SVAMC AI guidelines: what the seven rules require and how to adopt them

The short version

• The SVAMC AI Guidelines are the Silicon Valley Arbitration & Mediation Center’s (SVAMC) First Edition (2024) soft-law text on using artificial intelligence (AI) in arbitration: seven short principles, each with a commentary.
• They are not rules. They apply only where the parties agree, or a tribunal or institution adopts them, in whole or in part (Preliminary Provisions, p 8; Commentary, p 13).
• They keep the decision with the arbitrator. An arbitrator “shall not delegate any part of their personal mandate to any AI tool” (Guideline 6, p 12).
• They set no general duty to disclose AI use. Disclosure is decided case by case (Guideline 3, p 10).
• They give a one-sentence model clause for a procedural order (p 21), and they do not override mandatory law or professional rules (p 8).
• As of 30 October 2026 the First Edition is still the only edition. SVAMC said on 17 June 2026 that it “is currently developing an updated edition”.^1,4

What are the SVAMC AI guidelines, and are they binding?

The SVAMC AI Guidelines are a soft-law text, not rules. The Silicon Valley Arbitration & Mediation Center (SVAMC), a non-profit based in Palo Alto, California, published the First Edition on 30 April 2024 (cover, p 2). They bind only where the parties agree to them, or a tribunal or arbitral institution decides to adopt them, and they can be adopted in whole or in part (Preliminary Provisions, p 8; Commentary, p 13). The wider setting is in the pillar guide, AI in arbitration, and our AI rules tracker places them beside every other body’s text.^1,2

SVAMC is a non-profit foundation that “does not administer cases”, appoint neutrals or run its own arbitration rules. So it cannot apply the Guidelines itself. Adoption is a decision for the parties, a tribunal or an institution.^1,8,9

They are written for everyone in a case: arbitrators, parties, counsel, experts and institutions, in domestic or international arbitration (Introduction, p 5). They define AI as “computer systems that perform tasks commonly associated with human cognition, such as understanding natural language, recognising complex semantic patterns, and generating human-like outputs” (p 8). The commentary says this reaches both generative AI and evaluative or discriminative AI, the kind behind recommendation and classification tools, and leaves out rule-based automation such as spelling checkers and spam filters (pp 13 to 14).¹

The Guidelines “shall not derogate from any legal obligations, ethical duties, or rules of professional conduct, or any other binding rules” (p 8). They are a “supplementary international standard” (Commentary, pp 13 to 14), and where they clash with mandatory rules the mandatory rules prevail. The text is licensed under Creative Commons Attribution-NoDerivatives 4.0, which lets you copy it with attribution but bars a reworded version (p 2). The official text is the First Edition PDF.¹

Who wrote them, and what changed from the draft?

An SVAMC AI Task Force wrote them over about a year of drafting and public comment. A draft opened for comment on 31 August 2023, and the final text followed on 30 April 2024 after, SVAMC says, hundreds of comments from arbitrators, parties, advocates, universities and institutions (Foreword, p 4).^1,6

• SVAMC announced the project on 27 July 2023. It set up a Task Force of about 34 members and an eight-member Drafting Subcommittee, and planned to publish in 2023 in five languages, with a later “Institutional Guidance to the Guidelines” chapter. We found no translations and no such chapter.
• The draft opened for public consultation on 31 August 2023.
• The comment deadline was extended from 30 September 2023 to 15 December 2023 for the public, and to 15 February 2024 for institutions.
• An independent Review Committee scrutinised the draft over about two months. The Drafting Subcommittee then finalised it, and SVAMC’s Executive Committee approved it (Foreword, p 4).^1,5,6,7

From SVAMC’s own documents, the draft’s three Chapters were recast into seven numbered principles in three Parts. Disclosure (Guideline 3) became a principle for all participants, replacing the draft’s arbitrator-only “Protection and disclosure of records”. The final title dropped the word “International” that the draft and the July 2023 announcement had used, and a model clause appeared (p 21).^1,6

What are the seven SVAMC guidelines?

