AI document review in arbitration: production and deepfake evidence

The short version

• Parties already use AI, including technology-assisted review (TAR), to review documents in arbitration. No arbitral rule we checked (how we search) makes a party disclose it or get the other side’s agreement.
• The International Bar Association (IBA) Rules on the Taking of Evidence, which apply where adopted, do not mention AI. They leave the tribunal to decide what evidence to admit and how much weight it carries (art 9.1).
• The Silicon Valley Arbitration & Mediation Center (SVAMC) guidelines, which apply where adopted, bar using AI to falsify evidence. The Chartered Institute of Arbitrators (Ciarb) guideline lets a tribunal ask parties how AI output links to the record.
• Against a suspected fake, a tribunal can call for originals, appoint an expert, draw adverse inferences and award costs.
• The UK’s Upper Tribunal said in November 2025 that lawyers who put confidential documents into an open-source AI tool breach confidentiality and waive privilege.
• In May 2026 a US rules committee voted to withdraw its published draft of Federal Rule of Evidence 707 on machine-generated evidence. It would not bind a tribunal anyway.^2,3,5,6,7,8

How is AI used for document review and production in arbitration?

In arbitration, AI document review means software that picks out the relevant documents and summarises them. In the 2025 Queen Mary University of London survey, 59% of respondents had used AI for document review in the previous five years. Looking ahead, 90% expect to use it and 19% expect to use it “almost always” (Chart 18, printed p 28). The wider picture is in our guide to AI in arbitration.¹

• Predictive coding: a 2022 International Chamber of Commerce (ICC) Commission report describes it as machine learning trained on human coding of documents as relevant or not, then applied to the whole set once its precision is judged acceptable (section 5.2, p 29). Guidance for judges in England and Wales treats TAR as one of the approaches now used in e-disclosure, the English counterpart of US e-discovery (printed p 6).
• Production: a working group of in-house counsel on the Swiss Arbitration Association (ASA) User Council lists current or foreseeable uses: flagging overbroad requests, filtering out privileged material, removing duplicates and clustering documents by topic, with the tribunal supervising any TAR (pp 16 to 17). It doubts AI “will by itself reduce the time and cost of document production” (p 1).
• Tribunals: a tribunal of the International Centre for Settlement of Investment Disputes (ICSID) said in ICSID Case No. ARB/24/46 that it may use AI tools to prepare summaries and to search and organise the record, but will delegate no decision to them (s 28.1). How far arbitrators may go in summarising and drafting is in arbitrators drafting awards with AI.^2,9,10,11

Can parties use technology-assisted review (TAR) in arbitration, and must they disclose it?

Parties can use TAR in arbitration, and we found no rule that makes a party disclose it unless the parties agree or the tribunal orders it. ICC’s 2022 report found “no rules or guidance” on disclosing predictive coding, and none on whether it needs agreement or approval (section 5.2, p 29). It suggests raising it at the case management conference (CMC), the tribunal’s procedural meeting with the parties, perhaps in a protocol fixing search terms, date ranges or predictive-coding parameters (pp 29 to 30).²

• Disclosure of the review method: ICC counts transparency as a benefit of predictive coding only if its use and parameters are disclosed (p 29). Ciarb’s template agreement offers document review as one AI use the parties can permit (Appendix A para 6). The general standards on disclosing AI use are in disclosure of AI use in arbitration and the AI rules tracker.
• American Arbitration Association (AAA): the parties should try to agree, and the arbitrator may decide, “reasonable search parameters” for electronically stored documents, and the arbitrator may allocate the cost (Commercial Rules R-23(b)(iv), R-24(b), (c)).
• International Centre for Dispute Resolution (ICDR): requests for electronic documents “should be narrowly focused”, and the tribunal “may direct testing or other means of focusing and limiting any search” (art 24(6)).
• ICC’s 2026 Note suggests limiting document requests to what is relevant and material to the outcome, and using a schedule of document production (Part III para 6(j)). More is in ICC and AI in arbitration.
• Courts have approved TAR, though their rulings do not bind tribunals. A New York federal court held computer-assisted review, another name for TAR, acceptable “in appropriate cases” (Da Silva Moore v. Publicis Groupe, 24 February 2012, p 2), a ruling upheld on 25 April 2012. In England, Pyrrho Investments v MWB Property [2016] EWHC 256 (Ch) approved predictive coding the parties had agreed ([33], [34]), and Brown v BCA Trading [2016] EWHC 1464 (Ch) ordered it over one side’s objection, on cost grounds ([1], [9]).^{2,6,12,13,14,15,16,17,18}

What do the IBA Rules on the Taking of Evidence say about AI?

The IBA Rules say nothing about AI expressly, but their articles on electronic documents, originals, experts and inferences apply to AI-assisted review, document production and AI-generated material. They apply where the parties or the tribunal adopt them, or as guidelines (Preamble para 2). A Request to Produce may name “specific files, search terms, individuals or other means of searching” (art 3.3(a)(ii)), which leaves room for TAR. The IBA’s commentary says the default form of production “will generally not be the native format with full metadata”, meaning the original file and the data stored with it, such as who made it and when (p 14).^3,4

Does putting case documents into an AI tool risk confidentiality or privilege?

