Tahkimde yapay zekâ: Kurumlar gerçekte ne söyledi
Yayımlanmış beş metin, açılmış iki çalışma ve sözleşmelerde en çok adı geçen kurumlardan dördünde yazılı hiçbir şey. Bugün bir usul kararında neyi dayanak gösterebilirsiniz, neler hâlâ yalnızca görüş.
Bu içerik henüz çevrilmedi ve İngilizce gösteriliyor.
Who signs the award
Every published instrument agrees on one thing before it agrees on anything else: the mandate is personal and cannot be delegated. A tool may draft. A person is answerable for what issues.
Where this actually stands
Two years ago there was nothing to cite. Today there are five published instruments on the use of AI in arbitration, one institution that has skipped writing guidance and put an AI arbitrator into service instead, two work streams that have opened and not yet reported, and four of the institutions international contracts most often name — the LCIA, SIAC, HKIAC and DIAC — with no dedicated text of their own.
The gap matters in a specific way. Where the rules say nothing, the ground rules for AI in a proceeding are whatever the tribunal and the parties agree in the first procedural order. Somebody has to raise it, and the instruments below are what they will reach for when they do.
The ICC's 2026 Rules, in force this year, are silent on it. The ICC's answer is expected from its task force rather than from its rules, which means that for the single most named institution in international arbitration there is currently no text to cite.
What has been published
- Guideline
- A dedicated instrument addressed to arbitrators, parties and counsel.
- Note
- Shorter guidance, written to open the parties' discussion rather than settle it.
- Service
- AI the institution runs itself, under rules of its own.
- In progress
- A work stream opened; nothing published yet.
- Nothing published
- Checked, and there is no dedicated instrument.
| Body | Instrument | Since | What it is |
|---|---|---|---|
| SVAMCSilicon Valley Arbitration & Mediation Center | GuidelineGuidelines on the Use of Artificial Intelligence in Arbitration — primary document | Nis 2024 | The first of them, and the source of the formulation the later texts converge on: an arbitrator shall not delegate any part of the personal mandate to an AI tool. Principle-based, and deliberate in treating disclosure as a case-by-case judgement rather than a standing duty. |
| SCCSCC Arbitration Institute, Stockholm | NoteGuide to the Use of Artificial Intelligence in Cases Administered under the SCC Rules — primary document | Eki 2024 | Short and high-level, built on four considerations: confidentiality, the quality of the tool and the oversight it needs, the integrity of the proceedings, and non-delegation of the decision and of the reasoning leading to it. |
| CiarbChartered Institute of Arbitrators | GuidelineGuideline on the Use of AI in Arbitration — primary document | Mar 2025 | The most developed of them: four parts, updated in September 2025, and the only one that ships drafting material — a template party agreement and a template procedural order that can go on the record as they stand. Non-binding, and written for ad hoc and institutional proceedings alike. |
| VIACVienna International Arbitral Centre | NoteNote on the Use of Artificial Intelligence in Arbitration Proceedings — primary document | Nis 2025 | Six headings, non-binding, and framed explicitly as a way to get the parties discussing AI early rather than as a set of obligations imposed on them. |
| CIETACChina International Economic and Trade Arbitration Commission | GuidelineProvisional Guidelines on the Use of Artificial Intelligence Technology in Arbitration | Tem 2025 | In force 18 July 2025 and the first from an Asia-Pacific institution. High-level risk mitigation: AI does not substitute for human decision-making, and it does not move responsibility for the accuracy of a submission off the party that filed it. |
| AAA-ICDRAmerican Arbitration Association / International Centre for Dispute Resolution | ServiceAI Arbitrator — primary document | Kas 2025 | Not guidance but a service, and the reason this table has a third category. An AI arbitrator for documents-only construction cases, live since November 2025, opt-in by both parties and governed by its own rules. Trained on the institution's own awards; a human arbitrator reviews, revises where needed, and is the one who confirms and issues. |
| ICCInternational Chamber of Commerce | In progressTask Force on AI in International Dispute Resolution | 2025 | Established in 2025 under the Commission on Arbitration and ADR. The 2026 Rules say nothing about AI, and the ICC's position is expected from the task force rather than from the rules — so until it reports there is no ICC text to put in a procedural order. |
| UNCITRALUN Commission on International Trade Law, Working Group II | In progressWorking Group II — colloquium and report — primary document | Şub 2026 | Working Group II held a colloquium on AI in dispute resolution and on remote hearings at its 83rd session in February 2026 and adopted a report for the Commission's fifty-ninth session. The texts actually moving concern awards in electronic form, the Model Law and the Notes on Organizing Arbitral Proceedings — the infrastructure AI sits on, ahead of AI itself. |
| LCIALondon Court of International Arbitration | Nothing published— | — | Nothing dedicated. In practice the Ciarb guideline is what parties arbitrating under LCIA rules reach for, which is a workable answer and not the same as an institutional one. |
| SIACSingapore International Arbitration Centre | Nothing published— | — | Nothing dedicated. Its 2025 procedural rules accommodate the technology generally without naming it, which leaves the ground rules to the tribunal. |
| HKIACHong Kong International Arbitration Centre | Nothing published— | — | Nothing dedicated — while using AI itself: a 2025 partnership generates procedural decision summaries for its case digest. Using it and regulating it are separate decisions, and only one has been taken. |
| DIACDubai International Arbitration Centre | Nothing published— | — | Nothing dedicated. The same shape as HKIAC: an AI partnership announced in 2025, no instrument addressed to participants. |
Rows without a link are ones where we have not settled on a stable primary URL; they are compiled from the institution's own announcement and from law-firm reporting, and they are the first rows checked on each review. “Nothing published” means we looked and found no dedicated instrument — not that the institution is indifferent to the question.
