AI and Document Disclosure in International Arbitration

Artificial intelligence is often presented as a solution to one of international arbitration's long-standing challenges: the time and expense of document disclosure. AI-powered review tools can analyse vast collections of emails, messages and other electronically stored information (ESI) in significantly less time than a traditional manual review. Semantic search tools can identify documents by concepts and contextual relationships, not just keywords. The promise is straightforward: faster, cheaper and more efficient disclosure.

What has received less attention, however, is that AI does more than improve the efficiency of disclosure. It significantly changes the dynamics of disclosure, creating new risks and a new dimension of procedural and tactical dispute.

Below, we examine the key risks and explore how parties can manage them — or turn them to tactical advantage.

 

Expansion of the Scope of Disclosure

International arbitration has traditionally taken a more limited and targeted approach to document production. Unlike litigation, there is no automatic disclosure obligation. Instead, parties produce documents only in response to specific requests, typically submitted in a Redfern Schedule, which are assessed and decided by the tribunal. Requests must be narrowly tailored, relevant, material and proportionate; broad or speculative “fishing expeditions” are routinely rejected.

Historically, practical considerations reinforced these limits. Reviewing large volumes of documents was expensive, time-consuming and often difficult to justify. Parties and tribunals therefore had to focus on the categories of documents most likely to affect the outcome. AI may change that balance by reducing the practical burden of wider searches.

A review exercise that once required hundreds or even thousands of lawyer hours can now be completed in significantly less time using semantic search, predictive coding and AI-assisted workflows. Unlike traditional keyword searches, semantic search tools can identify documents based on conceptual relationships. A search for “pricing discrepancies”, for example, may identify documents referring to “unusual margins”, “rebate variances” or “departures from standard pricing policies”, even if none of those terms appears in the underlying documents. As the cost of searching and producing documents falls, tribunals may become less receptive to objections based on burden and expense alone.

The question may gradually shift from:

“Are the documents requested relevant and material, and are the costs of searching for and producing those documents proportionate to their evidential value?”

to:

“If the technology exists to conduct the search quickly and economically, why should the search not be undertaken?”

Such a shift may have significant consequences. Requests that would previously have been viewed as disproportionate may become harder to resist, increasing the range of custodians, repositories and document categories reviewed. This risks broadening the scope of disclosure beyond what is genuinely material.

 

Search Methodology: The New Battlefield

AI-assisted disclosure may create a new category of procedural dispute: disputes about how documents are identified and reviewed.

Traditional disclosure disputes typically concerned matters tribunals could readily understand and assess, such as custodians, repositories, date ranges and search terms. AI-assisted review is different. Modern review platforms may use semantic search, predictive coding and generative AI. The results can vary materially depending on the model, the instructions given to the system and the checks used to test the results.

These differences do not necessarily indicate error. They may reflect legitimate methodological choices. Choices about search instructions, retrieval settings and quality-control testing can affect which potentially responsive documents are identified for review.

As AI plays a larger role in document disclosure, parties may seek information about the process used by their opponents. They may ask what systems were used, how searches were configured and what checks were carried out.

Some transparency may be needed to maintain confidence in the integrity of the disclosure exercise. However, requiring production of search instructions, review settings or quality-control materials may expose counsel's legal strategy and case assessment, potentially compromising privilege.

The result could be what might be called “disclosure about disclosure”. Rather than focusing only on the substantive dispute, parties may find themselves arguing about search design, quality control and alternative methods.

AI could give rise to secondary disputes about the adequacy of the disclosure process itself. Without early agreement on methodology, these disputes risk adding cost and delay rather than reducing them.

 

Privilege and Confidentiality Risks posed by AI

A further significant risk arises where the use of AI itself threatens a loss of privilege. Traditionally, legal advice provided by lawyers to client are protected by privilege and therefore not subject to disclosure.

However, AI are not lawyers. When privileged or confidential materials are uploaded to an AI platform, the platform processes them through its underlying infrastructure. Questions arise about data retention, model training, access rights, security architecture and confidentiality safeguards. Not all AI tools provide the same protections. Particular caution is warranted with consumer-facing generative AI platforms that retain user inputs or process information outside dedicated, ring-fenced environments.

Recent case law shows that these concerns are not merely theoretical. In United States v Heppner, 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 17, 2026), Bradley Heppner, a criminal defendant, used consumer-grade Claude without any direction from his lawyers to research legal issues and prepare reports on his defence strategy. He then shared the AI-generated material with counsel. The court held that the material was protected by neither attorney-client privilege nor the work-product doctrine. It gave three central reasons. First, Claude is not a lawyer, so the communications were not between a client and an attorney. Second, Anthropic’s privacy policy allowed data to be shared with third parties, so the communications were not confidential. Third, the material was not privileged when created. It did not become privileged merely because Heppner later shared it with his lawyers.

The client-side lesson is important. If a party uses public AI to analyse its legal position or prepare strategy documents, those materials are unlikely to be privileged. The AI is not a lawyer and does not owe duties of confidentiality. The communication was never confidential. Sharing the material with lawyers later does not change that. An opposing party may therefore seek disclosure of those AI-generated strategy documents.

