Can Arbitration Remain Confidential in the Age of AI?

Why Confidentiality May Not Survive AI

Artificial intelligence is rapidly becoming part of everyday legal practice, and arbitration is no exception. It is already being used for research, document review, translation, drafting support, summarisation, and case management, while institutions and practitioners continue to test how far those tools can safely go.

The real question is no longer whether AI will enter arbitration, but whether arbitration can remain meaningfully confidential once it does. If confidential submissions, witness materials, and draft awards are being processed by systems outside the direct control of the parties and the tribunal, the foundation of confidentiality itself begins to look fragile.

My own experience confirms both the promise and the limits of these tools. AI can accelerate research, summarise complex materials, and assist with drafting in ways that would have seemed remarkable only a few years ago. But speed is not judgment, and arbitration depends on the latter.

AI has already entered the arbitration process in several ways. Counsel use it for legal research, document review, drafting, and case preparation. Arbitrators use it for summarisation and administrative support. Arbitral institutions are also beginning to explore AI for analytics, workflow support, and procedural efficiency.

The central issue is not whether AI should be used, but how it should be used. In arbitration, the critical concern is whether AI assists human decision-making or begins to displace it. Arbitration is built on trust in a designated human decision-maker. Parties appoint an arbitrator because they want that person’s analysis, reasoning, and experience, not the output of a machine.

This raises difficult questions. At what point does AI-assisted drafting become machine adjudication? Where is the boundary between administrative support and substantive decision-making? An award prepared primarily through undisclosed AI assistance may undermine confidence in the arbitral process. Parties expect a human being to evaluate the facts, law, arguments, and credibility of witnesses.

The principle of due process creates further complications. AI systems may introduce authorities, analyses, or factual points that were never argued by the parties. If an arbitrator relies on such material, questions may arise about whether the parties had a fair opportunity to address those issues. That can create grounds for challenging or setting aside an award on procedural fairness grounds.

Equality of arms is another concern. Better-resourced parties may have access to sophisticated AI systems capable of analysing vast quantities of material quickly, while less-resourced parties may not. That imbalance could distort the arbitration process, especially in document-heavy disputes. Arbitrators may need to address this through proportionality, transparency, and procedural safeguards.

Bias is also a serious issue. AI systems are trained on data, and both the composition of that data and the design of the model influence outputs. Some systems may reproduce biases or limitations embedded in the technology. There is also the problem of opaque reasoning: users may see a polished answer without being able to trace how it was produced. That lack of transparency is troubling, particularly when the result is used in a legal setting.

Hallucinations remain a known risk. AI systems can fabricate cases, quotations, authorities, and legal principles while presenting them with confidence. Several Courts have sanctioned lawyers who filed fictional authorities generated by AI. Arbitration presents additional difficulties because tribunals generally lack the same coercive powers available to Courts.

Confidentiality: the real fault line

Confidentiality in arbitration is not merely a contractual obligation; it is part of the bargain on which parties choose arbitration instead of litigation. The risk posed by AI is different in kind from the risk of a physical file being mishandled in a law office. When a confidential document is kept within a law firm or tribunal, access is limited to a known and identifiable group of people who owe duties of confidence. When the same document, or even prompts describing the dispute, the issues, or the desired outcome, is uploaded to an AI system, control can be significantly reduced in many environments.

In particular, externally hosted systems may leave users with limited visibility into storage, retention, access controls, and downstream use of data. Even where contractual protections exist, parties may not enjoy the same practical degree of control that traditionally accompanies confidential arbitral materials. In some settings, users may not know where the data is stored, who may access it, whether it is used to improve future models, or whether it could later influence outputs generated for others.

That is why the confidentiality question is so difficult. The level of risk depends heavily on the type of AI tool being used. Public or externally hosted systems may create serious confidentiality concerns because the user may have limited control over storage, retention, access, and downstream use of data. By contrast, internally controlled or enterprise-grade tools may reduce the risk, especially where strong governance, contractual protections, and technical safeguards are in place. Even so, any use of confidential arbitral material in AI systems requires caution, verification, and strict data controls.

If confidentiality cannot be meaningfully safeguarded in practice across the tools actually being used, there is a strong argument for rethinking the balance between privacy and transparency in some contexts, including the publication of awards.

It is difficult to justify a strong confidentiality narrative while allowing highly sensitive information to be processed by systems over which the parties and tribunal may have limited real control.

Disclosure as a double-edged issue

Even if a tribunal or a party discloses AI use, that disclosure may not fully resolve the problem. A disappointed party may argue that the AI model used was unlawful in the seat of arbitration or in another jurisdiction closely connected to the proceedings. For example, if an AI service is restricted in a given country and an arbitrator or party accesses it through a VPN or a subscription arranged elsewhere, the use may raise questions under local technology, data, or national-security laws. In that setting, a challenge may be framed not merely as a procedural irregularity, but as a public policy issue.