Putting confidential case documents into a public AI tool can breach confidentiality and waive privilege. The UK’s Upper Tribunal, speaking of lawyers in immigration cases, said an “open-source AI tool” places such documents in the public domain, while closed tools can summarise “without these risks” ([2026] UKUT 81 (IAC), headnote 4, para 21). In a criminal case, a New York federal court held that a defendant’s exchanges with a publicly available AI platform were neither privileged nor work product, the US shield for material prepared for litigation. It suggested the answer might differ if counsel had directed the use (United States v. Heppner, 17 February 2026, pp 4 to 12; p 7).^7,19

• In arbitration, the tribunal decides privilege under the rules it finds applicable, and may weigh waiver by “earlier disclosure” (IBA arts 9.2(b), 9.4(d)). Where parties face different privilege rules, an ICDR tribunal should, where possible, apply one rule to all, preferring the one giving “the highest level of protection” (art 25).
• SVAMC: no confidential information goes into an AI tool without appropriate vetting and authorisation (Guideline 2).
• AAA-ICDR guidance: arbitrators must not upload confidential or identifying case material into publicly available AI systems without contractual assurances against data reuse, model training or third-party access (p 2).
• The WIPO Arbitration and Mediation Center (WIPO) says pleadings and evidence should not be uploaded without adequate safeguards (section 4).
• Security: the cybersecurity protocol of the International Council for Commercial Arbitration (ICCA), the New York City Bar and CPR does not mention AI, but says information security should ordinarily be raised by the first CMC (Principle 10). The Singapore International Arbitration Centre’s information-security rule (rr 61.1 to 61.3) is in SIAC and AI in arbitration and arbitration in Singapore, and tool evaluation in AI legal tech.^{3,5,14,20,21,22,23}

How do arbitral rules treat AI-generated or deepfake evidence?

Arbitral rules treat AI-generated and deepfake material like any other evidence: the tribunal decides whether to admit it (AAA R-35(b)), and under the IBA Rules how much weight it carries (art 9.1). A deepfake is realistic text, images or video made or altered with AI. SVAMC’s guidelines say parties, their 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)” (Guideline 5). Its commentary says such fakes can be “virtually indistinguishable from authentic versions”, and lets a tribunal strike them, exclude them, draw adverse inferences or reflect the conduct in costs (p 19).^3,5,13,31

For AI-generated material a party relies on, Ciarb lets the tribunal ask the parties to explain “how inputs from the record are linked to machine-generated outputs” (para 6.6). When ruling on a tool’s use, it may consider the data behind the output, any bias and the tool’s quality, accuracy and security (para 6.4). Guidelines of the China International Economic and Trade Arbitration Commission (CIETAC) say AI use does not lessen a party’s responsibility (art 3.3, our translation). The table adds the SCC Arbitration Institute (SCC), the Vienna International Arbitral Centre (VIAC), AAA-ICDR, WIPO and JAMS.^6,25

How can a tribunal check that evidence is authentic and not a deepfake?

A tribunal tests authenticity, including a suspected deepfake, with tools already in the rules: originals, native files and metadata, its own expert, and adverse inferences or costs (IBA arts 9.6 to 9.8). It can ask for any original to be presented (IBA art 3.12(a)), and ICC’s 2022 report says it can permit inspection of originals or metadata when the authenticity of electronic copies is disputed (section 4.2.4). Who must prove a deepfake, and to what standard, is for the tribunal under its power over admissibility and weight (IBA art 9.1).^2,3

• ICC’s 2022 report says that, on anecdotal evidence, data integrity problems are rarely identified in arbitration, and that most parties do not make metadata part of an exhibit unless authenticity is “legitimately disputed” (section 4.6.2).
• The tribunal may appoint its own expert, including an AI expert, after consulting the parties (IBA art 6.1; ICC Rules 2026 art 26(3); Ciarb para 4.2).
• Ciarb says detection tools can spot deepfakes (para 1.8); SVAMC warns AI can make manipulation harder to detect (p 19).
• Hidden text: guidance for judges in England and Wales warns of “white text”, hidden prompts a computer reads but a person does not, and says judges must read the underlying documents (printed pp 6, 7).
• A memo in the US Evidence Rules agenda book for 15 October 2026 proposes noting, for draft Rule 901(c), facts that could support a deepfake challenge, such as missing or anomalous metadata, inconsistent timestamps, no native file and gaps in the chain of custody (pp 153, 159). None of it is approved.
• EU marking: from 2 August 2026, providers of generative AI must mark its output as detectable as AI-generated, and deployers must disclose deepfakes (EU AI Act, Regulation (EU) 2024/1689, arts 50(2), 50(4), 113). Providers of systems already on the market get until 2 December 2026 for art 50(2) (Regulation (EU) 2026/1744, art 1 point (39)(b)). These duties are not rules of evidence. The rest of the Act is in AI in arbitration.^{2,3,5,6,8,9,28,29,30}

What do the rules say about AI in expert reports and witness statements?