There are seven Guidelines in three Parts, each with its own commentary. SVAMC numbers them Part I, Part 2 and Part 3. Part I covers all participants (Guidelines 1 to 3), Part 2 covers parties and their representatives (Guidelines 4 and 5), and Part 3 covers arbitrators (Guidelines 6 and 7).¹

Guidelines 1 to 3: understanding, confidentiality and disclosure

Part I binds everyone who uses AI in a case. It asks them to understand the tools they use and to keep confidential material safe, and it treats disclosure as a case-by-case question.¹

Guideline 1 is about understanding. Users are “responsible for familiarising themselves with the AI tool’s intended uses” and “should make reasonable efforts to understand each AI tool’s relevant limitations, biases, and risks and, to the extent possible, mitigate them” (p 9). The commentary names the black-box problem, training-data limits, hallucination, the tendency of a tool to invent plausible but wrong answers, and bias. It suggests explainable-AI features where practical, and urges “extreme caution” when AI is used to search for or appoint arbitrators, experts or counsel, where biased data can entrench discrimination, and says participants should be mindful of the impact on diversity (pp 15 to 16). How shortlisting and conflict checks run is in AI and choosing arbitrators.¹

Guideline 2 is about confidentiality. Participants “should not submit confidential information to any AI tool without appropriate vetting and authorisation”, should use “only AI tools that adequately safeguard confidentiality”, should check a tool’s “recording, storage, and use of prompt or output histories”, and “where appropriate” should “redact or anonymise materials” (p 9). The commentary warns that some public tools keep what users enter, or claim rights over it (p 17). What confidentiality, privilege and data protection require when AI touches the file is in confidentiality and AI in arbitration.¹

Guideline 3 is about disclosure. “Disclosure that AI tools were used in connection with an arbitration is not necessary as a general matter”, and it is decided “on a case-by-case basis taking account of the relevant circumstances, including due process and any applicable privilege” (p 10). Where disclosure is appropriate, three details “may help reproduce and evaluate” an output: the tool’s “name, version, and relevant settings”; “a short description of how the tool was used”; and the “complete prompt (including any template, additional context, and conversation thread) and associated output” (p 10). The commentary says a general disclosure test “could create more controversy than it resolves”, and that disclosure disputes go to the tribunal under the applicable procedural rules (p 17). Who must disclose what, text by text, is in AI disclosure in arbitration.¹

Guidelines 4 and 5: competence and the integrity of evidence

Part 2 binds parties and their representatives. They keep their professional duties when they use AI, and they must not use it to corrupt the evidence or mislead the tribunal.¹

Guideline 4 is about competence and diligence. Representatives “shall observe any applicable ethical rules or professional standards” when using AI. Parties “shall review the output of any AI tool used to prepare submissions” for factual and legal accuracy. Parties and representatives on record “shall be deemed responsible for any uncorrected errors or inaccuracies” (p 11). The guideline sets no standard of its own; it points back to the professional rules that already bind the people involved. The commentary says the tribunal and opposing counsel “may legitimately question” how AI was used, and that a tribunal facing an AI-induced error can strike the submission, order a correction, disregard it, attach lower credibility to it or draw inferences (p 18). How tribunals treat AI-made evidence is in AI document review in arbitration.¹

Guideline 5 is about the integrity of the proceedings and the evidence. Parties, representatives and experts “shall not use any form of AI to falsify evidence, compromise the authenticity of evidence, or otherwise mislead the arbitral tribunal and/or opposing party(ies)” (p 11). The commentary singles out deepfakes as a sharper risk, and lists measures a tribunal can use: striking or excluding the evidence, adverse inferences, and cost consequences (p 19). In our reading, one detail is easy to miss: experts are named in Guideline 5, though not in Guideline 4.¹

Guidelines 6 and 7: non-delegation and due process

Part 3 binds arbitrators. They may use AI to help them, but they may not hand it the decision, and they may not rely on AI information from outside the record without telling the parties first.¹

Guideline 6 is about non-delegation. “An arbitrator shall not delegate any part of their personal mandate to any AI tool. This principle shall particularly apply to the arbitrator’s decision-making process. The use of AI tools by arbitrators shall not replace their independent analysis of the facts, the law, and the evidence” (p 12). The commentary does not ban AI. An arbitrator may use it to analyse and to help draft, but must review the output, take “full responsibility” for it, and “cannot simply reproduce the AI’s output”. The tribunal “will not be released from its duty to personally review the file and/or draft” its decision (pp 19 to 20). What arbitrators may and may not draft with AI is in AI and drafting awards.¹