Disclosure therefore cuts both ways. Non-disclosure may itself be advanced as a ground for challenge because the parties were denied the opportunity to regulate or object to AI use. Yet disclosure may also invite a separate challenge based on legality, data protection, or confidentiality concerns. The real problem is not simply transparency versus secrecy; it is that disclosure can destabilise the award from either direction.

The practical implications of mandatory disclosure may be significant. Once disclosure is required, parties may ask not only whether AI was used, but which model was used, when it was used, for what purpose, what prompts were entered, what data was provided, what outputs were generated, and whether an audit trail exists. A regime designed to promote transparency could therefore generate satellite disputes, confidentiality objections, jurisdictional challenges, and post-award litigation. In some cases, the disclosure obligation itself may become a greater threat to finality than the underlying use of AI.

The reputational harm from such disputes may also be severe. Even if an award survives, the parties and the arbitrator may face prolonged litigation, regulatory scrutiny, and damage to their standing in the market. In many cases, the reputational cost may outweigh any efficiency gained from using AI.

The Quebec case

The risks described above are no longer theoretical. In ARIHQ c. Santé Québec (CIUSSS du Centre-Sud-de-l’Île-de-Montréal), 2026 QCCS 1360 (22 April 2026), the Quebec Superior Court annulled an arbitral award after finding that the reasoning relied on non-existent authorities and that this seriously compromised the integrity of the arbitral process. The Court did not prohibit AI use in arbitration altogether, but it held that AI cannot replace an arbitrator’s independent reasoning and decision-making. The award, which concerned a healthcare dispute of roughly CA$1.2 million, was set aside and the parties were ordered to appoint a new arbitrator within 60 days.

The judgment makes clear that not every use of AI will invalidate an award. The key question is whether procedural integrity was compromised in a way that could affect the result. It also confirms that AI may be used for limited tasks such as summarising documents or translating material, provided that the final reasoning reflects the arbitrator’s own judgment and the sources are properly verified. In effect, the Quebec decision draws a bright line between permissible AI assistance and impermissible AI adjudication.

Other jurisdictions

The Quebec decision is not isolated. In the United States, a party filed a petition to vacate an arbitral award on the basis that the arbitrator had outsourced his adjudicative role to artificial intelligence: LaPaglia v Valve Corp, No. 3:25-cv-00833 (S.D. Cal., filed 8 April 2025). The petitioner alleged that the award contained factual assertions not in the record and bore the hallmarks of AI-generated drafting, and also alleged that the arbitrator had previously acknowledged using ChatGPT for writing. The petition was ultimately dismissed on jurisdictional grounds under the Federal Arbitration Act, and the Court did not reach the merits of the AI allegations. Even so, the case shows that parties are prepared to invoke AI misuse as a basis for challenging awards, and Courts will increasingly be asked to confront the line between permissible assistance and impermissible delegation.

Guiding principles

Emerging institutional guidance, including the CIETAC Guidelines on the Use of Artificial Intelligence in Arbitration (July 2025), rests on three core principles. CIArb adopts a similar approach, treating AI as an assisting tool and allowing arbitrators to regulate, limit, or prohibit its use where transparency, fairness, confidentiality, or data security are at risk.

· Party autonomy

· Assistance only

· Continuing obligations

These principles confirm that efficiency must never come at the expense of fairness, confidentiality, or human judgment.

Safeguards for practice

A cautious approach requires several disclosures, verifications, security measures, ownership checks, clear procedural rules, and a pre-deployment assessment before any AI tool is used.

From tool to agent

When many early guidance documents were drafted, AI was largely a passive tool that responded to individual prompts. Today, autonomous AI agents can read files, adapt strategies, and execute multi-step workflows with limited human input. The challenge is therefore no longer only how to use a tool, but how to supervise a digital actor.

That shift requires fresh calibration. The balance is no longer simply between efficiency and accuracy, but between managing semi-autonomous systems and preserving the human core of adjudication. Any framework must ensure that AI remains a powerful assistant to justice, never a quiet substitute for it.

There is also an ongoing debate about whether formal AI guidelines are necessary at all. Some argue that technology evolves too quickly for detailed regulation. Broad principles may be preferable to rigid rules that become outdated within months. Questions remain about disclosure obligations, including whether parties should disclose all AI use or only substantive applications involving drafting, analysis, or decision support.

Many practitioners already rely on AI as an assistant. They use it for research, summarisation, document analysis, presentations, and organisational tasks. However, they generally stop short of allowing AI to determine witness credibility, interpret evidence conclusively, or decide legal issues. Human judgment remains indispensable.

