AI Agents for Lawyers

Practice Areas

AI for Arbitration Practitioners: Submissions, Evidence and Tribunal Rules

AI for arbitration practitioners: what CIArb's 2025 guideline, tribunal procedural orders and disclosure rules mean for long records, witnesses and quantum.

By Dr. Niklas SchmidtUpdated 13 November 202611 min read
On this page
  1. What the soft law now says
  2. Raise it at the first CMC, before the tribunal does
  3. Long records: what a million tokens changes and what it does not
  4. Witness statements: the consistency check
  5. Quantum: experts and the Kohls lesson
  6. Confidentiality when there is no institution to lean on
  7. A Vienna seat
  8. Where this leaves counsel

Arbitration counsel can use AI on the record, in the drafting room and at the hearing, but since September 2025 the Chartered Institute of Arbitrators' template procedural order has let a tribunal require a party to identify "the specific portions of text" a generative tool drafted. The rules of the game are now written down, and they start at the first case management conference.

What the soft law now says

Arbitration got its AI rules before most court systems did, because institutions could write guidelines without waiting for legislation. As of September 2026 the shelf looks like this.

InstrumentDateWhat it does for counsel
CIArb Guideline on the Use of AI in Arbitration (2025)Published March 2025, updated September 2025Four parts (benefits and risks; general recommendations; parties' use; arbitrators' use) plus a template party agreement (Appendix A) and a template procedural order in short and long form (Appendix B)
SVAMC Guidelines on the Use of AI in Arbitration2024The Silicon Valley Arbitration and Mediation Center's earlier framework, widely cited by tribunals and in party agreements
VIAC note on AI2025The Vienna International Arbitral Centre's position paper for proceedings under its rules
SCC guideUndated in the sources reviewedThe Stockholm Chamber of Commerce's practical guide on AI in SCC proceedings

Gowling WLG's 2026 overview of these instruments is a useful map (Gowling WLG), and Freshfields' two-part series on principles for AI in arbitral proceedings covers the underlying due-process theory. The CIArb text itself is the one to read in full, because tribunals lift its language directly into procedural orders.

Four provisions do most of the work. Paragraph 3.4: unless the tribunal and parties expressly agree otherwise in writing, using an AI tool "shall not diminish" a participant's responsibility and accountability. Paragraph 4.7: if a party ignores a direction on AI, the tribunal may "draw any appropriate conclusion (including drawing adverse inferences, if appropriate)" and take the failure into account on costs. Paragraph 7.3: after consulting the parties, arbitrators may impose AI-disclosure obligations on the parties "including any party-appointed experts or factual witness", and by paragraph 7.4 the duty is continuous. Paragraph 8.2, for the tribunal itself: arbitrators "should avoid delegating any tasks to AI Tools, such as 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", and by paragraph 8.4 the arbitrator "shall assume responsibility for all aspects of an award, regardless of any use of AI". Paragraph 9.1 adds that if the parties disagree about the tribunal's proposed use of a tool, "the arbitrators should refrain from using the specified AI Tool."

Raise it at the first CMC, before the tribunal does

The guideline's Appendix A is a menu, and whoever fills it in first shapes the proceedings. Work through it before the first case management conference.

  1. Read the arbitration agreement. Paragraph 5.2 tells the tribunal to check whether the parties already provided for AI; some 2025–2026 clauses do.
  2. Decide your permitted-tools position. Appendix A offers four drafting options: a list of permitted tools, tools with stated characteristics, an exclusion list, or excluded characteristics. A list by name and version ages fast; characteristics ("enterprise tier, no training on inputs, retention not exceeding 30 days") age better.
  3. Decide the permitted uses. The template's own list is research, document review, search for inaccuracies, document formatting and language review. Drafting submissions is conspicuously absent from that list, and you should decide whether to add it or leave it to the private-use carve-out.
  4. Protect private use. Paragraph 4.5 and Appendix A paragraph 9 say a tribunal may not regulate private use of AI that does not interfere with the proceedings and would be allowed in the seat's courts. Keep that carve-out; it covers most of what your team does day to day.
  5. Choose short or long form for the PO. The short form simply orders that participants "shall be guided by the Ciarb AI Guideline". The long form defines "High Risk AI Use", requires written disclosure identifying the tool and, for generative drafting, "the specific portions of text that have been so drafted", plus a certification that no confidential information reached an unauthorised party, and gives the other side seven days to comment and seven to reply.
  6. Expect your prompts to be discoverable. Appendix A paragraph 8 has the parties agree to help the tribunal "reproduce and evaluate the outputs" of a tool "by disclosing the prompts used for the specific input". Keep a prompt log per matter from day one.
  7. Cover the experts now. Paragraph 7.3 reaches party-appointed experts. Tell your quantum and technical experts what the PO requires before they open a chatbot.

