Introduction
Confidentiality is widely regarded as a crucial advantage of international commercial arbitration, particularly in commercial disputes where parties seek to avoid disclosure of sensitive business information, trade secrets, and strategic material.[1] Confidentiality protections extend to hearings, documents, awards, and external disclosures, serving vital commercial interests including the protection of trade secrets and reputational capital.[2] Many arbitral regimes expressly protect this principle; Article 30 of the 2020 LCIA Rules and Rule 59 of the 2024 SIAC Rules, for example, impose confidentiality obligations, while some national laws remain silent and leave the matter to implication or party agreement.[3] Yet the rapid integration of generative Artificial Intelligence into legal practice now places this principle under unprecedented strain.
AI tools are increasingly used for document review, legal research, summarization, and drafting. Their efficiency is attractive, but their use in arbitration creates novel risks because confidential case materials may be processed by external systems beyond the parties’ and tribunal’s control. These concerns are no longer speculative. According to the 2025 White & Case international arbitration survey, while 91% of respondents expect to use AI for research and data analytics in the next five years, 47% identified confidentiality and data breaches as a primary obstacle to wider adoption.[4]
Although recent guidelines acknowledge these dangers, they remain non-binding and largely discretionary. That is insufficient in a process whose legitimacy depends heavily on party trust in confidentiality. This article argues that arbitral institutions must move beyond soft-law guidance and amend their formal rules to impose clear, binding obligations governing the use of AI in arbitral proceedings.
The Nature of the AI Threat
The confidentiality risks introduced by generative AI are qualitatively different from ordinary cybersecurity concerns. The problem is not merely unauthorized access by hackers, but the possibility that participants themselves may input confidential materials into third-party AI systems that store, reuse, or repurpose that information. Once entered into such systems, sensitive material may leave the effective control of the user altogether.
This risk arises first through the conduct of parties and counsel. In the interest of speed and efficiency, legal teams may use public large language model platforms such as ChatGPT to summarize evidence, organize submissions, or refine written arguments. In doing so, they may expose witness testimony, internal legal analysis, trade secrets, and litigation strategy to systems that may retain user inputs or use them for training. Such disclosure may also produce waiver consequences. The decision in Tremblay v. OpenAI emphasized that prompts reflecting counsel’s mental impressions may constitute protected work product, but that voluntary exposure of such information generally undermines this protection.[5] This risk is compounded in arbitration, where no clear institutional rule defines what constitutes ‘voluntary disclosure’ when AI tools are used, leaving parties uncertain whether their use of AI (even with contractual safeguards) may inadvertently waive confidentiality protections.
The danger is as serious when AI is used by arbitrators. Arbitrators may be tempted to use AI tools to summarize records, conduct legal research, draft procedural orders, or prepare portions of awards. Yet this creates a direct risk to the confidentiality of tribunal deliberations. The tribunal’s internal reasoning, preliminary views, and evaluative thought process form part of the protected secrecy of arbitration.[6] If such material is processed through external AI tools, the result is not simply a technical lapse but a potential intrusion into the adjudicative core of the arbitral process.[7]
Why Current Rules Are Insufficient
The problem is that AI introduces confidentiality threats qualitatively different from those contemplated by traditional rules. Standard confidentiality clauses prohibit unauthorized disclosure by participants, but they do not regulate what happens when confidential information is voluntarily entered into third-party AI systems that may store, republish, or use that data for training purposes. The SVAMC Guidelines warn that trade secrets can lose protection if entered into an AI model that “stores or republishes user inputs,” yet this warning remains advisory.[8] Similarly, the 2025 CIArb Guideline encourages parties and arbitrators to vet AI tools for privacy and security, distinguish between open and closed models, and prioritize transparency; but it does not prohibit risky conduct outright.[9]
These instruments are explicitly soft law. They are non-binding, discretionary, and lack enforcement mechanisms. The ICC Rules illustrate the broader problem: Article 22 grants tribunals discretionary authority over confidentiality, and Article 8 protects the ICC Court’s internal work, but the 2021 Rules still contain no default obligation of confidentiality applicable to all participants.[10] This discretionary approach is insufficient to create the robust and predictable protection that confidentiality demands, particularly when 91% of respondents expect to use AI for research and data analytics in the next five years. This inadequacy is not unique to AI. Scholars have observed that traditional confidentiality frameworks in arbitration were designed for analog practices and fail to account for the structural risks posed by digital technologies, which expose private information in ways that conventional security measures cannot adequately address.[11]
Current frameworks are reactive, not preventative. Breaches may support annulment under Article 34(2)(a)(ii) of the UNCITRAL Model Law, but that remedy comes too late, and after confidential material has already been compromised. The recent U.S. case LaPaglia v. Valve illustrates this problem: the court vacated an award due to the arbitrator’s undisclosed AI use and potential data leaks, but only after the arbitration had been concluded and any confidential information was already at risk.[12]
Proposed Reforms
The inadequacy of current frameworks calls for urgent institutional reform. Arbitral institutions must move beyond non-binding guidelines and amend their formal rules to impose clear, binding obligations on all participants regarding AI use.
