AI confidentiality and data protection in arbitration: what the rules say

The short version

• Putting confidential or privileged case documents into a public AI tool can breach confidentiality, waive privilege and bring data protection law into play.
• No arbitration rule we checked mentions AI, but the confidentiality and information-security duties in the LCIA, ICC, SIAC, HKIAC and AAA-ICDR rules still apply.
• The AI-guidance texts condition use, they do not ban it: vet the tool, bar reuse and model training, check retention, and redact before you upload.
• Courts have set no categorical rule. A public tool whose provider can use and disclose inputs can lose protection, but counsel-directed or closed-tool use may not.
• Data protection law treats uploading case data as processing. A third-party AI vendor is usually a processor needing an Article 28 contract, and a tool outside the EU needs a transfer basis.
• Settle it early. Raise AI, confidentiality and data security at the first case management conference (CMC) and record it in Procedural Order No. 1.^{1,2,3,5,6,7,8,12,14,17,22,23,24,32,33,34}

This page is the hub for keeping case material protected: confidentiality, privilege and data protection. Telling the tribunal you used AI is a separate question, in AI disclosure in arbitration. See also AI in arbitration and the AI rules tracker.

Can you put arbitration documents into an AI tool?

Not into a public tool without safeguards. With a vetted closed or enterprise tool, usually yes, if the rules, the parties’ agreement and data protection law allow it. In the 2025 Queen Mary University of London survey, 47% named confidentiality among their concerns about AI in arbitration (Chart 20).⁴¹

A UK tribunal put the risk plainly. Uploading confidential documents into an open tool is to place the information “in the public domain”, which breaches confidentiality and waives legal professional privilege; a closed tool that does not do so is “available for tasks such as summarising without these risks”.¹

The arbitration rules carry confidentiality duties, but none we read mentions AI, so uploading is governed by the general duty plus any tribunal order or applicable law. Whether an AI provider is even a permitted recipient is untested: it might be a “service provider” an LCIA party must bind (30.1), or a HKIAC “professional or any other adviser” (45.3(c)). The bar on arbitrators is the sharpest: the AAA-ICDR guidance says an arbitrator “must not upload any confidential or identifying case materials into publicly available AI systems that do not provide contractual assurances against data reuse, model training, or third-party access”.^8,14,22,23

For uploading the other side’s produced documents, US courts now set conditions. One barred putting documents marked confidential onto any AI platform. Another amended a protective order so confidential material may go into an AI tool only where the provider is contractually barred from storing or training on inputs and from disclosing them to third parties except as needed to deliver the service. The same order requires deletion on request and written documentation kept by the user. The document-production and IBA-privilege mechanics are in AI document review in arbitration; lawyers’ own duties are in AI legal tech.^3,4

What is the difference between public, enterprise and closed AI tools here?

The dividing line is what the provider’s terms let it do with your inputs: store them, train on them, or disclose them. Ciarb says, when evaluating an AI tool, to consider whether it has been vetted, its privacy and security standards, “and whether it is an open or closed model”.⁷

SVAMC says publicly available tools “could pose a risk of disclosing confidential information”, while business-oriented or privacy-oriented tools may offer safeguards, and tells participants to assess the tool’s confidentiality policies and how the information they provide is retained. An ICC technology report from 2022 warned that the terms of free online tools such as machine translation “may include a user license to the developer to use any content that is uploaded”, and that parties “should generally avoid” them for confidential or sensitive information.^6,20

This is also why the distinction bites on privilege. In the US ruling below, a public tool lost protection because the provider’s terms let it use and disclose the inputs. A closed or enterprise tool changes that calculus, but no court has held that a closed tool “preserves privilege”: the UK tribunal said only that it avoids the public-domain problem. How to evaluate a tool before buying it is in AI legal tech.^1,2

Does putting documents into an AI tool waive privilege or work product?

