix.
- Recast Article 44 functionally. Define assistance by what is done, not by who does it. Paragraph 210 of the Note forbids delegating the decisional function; Paragraph 211 permits a defined set of assistive functions. That boundary already turns on function, not the identity of the actor. A retrieval tool and a generative tool completing the same drafting task are not the same thing under this test, not because of what either is called, but because of what each actually produces. The ICC only needs to say so expressly and extend the same language to machines.
- Condition the unreasoned award, rather than abolishing it. Under the Highly Expedited Arbitration Provisions, unless the parties agree that no reasons are to be given, the award must state its reasons. I would not remove that waiver; speed has genuine value, and parties are entitled to buy it. But where reasons are dispensed with, the ICC should require a statement of authorities relied upon. The cost to speed is trivial, because the arbitrator has already consulted those authorities to reach the decision; listing them is disclosure, not new research. The gain, for reasons the rest of this piece explains, is not trivial at all.
A Word on Timing
Six recommendations that a Note update or a short Rules proviso can carry is not a call for hasty regulation. A real objection exists elsewhere in the commentary. AI-specific soft law tends to arrive before any settled practice exists to codify, prescribing rules for a technology nobody has finished learning to use. That objection lands against guidelines purporting to state emerging best practice. It does not land against a model clause, which asserts nothing about what practice has settled. It offers optional drafting language, adoptable or ignorable case by case, on the model of the Data Protection Clause the Note already provides. The difference is between telling tribunals what to think and giving them somewhere to start.
The Paradigm Shift, Stated Plainly
Here is what I think is actually happening, and it is larger than the ICC. Arbitration’s legitimacy rests on a personal mandate. The parties chose this arbitrator. Every instrument in the regulatory vocabulary (the personal mandate, the reasoned award, the secrecy of deliberation, and non-delegation) is built on the premise that the unit of adjudication is a discrete human mind, and that the question “who decided?” always has a clean answer. I develop those four doctrines at greater length in my Lex AI work.
Once cognitive labor is distributed across a human and a machine, that question stops having a clean answer. Not because the arbitrator has abdicated, but because deciding is no longer a single locatable act.
And this is the part I most want practitioners to sit with. Reasons were never the point in themselves. Reasons were the technology by which we verified that a mind had engaged. A reasoned award is evidence of cognition. That is why the flaw in ARIHQ was detectable at all: the fabricated citation was the tell, and the tell was only available because reasons were given and could be checked. But when reasons can be generated as readily as they can be reasoned, reasons stop being probative of engagement. The instrument survives. Its evidentiary value quietly collapses.
That is the paradigm shift, not the fact that arbitrators use software. The reasoned award no longer reliably performs the fact-finding and informational work it was built to perform, and the field has not yet built its successor.
Provenance, Not Prohibition
The successor, I think, is provenance, not prohibition, which will fail, and not a disclosure checkbox, which will be gamed. The durable answer is an auditable record of how the award was produced: a structured log generated contemporaneously, not reconstructed after the fact, capturing which tool touched which function, what it produced, what source it was checked against, and who signed off on the check. Borrow the concept from the law of evidence. A chain of custody does not ask whether evidence could have been tampered with. It asks whether the record shows, step by step, that it was not.
For the ICC, this requires no new technology, only a retention convention layered onto research tools already in use. The record sits in the case file the way a research memo does today. It need not be public. It needs only to exist, be time-stamped at creation and not assembled afterward, and be producible to a reviewing court if enforcement is challenged. That timing is what resists manipulation. A record built after an annulment challenge invites the same suspicion ARIHQ‘s fabricated citations did. A record contemporaneous with the drafting does not, because reconstructing a false one means fabricating timestamps across an entire log the tribunal has relied on internally, a materially harder act than fabricating a single citation. The statement of authorities proposed above for unreasoned awards is this same idea in miniature: even where the parties have bought their way out of reasons, a record of what was checked is the one thread of provenance the award can still carry.
This is the artifact around which the next generation of arbitration rules will be organized. The institution that builds it first sets the global standard, and does so with the enthusiastic support of every party that has to enforce an award under the New York Convention (1958). Whether AI-assisted awards survive that enforcement scrutiny is being actively debated.
The ICC scrutinizes every draft award in the world’s largest arbitration caseload. It is the natural author of that standard. Article 46 already tells it to make every effort to ensure the award is enforceable at law. Every effort now means adopting this one. If the 21st ICC New York Conference on International Arbitration gives the subject the floor it deserves, that would be a fitting place to start.
ABOUT THE AUTHOR
Prof. Barry Appleton, FCIArb, is Faculty Director of the ABA International Law Section’s TechCred credential, Co-Director and Distinguished Senior Fellow of the Center for International Law at New York Law School, and Managing Partner of Appleton and Associates International Lawyers LP.
*The views and opinions expressed by authors are theirs and do not necessarily reflect those of their organizations, employers, or Daily Jus, Jus Mundi, or Jus Connect.