Guideline 7 is about due process. “An arbitrator shall not rely on AI-generated information outside the record without making appropriate disclosures to the parties beforehand and, as far as practical, allowing the parties to comment on it.” And where a tool “cannot cite sources that can be independently verified, an arbitrator shall not assume that such sources exist or are characterised accurately” (p 12). A footnote keeps the principle of iura novit arbiter, that an arbitrator may apply law the parties did not cite, where that applies (footnote 3).¹

Guideline 6 has not been tested in court in the cases we found. A US petition that alleged an arbitrator outsourced the decision to AI, LaPaglia v. Valve Corporation, was dismissed in December 2025 for lack of subject-matter jurisdiction, so no court reached the point. The detail is in awards challenged over AI use.¹¹

What does the SVAMC model clause say, and how do you adopt the Guidelines?

The Guidelines can be adopted by a one-sentence model clause, placed in a procedural order or the arbitration agreement, in whole or in part, at the start of a case or later (p 21; Commentary, p 13).¹

The model clause reads: “The Tribunal and the parties agree that the Silicon Valley Arbitration & Mediation Center Guidelines on the Use of Artificial Intelligence in Arbitration (SVAMC AI Guidelines) shall apply as guiding principles to all participants in this arbitration proceeding” (p 21). The words that carry the weight are “guiding principles”. The clause makes the Guidelines a reference point, not binding rules, and it cannot override mandatory law (p 8). The Foreword hopes the clause “will be routinely adopted” (p 4). That is an aspiration, and we found no published tribunal order or institution that has in fact adopted it.¹

One caution. Do not confuse this with the SVAMC Arbitration Clause Insert published on svamc.org. That is a separate model clause for selecting a neutral from SVAMC’s Tech List, its list of technology neutrals, not the AI model clause. On that page SVAMC states again that it “does not administer cases, appoint arbitrators or mediators, or have its own detailed administering rules”.⁹

A tribunal that adopts the Guidelines can record in Procedural Order No. 1 which parts apply, then settle disclosure under Guideline 3 and confidentiality under Guideline 2 at the first case management conference, the tribunal’s first procedural meeting with the parties. Model wording for a clause or a first procedural order is in AI arbitration clauses.¹

How do the SVAMC Guidelines differ from the Ciarb Guideline?

Both are non-binding, and both keep the decision with the arbitrator. They differ most on disclosure and on the tools they give for adoption. SVAMC leaves disclosure case by case and offers a one-sentence clause. The Chartered Institute of Arbitrators (Ciarb) lets a tribunal require disclosure and supplies template documents. The Ciarb Guideline in full is in the Ciarb Guideline.^1,3

The SCC, VIAC, CIETAC, AAA-ICDR and WIPO texts are summarised in our AI rules tracker, and the SCC Guide has its own SCC explainer.¹

Is there a newer edition, and has anyone adopted them?

As of 30 October 2026 the First Edition (2024) is the only edition, and we found no institution that has adopted the Guidelines as rules. On 17 June 2026 SVAMC said it “is currently developing an updated edition of the Guidelines to address emerging issues and evolving best practices”. It gave no draft, no consultation and no date.^1,4

The live PDF is unchanged since publication, and SVAMC’s news archive shows no second-edition announcement. The Guidelines promise their own review: they “will be subject to continuing analysis and review” (Foreword, p 4). One oddity: SVAMC’s 2024 announcement still titles them as being “in International Arbitration”, though the PDF title drops “International” and the Introduction covers domestic arbitration too (p 5).^1,2

No published tribunal order we found applies them. Several bodies refer to them. The International Chamber of Commerce (ICC) uses them as its “AI in international arbitration” case study in a 2025 paper on responsible AI (published 13 November 2025). Other references, and the LaPaglia v. Valve Corporation challenge, sit on their own pages: the LCIA, JAMS, Hong Kong andawards challenged over AI use.^1,10

What is still open?

• Whether the updated edition changes the disclosure default or the definition of AI, and when it lands.
• Whether tribunals and institutions actually adopt the model clause. The Foreword’s hope is not yet matched by any published adoption we found.
• How “any part of” an arbitrator’s mandate under Guideline 6 maps onto AI-assisted drafting.
• Whether a court would treat an award that adopted the Guidelines differently from one that did not. No reported case we found has decided it.
• Whether “where appropriate” in Guideline 2 sets any floor on redacting or anonymising materials before they reach an AI tool, or leaves it wholly to the user.^1,4