Several broader concerns remain unresolved. Should parties disclose which AI systems they use? Who should bear the cost of verifying AI-generated work? Could widespread reliance on a small number of AI systems create systemic bias across the legal industry? If AI-generated awards become more common, should there be additional review mechanisms beyond traditional setting-aside and enforcement proceedings?

Different jurisdictions are already taking different approaches. Some Courts focus primarily on lawyer accountability, requiring lawyers to stand behind all submissions regardless of AI assistance. Others require disclosure of AI use in particular circumstances. Concerns are especially acute in relation to witness statements and expert reports, because AI-generated language may distort the evidence.

At the same time, it is important not to overreact. Lawyers have always relied on assistants, clerks, researchers, and junior colleagues to help prepare documents. In many respects, AI is another tool in a long tradition of delegated assistance. The key distinction is ensuring that responsibility, judgment, and accountability remain with the human professional.

Conclusion

The central issue is not whether AI will be used, but whether arbitration can remain meaningfully confidential if it is used widely. Hallucinations are a known risk, but they are not the only challenge. The deeper issue is that once confidential information, or even prompts revealing the nature of a dispute, is entered into many AI systems, the risk is no longer merely local; it can become broader, harder to control, and difficult to reverse

If the arbitration community is willing to accept that risk in the name of efficiency, it should also be willing to confront the consequences honestly. The future of arbitration may depend less on partnership with AI than on a candid choice: preserve confidentiality by restricting AI use in sensitive contexts, or accept that the traditional model of private arbitration is under increasing strain

In arbitration, the most important question about AI may not be how to use it, but when to refuse it.

Bibliography

1. LaPaglia Valve Corp, No. 23-cv-02600 (ND Call filed 14 March 2023)

2. Quebec Association of Intermediate Housing Resources (ARIHQ) v Quebec Health – Integrated University Health and Social Services Centre of South-Central Montreal (CIUSSS), 2026 QCCS 1360 (22 April 2026).

3. China International Economic and Trade Arbitration Commission (CIETAC), Guidelines on the Use of Artificial Intelligence in Arbitration (July 2025); and Chartered Institute of Arbitrators, Guidelines on the use of AI in Arbitration (2025), available at:https//www.ciarb.org/news-listing/ciarb-launches-guidelines-on-the-use-of-ai-in-arbitration/

Other works by the author (AI)

4. ‘2025 Changed Everything For AI Accountability’ (25 December 2025) LinkedIn https://www.linkedin.com/pulse/2025-changed-everything-ai-accountability-ahmed-ashfaq-dmqhc?utm_source=share&utm_medium=member_ios&utm_campaign=share_via

5. ‘The 2026 Agentic Accountability Crisis’ (1 January 2026) LinkedIn https://www.linkedin.com/pulse/2026-agentic-accountability-crisis-ahmed-ashfaq-l7v7c?utm_source=share&utm_medium=member_ios&utm_campaign=share_via

6. ‘Why the Sceptic Will Inherit the Courtroom’ (11 March 2026) LinkedIn https://www.linkedin.com/pulse/why-sceptic-inherit-courtroom-ahmed-ashfaq-utf6c?utm_source=share&utm_medium=member_ios&utm_campaign=share_via

7. ‘A Few Practical Insights on AI Tools’ (27 March 2026) LinkedIn https://www.linkedin.com/pulse/few-practical-thoughts-ai-tools-from-someone-who-has-used-ashfaq-ryxec?utm_source=share&utm_medium=member_ios&utm_campaign=share_via

8. ‘Artificial Intelligence Will Write the Pitch. People Will Win the Client. Exploring the Future of Legal Services Through AI, Data, Pricing, and Human Connection’ (11 July 2026) LinkedIn https://www.linkedin.com/pulse/artificial-intelligence-write-pitch-people-win-client-ahmed-ashfaq-rbefc?utm_source=share&utm_medium=member_ios&utm_campaign=share_via

9. ‘Reimagining International Arbitration: A New Architecture for Trust, Efficiency, and Human Judgment in a Disrupted World’ (2 July 2026) LinkedIn https://www.linkedin.com/pulse/reimagining-international-arbitration-a-new-architecture-ahmed-ashfaq; also available on the author’s newsletter https://newsletter.ahmedashfaqsolicitor.com/p/reimagining-international-arbitration?utm_campaign=post&utm_medium=web

Disclaimer:

This article is provided for general informational purposes only and does not constitute legal advice. While reasonable efforts have been made to check the accuracy of the content, no guarantee is given, and errors or omissions may remain. Readers should conduct their own research and seek independent professional advice before relying on any of the matters discussed. The views expressed are those of the author in a personal capacity and do not necessarily reflect the positions of any institution, tribunal, or client.