top of page
Search

Can Artificial Intelligence Replace Human Arbitrators? Research Says the New York Convention Can Accommodate AI Awards

shwetasabuji
4 days ago
10 min read

A research paper by Horst Eidenmüller, Professor of Commercial Law at the University of Oxford, and Faidon Varesis, then a PhD candidate at the University of Cambridge, argues that fully AI-powered arbitration will become technically feasible at some point and should be permitted by law. Their central claim is that nothing in the concept of an arbitration requires human control, governance or even input, and that the New York Convention of 1958 can be read to accommodate awards made by an AI system. The paper, titled What is an Arbitration? Artificial Intelligence and the Vanishing Human Arbitrator, is available on SSRN.

Paper at a Glance

  • Paper: What is an Arbitration? Artificial Intelligence and the Vanishing Human Arbitrator

  • Authors: Horst Eidenmüller (Professor of Commercial Law, University of Oxford) and Faidon Varesis (PhD Candidate, University of Cambridge, at the time of writing)

  • Published: 2020, SSRN Electronic Library, abstract 3629145

  • Question asked: Can an arbitration be conducted entirely by an AI system, and would the resulting decision qualify as an arbitral award?

  • Instruments examined: The New York Convention 1958, the UNCITRAL Model Law on International Commercial Arbitration, and domestic arbitration statutes

  • Credit: All ideas, analysis and arguments summarised below belong to the authors. This post is ILW's plain-language summary and commentary, and it is not a substitute for reading the original paper.

Why This Question Matters Now

Arbitration is a private and consensual method of dispute resolution that ends in a binding decision, the arbitral award. For most of the twentieth century, when modern international commercial arbitration took shape, human arbitrators were the only technologically possible option. The authors point out that digitisation, artificial intelligence and blockchain are now disrupting that format, and that the COVID-19 pandemic pushed the arbitration community towards online hearings and smart tools.

The authors also cite survey evidence that parties choose arbitration mainly for quality. In the survey they discuss, about 45 percent of respondents rated the neutrality and expertise of arbitrators and the free choice of arbitrators as very important, while only around 5 percent rated lower cost as very important. That matters because the paper's argument is that AI could eventually deliver better quality, and not only cheaper and faster proceedings.

The Anatomy of an Arbitration: Three Functions

To test whether a machine can run an arbitration, the authors break arbitration down into three functions and ask whether each can be performed without a human.

1. An Independent and Impartial Third Party

There is no arbitration without a neutral third party. The authors say an AI system can in principle be designed to be independent and impartial, and independence is the easier test because it rests on a limited set of facts, as the IBA Guidelines on Conflicts of Interest show. The harder problem is legal personality. A third party needs a legal entity status so that rights, obligations, compensation and liability have a reference point.

They discuss two routes. One is a new category of electronic personality for sophisticated robots, an idea the European Parliament pondered in a 2017 resolution. The other, which they find simpler, is a self-driving corporation managed exclusively by AI, allowed only with safeguards such as strict liability for algorithmic failure and calibrated mandatory insurance. In their view, an AI-managed corporation providing arbitration services would conceptually qualify as an independent third party.

They add an important limit. A system used by only one party to manage its own contracts is not an arbitration. But if one party's standard terms designate a system to handle future disputes, the process could be an arbitration provided the independence and impartiality criteria are met.

2. Management of the Process

An arbitration involves establishing facts, applying agreed procedural rules, applying the substantive law and handling many case management tasks. The authors see no conceptual reason why only humans can do this. The real limiting factor is not software but the availability of carefully labelled training data. Even a hearing, they argue, only needs to give each party a forum to be heard and to present its case, and an AI-powered hearing can do that functionally even if it looks nothing like a conventional hearing.

3. Rendering an Award

Rendering an award means applying law to facts and giving reasons. Civil law systems reason top-down from statutory rules, while common law systems reason bottom-up from precedent. The authors accept that explainability is the biggest hurdle. The UNCITRAL Model Law (Article 31(2)) requires an award to state its reasons unless the parties agree otherwise, and parties almost always want reasons. Probabilistic assessments will not be enough, and stakeholders will trust an AI award only if it offers detailed reasoning on facts and rules. The authors call this a question of technical capability and progress, not of principle.

Humans need not necessarily apply when it comes to rendering an award.

Where AI Stands Today in Arbitration

The paper separates AI that assists arbitrators from AI that would replace them.

