Debate: Are AI-Assisted Awards Enforceable Under the New York Convention?

When the Convention on the Recognition and Enforcement of Foreign Arbitral Awards (“New York Convention”) was adopted in 1958, the idea that an arbitral award might be assisted—or even generated—by a machine would have belonged to science fiction. Sixty-eight years later, that science fiction is real.

The topic of how artificial intelligence (“AI”) may influence the enforceability of arbitral awards has been a recurring theme on the Kluwer Arbitration Blog since 2020, most recently here. Lately, it has gained renewed urgency with initiatives like the AAA’s introduction of its AI Arbitrator—the first AI-assisted system launched by a major arbitral institution. Awards rendered under such systems will soon reach state courts, and when they do, the New York Convention will be put to a test its drafters never anticipated.

That moment is fast approaching. A Canadian court has recently become the first to set aside an award on the ground that the arbitrator relied on AI-generated hallucinations, although that case arose in a domestic context and did not engage the New York Convention.

This seems to us the right moment to take stock. We therefore invited three acclaimed practitioners and scholars recognised for their expertise in both arbitration and technology to engage—in this new debate format—with six questions at the heart of this matter.

Federico Ast is the founder and CEO of Kleros, the world’s first decentralised dispute resolution platform combining blockchain and AI. He holds a Ph.D. from IAE Business School and has lectured at Harvard, Stanford, and Oxford.

Sophie Nappert is dual-qualified in Quebec and England & Wales and acts as an independent arbitrator at 3 Verulam Building (3VB) in London. In 2021, she co-founded ArbTech, a community fostering dialogue on technology and dispute resolution.

Pietro Ortolani is Professor of Digital Conflict Resolution at Radboud University. He previously served as Senior Research Fellow at the Max Planck Institute Luxembourg and Law Research Associate at Queen Mary, University of London.

Ast and Ortolani begin from the same textual observation: the New York Convention does not expressly require that an award be rendered by a human. As Ast notes, its text mentions no human decision-maker—perhaps because, in the 1950s, that went without saying— requiring only a binding determination from a competent tribunal under a valid arbitral agreement, so that legitimacy flows from procedural integrity and party consent. Ortolani reaches the same point from a different direction: since some national laws allow legal persons to serve as arbitrators, it is hard to argue that arbitration entails any transnational principle confining tribunals to natural persons.

That textual freedom, Ortolani cautions, calls for a reality check. First, AI remains far from able to replace a human arbitrator in a complex dispute, so the real question is not whether it can act as an arbitrator but how it can assist one. Second, even where legal persons may serve as arbitrators, that possibility is exceedingly rare, and it is unclear whether it could extend to AI systems, which have no legal personhood. Third, many arbitration statutes clearly require a human arbitrator (e.g., Article 1023 of the Dutch Arbitration Act); there, a machine-rendered decision could be set aside—or treated as non-existent—opening the door to refusal under Article V(1)(e) of the New York Convention.

Nappert approaches the question on a more fundamental level, reading the New York Convention as the tangible recognition of international arbitration as an instrument for the guardianship and application of the rule of law. Historically, she argues, international arbitration belongs to the “human element” of governance, which a school of legal philosophy terms “thoughtfulness and the rule of law”—the idea that human beings want to be ruled thoughtfully, rather than rigidly and mechanically. AI, she notes, is deployed to streamline and optimise dispute resolution, with the laudable purpose of cheaper, more accessible justice. The valid question, for her, is how this affects the rule of law, weighing “thoughtfulness” against expedient, effective justice.

From this, she draws her counterpoint to the textual reading: there is a valid argument that the New York Convention conceives of the award as comprising an ineffable human element, perhaps enhanced by AI, but human, nonetheless. What, she asks, would become of the rule of law if we were to jettison its “thoughtfulness” in favour of the mechanistic approach currently inherent to AI?

Can enforcement of an award be refused where AI drafted the award and a human arbitrator merely reviewed and signed the output, rather than independently evaluating the merits?

For Ast, the personal-mandate doctrine is triggered not by AI per se but by deception. When a named arbitrator rubber-stamps an AI output without independently evaluating the merits, the parties get something other than what they bargained for—a signature, not the human judgment they chose. Since an award carries, for Ast, the implicit representation that its signatory actually thought about it, an AI-drafted award rubber-stamped by a human already engages Article V(1)(d) of the New York Convention as a procedural irregularity; the analysis would be no different had the arbitrator handed the decision to an intern and signed off. According to Ast, the question is not “AI or human?” but whether the process matched the promise.

Ortolani connects the concern to existing doctrine through two scenarios. In the first, the tribunal leans so heavily on AI drafts that doubts arise whether its adjudicative function was improperly delegated to a machine; the closest parallel is the tribunal secretary, on which national authorities differ. The threshold for refusal is very high, he argues, but some systems do not exclude that a wholesale de facto delegation could breach requirements on the arbitral mandate, the proper constitution of the tribunal, or public policy—so much depends on what the arbitrators actually did when “reviewing”. The second scenario is different in kind: where the arbitration agreement itself requires the tribunal merely to rubber-stamp an AI output and “transform” it into an award without altering its substance. Here, Ortolani draws a comparison to a consent award, with the twist that the content was not agreed by the parties but generated by an AI system.