First, institutional rules should establish a mandatory disclosure regime. Any party, counsel, or arbitrator intending to use AI tools must disclose this intention in advance, specifying the tool, its provider, the nature of tasks it will perform, and its data handling practices, and the other party should have the right object to such use. The opposing party shall have the right to object to such use; however, as with objections based on privilege or confidentiality, the ultimate determination should rest with the tribunal.
Second, rules must require express authorization before confidential material may be processed by AI systems. The default position should be a prohibition: no confidential hearing transcripts, witness statements, or submissions may be input into any AI tool (whether public or proprietary) without the informed, written consent of all parties and the tribunal’s approval. Alternatively, absent such consent and approval, parties should be required to redact all confidential information before using any AI system, thereby anonymizing the confidential data.
Finally, rules should provide enforceable sanctions. Undisclosed or unauthorized AI use should constitute a breach of procedural fairness, giving rise to challenges to awards under Article 34(2)(a)(ii) of the UNCITRAL Model Law, and potentially to professional disciplinary consequences for counsel or arbitrators.
These reforms would transform AI from a latent threat into a regulated tool, preserving the confidentiality and integrity that underpin international arbitration.
Although critics may argue that mandatory disclosure and authorization requirements impose excessive costs, slow arbitration proceedings, and risk stifling beneficial AI innovation. However, these concerns are overstated. First, disclosure need not be burdensome; a brief statement identifying the tool, provider, and task suffices and can be standardized by institutions. Second, the efficiency gains AI offers are meaningless if they come at the cost of confidentiality breaches that undermine the award’s enforceability or expose parties to liability. And finally, the proposed framework does not ban AI; it regulates its use to ensure compliance with arbitration’s foundational confidentiality obligations. The alternative which is waiting for breaches to occur and relying on post-hoc remedies like award challenges would be far more costly and damaging to the legitimacy of international arbitration.
References
[1] Gary Born, International Commercial Arbitration (3rd edn, Kluwer Law International 2021), §20.01.
[2] Margaret Moses, Principle and Practice of International Commercial Arbitration (2nd edn, Cambridge University Press 2012), p 3.
[3] Ibid §20.03 [D] [1].
[4] White & Case LLP, ‘2025 International Arbitration Survey: Adapting Arbitration to a Changing World’ (2025) https://www.qmul.ac.uk/arbitration/media/arbitration/docs/White-Case-QMUL-2025-International-Arbitration-Survey-report.pdf accessed 19 July 2026, p 28-30.
[5] Stephen Embry, 'Privilege In the Age of Gen AI: Lots of Questions' (TechLaw Crossroads, 10 December 2024) https://www.techlawcrossroads.com/2024/12/privilege-in-the-age-of-gen-ai-lots-of-questions/ accessed 19 July 2026.
[6] Article 39.1 of the SIAC Rules.
[7] Mark‑Silas A. Malekela, ‘AI and Confidentiality protection in International Commercial Arbitration: Analysis of the existing legal framework’ (2025) 5:83 Discover Artificial Intelligence, p 6.
[8] Silicon Valley Arbitration and Mediation Center, 'Guidelines on the Use of Artificial Intelligence in Arbitration' (2024) https://svamc.org/wp-content/uploads/SVAMC-AI-Guidelines-First-Edition.pdf accessed 19 July 2026, p 9.
[9] Chartered Institute of Arbitrators, 'Guideline on the Use of AI in Arbitration' (2025) https://www.ciarb.org/media/m5dl3pha/ciarb-guideline-on-the-use-of-ai-in-arbitration-2025-_final_march-2025.pdf accessed 19 July 2026, p 5.
[10] Except for the ICC court itself in its internal rules, Appendix II.
[11] Nobumichi Teramura & Leon Trakman, ‘Confidentiality and privacy of arbitration in the digital era: pies in the sky?’ (2024) 40 Oxford Arbotration International, p 303.
[12] Aceris Law, 'When Arbitrators Use AI: LaPaglia v. Valve and the Boundaries of Adjudication' (Aceris Law, 19 April 2025) https://www.acerislaw.com/when-arbitrators-use-ai-lapaglia-v-valve-and-the-boundaries-of-adjudication/ accessed 19 July 2026.