  • Case management: Smart scheduling assistants that read emails and find meeting slots can help organise hearings across time zones.

  • Fact gathering and analysis: Document review, e-discovery, clustering and automated transcription tools reduce the workload of reviewing large volumes of material.

  • Decision support: Predictive analytics tools such as ArbiLex, Lex Machina and Solomonic map risk factors and estimate outcomes or timing, although they are data-hungry and depend on large labelled datasets.

As for replacement, the authors say that at the time of writing no system could perform all the tasks of an arbitrator, which include conflict checks, fact analysis, assessing witness credibility, and a reasoned final decision. They expect AI arbitrators to arrive first for simple disputes such as straightforward money claims or tax disputes, where outcomes rest on quantifiable variables. Disputes with hidden social or economic variables remain much harder, and merging different AI tools into one architecture is a formidable problem.

Can the New York Convention Accommodate an AI Award?

International enforceability is a major reason parties choose arbitration, so the authors treat the New York Convention as the gateway. States can change their own arbitration laws, but they cannot unilaterally change the Convention. The authors' answer is that the Convention is a living instrument that can be interpreted teleologically.

  • Gap, not exclusion: The Convention says nothing about the nature of the arbitrator, although its drafters assumed humans. For example, the requirement of an authenticated award has been read to need the signatures of the arbitrators as natural persons, as the Austrian Supreme Court held in 2008.

  • Meaning of an award: Applying Articles 31 and 32 of the Vienna Convention on the Law of Treaties, the authors say a decision is an award if it finally and bindingly resolves a dispute, in whole or in part. An AI decision that does so should qualify.

  • Article I(2): The reference to permanent arbitral bodies was added at the request of the USSR and Czechoslovakia, and the authors find no sign it was meant to exclude legal persons. A self-driving corporation delivering a final and binding decision would meet the Convention's purpose.

  • Precedent for adapting to technology: Article II(2) speaks of letters and telegrams, yet a 2006 UNCITRAL recommendation says its list is not exhaustive, which supports a functional reading that covers electronic communication.

  • Public policy, Article V(2)(b): This is the real risk, because public policy is judged by the enforcement forum. The authors note that France requires arbitrators to be natural persons in domestic arbitration, so an AI award would likely face refusal there, while England and other Model Law countries are likelier to look at the fairness of the process. They argue AI awards should get the same scrutiny as human awards and no higher bar.

What National Laws Say About Non-Human Arbitrators

The authors survey the position at the time of writing and find a mixed picture, with most jurisdictions silent.

  • Legal persons permitted: Greece (Article 871(1) of the Code of Civil Procedure), Iran (its 1997 international arbitration law refers to natural persons or legal entities), and a Spanish court that allowed accounting firms to act as arbitrators in accounting disputes.

  • Natural persons required: France for domestic arbitration, the Netherlands (Article 1023), Scotland (only an individual may act), and England, where Crowther v Rayment [2015] EWHC 427 (Ch) held that only a natural person can be an arbitrator. The authors still think the English Arbitration Act 1996 is open to a purposive reading.

  • Model Law and party autonomy: Article 19 lets parties agree on procedure, subject to Article 18, which guarantees equal treatment and a full opportunity to present one's case. The authors say this should be read broadly enough to let parties choose AI-assisted or even fully autonomous arbitration.

  • Current position: To the authors' knowledge, no jurisdiction then explicitly allowed a legal person without human directors, or an unincorporated AI system, to run an arbitration.

Regulatory Competition and the Common Law Advantage

States compete for arbitration business, and the authors predict that this competition will push them to modernise their laws for AI. They describe it as a race to the top rather than to the bottom, because arbitration rests on consent between usually sophisticated parties and has no obvious third-party externalities. The public policy exception acts as a brake on extreme experiments, and states that want to attract technology-driven arbitrations will also have to use that exception with restraint when enforcing foreign AI awards.

The authors also expect common law jurisdictions to have an edge. Machine learning of the supervised kind needs carefully labelled data, and precedent-based reasoning and publicly reported cases supply more of it than top-down statutory reasoning does. Indeed, they note that the existing AI litigation analytics tools were developed in and for common law jurisdictions.

What the Authors Recommend for Cautious States

  1. Review unnecessary mandatory rules: Arbitration laws should be enabling rather than mandatory. For example, they should not require in-person hearings or in-person witness testimony, except where equality and the right to be heard demand it.

