Wer hat entschieden — der Schiedsrichter oder das Werkzeug?
Nicht ob KI eingesetzt wurde, sondern wie weit sie in die Begründung hineinreichte. Sieben Einsatzformen auf einer Skala, die drei Gründe des New Yorker Übereinkommens, auf die eine Anfechtung tatsächlich gestützt würde, und fünf Punkte für die erste Verfahrensverfügung.
Dieser Eintrag ist noch nicht übersetzt und wird auf Englisch angezeigt.
On which side of the rail
A tool can find the passage, line up the chronology and put a draft on the bench. What it cannot do is cross to the other side and sign. The interesting question is not whether it was used, but how close it came.
The wrong question, and the right one
An arbitrator takes on a heavy case and uses AI to work through it: the system lines up the exhibits, flags where two witness statements do not sit together, and produces an analysis that ends up shaping how the arbitrator thinks about the dispute. The parties never hear about any of it. Some months later the losing side is reading the award and wondering how a point it never argued came to be decided against it.
Asked whether AI was used, that scenario is unremarkable — a great deal of what a tribunal's staff already do is the same work under a different name. Asked how far the tool reached into the reasoning, it is a different case entirely. The first question sorts nothing; the second one sorts everything, and it is the one an enforcement court would eventually have to answer.
The 1958 New York Convention says nothing about artificial intelligence, and there is no reason it should — it was drafted for a problem that had not been invented. What it does say about the opportunity to present a case and about agreed procedure is capable of carrying most of this, which is why the challenge, when it comes, is unlikely to be framed as a technology argument at all. What the institutions have actually published on the subject is surveyed separately; this page is about the line those texts all draw and nobody has yet had to apply.
How far the tool reached
Seven things an arbitrator might plausibly do with a model in a live case, ordered by how close each one comes to the decision. The step that matters is not from simple to sophisticated — it is from finding material to weighing it, and it happens between the third row and the fourth.
- Clear of it
- The same work if a clerk or a paralegal had done it. Nothing about the outcome turns on the tool having been the one to do it.
- At the line
- It changes what the tribunal has in front of it, without deciding anything. Harmless if the parties get to see and answer what it surfaced; not otherwise.
- Across it
- It produces reasoning, or the finding itself. This is the mandate, and the mandate is personal — every published instrument on AI in arbitration says so before it says anything else.
| The use | How far it reaches | On the record? |
|---|---|---|
| TranslationWorking translations of exhibits and correspondence | Clear of it A reading aid. The certified translation on the file is still the certified translation, and nothing is decided off the draft. | Not disclosed No. It is a comprehension tool, not a source. |
| Chronology and indexingBuilding a timeline, an index, a table of contents | Clear of it Organising material the parties filed. Every entry points back at a document already in the record. | Not disclosed No, unless it is being relied on as a finding rather than as a map. |
| Finding the passageLocating what a party said, across a file too large to read twice | Clear of it A better search. The tribunal still reads the passage and forms its own view of it. | Not disclosed No. Nobody discloses their use of a search box. |
| Legal researchSurfacing authorities, commentary, comparative material | At the line Fine where it points at authority the tribunal then reads and, where it matters, puts to the parties. Not fine where the summary is what gets relied on. | Raise it Raise it if an authority the tool surfaced does real work in the reasoning — and put the authority itself to the parties, which is the older rule and the one that actually protects the award. |
| Cross-examining the fileFlagging that two statements, or a statement and a document, do not agree | At the line This is the row that looks like assistance and behaves like something else. Finding the contradiction is analysis; what the tribunal does with it decides the case. | Raise it Raise it. Not the tool — the contradiction. A party is entitled to explain an inconsistency before it counts against them, however it was spotted. |
| A point nobody arguedAn issue, characterisation or connection that appears in the award and in no submission | Across it The problem here has nothing to do with AI. A tribunal that decides on a ground the parties never addressed has a due process problem whatever put the ground in front of it. | On the record Yes, and before the award. Put the point to the parties and let them answer it; the alternative is a challenge you cannot answer. |
| Drafting the analysisReasoning, findings, or the operative part produced by the model | Across it The mandate itself. A tribunal may use a tool to put words on a page it has already decided; it may not use one to decide what the page should say. | On the record Yes — but disclosure is not what saves this one. If the tool did the deciding, saying so on the record does not make it the arbitrator's decision. |
4 of the 7 uses are worth putting on the record; the other 3 are not, and treating them as though they were is its own kind of failure. A tribunal that discloses every spell-check trains the parties to stop reading the disclosure, which is exactly the wrong habit to have built by the time something material happens.
What a challenge would actually be built on
There is no AI ground, and there is unlikely to be one. A party resisting enforcement works with the Convention it has, and on the facts above three articles do the work — one of them well, one narrowly, and one hardly at all. Knowing which is which is the difference between a challenge worth bringing and a challenge worth ignoring.
Unable to present its case
The strongest of the three, and the only one that fits the facts without being stretched. Where analysis the parties never saw carries real weight in the award, the complaint is not that a machine produced it. It is that something material entered the reasoning that nobody had a chance to answer — which is the oldest complaint in arbitration, arriving in new clothes.
Was there a point they never got to answer?Procedure not as agreed
Narrow, and entirely within the parties' gift. If the procedural order or the applicable rules said something about AI and the tribunal did otherwise, this article is squarely engaged. If nothing was ever agreed, there is no agreed procedure to depart from, and the argument has very little to grip.
Did anyone put it in the procedural order?Contrary to public policy
The ground most often reached for and least often granted. Courts read it narrowly by design, and 'an arbitrator used software' is not, on its own, the kind of offence to the forum's basic notions of justice it exists for. A delegated mandate might be one day. Nothing decided so far says that it is.
Would this shock the forum, or merely annoy it?The first fight would not be about enforcement at all. Annulment at the seat comes first, and in the many jurisdictions that follow the UNCITRAL Model Law the grounds in Article 34(2) track these three closely enough that the analysis carries across — the difference is who is asked and what they are being asked to do about it.
Five things to settle at the first conference
None of this is difficult if it is dealt with early, and all of it is difficult afterwards. The point of settling it in the first procedural order is not to police anybody — it is that agreed procedure is the one thing on this page a losing party cannot later characterise as a surprise.
Say who this covers
The tribunal, the secretary, counsel and the institution's registry are four different users with four different exposures. A term that only binds the tribunal leaves the file passing through tools nobody agreed to.
Whose use are we actually agreeing about?Name the line, not the tools
A list of permitted products is out of date the week it is written. What holds is the distinction the instruments already draw: the tool may prepare, organise and find; the tribunal decides, and cannot pass that off.
Does the wording survive the next model release?Agree what gets raised
Not a duty to log every use — that produces a document nobody reads. A short standard instead: anything the tribunal would rely on, or that materially shapes the reasoning, goes to the parties before it goes in the award.
Would a losing party learn of this for the first time from the award?Say where the submissions may go
Which deployment, in which jurisdiction, and whether anything the parties filed can be retained or used for training. This is the term most likely to matter in practice and the one least likely to be raised, because it is the least dramatic.
Where does the file sit when it is pasted in?Keep responsibility where it is
Every citation opened, every quotation checked against the source. Fabricated authority has already reached courts in ordinary litigation, and responsibility for what is filed or issued stays with whoever filed or issued it.
Has someone actually opened the case?
The parties agreed on a person. That is not a formality about who signs; it is the thing they bargained for, and it is what enforcement is protecting.
Which is why the line belongs in the procedural order, long before it is worth arguing about in court.