The Ciarb AI Guideline: what it requires and how its templates work

The short version

• The Ciarb AI Guideline (2025) is guidance from the Chartered Institute of Arbitrators (Ciarb), also styled CIArb, a professional body and appointing authority rather than an institution that administers cases.
• It is soft law. It binds no one unless the parties adopt it by agreement (Appendix A, para 11) or a tribunal orders it (Appendix B).
• A tribunal may direct and rule on the parties’ AI use, require disclosure, and draw adverse inferences or reflect a breach in costs (paras 4.1 to 4.7, 7.1 to 7.7).
• Arbitrators may use AI to help, but must not hand over the decision and stay responsible for the whole award (paras 8.2 to 8.4); if the parties disagree on a tool, the arbitrators should refrain (para 9.1).
• It carries two templates: an Agreement (Appendix A) and a Procedural Order with a short form and a long form (Appendix B).
• First published 19 March 2025, re-issued with editorial corrections in September 2025. Ciarb invited feedback until 26 May 2026 and had published no revision by 30 October 2026.^1,3,4,5,7

What is the Ciarb AI Guideline, and is it binding?

The Ciarb AI Guideline is Ciarb’s soft-law guidance for using AI in an arbitration, and on its own it binds nobody. Ciarb, the Chartered Institute of Arbitrators, also styled CIArb, describes itself as an “independent, charitable membership organisation incorporated by Royal Charter” (RC000099; registered charity 803725). It is a professional body and appointing authority, not an institution that administers cases, though its August 2026 EU-letter news lists it among “arbitral institutions”.^7,8,11

Its stated purpose is to let dispute resolvers, parties and others take the benefits of AI while mitigating risk to “the integrity of the process, any party’s procedural rights, and the enforceability of any ensuing award or settlement agreement” (Introduction). It is “not intended to be a manual” for particular AI Tools, and sits with, not above, applicable laws and institutional rules. It is Guideline 18 of Ciarb’s International Arbitration Professional Practice Guidelines, alongside Guideline 15, the 2021 Framework Guideline on the Use of Technology.^1,6,9

It takes effect only once adopted: by the parties’ agreement under Appendix A (para 11.1, where it serves as “guiding principles to all participants”) or by a tribunal’s procedural order under Appendix B. It can even override the agreement that brings it in: where an Appendix A agreement conflicts with the Guideline, the Guideline “shall prevail” (Appendix A, para 11.2). The pillar page, AI in arbitration, sets it next to the other institutional texts; the AI rules tracker keeps a one-line summary.¹

What does the Guideline say are the benefits and risks?

Part I lists benefits and risks, and says each risk’s weight depends on the tool, the nature of the arbitration and how the tool is used (para 2.9).¹

• Benefits (paras 1.1 to 1.10): legal research (1.2); data analysis (1.3); translation, including of witness examinations (1.6); transcripts “at a fraction of the cost of human stenographers” (1.7); tools that “detect deep fakes” (1.8); “predictions of potential case outcomes” (1.9); and remedying “inequality of arms” (1.10).
• Risks (paras 2.1 to 2.8): transparency is “of paramount importance” (2.1); confidentiality is “a fundamental prerequisite”, since a third-party tool may store input or use it “for machine learning purposes”, so check whether a tool “has been vetted” and “whether it is an open or closed model” (2.2); cybersecurity is “a vital concern” (2.3); and bias includes “algorithmic bias” and “affirmational authority bias”, with searching the case documents “a different (and likely far lower) risk” than using AI “to decide an issue in dispute” (2.4).
• More risks: due process, where the tribunal must weigh the parties’ arguments and exhibits, not “a limited selection determined by an AI Tool” (2.5); the “black box” problem and hallucination, including invented “academic sources and pieces of legislation or evidence” (2.6); enforceability, since tools “may be banned or restricted in certain jurisdictions” (2.7); and energy and the environment (2.8).¹

For its definition of AI the Guideline adopts the OECD definition, a machine-based system that infers from its input how to generate outputs, instead of writing its own. Its other defined terms include AI Tool, GenAI Tool, Hallucination, and Mandatory rule, a rule of law or public policy that “may not be waived”. Part II then asks parties and arbitrators to “make reasonable enquiry” about a prospective tool (3.1); unless they “expressly agree to the contrary in writing”, using AI “shall not diminish their responsibility and accountability” (3.4).^1,10

Can a tribunal regulate or ban the parties’ use of AI?

