The Robot Arbitrator Can’t Issue Confirmable Awards

The red eye of the HAL-9000, an artificial intelligence from the 1960s. Could HAL be a robot arbitrator? — Credit: Pixabay (CC0)

Introduction: How to fight the robot arbitrator

There are a lot of reasons why I am pretty strongly opposed to allowing AI systems—what I like to call the “robot arbitrator”—to decide real disputes in the real world. I’ve written about them here at Letters Blogatory before. Here is a post that suggests that the robot arbitrator is incompatible with the social institution of dispute resolution, just like the “automated ball-strike system” is incompatible with baseball as a social institution. Here is a post that takes a similar line. Here is one that attacks a startup that is offering the robot arbitrator now, today, and whose founders think their system is “rooted in the modern enlightenments of formalism and free markets where efficiency is God and the machine is its vicar on Earth.” But I can see the writing on the wall. Major League Baseball is using the robot umpire to call balls and strikes. The world really is run, or it seems to be, by people who write things like “efficiency is God and the machine is its vicar on Earth.” It’s time to get more serious about the problem of the robot arbitrator, because it’s coming faster than we think.

One approach would be philosophical. I could noodle about what “judgment” means, and the ways in which we know the human minds has capabilities that a computer system that runs a formal algorithm cannot in principle have, and what distinguishes a “person” from a machine on the one hand or a nonhuman animal on the other,¹ and what an algorithmic decisionmaker means for access to justice. But I think there’s a better approach. A more practical approach, and one that, if accepted, has the potential to stop the robot arbitrator in its tracks.

Here is the approach. Arbitration works as a system only because the state undertakes to enforce arbitral awards. Yes, we hope that parties voluntarily comply with awards. But it’s hard to imagine sophisticated parties agreeing to arbitrate or including arbitration agreements in their form contracts if a court will not enforce the arbitrator’s decision. So I looked at the Federal Arbitration Act, and it seemed fairly obvious to me, for reasons I’m about to explain, that a robot arbitrator’s award cannot be confirmed under the statute.

My goal for the argument is to put forth the most boring, technical, almost trivial argument against the confirmability of a robot arbitrator’s award that I can muster. But I also want to put forth an argument that, if it’s right, is going to be really hard for the robot arbitrator people to overcome, because it’s going to require amendment of a statute that hasn’t really been amended in a fundamental way in a century. In other words, I want to put a pebble in the path of a big stream of AI boosterism, but a special kind of pebble that can dam the stream in a way that only the US Congress can break.

I want to make one acknowledgement, to my friend and colleague David Evans. He and I started out working together on an article that explained why awards generated by AI systems without a human in the loop are not confirmable under the FAA. But while we agreed about that basic point, it became clear we had some disagreements about methodology, about some of the underlying issues, and maybe about the future of AI in arbitration. So I owe a lot to discussions with David, and we decided to each pursue this in our own ways.

How to interpret a statute

If the law worked the way I wish it worked, we wouldn’t have theories or even ideologies of statutory interpretation. Judges with a broad experience of the world and of the law would exercise their judgment in a way that makes good sense in the historical, legal, moral, and practical context they’re working in. We haven’t lived in that world for a long time, especially in the United States. Our Supreme Court has a very well-defined theory and ideology of what statutory interpretation is about. Here is a recent restatement, from Chief Justice Roberts’s opinion in Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), the case overturning the doctrine of Chevron deference to agency interpretations of statutes:

[Ambiguous] statutes, no matter how impenetrable, do—in fact, must—have a single, best meaning. That is the whole point of having written statutes; every statute’s meaning is fixed at the time of its enactment.²

To find this “single, best meaning” that was “fixed at the time of … enactment,” the Court looks at contemporary dictionaries,³ and at the context in which the phrase in question is used within the statute.⁴

You can see where this is going. The FAA was enacted in the 1920s, before Gödel and Turing, before the ENIAC, when Hobbes’s idea that “REASON … is nothing but Reckoning (that is, Adding and Substracting) of the Consequences of generall names agreed upon, for the marking and signifying of our thoughts” still had the power to shock. Let’s look at the language that the FAA uses—at how its words were understood generally a hundred years ago and at how they fit in to the context and structure of the statute to see whether, at the time, the statute could have been read to mean that a court can confirm an award that is the output of a vastly complicated computer program that has taken the parties’ submissions as input.

Interpreting the FAA

Here is the relevant part of 9 U.S.C. § 9, the FAA’s provision on confirmation of awards:

Award of arbitrators; confirmation; jurisdiction; procedure

If the parties in their agreement have agreed that a judgment of the court shall be entered upon the award made pursuant to the arbitration, and shall specify the court, then at any time within one year after the award is made any party to the arbitration may apply to the court so specified for an order confirming the award, and thereupon the court must grant such an order unless the award is vacated, modified, or corrected as prescribed in sections 10 and 11 of this title. …

Notice that the text of the statute itself does not use the word “arbitrator.” It speaks of “the award made pursuant to the arbitration.” Section 2, the statute that puts agreements to arbitrate on an equal footing with other contracts, is similar: it speaks of agreements “to settle by arbitration a controversy,” but it does not use the word “arbitrator.” The first step in my argument means to show that the statute requires that the award be made by an arbitrator. Once we get that far, we’ll show that by “arbitrator,” the statute means a natural person.

