We Still Need to Talk About the EU AI Act – and Before 23 July Now the Draft High-Risk Guidelines Are Here

A little over two years ago – shortly after the EU AI Act of June 2024 (the “Act”) was adopted by the European Parliament – I wrote a post on this blog entitled “We Need to Talk About … the EU AI Act!”.  My point was that the international arbitration community had paid surprisingly little attention to what was then the world’s most comprehensive AI legislation.  That neglect was striking: the Act’s “high-risk” category could potentially capture arbitrators’ activities, while unresolved questions about its material, personal and territorial scope carried significant implications for arbitration and for the Act’s application to it.

Since then, both the Act and the broader regulatory landscape have attracted greater attention, and much of the Act’s machinery has come into force.  Most importantly for present purposes, on 19 May 2026 the European Commission published its long-awaited draft guidelines on the classification of high-risk AI systems (the “Draft Guidelines”).  While the consultation process remains open until 23 July, this post identifies the key takeaways and unresolved questions that, in my view, deserve close attention from the arbitration community.

Brief Recap of the High-Risk Framework

The Act takes a risk-based approach, sorting AI systems into four tiers according to their potential for harm: unacceptable, high, limited, and minimal risk.  Systems posing an unacceptable risk are prohibited outright.  High-risk systems, by contrast, are permitted, but only subject to the most demanding obligations in the Act.  Before such a system reaches the market, it must satisfy ex ante requirements including possibly risk-management and testing procedures, technical documentation and traceability, human oversight, and a conformity assessment.

The gateway for dispute resolution is Annex III, point 8(a) of the Act, which classifies as high-risk “AI systems intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution.”  Recital 61 of the Act adds that AI systems used by ADR bodies should be considered high-risk “when the outcomes of the alternative dispute resolution proceedings produce legal effects for the parties.”

Article 6 of the Act required the Commission to publish guidelines on the practical implementation of this classification.  The deadline was 2 February 2026, but the Draft Guidelines eventually were only published on 19 May 2026.  Because of that delay, among other things, the obligations attaching to high-risk systems have themselves been deferred and now apply only from either 2 December 2027 (for stand-alone high-risk AI systems) or 2 August 2028 (for high-risk AI systems embedded in products).  In other words, the community still has runway.

Material Scope: Arbitration Is In

In my earlier post I suggested that the reference to ADR in Annex III, point 8(a) of the Act “is likely to include international arbitration.”  The Draft Guidelines remove the doubt.  Annex III, paragraph 426 of the Draft Guidelines[i] defines ADR as “adjudicating a dispute out of court with the assistance of an impartial ADR body,” and offers as examples “arbitration, mediation, conciliation, ombudsmen and complaints boards.”  Arbitration is thus squarely on that list.

The Draft Guidelines also clarify the Act’s Recital 61 “legal effects” requirement.  Paragraphs 430–431 of the Draft Guidelines state that high-risk classification depends on whether the outcome produces legal effects for the parties, a question to be determined under the national law of the relevant Member State.  On that basis, arbitration will usually qualify: an award has legal effects and is enforceable, including under the New York Convention.  Mediation and conciliation, by contrast, will typically not qualify, because their outcomes do not, in themselves, bind the parties.  This is a sensible distinction and answers one of the questions I had raised: the high-risk label is triggered not simply because a process is “ADR,” but because its outcome has legal force.

Material Scope: What Kind of Activities Are Covered

Here the Draft Guidelines are at their most useful, because they translate the abstract language of Annex III, point 8(a) of the Act into examples.

“Researching and interpreting facts and the law” is understood to mean the examination and assessment of factual circumstances; the identification and interpretation of the relevant legal provisions and jurisprudence; and the assigning of meaning, the resolution of ambiguity, and the drawing of legal conclusions.  The mere organisation of documents, or the retrieval of information through advanced search tools, does not fall under these terms.  “Applying the law to a concrete set of facts” is the subsumption of the facts under the applicable law.