Putting documents into a public AI tool can waive privilege or work-product protection, but it is not automatic, and arbitration adds the question of which law governs privilege. In the US, on what the court called a question of first impression, a defendant’s written exchanges with a public AI platform were protected by neither attorney-client privilege nor the work-product doctrine: the platform is not a lawyer and its privacy policy let the provider use and disclose the inputs, and the exchanges were not prepared at a lawyer’s direction and were shared with a third party in any event. The court left open exchanges made at a lawyer’s direction.²

There is no categorical rule the other way. A later court, reading an earlier decision, recognised that AI-assisted litigation preparation “may qualify for work-product protection” and that using a third-party AI tool “does not necessarily result in waiver”. Whether protection holds turns on “the nature, purpose, and circumstances” of the materials, and AI communications can still be discoverable. In the UK, by contrast, uploading to an open tool breaches confidentiality and waives privilege, which the solicitors’ regulator warns “may be permanently waived and unable to be recovered”, with paid tools no safer by default.^1,3,5,40

In arbitration itself we found no award or court ruling on AI and privilege. The tribunal decides privilege under the rules it finds applicable. The IBA evidence rules let a tribunal exclude a document for a legal impediment or privilege and weigh any waiver, and the ICDR procedures tell a tribunal, where the parties would be under different privilege rules, to apply one rule to all and prefer “the rule that provides the highest level of protection”. Whether the prompt itself is privileged, and the disclosure duties, are in AI disclosure in arbitration; the decisions are tracked in AI arbitration case law.^25,31

Do the arbitration rules cover AI and confidentiality?

No rule mentions AI. The live question is whether the confidentiality duty binds the parties at all, or only the institution and the tribunal. Party-level duties exist under LCIA 30.1, SIAC 59.1, HKIAC 45.1, the JAMS International Rules 16.1 and the JAMS AI Disputes Rules 26(b).^{14,22,23,27,28}

Under the ICC, AAA Commercial, ICDR, JAMS Comprehensive and SCC rules, the confidentiality duty binds only the institution and the tribunal; the parties are bound only by their agreement, a tribunal order or the applicable law (our reading).^{17,24,25,26,30}

The 2026 ICC Rules add one duty that was not in the 2021 Rules: arbitrators “shall keep confidential all matters relating to” the arbitration (Art 12(8)), and the tribunal may make confidentiality orders and protect trade secrets (Art 23(3)). The 2021 Rules gave the tribunal the ordering power (then Art 22(3)) but no general confidentiality duty on the arbitrators. Each institution’s own text is covered in ICC, the LCIA, SIAC, AAA-ICDR and JAMS, and summarised in the AI rules tracker.^17,19

What do the AI guidelines require before you upload?

Every AI-specific text points one way: vet the tool, keep inputs out of training, check retention, and redact. SVAMC bars confidential information from any AI tool “without appropriate vetting and authorisation”. Ciarb tells a tribunal to preserve input confidentiality, “including by redacting or anonymising”, and treats a confidentiality breach as a “High Risk AI Use”. The table sets out each text; detail is in SVAMC Guidelines, Ciarb Guideline, SCC and AI and VIAC and AI.^6,7,8,9,12

How does data protection law apply when case data goes into an AI tool?

Uploading personal data to an AI tool is “processing” under the GDPR. The uploader is usually a data controller and the AI vendor a data processor, which needs a written processor contract under Article 28. The ICCA-IBA roadmap says arbitral participants “will generally be data controllers”, and that a processor must be bound by a compliant agreement.^33,34

The lawful basis is usually the legitimate interest of the participants in running the arbitration; special-category data should be processed only as far as necessary to establish, exercise or defend legal claims. The ICC model clause sets out exactly this split (GDPR Articles 6(1)(f), 9(2)(f)).^21,34

Security has to match the data. Article 32 requires “appropriate technical and organisational measures”, “taking into account the state of the art”, including pseudonymisation and encryption; the ICC note echoes the “state of the art of the technology”. Data minimisation cuts against feeding a model everything: redact or pseudonymise, and upload only what is necessary (Article 5(1)(c)).^18,33,34

