The New ICC Rules Are Not Silent About Their Own AI Silence: They Already Govern AI

The 2026 ICC Arbitration Rules (the “Rules”) never mention artificial intelligence (“AI”). They already govern it. Articles 20, 46, 23(2), and 37(3) build the whole architecture for AI governance, and the tribunal-secretary provisions quietly state the test. The ICC has done more than it knows.

The Rules entered into force on June 1, 2026, the most substantial revision of the ICC framework in a decade. They say nothing about AI.

They arrived the same season that the Québec Superior Court, in ARIHQ v. Santé Québec, set aside an arbitral award because the authorities on which the arbitrator's reasoning rested had been fabricated by a generative AI tool and did not exist. That case is the backdrop to everything that follows.

The reflexive reading is that the ICC has failed to address AI. I want to argue the opposite. The ICC's decision to leave AI out of the rule text is defensible, deliberate, and probably correct. A rulebook revised every five years is the wrong instrument for a technology that changes every five months. The ICC said as much by removing the fixed list of case management techniques from the Rules altogether.

But a deliberate allocation still has costs, and those costs fall now on the parties arbitrating under these Rules, on the tribunals deciding what belongs in Procedural Order No. 1, and on the enforcement courts that will review awards. The responsibility sits with the arbitrator, not the institution, as it always has.

The good news is that the ICC has built the machinery to discharge that responsibility. Four provisions do the work. None mention AI. All of them govern it.

The Architecture the ICC Already Has

Article 20 is the first provision: where the Rules are silent, the proceedings are governed by rules that the parties agree upon, or, absent their agreement, such rules as the tribunal determines appropriate. That silence delegates authority to party autonomy and, failing that, to tribunal discretion.

Article 46 is the second, and it is the foundational one. 

“In all matters not expressly provided for in the Rules, the Court and the arbitral tribunal shall act in the spirit of the Rules and make every effort to ensure that the award is enforceable at law.”

Read that against the AI question. AI is a matter not expressly provided for, so Article 46 applies to it by its own terms, and what Article 46 commands is enforceability. An arbitrator whose undisclosed and unverified use of a generative AI creates a live annulment risk is acting against the standard the Rules set for her. No amendment is required for that to be true today.

A careful reader will object that Article 46 is boilerplate, and that reading a verification duty into its enforceability limb is overreach. But Article 46 is the operative general rule for every matter the Rules do not expressly address, and AI is such a matter.

Article 23(2) is the third. Procedural measures adopted by the tribunal may include case management techniques described in the guidance section issued by the Secretariat, taking into account the work of the Commission on Arbitration and ADR, which has established a task force on AI in dispute resolution. The rule text routes any resulting guidance straight into the tribunal's procedural powers, though nothing has yet flowed through it.

Article 37(3) is the fourth, and it is the ICC's distinctive asset. When the Court scrutinizes draft awards, it considers, to the extent practicable, the validity, enforceability, and mandatory-law requirements at the seat. Few institutions apply as exacting a standard of scrutiny, and that scrutiny is an inquiry into the same register in which the AI question lives. That is enough to bring AI-driven enforceability risk inside the ICC's existing mandate.

The framework is complete. It is simply empty.

Where the ICC Has Already Written the Right Test, Without Noticing

The most interesting drafting in the Rules concerns tribunal secretaries, not AI.

Article 44 permits the tribunal to appoint a secretary to work under its direction and control, without delegating decision-making authority. It also requires the secretary to meet the same independence, impartiality, and confidentiality standards as an arbitrator. The ICC's Note to Parties and Arbitral Tribunals on the Conduct of the Arbitration (the "Note") goes further. The tribunal may never delegate its decision-making functions, and entrusting a secretary with notes or memoranda does not relieve it of the obligation to personally review the file and draft its own decisions.

The Note then lists what a secretary may permissibly do: prepare drafts of procedural orders and the factual portions of an award, subject to the tribunal's own review; conduct legal research; and check citations, dates, and cross-references.

Whether AI is literally a tribunal secretary is not the interesting question, since obviously it is not a person. The real question is whether it can assume the secretary's core risk, substituting its output for the arbitrator's own reasoning. Framed that way, the "fourth arbitrator" analogy misleads, because it invites a status test rather than a functional one.

The better reading is that the ICC has already written a functional test without noticing it describes functions, not persons. The tribunal-secretary provisions do not ask who performs a task; they ask what task is performed. Paragraph 210 draws an absolute line against delegating the decisional function; paragraph 211 identifies a defined set of assistive functions that may be delegated. Because the boundary turns on the function, not the identity of the actor, it applies to a machine as readily as a human. Applied to machines, this yields three tiers.

Tier 1, mechanical. Search, retrieval, translation, transcription, formatting, verifying a citation against a real database. The tool locates; it does not compose. No disclosure obligation. This is what a legal database has always done, and treating it otherwise would be absurd.

Tier 2, generative-assistive. Summarizing the record, drafting chronologies and factual narrative, producing memoranda on the law for the tribunal's independent evaluation. The system now composes propositional content capable of migrating into the award. Paragraph 211 permits exactly this from a human. When performed by a machine, it should carry the same conditions: disclosure to the parties, verification of every output against the source, and the tribunal's personal-review floor.

Tier 3, decisional. Weighing evidence, evaluating arguments, generating reasoning or conclusions that become the award's own. Paragraph 210 forbids this absolutely. It is non-delegable and not curable by party consent, because it goes to the mandate that the parties conferred on the arbitrator.

The dividing line runs through function, not personhood, turning on whether the system generates propositional content that can enter the award. Retrieval locates; generation composes. A modern research tool can do both in a single query, yielding a Tier 1 output and a Tier 2 output at once. The obligation attaches to whichever output could migrate into the award, whatever the tool that produced it. The place to fix that is the first procedural order. The ICC does not need to invent this test; it needs to recognize that it has written one.

The EU Has Just Drawn the Same Line

This is not only an ICC insight. The EU has drawn the identical line. The EU AI Act's Annex III point 8(a) classifies as high-risk any AI used by a judicial authority to research and apply the law, or used similarly in dispute resolution. Arbitration is covered, and the classification attaches now, even though compliance is not required until later. The Digital Omnibus on AI, adopted this June, deferred the Annex III high-risk obligations from August 2, 2026 to December 2, 2027. The question is not whether arbitration falls within scope, but which uses do, and that runway is the window in which an institution should build.

The Commission's draft classification guidelines closed for targeted consultation on July 23, 2026, with final guidelines expected by year-end, and they do the ICC's work for it. The worked examples place decision-making, judgment drafting, and precedent selection inside the classification, and place transcription, anonymization, case allocation, and retrieval-only search outside it. That is Tier 3 and Tier 1, drawn by function, not status. Article 6(3) then exempts a system that does not materially influence the outcome, the same materiality line the Québec court reached through the law of annulment in ARIHQ.

I have argued this function-over-status point to the Commission directly, and set it out on this Blog. It matters because it dispels the objection that institutions escape regulation by merely administering. Under a functional reading, the institution is the provider, the reviewing human is the deployer, and the obligation attaches where the risk sits. The AAA-ICDR AI Arbitrator tests this rule directly; status-based drafting would exempt the very case the classification exists to reach. The ICC's tribunal-secretary test and the EU's classification test are the same instrument. An institution that recognizes this is drafting with, not against, the mandatory law under which its awards will be enforced.

The author thanks Arsenio Moussa for his research assistance and gratefully acknowledges the support of the Center for International Law at New York Law School.