The tribunal tests and weighs AI-prepared expert reports and witness statements like any other evidence. An IBA expert report must describe “the methods, evidence and information used” (art 5.2(e)), and SVAMC says the tribunal and opposing counsel “may legitimately question a party, witness, or expert” about AI use (p 18). No arbitral text we checked bars AI help with a witness statement. Unless the parties agree otherwise, Ciarb lets a tribunal extend AI disclosure duties to party-appointed experts and fact witnesses (para 7.3); when AI use must be disclosed is covered in disclosure of AI use in arbitration.^3,5,6

• Courts, for comparison: in England and Wales, the Civil Justice Council’s February 2026 interim report proposed that experts explain AI use other than transcription (para 8.2) and that witness statements declare AI did not generate their content (paras 7.15, 7.16). Its June 2026 update kept witness statements under review and expects a final report later in 2026. The Supreme Court of New South Wales bars generative AI for the content of affidavits and witness statements (Practice Note SC Gen 23, para 10).
• In Crypto Open Patent Alliance v Wright [2024] EWHC 1198 (Ch), the court found a witness statement passage was concocted, probably with a generative AI chatbot ([537]).
• Translations must be marked as such (IBA art 3.12(e)). ICC’s 2022 report calls AI translations cheaper but “potentially less accurate”, and suggests agreeing their use for exhibits, with “a protocol for a supplemental human-generated translation” (section 5.2). Live interpretation and transcripts are in AI transcription and interpretation at hearings.^{2,3,31,32,33,39}

What have courts and court rule-makers said about machine-generated evidence?

No US federal rule of evidence on machine-generated or deepfake evidence is final, and none would bind an arbitral tribunal. As published in August 2025, proposed Federal Rule of Evidence 707 would have applied the Rule 702 reliability test for expert evidence to machine-generated evidence offered without an expert (PDF p 109). The Advisory Committee on Evidence Rules’ agenda for 15 October 2026 says that on 7 May 2026 it voted to withdraw that version and study a modified proposal (agenda book pp 5, 13).^8,34

• That agenda scheduled two expert panels, on deepfakes and on Rule 707 (pp 5 to 6). We found no published record of any decision taken at that meeting as of 29 October 2026.
• The working draft of Rule 707, tentatively approved in May 2026, says the party offering the evidence must “ordinarily” provide an expert to explain how the AI system reliably produced it, but in “exceptional circumstances” the court may consider other proof (pp 173 to 174).
• Draft Rule 901(c) on deepfakes is “prepared, but not finally approved” (p 5). Under the working draft, a party claiming an item was fabricated by generative AI would first give evidence sufficient to support a finding of fabrication; the party offering it would then have to show it is more likely than not authentic (p 150).
• Of the 931 federal trial judges who answered a Federal Judicial Center survey, 15 reported having encountered a deepfake challenge (draft minutes of 7 May 2026, agenda book p 20).
• Arbitration differs: AAA’s Commercial Rules say “Conformity to legal rules of evidence shall not be necessary” (R-35(a)).^8,13

AI in law covers court sanctions for AI-invented citations, or hallucination. Awards challenged over AI use explains the Québec annulment of an award whose authorities did not exist or did not support it. English courts are covered in arbitration in London, and an Indian Supreme Court ruling on unverified AI-generated precedents, which does not mention arbitration, in arbitration in India.^35,38

What should tribunals and parties settle early on AI review and evidence?

• If either side will use TAR or another AI review tool, raise its parameters at the first CMC (ICC 2022 report section 5.2) and record the decision in a procedural order such as Procedural Order No. 1 (Ciarb para 4.6).
• For wording, ICC’s 2022 report has sample procedural language on search parameters and production format (Appendix B, Example III.8, p 54), and Ciarb a template procedural order on AI (Appendix B).
• Keep requests narrow and proportionate whatever the tooling (IBA arts 3.3(a)(ii), 9.2(g); ICDR art 24(6)).
• Settle the format of production early, with native files and metadata for important exhibits whose authenticity may be disputed (IBA art 2.2(c); IBA commentary p 14).
• Adopt a rule against AI-falsified evidence, such as SVAMC Guideline 5.
• Keep produced documents out of tools that retain or reuse inputs: the IBA Rules require them to be kept confidential and used only for the arbitration (art 3.13).
• Tribunals: cross-check AI summaries and citations against primary sources, and verify “every automatically generated citation link” (AAA-ICDR guidance pp 1, 3).^{2,3,4,5,6,14,20}

What is still open on AI review and deepfake evidence?

• Whether a party must disclose that it used TAR, or the checks it ran on the results, without an order or agreement.
• Whether putting privileged material into an AI tool waives privilege in a given arbitration (IBA art 9.4(d); Heppner pp 6 to 8 and n 3).
• Who must prove a deepfake in arbitration, and to what standard.
• What ICC’s AI task force, Ciarb after its consultation and the London Court of International Arbitration’s rules review will say next. Our explainers on the LCIA, AAA and JAMS give each institution’s current position, and AI in dispute resolution covers mediation.^{3,19,36,37,40}