Similar concerns have arisen in England, but Munir v Secretary of State for the Home Department [2026] UKUT 81 (IAC) is primarily about AI hallucinations, professional standards and supervision. The Upper Tribunal examined lawyers’ use of non-specialist AI, including Google AI, for legal research and drafting. That use generated fake case citations and exposed failures to check legal work and supervise junior staff. The privilege point was ancillary. The Tribunal observed that uploading confidential case materials to an open-source AI tool such as ChatGPT places the information in the public domain, breaches client confidentiality and waives legal privilege. It distinguished open-source tools from closed-source tools such as Microsoft Copilot, which do not carry the same risk. Munir should therefore be read mainly as a warning about hallucinations and professional standards, with the privilege observation as an additional warning.

The lesson is not that AI should be avoided. Rather, AI should be deployed with proper lawyer oversight and clear governance frameworks.

For parties to an arbitration, the practical risk is clear. AI-generated analysis of legal exposure, claims, defences or strategy may not attract privilege — and an opposing party could seek its disclosure. In-house teams and employees should be made aware of this risk.

Parties and their in-house teams should treat AI as a tool for factual research, document organisation and information retrieval, not as a substitute for legal advice. Where a question involves legal analysis, litigation strategy or an assessment of legal risk, they should consult their lawyers rather than an AI platform. If AI is used to support legal work, it should be used under the direction of counsel and within an enterprise-grade platform that provides contractual confidentiality protections. This distinction matters: using AI to search for documents is different from using it to assess whether those documents help or harm the case.

 

Strategies and Practical Safeguards in AI-Assisted Disclosure

AI is changing the mechanics of document disclosure. The systems selected and the rules governing their use may have a significant impact on the evidentiary record before the tribunal.

There is no one-size-fits-all approach. What benefits one party may disadvantage the other. Parties should think strategically about AI from the outset — not just whether to use it, but how.

To avoid surprises and disputes during the arbitration, parties should address the following issues in the initial procedural order.

• First: which AI platforms may be used?

Parties should decide whether to limit disclosure to enterprise-grade review platforms or allow other AI tools. Agreeing this at the outset can reduce disputes about confidentiality, privilege and the integrity of the review.

• Second: what evidence is sufficient to show compliance with disclosure obligations?

This may be the most important strategic issue in AI-assisted disclosure.

A party seeking certainty and finality may argue that a signed verification should provide initial evidence of compliance. It would confirm that the agreed or tribunal-ordered search instructions were run on an approved AI platform and that all responsive documents identified were produced.

By contrast, a party seeking broader disclosure may argue that an AI-assisted search should not be treated as conclusive. It may seek information about how the search was configured and what checks were used to assess the results. If the documentary record contains unexplained gaps, it may request supplementary searches.

The strategic debate is therefore not whether AI was used, but whether the resulting search should be treated as conclusive.

• Third: how should inadvertent disclosure of privileged material be addressed?

Because AI-assisted review may misclassify privileged material, parties should consider enhanced claw-back protections in their disclosure protocol.

The protocol can provide that inadvertent disclosure does not waive privilege and set out procedures for the prompt return and non-use of affected materials. The scope of these protections may itself become strategically important, balancing efficiency against adequate quality control.

• Fourth: what standard should apply to disclosure requests?

A producing party is likely to argue that traditional legal standards — relevance, materiality and proportionality — still apply. AI may make review more efficient, but it should not be used to justify broad, fishing-expedition-style requests.

Conversely, a requesting party may argue that AI reduces the time and cost of disclosure. The tribunal may therefore take those efficiencies into account when assessing proportionality. Searches once considered unduly burdensome or disproportionate may now be justified because AI tools are available.

Beyond formal protocols, parties should take practical steps to reduce risk and strengthen their position. They should avoid non-enterprise AI tools, clearly label privileged material to reduce the risk of inadvertent disclosure, and ensure that privilege decisions and quality-control reviews remain subject to proper human oversight. Parties should also ensure that employees and in-house teams do not use public AI tools to analyse legal exposure or prepare litigation strategy, as this may compromise privilege.

When framing disclosure requests, legal representatives should be precise. They should specify the search instructions, confidence and relevance measures, repositories to be searched and other relevant parameters. This can reduce the producing party’s scope for argument and limit unnecessary disputes.

 

Conclusion

Ultimately, AI-assisted disclosure should be viewed not merely as a question of technology and efficiency, but as a question of strategy. Parties should seek to agree an AI protocol early in the procedural timetable and seek specialist advice where the technology, privilege or confidentiality risks warrant it. In particular, parties should be clear about what AI can and cannot safely be used for: factual research and document management carry different risks from legal analysis and strategy, and the latter should remain with qualified lawyers. Those that engage with the risks and opportunities early will be best placed to capture the benefits and shape the disclosure process to their advantage. The landscape of disclosure is shifting; the question now is who is prepared to adapt.