A tool hosted outside the EU brings a cross-border transfer, which needs a Chapter V basis: adequacy under the EU-US Data Privacy Framework, but only where the vendor is certified under it, or standard contractual clauses, or a derogation. The legal-claims derogation in Article 49(1)(e) is for transfers that are “occasional” and have “a close and substantial connection” to the claim, so routine uploading for convenience is hard to fit (our reading). The LCIA’s data privacy notice is a worked example: it names the third-party providers that host proceeding data and the transfers it makes outside the EEA (paras 4.2.3, 5.1).^15,34,35,36

There is a regulatory layer on top. The EU AI Act can treat AI that helps decide an arbitration or other ADR with legal effects as high-risk under Annex III point 8(a); the 2 December 2027 start date under the AI Act digital omnibus, and who counts as a provider or a deployer, are in the EU AI Act in dispute resolution. The separate GDPR question, whether an AI-assisted ruling is a solely automated decision a party can object to, is in AI in dispute resolution in Europe.^37,38

What should Procedural Order No. 1 say about AI, confidentiality and data security?

Settle AI use, confidentiality and information security at the first CMC and record it in PO1. No rule forces this, but the confidentiality and information-security provisions are the hooks. The cybersecurity protocol says to raise information security “as early as practicable”, ordinarily by the first CMC; the SIAC rules have the tribunal discuss measures there, including cyber resilience; and the HKIAC, LCIA, ICDR and AAA rules each give the tribunal a place to address cybersecurity, privacy and data protection.^{14,22,23,24,25,32}

The ICC model data protection clause is drafted for PO1: a legitimate-interest basis, special-category data only for legal claims, transfers outside the EU on adequacy or standard contractual clauses, and a line telling the parties not to do anything contrary to it, “for example seeking consent, without first raising the issue with the tribunal”.²¹

Ciarb offers a template agreement and procedural order that preserve the confidentiality of inputs and flag High Risk AI Use, and VIAC suggests agreeing confidentiality provisions in PO1. The clause-level drafting lives in AI arbitration clauses.^7,11

The US model wording above travels: bar confidential material from any AI tool unless the provider is contractually barred from storing, training on and disclosing inputs and allows deletion, with written documentation kept.^3,4

One thing PO1 should not confuse: the JAMS AI Disputes Rules, with their default AI Disputes Protective Order and their secured-environment, expert-only inspection of AI systems (Rule 16.1(b)), are a different instrument, for disputes about AI systems rather than a party’s own use of an AI tool. See JAMS and AI.^28,29

A checklist before case documents go into an AI tool

• No training on inputs: the provider is contractually barred from using them to improve its model.
• Retention of prompts and outputs is limited, and deletion on request is available.
• No third-party access or disclosure, and no onward sub-processing without Article 28 terms.
• You know where the data is held and have a lawful cross-border transfer basis if it leaves the EU.
• You are using a closed or enterprise version, not a free consumer one.
• You have redacted, anonymised or pseudonymised the material before upload.
• You have read the terms of use and the privacy policy; engagement-letter wording is not enough. Deeper lawyer duties are in AI legal tech.
• You keep a written record of the tool used and can disclose it if ordered. The disclosure standards are in AI disclosure in arbitration.^{3,6,7,8,15,20,33,34,36,39,40}

Open questions on AI, confidentiality and data protection

• Whether an AI provider is a permitted recipient under HKIAC 45.3(c) or an LCIA 30.1 service provider.
• Whether AI prompts are privileged in arbitration, and which law governs privilege, where the IBA and ICDR texts set the method but no decision we found applies it to AI.
• Whether closed or enterprise tools preserve privilege: so far there is only dicta and regulator guidance, not a holding.
• How the parties’ confidentiality duty arises where the rules bind only the institution and the tribunal, as under the ICC, AAA, ICDR, JAMS Comprehensive and SCC rules.
• Security of the record itself, such as prompt injection or poisoned inputs, which no rule we found addresses beyond the general information-security duty.
• What the LCIA’s rules-revision committee on electronic communications, AI and information security, and the next Ciarb text, will say.^{1,7,14,16,17,23,24,25,26,30,31,32,40}