The confidentiality certification in step 5 is the one that bites: you can only certify that nothing confidential went to an unauthorised party if you know which tier of which tool your whole team used. The confidentiality checklist lists the current terms vendor by vendor.

Long records: what a million tokens changes and what it does not

A construction or investment arbitration record runs to tens of thousands of pages. Claude Fable 5.1, Opus 5 and Sonnet 5 all take a 1M-token context window, which Anthropic puts at roughly 555,000 words, so a substantial part of a record now fits in one pass; GPT-6 Astra is at about 1.05M. That removes the chunking problem but not the selection problem. CIArb paragraph 2.5 names the due-process risk precisely: a tribunal considering "a limited selection determined by an AI Tool" rather than all of the parties' arguments and exhibits. For counsel the same risk runs the other way; a model that "reads" 40,000 pages and returns twelve key documents has made an editorial decision you cannot see.

The answer is to force structure. Tabular extraction against a fixed question set, with a source citation in every cell, is what the platforms have been built for: Legora's Tabular Review handles up to 100,000 documents in parallel, Thomson Reuters' rebuilt CoCounsel Legal runs Tabular Analysis at 10,000 documents against 100 questions (August 2026), and Legora bought Wexler, a litigation fact-analysis tool, on 29 July 2026. The independent evidence supports the extraction use: in the Vals Legal AI Report of February 2025, Harvey scored 94.8% on document Q&A against a 70.1% lawyer baseline, and chronology building was a tie (Vals). The evidence also supports the gate: on Harvey's strict all-criteria-pass LAB metric the best model completes 13.3% of long legal tasks flawlessly as of 18 September 2026, with criteria-level pass rates around 83–88% (LAB leaderboard). The benchmarks guide explains why those two numbers describe the same tools.

Two record tasks the guideline singles out are cheaper than they were. Paragraph 1.7 notes that AI generates hearing transcripts "at a fraction of the cost of human stenographers"; the transcript workflow applies to hearing bundles as well. Paragraph 1.6 covers translation and simultaneous interpretation; the legal translation page sets out where machine translation is acceptable for a working bundle and where a certified translation is still needed for an exhibit.

Witness statements: the consistency check

CIArb paragraph 1.3 lists "the detection of conflicting statements within text" as a core benefit, and it is the single most useful thing a model does for arbitration counsel, on both sides of the file: testing your own witnesses before they sign, and preparing cross-examination of the other side's. The prompt:

Role: Junior counsel in an ICC arbitration. Work ONLY from the attached
materials. Do not use outside knowledge of the parties or the industry.

Attached:
 A. Witness statement of [name], paragraphs 1–84
 B. Contemporaneous documents cited in A, by exhibit number (C-1 to C-63)
 C. The same witness's earlier statement in the [related] proceedings
 D. Agreed chronology

Task:
 1. For each factual assertion in A, find every document in B, C or D
    that supports, qualifies or contradicts it. Cite exhibit number and
    page. If a document is illegible, say [ILLEGIBLE], never guess.
 2. Classify each tension: DATE / AMOUNT / SEQUENCE / WHO-SAID-WHAT /
    OMISSION (a fact in B or C that A does not mention).
 3. Output a table: A para | Assertion | Exhibit | Tension type |
    What the document actually says (quote, max 40 words).
 4. Then list the ten assertions in A that rest on no document at all.
 5. Do not draft cross-examination questions. Do not characterise the
    witness's credibility. Do not summarise A.