No arbitration without an arbitrator

First, look at the title Congress gave to § 9. While the body of the statute doesn’t mention the arbitrator, the title does. As the Supreme Court noted in 2023, it “has long considered that the title of a statute and the heading of a section are tools available for the resolution of a doubt about the meaning of a statute.”⁵ Next look at § 5 of the statute. That section explains the role of courts in appointing an arbitrator when the parties have not agreed, or when there is some vacancy in the office. The last provision in § 5 is particularly relevant. It provides: “unless otherwise provided in the agreement the arbitration shall be by a single arbitrator.” Last, look at § 10, the part of the statute that explains when a court can vacate an award. Section 10 will be important later in the argument, too. Among the grounds for vacatur are “evident partiality or corruption in the arbitrators,” misconduct or misbehavior by “the arbitrators,” or “the arbitrators” exceeding their powers.

So let’s take it as given that in an arbitration, there has to be an arbitrator to do the arbitrating. Is it just as clear that the arbitrator has to issue the award? The title of § 9 supports that view, as does Section 11, which explains when a court can modify or correct an award. The statute allows modification “Where the arbitrators have awarded upon a matter not submitted to them.” All the textual evidence supports the view that you need an arbitrator in order to have an arbitral award. I don’t think that point is controversial.

The arbitrator must be a natural person

Now we get to the heart of the argument. Suppose the parties’ contractor designates a particular AI system as the arbitrator. Is that use of the term “arbitrator” consistent with the statute?

Here is what we shouldn’t do. We shouldn’t beg the question by saying that “a human being can reason, exercise moral judgment, and decide what evidence to credit and what evidence not to credit. An AI system can’t do those things, therefore, an AI system can’t be an arbitrator.” Don’t mistake me. I certainly think that human beings have capacities that an AI system cannot, in principle, have. But there are a lot of people who disagree, and trying to make these arguments gets us into very deep and very contested waters quickly. One of the hazards of thinking about this stuff is that you start to wonder whether human beings really have the wonderful abilities and capabilities we say distinguish them from AI systems.⁶ First, we’ll look at some contemporary dictionaries. But then, we’ll see that the FAA provides us with a really clever way to do this. Instead of focusing on human beings’ strengths and capabilities relative to AI systems, let’s focus on our weaknesses and vulnerabilities relative to AI systems.

Here are what the old dictionaries have to say:

Black’s Law Dictionary 83-84 (2d ed. 1910). ARBITRATOR. A private, disinterested person, chosen by the parties to a disputed question, for the purpose of hearing their contention, and giving judgment between them; to whose decision (award) the litigants submit themselves either voluntarily or, in some cases, compulsorily, by order of a court.
Webster’s New International Dictionary (1913). Arbitrator. 1. A person, or one of two or more persons, chosen by parties who have a controversy, to determine their differences.

It can hardly be surprising that both of these dictionaries define an arbitrator to be a person. No one in 1925 could really have conceived of anything else, except in a literary sense.⁷

If we turn to the statute but leave aside § 10 for the moment, we see some hints. One hint is in § 7, the provision on summoning witnesses. The arbitrators “may summon in writing any person to attend before them or any of them as a witness.” The summons “shall issue in the name of the arbitrator or arbitrators, or a majority of them, and shall be signed by the arbitrators.” I don’t think it makes sense to imagine a summons that issues in the name of an algorithm, or that it makes sense to imagine that an algorithm can “sign” a document even though it can output text. But the stronger point is that it isn’t sensible to imagine a witness “attending before” an algorithm. The statute presupposes that the witness will testify in the presence of the arbitrators (today, we would include the virtual presence, given the wide use of videoconferencing technology in arbitration).

But § 10, which I referenced earlier, gives the clearest indication. Here is the relevant part of the statute:

In any of the following cases the United States court in and for the district wherein the award was made may make an order vacating the award upon the application of any party to the arbitration—

where the award was procured by corruption, fraud, or undue means;

where there was evident partiality or corruption in the arbitrators, or either of them;

where the arbitrators were guilty of misconduct in refusing to postpone the hearing, upon sufficient cause shown, or in refusing to hear evidence pertinent and material to the controversy; or of any other misbehavior by which the rights of any party have been prejudiced; or

where the arbitrators exceeded their powers, or so imperfectly executed them that a mutual, final, and definite award upon the subject matter submitted was not made.