Two important clarifications are contained in the Draft Guidelines.

First, Annex III, point 8(a) of the Act captures “AI systems intended to be used […] in researching and interpreting facts and the law and in applying the law to a concrete set of facts […].”  Read literally, that emphasised “and” might suggest that the two limbs are cumulative, meaning that a system is covered only where it both researches and interprets the facts and the law and then applies the law to the facts.  However, paragraph 417 of the Draft Guidelines rejects that reading: in fact, the “and” is to be understood as “or.”  The two limbs are alternative conditions, so an AI system that performs either of them – researching and interpreting the facts and the law or applying the law to a concrete set of facts – is covered, without having to do both.  This materially widens the net, including in arbitration settings: a tool need not carry a matter all the way from legal research through to a worked application of the law in order to fall within the high-risk category.

Second, the Draft Guidelines underline that an AI system need not itself decide anything to be covered.  Paragraph 418 of the Draft Guidelines makes clear that the use case does not require the AI system to actually perform the adjudicative tasks; it is enough that it “assist[s]” a judicial authority in a manner relevant to its decision-making.  An AI system can therefore be high-risk even where it merely helps apply the law to the facts without going so far as to recommend a decision.

The Draft Guidelines also give concrete illustrations of the above.  AI systems that generate a draft decision, that support adjudicators in handling identical or similar cases by extracting and clustering the relevant facts, or that select relevant precedents and suggest how to apply them to the facts, fall within Annex III, point 8(a) of the Act.  By contrast, systems that perform speech-to-text, that facilitate communication with the public, that assign cases, that manage and search large volumes of evidence and produce a chronology, or that carry out ancillary administrative tasks such as anonymising decisions, fall outside.

Equally important is Article 6(3) of the Act which works as a “filter” and exempts otherwise high-risk systems that perform only a narrow procedural task, that merely improve the result of a previously completed human activity, or that detect patterns or deviations without being meant to replace or influence the human assessment absent proper human review.  The Draft Guidelines illustrate this category too.  Are exempt from high-risk for instance: pre-classification of a matter by subject (contract, inheritance, and so on); metadata extraction (identifying procedural roles such as parties, counsel and experts); advanced search; proof-reading and language editing that does not change the content; and tools that suggest factual questions, for instance, checking, after a decision has been drafted, whether anything in the file has been overlooked.

For arbitrators, this amounts to a practical map.  A tool that drafts reasoning or proposes how the law applies to the facts sits in the high-risk zone; a tool that transcribes a hearing, organises exhibits, or polishes language does not.  The grey area – and there is one – lies in the tools that hover between research and analysis, and arbitrators will need to look closely at what their tools actually do, rather than at how they are marketed.

Finally, one caveat is important.  The Draft Guidelines are framed around AI systems used by, or on behalf of, judicial authorities (i.e., courts and judges etc) rather than arbitral tribunals.  The examples therefore refer to “judges” and “judicial decisions.”  Paragraph 434 of the Draft Guidelines, however, indicates that those examples may also be relevant to ADR, stating that “[p]ractical examples of AI systems falling within the scope of this use case can be derived from the examples provided for AI systems intended to be used by judicial authorities above.” In my view, that analogy is justified.  Annex III, point 8(a) of the Act expressly extends to ADR where AI systems are “used in a similar way,” and the Draft Guidelines themselves recognise arbitration as a form of ADR, as discussed above.  Even so, the particular features of arbitration may not always be fully captured by the examples given in the Draft Guidelines.

Personal Scope: Arbitrators, Arbitral Institutions, Counsel and Experts

On personal scope, my earlier post concluded that arbitrators, as natural persons using AI for a professional activity, are “deployers” within Article 3(4) of the Act.  The Draft Guidelines do not disturb that conclusion.  They go further, however, on three points left open in the Act itself: arbitral institutions, counsel and experts.