If two documents in B conflict with each other, flag [DOC CONFLICT] and
show both.

Step 5 matters. The model is good at finding tensions and bad at judging which ones matter; the questions are yours. Opus 5 "verifies its own work without being told to" as of July 2026, so drop any trailing instruction to double-check; it produces over-verification. For the adversarial version, where the model plays a hostile tribunal against your own submissions, see the red-team playbook.

Quantum: experts and the Kohls lesson

Expert evidence is where AI has already cost a party its witness. In Kohls v. Ellison (D. Minn., January 2025) a Stanford professor's declaration in support of a deepfake statute was struck because ChatGPT had invented the studies it cited (tracker entry). Deloitte Australia refunded the final instalment of a AU$440,000 government report in October 2025 after it was found to contain a made-up court decision and non-existent academic papers (The Register). Neither was an arbitration; both are the kind of document a quantum or delay expert files, and CIArb's disclosure powers reach party-appointed experts by name.

The workable division of labour: the expert's team may use models to build the document trail behind a damages model (which invoice supports which line), to draft the narrative around a sensitivity analysis, and to check internal consistency between the report's tables and its text. The model does not select the method, does not generate the literature, and does not write the paragraph that says what the loss is. Anthropic's Opus 5 launch note claims financial-modelling accuracy rose nine points with a third fewer tool calls and that the largest gains on legal agent work came "in practice areas like corporate governance and arbitration" (Anthropic, 24 July 2026). Those are vendor claims. The public leaderboards report aggregate scores rather than an arbitration line, so there is no independent number to set against them as of September 2026.

Confidentiality when there is no institution to lean on

CIArb paragraph 2.2 calls inputting confidential data into a third-party tool a source of "substantial risks" and asks parties to consider whether a tool "has been vetted, has sufficient privacy and security standards, and whether it is an open or closed model". In institutional proceedings the institution's own platform and rules set a floor. In ad hoc proceedings there is no floor; counsel's tool choice is the whole confidentiality control, and the certification in the long-form PO is signed on that basis.

TierTraining on your inputsRetentionFit for an arbitration record
Consumer chatbots (ChatGPT Free/Plus/Pro, Claude Free/Pro/Max, consumer Gemini)ChatGPT and Gemini train unless opted out; Claude consumer users must choose (opted-in data kept 5 years)Gemini 18 months by default, with human review of a subsetNo. UK v SSHD [2026] UKUT 81 (IAC) treated material pasted into a public tool as being "in the public domain"
Business and enterprise chat (ChatGPT Business/Enterprise, Claude Team/Enterprise, Microsoft 365 Copilot)No training by defaultContractual; Claude Team and Enterprise chat are not eligible for zero data retentionYes, subject to the PO and the arbitration's confidentiality regime
API under a zero-data-retention arrangementNoNone, except Claude Fable 5.1, which requires 30-day retention and is not ZDR-eligible without express authorisationYes; note that the strongest Claude model is the least private, per Anthropic's published retention policy
Legal platforms (Harvey, Legora, CoCounsel, Lexis+ Protégé)NoContractual; Lexis offers customer-held encryption keysYes, and easiest to describe in a permitted-tools clause

The hallucination record is the other half of the confidentiality conversation, because a tribunal that has read Ayinde v Haringey [2025] EWHC 1383 (Admin) will ask. The Divisional Court held that "freely available generative AI tools ... are not capable of conducting reliable legal research" (judgment), and Damien Charlotin's database stood at 2,041 decisions worldwide on 14 September 2026 (database). Arbitral awards are rarely published, so the arbitration share of that number is invisible; assume tribunals read the court cases and apply the same standard. The citation-checking protocol is the answer for any authority in a memorial.