The statute shows the kinds of failings that an arbitrator can have. They are human failings. An arbitrator can be guilty of partiality or corruption. An arbitrator can misbehave in was that prejudice the rights of a party. These failings are different from the kind of failings we can attribute to an AI system. The robot arbitrator may exhibit “algorithmic bias,” but it is not sensible to think of that as “partiality or corruption.”⁸

There is an irony in distinguishing human beings from AI systems by focusing on the ways in which human beings fail or underperform. Still, both the usage of the word “arbitrator” explained in dictionaries from the time of the FAA’s enactment and evidence from the statute itself show strongly that the word “arbitrator,” certainly in 1925, meant a human being who carries out the arbitrator’s tasks.

What about party autonomy?

But what about the parties’ freedom to decide how their disputes should be resolved? Isn’t that a cornerstone of arbitration? We know from the Hall Street case that the parties cannot modify the FAA by agreement. (In Hall Street the question was whether the parties could expand the statutory grounds on which a court could vacate or modify an award). The same is true here. If, as I argue, an “arbitrator,” as the term is used in the FAA, is a human being, then it is for Congress, not the parties to change the statutory rules. That’s not to say the parties can’t agree to resolve their disputes via a robot arbitrator. It’s just to say that if one party or the other decides it doesn’t like the outcome, no one can point to the FAA and ask the state to step in and enforce the agreement. There are limits to party autonomy.

I acknowledge, though, that courts have occasionally overweighted party autonomy in the interest of promoting arbitration. The example that is most familiar to me is one I had some personal involvement in, the Rockefeller Technology Investments v. Changzhou Sinotype Technology Co. case.⁹ I wrote the (unsuccessful) cert. petition seeking review of the decision. The issue in the case was whether the parties could contract around the Hague Service Convention, which allows states to refuse to permit service of process by postal channels in their territory. But the lawsuit in which the issue arose was for confirmation of an arbitral award, and so the case got a lot of attention from arbitration scholars and practitioners, mainly concerned with whether a decision not to allow “contracting around” would disadvantage parties to arbitrations. My own view, which might or might not be right, is that although the technical issue in the case didn’t have anything to do with arbitration, the courts’ policy favoring arbitration was important to the outcome. So I am not too comfortable predicting that the limits to party autonomy I’ve noted above would carry the day if litigated.

Conclusion

Awards issued by AI systems are not awards of “arbitrators” and therefore are not awards that a court can confirm under the FAA. Proponents of AI decision-making in arbitration face a significant hurdle to their plans. They must persuade Congress to amend the FAA if they want to see wide-scale adoption of their ideas.

For opponents of robot arbitration, the FAA’s current limitations are a strong wall around the centrality of human decision-making in arbitration. But they are not impregnable, and it behooves us to think hard about how to persuade others that robot arbitrators are generally not a good idea, just in case the wall crumbles.

1
Have a look at my archive of posts on the “animals are persons” movement, which I regard as absurd.
2
Why does a Court that has long been dominated by Catholic and (until recently) Jewish jurists have a theory of how to read a text that seems, to me at least, so archetypically Protestant?
3
See, e.g., Taniguchi v. Kan Pac. Saipan, Ltd., 566 U.S. 560, 566-68 (2012).
4
See, e.g., FDA v. Brown & Williamson Tobacco Corp., 529 U.S. 120, 132-33 (2000).
5
Dubin v. United States, 599 U.S. 110, 120-21 (2023) (citation and internal quotation marks omitted).
6
A similar thought occurs when you start to wonder what distinguishes human beings from intelligent nonhuman animals.
7
As in Shakespeare: “But now the arbitrator of despairs, / Just Death, kind umpire of man’s miseries, / With sweet enlargement doth dismiss me hence.” I think Shakespeare might have used “arbiter” instead, except that “arbitrator” has one more syllable and fits the meter.
8
Nor is it sensible to say that an algorithm can be “guilty of misconduct” or of “misbehavior,” though I don’t emphasize that point, because it runs the risk of begging the question that I mentioned earlier.
9
460 P.3d 764 (Cal. 2020).
10
See Guillermo Argerich et al., Could an Arbitral Award Rendered by AI Systems be Recognized or Enforced? Analysis from the Perspective of Public Policy, Kluwer Arbitration Blog (Feb. 6, 2020). On the international arbitration question, David Evans has usefully observed that several foreign states’ arbitration laws specify, in a way the FAA does not, that the arbitrator must be a natural person.

Fediverse Reactions

December 13, 2025

Ted Folkman

arbitration, artificial intelligence

2 responses to “The Robot Arbitrator Can’t Issue Confirmable Awards”

0xPsi

December 16, 2025

love how you turn the court’s own frozen-in-1925 textualism against robot arbitrators. even if congress doesn’t move, i worry providers will just slap a nominal human signer on an ai‑drafted award and claim that satisfies the faa’s “arbitrator” requirement.
1. Ted Folkman
  
  December 16, 2025
  
  @oxpsi I do like using the Court’s approach to statutory interpretation in this way!
  
  I’m not sure that a “human-in-the-loop” AI arbitration system is a bad thing for document-intensive cases that fall into repeatable patterns (except in the case of consumer arbitration, where I think it is a bad thing for other reasons).
  
  Reply

Letters Blogatory

The Blog of International Judicial Assistance
By Ted Folkman