AI in Arbitration: A Chance to Lead, Not Only a Risk to Manage

The public consultation for which AI uses should classify as “high-risk” under the EU AI Act is closed. On 23 July 2026, the window shut on the European Commission's targeted consultation on its draft classification guidelines. Arbitration showed up. Leading arbitral institutions filed a joint submission coordinated by Professor Maxi Scherer, whose earlier post on this blog put the issue in front of the field. Individual practitioners and academics filed their own.

What we filed now matters less than how it is read. Read as a compliance grievance, these submissions are one more regulated sector asking to be left alone. Read as drafting assistance, they are something better. A field that helps a regulator draw a workable line ends up governed by one. A field that only objects gets whatever is drafted without it.

From Identity to Function

Annex III, point 8(a) of the EU AI Act brings within the high-risk category AI used by a judicial authority to research and apply the law, or AI used in a similar way in alternative dispute resolution (“ADR”). The draft guidelines leave little doubt that arbitration is included. This is common ground, and it should be accepted rather than resisted. The productive question is not whether arbitration is covered but which uses are.

For judicial authorities, the European Commission has already distinguished covered from uncovered uses. The worked examples at paragraph 423 place decision-making and judgment drafting, as well as precedent selection that analyses the facts of the case, inside the classification. They place transcription, anonymization, case allocation by workload, and search functions that retrieve information without assigning meaning outside the classification.

Paragraph 434 extends this classification to ADR proceedings whose outcome produces legal effects. In her blog post, Professor Scherer considered this analogy justified. I would add one caveat, however, offered with respect for Professor Scherer and for the institutions, and in their interest as much as anyone's. The line must be drawn by function, not by the status of the actor.

Two distinctions in the EU AI Act support this claim, on different axes. The first separates providers from deployers, sorting entities along the value chain. The second, from Recital 61, separates proceedings producing legal effects from ancillary administrative activity, sorting activities rather than entities. Collapsing them is the drafting risk, because an institution is not one thing doing one kind of work.

Rémy Gerbay's study of more than forty arbitral institutions supplies the evidence. The assumption that institutional acts are always administrative does not survive contact with what institutions do, and some acts are outcome-determinative. Gerbay draws a due process conclusion from that. The classification conclusion is mine, and it follows the same logic. Institutional activity should fall within point 8(a) to the extent it is not purely administrative, and outside it to the extent it is.

Take the case that tests the rule. An institution can build a system that itself drafts the award. The AAA-ICDR launched one in November 2025, trained on construction awards, drafting for a trained human to review. Read by status, the institution is an institution, institutions administer, and the system escapes. Read by function, the institution is the provider, the reviewing human is the deployer, and the obligations attach where the risk sits. Status-based drafting would exempt the case point 8(a) exists to reach.
 

From Burden to Enabler

The safeguards the EU AI Act requires are easy to caricature as friction. Seen correctly, they make adoption safe. A personal and non-delegable mandate, genuine judgment rather than nominal review, and disclosure calibrated to what a tool did, are the conditions under which a party can trust an AI-assisted award.

The Quebec Superior Court's annulment in ARIHQ v Santé Québec is the clearest illustration available, and it is routinely misread. The judge opens his treatment of AI by endorsing the view that technology capable of improving access to justice deserves to be welcomed and regulated rather than proscribed and stigmatized. The Court then holds that there is nothing inherently objectionable in using a reliable AI tool, that the rules impose a duty of verification on counsel and parties, and that the lawyer who skips the check evades responsibility. The next sentence is the one arbitrators should copy out: the pitfalls for decision-makers are greater still.

What the Court prohibited was narrow. The arbitrator's authority had been delegated, and he had abdicated his role of reviewing the result. What it preserved was broad. Research assistants, clerks, translation and citation help, and reliance on authorities compiled by counsel all survive, provided responsibility for the reasons stays with the decision-maker. The award fell because fabricated authorities carried the entire legal foundation of the reasoning, not because a machine touched the file.

The Court went further. Not every award citing faulty references, or using AI as a drafting tool, deserves the same fate. The use may have been minimal, or bear on a lesser issue. What governs is a weighing of the breach, whether it touches the integrity of the procedure, and its effect on the award. That is a materiality standard reached through the law of annulment.

Article 6(3) of the EU AI Act arrives near the same place through classification. It allows a system falling within Annex III to escape high-risk classification where it poses no significant risk of harm to health, safety, or fundamental rights, and in particular where it does not materially influence the outcome of decision-making. Its conditions track the line the court drew in ARIHQ v Santé Québec: a narrow procedural task, the improvement of completed human work, a preparatory task. One objection can be met now. A drafting system does not escape by having a reviewer attached to it. Paragraphs 70–71 of the guidelines state that human involvement cannot change what a system is intended to do, and a provider cannot reclassify a system as low risk merely by requiring it. Article 6(3) is no loophole. It is the proportionality filter that leaves the ordinary work of arbitration alone while catching the uses that matter.
 

From Anxiety to Advantage

Arbitration has spent several years worried that AI threatens its legitimacy. The better view is that its core commitments are its advantage in an AI age. Party consent, a personally accountable decision-maker, and an enforceable award are what automation alone cannot offer. The recent debate on this blog between Federico Ast, Sophie Nappert, and Pietro Ortolani pressed that proposition from three directions. In my view, this debate confirms that the personal mandate is not a constraint to apologize for. It is what distinguishes an award from an output.

A functional rule asks something of the people who apply it, and something of those who build the tools. The line between assisting judgment and substituting for it is determinate, but it is not printed on a product page. Seeing it means knowing what a tool does to the reasoning, which means the provider has to disclose. Transparency by the provider and oversight by the deployer are one mechanism, not two. An arbitrator who cannot draw the line cannot discharge the oversight the EU AI Act requires. Counsel who cannot draw it cannot discharge the verification duty ARIHQ v Santé Québec describes.

Regulation and competence meet here, and that is why sound AI policy strengthens the arbitration community rather than burdening it. A rule drawn by function gives practitioners something to be trained against, and training turns a legal standard into daily practice. The paradigm has shifted. Judgment has not become less important. It has become harder to exercise, and therefore worth more, and it must now be exercised over tools that produce fluent prose with no idea whether it is true.
 

What the Final EU Text Will Reveal

The credible course was never to seek a place outside a classification that plainly fits. Point 8(a) fits. What arbitration can offer is precision about where inside it the risk sits, and it sits with function, not status. The consequences run past Brussels. These guidelines will be read by regulators drafting elsewhere, and by courts enforcing awards under the New York Convention, where what the arbitrator did will be litigated on facts rather than labels. A rule keyed to function gives those courts something workable. One keyed to status gives them a formality any careful institution satisfies and any careless one hides behind.

Final guidelines are expected by the end of 2026, and whether arbitration shaped them will be visible in one feature of the text. Carve out institutional activity by what institutions are, and the field asked for shelter and received it. Carve out by what the activity does, and the field helped draw a line that will hold. That is the difference between being regulated and helping to write the rule.

Prof. Barry Appleton, FCIArb is Faculty Director of the ABA TechCred Program and Co-Director and Distinguished Senior Fellow at the New York Law School Center for International Law. He filed a submission in the consultation discussed above, in an individual academic capacity. The views expressed are his own.