A Vienna seat

For VIAC proceedings the institution's 2025 note sits alongside the CIArb text, and Austrian counsel carry the bar's rules on top. ÖRAK's four practice rules, set out at its first KI-Tag on 19 March 2026, fit a permitted-tools clause almost verbatim: client data stays on EU servers, AI use is disclosed to the client, the lawyer keeps final responsibility, and no client data is used for training (Anwaltsblatt). The Austrian courts have already shown what personal responsibility means in practice: on 7 October 2025 the OGH (14 Os 95/25i) dismissed a Graz defence lawyer's AI-drafted Nichtigkeitsbeschwerde without reaching the merits because it cited invented OGH decisions, pointing to § 9 Abs 1 RAO (Der Standard). A tribunal seated in Vienna will know that decision. The Austrian and German bar guidance page covers ÖRAK, BRAK and the CCBE technical guide in detail.

Where this leaves counsel

Three habits cover most of the risk. Keep an AI log per matter from the first day: tool, version, tier, use, who verified the output and when; a long-form PO may require exactly that, and an ad hoc tribunal may ask without warning. Bring the Appendix A menu to the first CMC with your positions filled in; the party that proposes the protocol usually gets most of it. And treat the record-facing tasks and the filing-facing tasks as different activities with different gates: the chronology, the extraction grid and the consistency check stay inside the team and are checked by the lawyer who would have done them by hand; anything that goes into a memorial passes through the same read-every-authority rule that governs litigators. Module 2 of the AI Lab for Lawyers course builds the witness-consistency workflow above on a sample record, and Module 6 covers the disclosure and confidentiality rules that a procedural order will hold you to.

Frequently asked questions

Do I have to disclose AI use to an arbitral tribunal?

Not automatically, but the tribunal can require it. CIArb's 2025 guideline says disclosure may be required where AI use could affect the evidence or the outcome, lets tribunals impose continuing disclosure duties on parties, experts and witnesses, and allows adverse inferences and costs consequences for non-compliance. Check the arbitration agreement and any procedural order first, and raise the question at the first case management conference rather than waiting to be asked.

Can arbitrators use AI to draft an award?

CIArb's guideline says arbitrators may use AI to process submitted information but should not delegate legal analysis, interpretation of facts or application of law to the facts where that could influence a decision, must verify anything AI produces, and remain responsible for every aspect of the award. If the parties disagree about a tribunal's proposed use of a tool, the guideline says the tribunal should refrain from using it.

Is it safe to upload an arbitration record to Claude or ChatGPT?

Only on a business, enterprise or API tier whose terms exclude training and specify retention, and only where the arbitration's confidentiality regime and any procedural order allow it. Consumer tiers of the main chatbots train on conversations unless opted out, and a UK tribunal has said that pasting client material into a public AI tool puts it in the public domain. In ad hoc proceedings there is no institution to fall back on; your tool choice is the entire control.

Which AI tasks give arbitration counsel the biggest return?

Record-facing work: building a cited chronology from the pleadings and exhibits, tabular extraction across thousands of documents, cross-checking witness statements against contemporaneous documents and prior statements, hearing transcription and first-pass translation. These stay inside the team and are checked before anything is filed. Drafting submissions and legal research carry the hallucination risk and need the same verification discipline as court litigation.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Practice Areas

AI for Litigators: Research, Drafting, Discovery and the Sanctions Trap

AI for litigators pays off in chronologies, discovery and deposition prep; it costs $110,204 when a brief cites invented cases. Where to gain and the one rule.

8 min read

Playbooks

How to Verify AI Legal Research: A Citation-Checking Protocol

Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.

11 min read

Guides

Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers

Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.

11 min read

Use Cases

AI Legal Translation: Accuracy, Confidentiality and Workflow

AI legal translation gives reliable working drafts of contracts and pleadings when a glossary, system prompt and bilingual review surround it. German-English example.

10 min read

Regulation & Ethics

AI Rules for Austrian and German Lawyers: ÖRAK, BRAK and CCBE Guidance

ÖRAK's four AI rules and § 40 Abs 3 RL-BA bind Austrian lawyers; BRAK's § 43a BRAO line binds German ones; CCBE guides cover both. With a bilingual checklist.

11 min read

Playbooks

Red-Team Your Argument with AI: Adversarial Prompts for Litigators and Negotiators

Red-team your legal argument with AI in five rounds: prompts that make the model argue the other side, play a sceptical judge and score your case before filing.

12 min read