Yes. Regulating the parties’ AI use falls within the tribunal’s general power to conduct the proceedings, subject to party agreement and mandatory law or institutional rules (para 4.1).¹

• Appoint AI experts (4.2); regulate use to “preserve the integrity” of the proceedings and “ensure the validity and enforceability” of awards (4.3); and require disclosure (4.4).
• Record any decision in a procedural order, revisit it, and address contentious use in the award (4.6); on non-compliance, draw “adverse inferences” or reflect the failure in costs (4.7).¹

There is a limit. A tribunal “may not regulate the private use of AI by parties” where that use “does not interfere with the proceedings” and is “generally allowed in litigation in the domestic courts” of the relevant place (4.5). The 2021 technology guideline already treated some private use as beyond the tribunal’s reach (paras 3.6 to 3.7).^1,9

Part III also sets the parties’ side. They may agree “whether and how AI may be used”, including naming tools (5.1); if the agreement is silent, the arbitrators invite views “either at the first case management conference or later” (5.3). Where the parties disagree, the tribunal rules (6.1), or “of their own motion after consulting the parties” if integrity is jeopardised (6.2). To admit AI-generated material it may ask “how inputs from the record are linked to machine-generated outputs” (6.6), and in ruling it is guided by the law of the seat, the law and rules governing the proceedings, the national laws of the parties and any applicable ethical rules (6.7). How the rules treat AI-made and deepfake evidence and bias in arbitrator selection are covered separately.¹

Must parties disclose that they used AI?

Not automatically. Disclosure “may be required” where a tool’s use could affect the evidence or the outcome, or involves delegating an express duty (para 7.1). The rationale is transparency, so other participants understand how AI has been used (7.2).¹

“Unless otherwise agreed by the parties, and after consulting them”, a tribunal may impose disclosure duties on the parties, “including any party-appointed experts or factual witness”, setting the type of AI, the circumstances, to whom and by when (7.3). Any such duty is “of a continuous nature” (7.4) and is weighed against any “duty of confidentiality or legal impediment” (7.5). On a failure to disclose, the tribunal may inquire, then apply remedial measures, adverse inferences or costs (7.6, 7.7). Who must disclose what, text by text, is in AI disclosure in arbitration; the arbitrator’s duty of disclosure is a separate, older idea.¹

Can arbitrators use AI themselves?

Yes, to help with efficiency and the quality of their decision-making, but they must not hand over the decision and stay responsible for the whole award (paras 8.1, 8.2, 8.4).¹

Part IV says arbitrators “should not relinquish their decision-making powers to AI”, and should avoid delegating “legal analysis, research and interpretation of facts and law, or application of the law to the facts, if such use could influence procedural or substantive decisions” (8.2). They should “independently verify the accuracy and correctness” of output, guard against “confirmation bias”, and “conduct their own research” (8.3), and “shall assume responsibility for all aspects of an award” (8.4). This non-delegation rule runs through every AI text; what an arbitrator may and may not draft with AI is in AI award drafting.¹

On transparency, “unless otherwise agreed by the parties”, arbitrators are encouraged to consult the parties, and “if the parties disagree on the use of AI by the arbitrators, the arbitrators should refrain from using the specified AI Tool” (9.1); a co-arbitrator consults the rest of the tribunal (9.2). That “refrain” rule sits against Appendix B Option 2 para 4.4, under which a tribunal may, after written notice and a hearing, affirm its intent to proceed with its own high-risk use. Awards later challenged over AI use are in awards challenged over AI use.¹

How do the Ciarb AI templates work?

The Guideline supplies two drop-in texts: an “Agreement on the Use of AI in Arbitration” (Appendix A) and a “Procedural Order on the Use of AI in Arbitration” in a short and a long form (Appendix B). The full templates are in the PDF.¹

Appendix A is an Agreement that can stand alone or go into the arbitration agreement. It lists “Permitted AI Tools”, defined in one of four ways (para 5), for a menu of “Permitted Use”: research, document review, searching for inaccuracies, document formatting and language review (para 6). A user must disclose the tool and its version (App A, 7.3), mitigate its “biases and/or limits” (App A, 7.4), verify output “from a factual and legal standpoint” (App A, 7.5), not mislead by “fabrication or falsification of any evidence” (App A, 7.6), and protect confidentiality “by redacting or anonymising” inputs (App A, 7.7). The tribunal may investigate, with parties helping to reproduce outputs “by disclosing the prompts” (App A, 8), and must disclose its own use “in advance” while keeping “paternity and responsibility” (App A, 10).¹

Appendix B is a Procedural Order. Option 1, the short form, simply orders that the tribunal and participants “shall be guided by the Ciarb AI Guideline”. Option 2, the long form, defines “High Risk AI Use” as a use that risks confidentiality or privacy, compromises “data security and integrity”, has “the potential to materially undermine the procedural integrity” of the arbitration, or “to assert a non-human influence on the award”, with the list itself left as “[INSERT LIST HERE]”. A participant engaging in such use must “promptly disclose” it in writing, naming the tool, the document and any portions a GenAI Tool drafted, with a certification that no confidential information reached “any unauthorised party” (para 2).¹