Can enforcement of an award be refused on the ground that the applicable rules require the award to be reasoned and its reasons were generated by AI?

Nappert treats this question together with the notion of an award, since for her the two turn on the same point. Legal reasoning, she explains, is closely linked to knowledge: epistemology tells us that knowledge requires someone who knows — historically a human being. Knowledge also “has the function of focusing our attention on what we do not know”, and legal knowledge in particular is “an activity of mind […] not reducible to a set of directions or any fixed description.” This, Nappert concludes, begs the question whether AI “knows” anything at all and, since reasoning is so closely linked to knowledge, whether the concept of reasoning comprises an essentially human element.

Ast takes the opposite view. The real problem with AI-generated reasoning, he argues, is not that it is machine-authored but that it may hallucinate—producing reasons unsupported by evidence; yet that risk is the same whether the error came from a senior partner or an LLM. A categorical rule against machine authorship would be unworkable anyway: where every word processor now offers AI drafting, what counts as “machine-authored”, and would using AI for style or research taint the result? Procedure should be about outcomes, not origins—the standard is the integrity of the reasoning, not the identity of the reasoner.

Can enforcement of an award be refused on the ground that AI systems carry embedded biases that cannot be addressed through the conventional mechanisms of disclosure and challenge concerning arbitrators?

Nappert’s answer is doctrinal: the lack of impartiality of an arbitrator or tribunal—human or algorithmic—is not, in her view, as such a ground for resisting enforcement under the New York Convention.

As Ast sees it, the disclosure-and-challenge mechanism was built for human arbitrators, whose conflicts are relational and surface by investigating prior relationships; it presupposes a person with a biography and translates poorly to algorithmic systems, where bias is structural rather than personal. That mismatch is a reason to build better procedures, not to treat all AI involvement as a ground for refusal.

The solution, for Ast, is a rigorous procedural audit trail of the specific dispute. He proposes a dual framework: first, transparency of deployment—disclosing the tools used, the purpose, the prompts, and the human verification, shifting the focus from a model’s provenance to its actual function in the decision; second, architectural diversity—running multiple independent models and requiring their convergence to cancel out individual biases, following the same logic that has long made a three-member tribunal a safeguard against a sole arbitrator’s bias.

If both parties expressly agree to have their dispute substantially shaped or even decided by an AI system and waive any challenge based on that fact, would that consent neutralise all enforcement risks?

Here, all three agree that consent can do much, but not everything. Nappert puts it most concisely: such an agreement is valid, subject to mandatory laws, but can only ever override Article V(1), not Article V(2) of the New York Convention, which concerns public policy.

Ast agrees that consent is decisive but draws the line elsewhere. He argues that genuine, informed, bilateral consent from sophisticated parties, given before the dispute, should not be second-guessed, and the public-policy exception should not become a paternalistic override. But where parties cannot understand what they accept, Article V(2)(b) of the New York Convention remains available. The line, for him, is not AI versus human but genuine versus manufactured consent.

For Ortolani, much depends on the governing framework: under the EU AI Act, this use would likely count as “high-risk”, triggering obligations on risk management, data governance, transparency, and human oversight, and mandatory rules may simply bar parties from delegating the adjudicative function wholesale to a machine. Ortolani also flags that, under Article 25 of the EU AI Act, a tribunal that uses a general-purpose LLM to research and apply the law might itself be treated as the “provider” of a high-risk system although it never developed the model. In the gravest cases, he explains, this could be a violation of the EU AI Act that consent cannot cure, leaving open, as uncharted territory, whether enforcement might then be refused under Article V(2)(b) of the New York Convention.

In light of current AI developments, does the New York Convention remain adequate as a living instrument capable of being interpreted to meet these challenges, or do they call for targeted amendments?

On the final question, the three are, in substance, of one mind: the New York Convention does not need rewriting.

Nappert calls it a cornerstone of international arbitration in the twenty-first century and locates the more fruitful reform in clarifying the tenets that underpin it—what we mean by an award, a properly constituted tribunal, and public policy in an era of AI-assisted arbitration.

Ast agrees: the New York Convention was drafted to be deliberately flexible from the outset; it absorbed the early digital revolution without amendment, and its core concepts retain enough elasticity to accommodate AI without touching the text. The enforcement analysis, he argues, should focus on procedural transparency, auditability, and meaningful consent. Ast sees a more urgent need in improving practitioners' understanding of how these systems work and developing soft law to guide AI disclosure. That, in his phrase, is a training, not a drafting problem.

Ortolani ends on a pragmatic note. Even if amendments were warranted in principle, he asks whether they are feasible. In his view, the recent work of UNCITRAL Working Group II suggests little appetite for amending the New York Convention, whether by revising the 1958 text or adding a protocol. He adds that the years needed for wide ratification could split the Contracting States into those bound by the new version and those bound only by the original. The realistic path, he suggests, is a recommendation—albeit non-binding—on how to interpret the existing text.

In summary, this debate has shown that the New York Convention can withstand the emergence of AI. Its existing grounds for refusing enforcement already let a court test an AI-assisted award without a word being added to the text of the Convention. However, no treaty can decide how much human judgment we are willing to automate in the name of faster and cheaper justice— that rests with the arbitrators who reach for these tools, the parties who agree to their use, and the courts asked to enforce what they produce.