
Peter Malanczuk(Germany)
Former Member of the International Law Association (ILA) Committee on International Commercial Arbitration
Member of the Academic Advisory Council of the University of Heidelberg
Distinguished judges, experts and colleagues, ladies and gentlemen.
The use of artificial intelligence in arbitration raises a fundamental question: who decided the case?
My remarks concern the relationship between the EU Artificial Intelligence Act, adopted in 2024 as the world’s first comprehensive and binding legal framework for AI, and the enforcement of AI-assisted awards under the 1958 New York Convention.
I shall focus on an award rendered outside the European Union and later presented for enforcement within it.
The AI Act treats certain systems used in the administration of justice and alternative dispute resolution as high-risk. Recital 61 states that AI may support adjudication, but final decision-making must remain human-driven.
The Commission’s 2026 Draft Guidelines seek to clarify this classification. A joint submission by sixteen arbitral institutions emphasises an essential distinction: arbitral institutions administer proceedings, but arbitral tribunals decide disputes.
There is also an important temporal qualification. Although the Act’s territorial-scope provisions already apply, the principal high-risk requirements for Annex III systems have been postponed until 2 December 2027. An award produced in 2026 ordinarily cannot violate requirements that are not yet applicable.
The extraterritorial issue arises from Article 2(1)(c). It applies to providers and deployers outside the EU where output produced by an AI system is used in the Union.
Suppose a tribunal seated in Beijing or Singapore uses AI to research the law or prepare an award, and the successful party later seeks enforcement in an EU Member State. It might be argued that the AI output is thereby used in the Union.
That conclusion is not automatic. The final award may not be identical to the AI-generated output. The enforcing party may never have operated the system. The court uses the award as a legal instrument, not necessarily the underlying AI output. Moreover, EU enforcement may not have been intended when the output was produced.
What, then, are the consequences under the New York Convention?
Where no EU court or EU AI Act is involved, AI use is not an independent ground for refusing enforcement. The objection must be connected to Article V.
Article V(1)(b) may apply if the tribunal relies on AI-generated facts, authorities or reasoning not put to the parties, depriving a party of a meaningful opportunity to present its case.
Article V(1)(d) may apply where delegation of the arbitrators’ personal mandate makes the tribunal’s composition or procedure inconsistent with the parties’ agreement or the law of the seat—for example, if AI determines liability or damages without meaningful human review.
Article V(2)(b) may apply in an extreme case where the AI system, rather than the tribunal, effectively decided the dispute. Whether that violates public policy depends on the standards of the enforcement forum.
Before an EU court, the Convention must be applied together with any applicable requirements of EU law. The AI Act presently creates no independent ground for refusing enforcement.
An objection based on the AI Act requires three links: territorial application; breach of an applicable obligation; and an existing Article V ground.
EU rules of fundamental importance may, however, enter the Convention through Article V(2)(b) as part of EU public policy. Only if EU law were interpreted as requiring refusal beyond Article V would a genuine conflict arise between EU-law primacy and the Member State’s international obligations under the Convention, subject, where applicable, to Article 351 TFEU.
The decisive distinction is between assistance and delegation.
If the arbitrators evaluated the evidence, determined the law, reached the conclusions and critically reviewed any AI-generated text, AI remained a tool.
If the system materially determined the reasoning or outcome and the arbitrators merely adopted its product, the question becomes unavoidable: who, in substance, decided the case?