First, as to institutions, paragraphs 416 and 428–429 of the Draft Guidelines note that the Act does not define which entities constitute “ADR bodies” and then include within that notion “commercial arbitration institutions (e.g. national and international arbitration bodies) or investment dispute bodies that resolve disputes between businesses.”  This is striking and, I would suggest, problematic.  It appears to rest on a misunderstanding of how arbitration works.  Arbitral institutions administer proceedings; they do not, as a rule, decide the merits of the dispute.  That decision is the tribunal’s, of course.  To treat the institution as the relevant ADR “body” for the purposes of the high-risk classification therefore risks attaching the obligations to the wrong actor, at least in most cases.

Importantly, concerning judicial activities, the Draft Guidelines stress repeatedly that judicial administration (i.e., “institutions or bodies that are responsible for the management, governance, and support of the judiciary, rather than for the adjudication of cases themselves”) is not covered by the Act.  The same should apply to arbitral institutions.  This is one area in which the draft would clearly benefit from refinement before it is finalised.

Relatedly, the statement that “investment dispute bodies […] resolve disputes between businesses” also appears inaccurate.  Investment tribunals resolve disputes involving not only businesses, but also sometimes individuals as claimants and, most importantly, always States as respondents.  This is another point on which the Draft Guidelines would benefit from further reflection.

Second, as to counsel or legal representatives, the Draft Guidelines are clear and, in my view, correct.  Paragraph 416 provides that AI systems intended to be used by parties and their legal representatives do not fall within Annex III, point 8(a) of the Act, because parties and their representatives are not acting “on behalf of a judicial authority.”  The Use of AI tools by counsel to prepare submissions is therefore not a high-risk activity under the Act.  That is a sensible boundary: the high-risk regime targets the adjudicative function, not advocacy.

Third, the position of experts also requires distinction.  Tribunal-appointed experts may fall within Annex III, point 8(a) of the Act.  Paragraph 415 of the Draft Guidelines states that court-appointed experts act “on behalf of a judicial authority” when they prepare reports requested by that authority and intended to assist it in exercising its adjudicative functions.  The same reasoning would apply to tribunal-appointed experts; it would not, however, extend to party-appointed experts in arbitration proceedings.

Territorial Scope: Left Unanswered

Some of the hardest questions from my earlier post, on the territorial scope, are untouched.  Article 2 of the Act notes that it covers deployers established or located in the EU, and deployers outside the EU “where the output produced by the AI system is used in the Union.”  For an arbitral tribunal, that raises questions the Draft Guidelines do not address: is the relevant place the respective arbitrators’ habitual residences, or the seat of the arbitration?  Is it enough that one party is located in the EU, or that an award might ultimately be enforced against assets in the EU?  The potential for significant extraterritorial reach that I described two years ago – reaching even arbitral tribunals seated, and arbitrators based, outside the EU – remains.

We Need to Talk More than Ever!

When I ended my earlier post, I noted that the arbitration community still had time, but that it “need[ed] to engage in the debate.”  That remains even truer today: the issues are now clearer, and the window for contributing is closing.

The Draft Guidelines answer several of the questions I raised: arbitration is ADR and high-risk turns on the legally binding effects of the outcome (i.e., the award).  The adjudicative tasks that trigger the high-risk classification, and the administrative tasks and “filter” exceptions that do not, are spelt out with welcome examples.  Arbitrators as deployers are covered by the Act and counsel are not.  But the Draft Guidelines also raise new questions, most obviously the list of examples and treatment of arbitral institutions.

Crucially, the guidelines remain in draft.  Arbitration practitioners, arbitral institutions and other stakeholders should therefore scrutinise them now and submit their comments by 23 July 2026, rather than wait until the rules enter into operation.  The arbitration community still needs to talk about the EU AI Act, now more than ever!

I wish to thank Chiann Bao (ArbBoutique), Emily Hay (ArbBoutique), Kevin Nash (LCIA) and Veronika Pavlovskaya (LCIA) for comments on an earlier draft of this blog post.  All errors and opinions are mine only.