The enforceability of arbitral awards shaped or drafted by artificial intelligence has moved from theory to imminent practice, and the 1958 Convention on the Recognition and Enforcement of Foreign Arbitral Awards (the New York Convention) is set to face a challenge its drafters never contemplated. The debate has sharpened following the American Arbitration Association's launch of its AI Arbitrator, described as the first AI-assisted system introduced by a major arbitral institution. Awards produced through such systems will eventually reach national courts, where their treatment under the Convention remains untested.
One marker already exists: a Canadian court recently set aside an award on the ground that the arbitrator relied on AI-generated hallucinations. That case, however, arose in a domestic setting and did not engage the New York Convention.
Does an "award" require a human decision-maker?
Commentators note that the Convention nowhere expressly requires an award to be rendered by a human—it demands only a binding determination from a competent tribunal under a valid arbitration agreement, with legitimacy flowing from procedural integrity and party consent. Because some national laws permit legal persons to serve as arbitrators, it is difficult to argue for a transnational principle confining tribunals to natural persons.
That textual latitude, however, meets several practical limits. AI cannot yet replace a human arbitrator in a complex dispute, so the operative question is how it may assist rather than substitute. AI systems lack legal personhood. And several statutes plainly require a human arbitrator—Article 1023 of the Dutch Arbitration Act among them—so that 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. A further view reads the Convention as embodying an "ineffable human element," tying legal reasoning to knowledge that presupposes a human who knows.
Rubber-stamping and delegation
Where a named arbitrator signs an AI-drafted award without independently evaluating the merits, the concern is characterised not as "AI versus human" but as a mismatch between process and promise—a procedural irregularity capable of engaging Article V(1)(d). The parallel drawn is the tribunal secretary, on which national authorities differ, and the threshold for refusal is high; much turns on what arbitrators actually did when "reviewing." A distinct scenario arises where the arbitration agreement itself requires the tribunal merely to transform an AI output into an award, likened to a consent award whose content was generated by a machine rather than agreed by the parties.
Bias, consent and the limits of party autonomy
Algorithmic bias is structural rather than relational, so the conventional disclosure-and-challenge mechanism—built around an arbitrator's biography—translates poorly. One proposed response is a dual framework of transparency of deployment (disclosing tools, prompts and human verification) and architectural diversity (running multiple independent models to cancel out individual biases).
On consent, commentators broadly agree it can achieve much but not everything: party agreement may override Article V(1) but cannot displace the public-policy exception in Article V(2). Under the EU AI Act, adjudicative use of AI would likely be classified as "high-risk," triggering obligations on risk management, transparency and human oversight; Article 25 could even treat a tribunal deploying a general-purpose model as a "provider." In the gravest cases, that could produce a violation consent cannot cure.
No rewrite needed
On whether the New York Convention requires amendment, the assessment is consistent: the instrument's core concepts—award, properly constituted tribunal, public policy—retain enough elasticity to absorb AI without touching the text. The more fruitful work lies in clarifying those concepts, improving practitioner understanding of the technology, and developing soft law on AI disclosure.