What they agree on
The published texts differ in length, in tone and in how far they go, but they converge on five points. These are common ground rather than competing features, which is why they are set out as points and not as a matrix — drawing them as filled and empty cells would invent a disagreement that is not in the texts.
Understand the tool
Every text carries some version of it: do not use a system whose behaviour you cannot describe. It is a competence duty, and it falls on arbitrators, counsel and the institution's own registry staff alike.
Could you explain, on the record, what it did?Raise it early, on the record
None of the instruments imposes a standing duty to disclose every use; SVAMC expressly treats disclosure as a case-by-case judgement. All of them expect the question to be settled between the parties in the first procedural order rather than discovered in a challenge.
Is it in the procedural order?The submissions do not leave
The worry is not principally that a model is wrong. It is that party material becomes training data, or comes to rest on a server nobody named. The SCC guide puts it as a duty to know how input is used and where it is deployed.
Where does the text go when it is pasted?Nothing is cited unchecked
Fabricated authority is the failure that has already reached the courts, in litigation ahead of arbitration. Responsibility for the accuracy of a submission stays with the party that filed it, whatever produced the text.
Was every citation opened?The mandate is personal
The one point on which no instrument hedges: an arbitrator may not delegate any part of the personal mandate, and least of all the decision and the reasoning that leads to it. Even AAA-ICDR's own AI arbitrator is built around a human who reviews, confirms and issues.
Who is answerable for the award?
The first time a court was asked
In LaPaglia v. Valve Corporation, the losing party asked a United States federal court to vacate an award on the ground that the arbitrator had relied on AI so heavily that he had outsourced the adjudicative role — a twenty-nine page award issued fifteen days after final briefing, over a holiday period, said to contain findings that were not in the record.
In December 2025 the court granted a motion to dismiss the amended petition, so the question was never reached on its merits. That is where it stands: no court has held that AI authorship alone is a ground for vacatur, and none has held that it is not.
No court has yet held that an AI-drafted award is unenforceable. No court has held that it is safe, either.
Which is why it belongs in the procedural order, not in the challenge.
What this asks of the system, rather than of the arbitrator
Most of the obligation above is professional and no software discharges it. Two of the five points are structural, and a system either supports them or quietly defeats them.
Confidentiality is the first. If case material has to leave the institution's control in order to be processed, then the duty to know how party data is used and where it goes cannot be met by anyone — not by the tribunal, not by the registry. That is an architectural question and it is settled before any tool is chosen: the plane holding the documents runs where the institution's record is allowed to live, and a model never becomes the place the record lives.
The record itself is the second. “Raise it early, on the record” and “nothing is cited unchecked” both assume a proceeding in which what was done, by whom, and when, is recoverable afterwards. That is an audit-trail requirement, and it is the difference between an institution that can answer a challenge and one that can only deny it.
Those are the two questions worth putting to any vendor, this one included.
How this page is maintained
Reviewed quarterly and whenever a body publishes. Corrections are made in place and the revision in the margin is raised, so a link to this page stays a link to the current text rather than to a snapshot of it.
The rows most likely to change are the four with nothing published and the two open work streams — the ICC task force and UNCITRAL Working Group II — either of which could move this page substantially. If a row is out of date, tell us and we will correct it.