  2. Help label legal data: Courts should publish judgments in accessible form, including in English, and institutions and tribunals should publish awards in anonymised form by default. The authors admit this conflicts with the commercial preference for confidentiality, and describe reduced confidentiality as the possible price of a better service.

ILW Commentary: Reading the Paper From an Indian Perspective

The paper does not analyse Indian law, so the following points are ILW's own observations and not the authors'. India's Arbitration and Conciliation Act, 1996 is modelled on the UNCITRAL Model Law and, like the Model Law, it was written with human arbitrators in mind. Section 2(1)(d) defines the arbitral tribunal as a sole arbitrator or a panel of arbitrators, Section 12 requires a person approached for appointment to disclose circumstances giving rise to justifiable doubts, and Section 31 requires awards to be signed by the members of the tribunal and, unless the parties agree otherwise, to state reasons. Whether such provisions can be read purposively to cover an AI system, or whether Parliament would need to legislate, is an open question.

Two cautions are also worth keeping in mind. First, the paper is from 2020 and is expressly forward-looking, and the authors themselves say no fully autonomous arbitrator existed then. Generative AI tools have entered legal practice since, so the timeline in the paper deserves fresh scrutiny. Second, the argument is normative. Concerns about bias in training data, opacity of reasoning, accountability when a system errs, and whether parties would genuinely consent to machine decision-makers all remain live debates.

What This Means for Law Students and Lawyers

For now, the realistic near-term impact of AI in arbitration is assistance, not replacement. Document review, transcription, scheduling, research and draft preparation are already changing how arbitration teams work. Lawyers who understand both the arbitration framework (seat, party autonomy, enforcement, public policy) and these tools will be better placed than those who understand only one.

If you want to build a practice in arbitration, ILW's Advanced Certificate Course in Arbitration and Dispute Resolution Practice covers the framework that this paper builds on, from arbitration agreements and tribunal procedure to awards and enforcement, taught by practising professionals.

Frequently Asked Questions

Can artificial intelligence act as an arbitrator?

According to the authors, no system could perform every arbitrator function when the paper was written. They argue that nothing in the concept of arbitration requires a human, that simple disputes may be the first candidates for AI arbitrators, and that the limits are technological and legal rather than conceptual. Most national laws still assume that arbitrators are humans.

Would an award made by an AI system be enforceable under the New York Convention?

The authors argue that it can be, if the Convention is interpreted according to its object and purpose, because a final and binding decision on a dispute performs the function of an award. The main obstacle is the public policy exception in Article V(2)(b), which each enforcement country applies under its own law. The point is untested, and courts in countries that require natural-person arbitrators, such as France, are more likely to refuse enforcement.

Which countries allow non-human arbitrators?

As discussed in the paper, no jurisdiction then explicitly allowed an unincorporated AI system or a legal person without human directors to run an arbitration. Greece and Iran provide for legal persons or entities to be appointed, and a Spanish court has accepted accounting firms as arbitrators. France (domestic arbitration), the Netherlands, Scotland and England require natural persons.

What is the difference between AI-assisted and AI-powered arbitration?

In AI-assisted arbitration, human arbitrators use AI tools for scheduling, document review, transcription or outcome prediction, and the award remains the product of a human. In AI-powered arbitration, an autonomous system manages the process and renders the decision without human involvement. The authors say the first raises no real conceptual issue, while the second requires a closer legal analysis.

Why might common law countries have an advantage in AI arbitration?

Supervised machine learning needs large, carefully labelled datasets. Common law systems reason from individual cases and precedents, and their decisions are publicly reported, so more usable training data exists. Civil law systems reason top-down from statutes, which the authors say makes data labelling harder.

Keywords

AI arbitrator, artificial intelligence in arbitration, AI-powered arbitration, robot arbitrator, New York Convention, UNCITRAL Model Law, enforcement of arbitral awards, public policy exception, Article V(2)(b), self-driving corporation, regulatory competition in arbitration, legal technology, Horst Eidenmüller, Faidon Varesis

Source and Credits

Horst Eidenmüller and Faidon Varesis, What is an Arbitration? Artificial Intelligence and the Vanishing Human Arbitrator (2020), SSRN: https://ssrn.com/abstract=3629145. Full credit for the research and arguments goes to the authors. Into Legal World has summarised the paper for educational purposes and has added clearly marked commentary of its own.

 
 
 

Recent Posts

See All

Comments

bottom of page