The Deadline Moved. Six Things Did Not.
Europe has just given international arbitration sixteen more months. It has not given you a reprieve, and the decision that will govern your practice was taken while everyone was watching the calendar
The Deadline Moved. Six Things Did Not.
Europe has just given international arbitration sixteen more months. It has not given you a reprieve, and the decision that will actually govern your practice was taken while everyone was watching the calendar.
Prof. Barry Appleton, New York Law School | TechCred | 28 July 2026
Where this stands as at 28 July 2026. The Digital Omnibus on AI was adopted by the Council on 29 June 2026, following the European Parliament’s endorsement on 16 June. It defers the stand-alone high-risk obligations under Annex III of the AI Act from 2 August 2026 to 2 December 2027. Arbitration sits in Annex III. If you are reading this later, that date is the one to check first.
For the past year, arbitration practitioners who were paying attention had 2 August 2026 circled. That was the day the high-risk regime under the EU AI Act was to apply, and with it Article 86, which entitles a person affected by a decision taken on the basis of a high-risk system to a clear explanation of the role that system played.
That date has moved to 2 December 2027.
Over the next fortnight you will hear a great deal of relief about this. Most of it will be misplaced, and acting on it will cost some practitioners a considerable amount of avoidable difficulty.
What moved, and what the move did not touch
The Omnibus deferred the compliance date. It did not defer the classification date.
On 19 May 2026 the European Commission published its draft guidelines on the classification of high-risk AI systems. The consultation closed on 23 July 2026, five days before this post. The guidelines will be finalized long before December 2027, and they will settle the two questions that matter more than the deadline ever did: which uses of these systems in arbitration are caught, and who carries the obligation when they are.
Whether arbitration is in at all is no longer seriously contestable. Paragraph 426 of the draft defines alternative dispute resolution as “adjudicating a dispute out of court with the assistance of an impartial ADR body,” and gives arbitration as an example. Recital 61 confines the limb to mechanisms producing legal effects for the parties. An arbitral award produces legal effects. Arbitration is in, and the Commission is right that it is in.
Everything interesting is downstream of that.
Six things that did not move
The personal mandate. An arbitrator may take help. An arbitrator may not transfer the decision. The Silicon Valley Arbitration and Mediation Center guidelines say the mandate cannot be delegated. The Chartered Institute of Arbitrators guideline says the same. None of this depends on the AI Act, and none of it has a commencement date in 2027.
ARIHQ c. Santé Québec, 2026 QCCS 1360. In April 2026 the Superior Court of Quebec annulled an arbitral award because artificial intelligence had generated authorities that did not exist and the award was built upon them. The vice was the delegation of the decision, not the use of a tool. Binding in Quebec today and persuasive everywhere. The Court was careful to limit itself: it said expressly that not every award citing erroneous authority, or using these tools in drafting, should meet the same fate. What made this breach fatal was that the fabricated authorities were the only legal support the award had.
Lnu v. Blanche (9th Cir. 3 June 2026). While Brussels was deferring, a United States court of appeals was suspending. Two attorneys who filed briefs citing decisions that never existed were sanctioned, suspended from practice before the court for six months, and their firm placed under a two-year obligation to certify in every filing whether generative tools were used, which tool, and that the signing lawyer personally verified every citation and quotation.
ABA Formal Opinion 512. Live since 29 July 2024. It requires American lawyers to apply an appropriate degree of independent verification to generative output, and requires supervising lawyers to train those they supervise in the capabilities and limitations of the tools. Most American arbitrators are American lawyers.
Article 50 transparency and Article 4 literacy. The Omnibus left the Article 50 transparency obligations on their original schedule, and the AI literacy obligation in Article 4 has applied since February 2025. Add to that data protection law, breach notification, your own professional conduct rules, and the institutional notes on AI use. A deferral of one chapter is not a holiday from the rest.
Your file. Which brings us to the trap.
The trap: arbitrations that straddle the deferral
This is the point that the relief is going to cost people.
An arbitration commenced this autumn will, on any ordinary timetable, still be running in December 2027. Many will be. A case that begins under a deferred regime and ends under a live one is not protected by the deferral, and the obligations that arrive in December 2027 will arrive in respect of a proceeding whose record was built in 2026.
You cannot construct a contemporaneous record retrospectively. That is what “contemporaneous” means. If an explanation is ever demanded of what a tool did and what you did, the answer is easy to give if you noted it at the time and impossible to reconstruct afterward.