• After a disclosed high-risk use in a submission, the other side has 7 days to comment and a further 7 days to reply (para 3).
• Before its own high-risk use, the tribunal gives written notice; the parties have 14 days to brief it and may request a hearing (para 4).
• The tribunal decides within 14 days of the hearing, failing which it “shall refrain” from the use (para 4.4).
• A provision breaching a mandatory rule is amended or severed, and “all deadlines stated herein shall be strictly enforced” (para 5).¹

The appendices carry small drafting slips, for instance Appendix A para 9 refers to “this procedural order” inside an agreement, so edit the wording rather than paste it. Model wording for a clause or procedural order is in AI arbitration clauses; keeping case material confidential when AI is used is in confidentiality and AI in arbitration.¹

How does the Ciarb Guideline compare with SVAMC’s?

The two agree the arbitrator decides, but they differ most on disclosure. Ciarb lets a tribunal require the parties to disclose AI use (7.3), while the SVAMC Guidelines say disclosure “is not necessary as a general matter” (Guideline 3). On the arbitrators’ own use, Ciarb says consult the parties and refrain if they disagree (9.1), while SVAMC asks arbitrators to disclose reliance on AI material outside the record (Guideline 7). Both carry a near-identical model clause: Ciarb Appendix A para 11.1 and the SVAMC model clause (p 21), each making the text “guiding principles to all participants”.^1,15

On disclosure the other institutional texts sit between the two: the SCC Guide (October 2024) encourages a tribunal to disclose AI it uses to research facts and law, the VIAC Note (April 2025) lets arbitrators name the tools they plan to use (para 5.2), and the AAA-ICDR guidance (January 2026) asks arbitrators to disclose generative-AI use that materially affects the process or the reasoning.^12,13,14

The SVAMC instrument, rule by rule, is in the SVAMC AI Guidelines. The tracker carries the SCC Guide, the VIAC Note, the AAA-ICDR guidance and the full line-up.¹

What changed in September 2025, and is Ciarb updating it?

The September 2025 re-issue made editorial corrections only; Ciarb then ran a feedback call that closed 26 May 2026 and had published no revision by 30 October 2026.^1,2,4,5

The Guideline was first published on 19 March 2025 and re-issued on 5 September 2025, with every page now marked “Updated September 2025”. The paragraph numbering did not change, so a paragraph cite is valid for either version.^1,2,3,4

On the review, Ciarb opened a call for feedback on 27 March 2026, timed to the Guideline’s first anniversary, with a deadline of 26 May 2026 and a note that “all feedback is confidential and anonymous”. No revised text or summary of feedback had appeared by 30 October 2026, so the September 2025 edition is still the current one.⁵

Who is using the Ciarb Guideline, and what is still open?

It is soft law, so uptake is voluntary, but its template wording is already turning up in procedural orders, and a national law-reform review has cited it as best practice.^16,18

Two ICSID first procedural orders (ARB/25/42, 19 June 2026; ARB/25/43, 8 June 2026) carry AI sections whose party-duty wording, in our reading, closely tracks Appendix A paras 7.5 to 7.7 and 8 (and the compliance duty in 7.4), including, in ARB/25/42, cooperation “by disclosing the prompts”. Both also bar the tribunal from using AI to draft the reasoning part of its rulings and awards. Neither order names Ciarb, so this is our inference, not a stated adoption.^16,17

A September 2026 Hong Kong Department of Justice consultation paper describes Ciarb as “a professional body for alternative resolution practitioners”, cites the Guideline as best practice, and recommends a light-touch approach with “no immediate need to amend” the Arbitration Ordinance. How AI sits in Hong Kong arbitration more widely is in arbitration in Hong Kong.¹⁸

• Whether and how the 2026 review changes the Guideline.
• How its soft-law standard meets hard law, such as the EU AI Act’s high-risk treatment of AI used in the administration of justice and ADR (Annex III point 8(a)); the Guideline itself names no AI statute, judgment or institutional rules, referring only to “applicable laws” and “institutional rules” in the abstract.
• How the para 9.1 “refrain” rule squares with the Appendix B Option 2 process for a tribunal’s own high-risk use.^1,5,20

The EU AI Act, and the 23 July 2026 joint letter Ciarb signed with 15 other institutions asking that the high-risk label fall on AI a tribunal uses to decide, are covered in the EU AI Act and dispute resolution and AI in dispute resolution in Europe.^11,19