The deferral gives you sixteen months to build good habits on cases where the stakes of getting it wrong are lower. It does not give you sixteen months of files you will never have to explain.
The real question: which uses, and whose obligation?
Paragraph 428 of the draft guidelines observes, correctly, that the Act does not define which entities constitute “ADR bodies.” Paragraph 429 then includes commercial arbitration institutions within that notion.
On 15 July 2026 a group of ten leading arbitral institutions with strong links to the Union filed a joint letter in the consultation, coordinated by the LCIA under the leadership of Professor Maxi Scherer. Their point is narrow and, in my view, plainly right: arbitral institutions administer proceedings. They do not decide the merits. Treating the institution as the relevant ADR body for its administrative services attaches the obligations to the wrong actor.
Professor Scherer has been telling this community to pay attention to the AI Act since 2024, when almost nobody was, and she pressed practitioners to file in the consultation rather than wait. The profession is in her debt on this subject. I supported their request in my own filing and said the guidelines should grant it.
What I filed, and why
I lodged comments with the European AI Office on 17 July 2026, in an individual academic capacity. Readers are entitled to know what I argued so they can discount for it.
I wrote in support of the draft guidelines, and to offer refinements. I did not ask for arbitration to be excluded. The credible course for arbitration is not to seek a place outside a classification that plainly fits. It is to help the Commission draw the line well.
Classify by function, not by label. The hard question is not whether arbitration is covered. It is which uses are, and the answer is functional. A system that drafts the reasoning, applies the law to the facts, or identifies authorities and proposes how to apply them operates in the high-risk zone. A system that extracts metadata, searches, proofreads or translates does not. That distinction is already in the Act: paragraph 70 of the draft guidelines makes clear that human involvement does not change what a system is intended to do, and Article 6(3) filters out systems performing a narrow procedural task or merely improving a completed human activity. A general-purpose assistant does not become high-risk because an arbitrator could open it. It becomes high-risk when it is put to, and relied upon in, the adjudicative function.
A United States federal appellate court reached the same conclusion about its own rules six weeks before the consultation closed. In Lnu v. Blanche, the Ninth Circuit explained that its procedural and ethical rules “are not violated at the point of research and drafting, but at the point of signing and filing,” that it does not matter whether a false citation came from an artificial intelligence tool “or from his own natural intelligence,” and that the rules apply “however legal papers are prepared, and however legal technology develops.” (Professor Eugene Volokh’s summary is the quickest way in.) Rules written to the function survive the technology. Rules written to the tool do not, and every list of tool categories drafted in 2026 will be obsolete before the obligations apply in December 2027.
The deployer is the arbitrator. The person who uses the system in the adjudicative function is the arbitrator, and the arbitrator therefore carries the deployer obligations. Where an institution provides administrative and organizational services, case management, scheduling, the logistics of appointment, it is not exercising the adjudicative function and should not carry deployer duties for those services.
One case needs care, and it is the constructive one. An institution can do more than administer. It can build and operate a system that itself performs the adjudicative function. The AAA-ICDR AI Arbitrator, launched in late 2025 outside the Union, drafts awards for a trained human to review before signature.
That is a careful design and the human-in-the-loop architecture is the right one. My point in the filing was not about that institution but about the drafting. Where an institution builds such a system it is the provider, and the human who adopts the output is the deployer. Each carries its own duties along the value chain, which is exactly how the Act is built to work. A classification keyed to the label of the actor rather than to the function performed would sort that case wrongly, and would do so in the one situation where the Act most needs to operate cleanly. Institutions that innovate with care should be recognized for it, and only a function-based line can recognize them.
Territorial scope has a gap, and I proposed filling it. The Act reaches deployers established in the Union, and deployers outside it where the output is used in the Union. For arbitration the guidelines do not say which connecting factor governs. My proposed answer: the relevant place for deployer status is the arbitrator’s establishment or habitual residence, not the seat, because the seat is a legal construct and not the place where a person uses a tool. The Commission should resolve that rather than leave it to be litigated.
Counsel’s tools are outside point 8(a). The provision reaches artificial intelligence used in the adjudicative function. It does not reach the tools counsel and parties use to present their case. That use is governed by professional responsibility rules and by the courts, as Lnu v. Blanche demonstrates. Where it bears on the fairness of the proceeding, the answer is the tribunal’s own duty of oversight, not treating counsel’s software as high-risk.
And nominal human involvement should not defeat the classification. Genuine judgment is required, not a signature on a machine’s output. ARIHQ is the illustration: the vice was the delegation of the decision. A clear functional line protects the enforceability of awards. It does not threaten it.
What this means for your next case
Article 3(4) defines a deployer as a natural or legal person using an AI system under its authority, other than in a personal, non-professional capacity. As Professor Scherer put it in her 11th EFILA Annual Lecture at Queen Mary in November 2025, an arbitrator using these tools professionally is a deployer of them.
Then notice what Article 2 does not contain. There is no criterion anywhere in it keyed to the seat.
Take a Delaware corporation arbitrating against a Texas corporation, New York seat, New York law, American counsel on both sides. Now add one ordinary fact: the claimant’s party-appointed co-arbitrator lives in Dublin. Nothing else about the case touches Europe. That is enough to put a deployer on your tribunal.
And a three-member tribunal deliberates jointly. Whether one member’s obligation reaches the tribunal’s collective process, or only her own contribution, is genuinely unsettled and will be argued. An arbitrator who has to argue it has already lost the year.
The tribunal secretary is a gap I did not raise in my filing, and on reflection I should have. The secretary is very often the person actually operating the tools. Some of that work, transcription, chronologies, formatting, is plainly outside the adjudicative function. Some of it, particularly research that identifies authorities and proposes how they apply, sits much closer to the line. The secretary works under the tribunal’s authority, which is the language of the deployer definition. If your secretary’s use is not recorded, the tribunal cannot give an account of it later even if it wants to, because nobody knows what was done.
What a careful tribunal does now
Four practices. Each is right on every reading of every open question above, which is the point of choosing them.
Run a connecting-factor audit at the first procedural conference. Record the establishment or habitual residence of each arbitrator and the secretary, and whether any party, witness or affected person is located in the Union. Five minutes. Almost nobody does it. You cannot apply a connecting factor you have never asked about.
Draft Procedural Order No. 1 as a tribunal undertaking, not a party permission. A permission invites the argument that what was granted could have been withheld, and therefore that what was withheld could have been granted. If the obligation turns out to be non-derogable, drafting it as a party permission concedes the case in your own document.
Bring the secretary inside the discipline, in writing, with a contemporaneous note of any tool used and for what.
Sort your own uses into the two categories now. Assistance that does not enter the reasoning, and use that supplies or shapes it. Once you have made that sort for your own practice, tiered disclosure becomes straightforward and the classification debate becomes somebody else’s problem.
The competency point, which is why this matters to you
Anyone can read a deadline off a calendar. What distinguishes a technologically competent arbitrator in 2026 is harder: the ability to hold an unsettled regulatory position and still run a procedure that is defensible whichever way it settles.
There is a second half to it, and the consultation is the illustration. Competence in a field whose rules are still being written is not only the capacity to comply. It is the capacity to participate. The rules governing how arbitrators use these systems are being drafted this year by people who will decide the question with or without us. Reading draft guidelines closely enough to find the gap at paragraph 429, and then saying so on the record, is the same skill as reading a procedural order closely enough to find the gap that will cost your client. The TechCred faculty is in that conversation because that is where the subject currently is, and I have told the Commission I remain at its disposal.
Several things remain genuinely open. Whether Article 86 travels with the Annex III deferral or applies on its own footing, given that it sits in a different chapter of the Regulation, is arguable. Whether one arbitrator’s obligation reaches a joint deliberation is arguable. Whether an award enforced in Frankfurt is “output used in the Union” for a New York arbitrator is arguable, and I have asked the Commission to settle it rather than leave it to litigation.
None of that changes what you do on Monday. Every path through every one of those questions ends at the same four practices, which is how you know they are the right four.
The institutions have made their submission and made it well. What happens to a plural tribunal when only one of its members is a deployer, and where the secretary sits in all of this, are questions the consultation has not yet reached. I will have more to say about them shortly, and in a venue better suited to the argument.
Barry Appleton, FCIArb, is Co-Director and Distinguished Senior Fellow of the Center for International Law at New York Law School, Faculty Director of the ABA International Law Section TechCred Program, and a Scholar and Fellow of the Balsillie School of International Affairs. He filed comments with the European AI Office on 17 July 2026 in the consultation on the draft guidelines for the classification of high-risk AI systems, in an individual academic capacity. Provisions cited are verified against the Official Journal text of Regulation (EU) 2024/1689. This post draws on a forthcoming article, “The Law Applicable to the Arbitral Regime.” For the institutional perspective, see the LCIA’s account of the cross-institutional initiative.
REGISTER NOW FOR TECHCRED 2026
 | A guest post by
|